Browse the corpus

Introduction Amyotrophic lateral sclerosis is a neurodegenerative illness causing progressive weakness and wasting of muscles controlling movement, breathing, and swallowing.1 Dysphagia is a common problem in patients with amyotrophic lateral sclerosis and causes difficulties in maintaining a safe and adequate oral intake of nutrition and fluids.2 Patients with severe dysphagia often experience weight loss, choking, and coughing on attempting to swallow, episodes of aspiration, and prolonged and effortful mealtimes.3–6 Gastrostomy feeding is recommended to provide long-term nutritional support for patients with amyotrophic lateral sclerosis with severe dysphagia.7 Three main methods of gastrostomy insertion are currently used in patients with amyotrophic lateral sclerosis: percutaneous endoscopic gastrostomy, radiologically inserted gastrostomy, and per-oral image-guided gastrostomy.8 However, with little evidence available,9,10 current practice in relation to choice of method and timing of gastrostomy insertion is largely based on consensus and expert opinion.8 Gastrostomy could be beneficial for the survival, quality of life, and nutritional outcome of patients with this disease, but there is a paucity of high-quality evidence relating to these aspects of the intervention.9,11–13

of method and timing of gastrostomy insertion is largely based on consensus and expert opinion.8 Gastrostomy could be beneficial for the survival, quality of life, and nutritional outcome of patients with this disease, but there is a paucity of high-quality evidence relating to these aspects of the intervention.9,11–13 In response to the paucity of evidence and calls by organisations such as the American Academy of Neurology and the European Federation of Neurological Societies for more evidence to guide clinicians and optimise standards of care,7,14 we aimed to identify the optimum gastrostomy timing and insertion method in terms of safety and clinical outcomes.

evidence and calls by organisations such as the American Academy of Neurology and the European Federation of Neurological Societies for more evidence to guide clinicians and optimise standards of care,7,14 we aimed to identify the optimum gastrostomy timing and insertion method in terms of safety and clinical outcomes. Methods Study design and participants In this large, multicentre, longitudinal, prospective cohort study (ProGas), we enrolled patients with a diagnosis of definite, probable, laboratory supported, or possible amyotrophic lateral sclerosis (as defined by the El Escorial criteria),15 who had agreed with their clinicians to undergo gastrostomy at one of 24 motor neuron disease care centres or clinics in the UK (21 in England, two in Scotland, and one in Northern Ireland). Patients who had been diagnosed with a disorder characterised by cognitive impairment, such as frontotemporal dementia, were excluded. Patients were approached and invited to take part in the study by a member of the research team when a decision had been made to refer the patient for a gastrostomy insertion. Ethical approval was granted by the National Health Service NRES Leeds (Central) Research Ethics Committee and applied to all participating care centres or clinics. Informal carers, such as family members, of patients who had accepted to take part in the study were also invited to participate. All participants who agreed to take part in the study provided written informed consent before data collection. Research in context Evidence before this study

Methods Study design and participants In this large, multicentre, longitudinal, prospective cohort study (ProGas), we enrolled patients with a diagnosis of definite, probable, laboratory supported, or possible amyotrophic lateral sclerosis (as defined by the El Escorial criteria),15 who had agreed with their clinicians to undergo gastrostomy at one of 24 motor neuron disease care centres or clinics in the UK (21 in England, two in Scotland, and one in Northern Ireland). Patients who had been diagnosed with a disorder characterised by cognitive impairment, such as frontotemporal dementia, were excluded. Patients were approached and invited to take part in the study by a member of the research team when a decision had been made to refer the patient for a gastrostomy insertion. Ethical approval was granted by the National Health Service NRES Leeds (Central) Research Ethics Committee and applied to all participating care centres or clinics. Informal carers, such as family members, of patients who had accepted to take part in the study were also invited to participate. All participants who agreed to take part in the study provided written informed consent before data collection. Research in context Evidence before this study We searched PubMed, Embase, The Cochrane Library, and ISI Web of Knowledge for reports published before July 1, 2010, combined with citation searching and reference chaining, using the keywords: “motor neuron* disease” or “MND”, “amyotrophic lateral sclerosis” or “ALS”, “gastrostomy”, “percutaneous endoscopic gastrostomy” or “PEG”, “radiologically-inserted gastrostomy” or “RIG”, “per-oral image-guided gastrostomy” or “PIG”, “timing”, “mortality”, “safety”, “nutritional outcome”, “benefits”, and “quality of life”. We identified several studies reporting mortality data after gastrostomy insertion in patients with amyotrophic lateral sclerosis, but only a handful directly compared survival time or 30-day post-procedure mortality after different methods of gastrostomy. In a meta-analysis of the data of the four studies that allowed within-study comparisons of percutaneous endoscopic gastrostomy versus radiologically inserted gastrostomy or per-oral image-guided gastrostomy, the difference in 30-day mortality was increased by 2·1% for percutaneous endoscopic gastrostomy compared with the other insertion methods. However, the results of the meta-analysis did not provide robust evidence to indicate which method is safer because of an absence of within-study comparisons, differences between populations, small sample sizes, and low event rates. The urgent need for prospective clinical trials in relation to the optimum method and timing for gastrostomy insertion, as well as the nutritional outcome for the patients, was highlighted in a Cochrane review on enteral tube feeding for amyotrophic lateral sclerosis and echoed in calls for more robust evidence by multiple organisations, such as the American Academy of Neurology and the European Federation of Neurological Societies.

on, as well as the nutritional outcome for the patients, was highlighted in a Cochrane review on enteral tube feeding for amyotrophic lateral sclerosis and echoed in calls for more robust evidence by multiple organisations, such as the American Academy of Neurology and the European Federation of Neurological Societies. Added value of this study To our knowledge, ProGas is the first large, multicentre, longitudinal, cohort study to assess and compare the different methods of gastrostomy and explore the issue of optimal timing for gastrostomy insertion in patients with amyotrophic lateral sclerosis. Implications of all the available evidence In the absence of data from randomised trials, our findings might help neurologists, patients with amyotrophic lateral sclerosis, and the carers of patients with amyotrophic sclerosis to make decisions about the timing and method of gastrostomy. The next steps in building the evidence base must be to understand further the nutritional requirements of patients with amyotrophic lateral sclerosis, particularly the quantity and quality of nutritional support that patients receive after gastrostomy, and to explore the factors that can lead to continuing weight loss after the procedure.

lding the evidence base must be to understand further the nutritional requirements of patients with amyotrophic lateral sclerosis, particularly the quantity and quality of nutritional support that patients receive after gastrostomy, and to explore the factors that can lead to continuing weight loss after the procedure. Procedures Adhering to the study protocol and the National Institute for Health Research guidelines for good clinical practice, data collection was carried out by experienced members of the local research teams. Data were collected at four timepoints: at the time of recruitment (baseline), at the end of the gastrostomy procedure, at 3 months after gastrostomy, and at 12 months after gastrostomy. At baseline, we collected the following information: demographic characteristics; clinician opinion on indication, timing, potential benefits, and preferred type of gastrostomy; patient's influence on the timing of gastrostomy; measures of respiratory function; and indices of disease progression. At baseline, 3 months, and 12 months we collected the following information: demographic characteristics, weight, height, and score on the revised amyotrophic lateral sclerosis functional rating scale (ALSFRS-R).16 Data related to the operation itself such as gastrostomy equipment, type of gastrostomy tube used, procedure length, and details of any complications were collected at the end of the gastrostomy procedure. At baseline and 3 months after gastrostomy insertion, patients who gave consent were asked to complete a questionnaire assessing quality of life (MQOL)17 and a questionnaire assessing the strain of caregiving activities was completed by consenting informal carers (MCSI).18

ected at the end of the gastrostomy procedure. At baseline and 3 months after gastrostomy insertion, patients who gave consent were asked to complete a questionnaire assessing quality of life (MQOL)17 and a questionnaire assessing the strain of caregiving activities was completed by consenting informal carers (MCSI).18 Outcomes The primary outcome of the study was 30-day mortality after gastrostomy.8 The secondary outcomes were perigastrostomy and post-gastrostomy complication rate (defined as complications that occurred during the gastrostomy procedure and those that occurred at any timepoint in the first 3 months after completion of the gastrostomy insertion procedure, respectively), median survival time from gastrostomy placement, nutritional status change, self-perceived quality of life changes after gastrostomy, and changes in carer strain after gastrostomy.

dure and those that occurred at any timepoint in the first 3 months after completion of the gastrostomy insertion procedure, respectively), median survival time from gastrostomy placement, nutritional status change, self-perceived quality of life changes after gastrostomy, and changes in carer strain after gastrostomy. Statistical analysis Assuming a 30-day mortality rate of 5%, based on a meta-analysis of the available literature,8 to estimate mortality within greater or less than 2·5% (ie, 95% CI 2·5–7·5) would require 30-day mortality data for a minimum of 300 patients with amyotrophic lateral sclerosis. Current European Federation of Neurological Societies guidelines recommend gastrostomy after weight loss of at least 10% from premorbid weight.14 This threshold was used in our study to classify patients into weight loss subgroups for subsequent analyses. Continuity corrected χ2 tests were done to determine the difference in the 30-day mortality and the complication rates after gastrostomy in patients who underwent percutaneous endoscopic gastrostomy, radiologically inserted gastrostomy, or per-oral image-guided gastrostomy. Kaplan-Meier survival curves were used to determine the median survival time from placement and disease onset for the treatment groups. Cox proportional hazards regression analysis was done to determine predictors of survival from the time of gastrostomy insertion and from the time of disease onset (to take into account variables that have an effect on survival over the whole course of the disease). Our rationale for inclusion of covariates in the Cox regression analysis was based on well known factors that might affect survival in patients with amyotrophic lateral sclerosis as reported previously, and on our clinical judgment of other probable factors that might affect survival post gastrostomy. Continuity corrected χ2 tests were used to determine changes in nutritional status in the patients who underwent percutaneous endoscopic gastrostomy, radiologically inserted gastrostomy, or per-oral image-guided gastrostomy. Cox proportional hazards regression analysis was also done to examine the effect of nutritional status at 3 months post gastrostomy on subsequent survival. Cronbach's α coefficients were determined for the quality of life and strain measures used in the study to assess their internal consistency.

ge-guided gastrostomy. Cox proportional hazards regression analysis was also done to examine the effect of nutritional status at 3 months post gastrostomy on subsequent survival. Cronbach's α coefficients were determined for the quality of life and strain measures used in the study to assess their internal consistency. A paired samples t test was used to determine differences in the self-perceived quality of life of the patients and the strain of caregiving activities of carers. We obtained complete mortality data for all patients who underwent gastrostomy. Initially, complete case analysis was done—ie, patients who had one or more missing values in the variables being analysed were omitted from the analysis pairwise. To compensate for missing data, post-hoc multiple imputation was done for the covariates of interest in our multiple regression analyses. Because ProGas was not a randomised controlled trial, we addressed the issue of treatment indication bias by undertaking a post-hoc propensity score analysis (appendix p 3). Data were managed and analysed with SPSS Statistics for Windows version 21.0.

A paired samples t test was used to determine differences in the self-perceived quality of life of the patients and the strain of caregiving activities of carers. We obtained complete mortality data for all patients who underwent gastrostomy. Initially, complete case analysis was done—ie, patients who had one or more missing values in the variables being analysed were omitted from the analysis pairwise. To compensate for missing data, post-hoc multiple imputation was done for the covariates of interest in our multiple regression analyses. Because ProGas was not a randomised controlled trial, we addressed the issue of treatment indication bias by undertaking a post-hoc propensity score analysis (appendix p 3). Data were managed and analysed with SPSS Statistics for Windows version 21.0. Role of the funding source This study was supported jointly by the Motor Neurone Disease Association of England, Wales, and Northern Ireland and the Sheffield Institute for Translational Neuroscience. Both funding bodies were consulted regarding the study design, and the decision to submit the report for publication fulfils their requirement for dissemination of the findings. However, the funding sources were not involved in data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all of the data and CJM had final responsibility for the decision to submit the report for publication.

ils their requirement for dissemination of the findings. However, the funding sources were not involved in data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all of the data and CJM had final responsibility for the decision to submit the report for publication. Results Between Nov 2, 2010, and Jan 31, 2014, 330 patients underwent gastrostomy and were included in the analysis for the primary outcome (figure 1). Table 1 shows their baseline characteristics. 163 (49%) patients underwent percutaneous endoscopic gastrostomy, 121 (37%) underwent radiologically inserted gastrostomy, 43 (13%) underwent per-oral image-guided gastrostomy, and three (1%) underwent surgical gastrostomy. Table 2 summarises the differences across the three gastrostomy groups. Data for criteria used for gastrostomy method selection, indication, and predicted benefits on influence of patients on timing of gastrostomy, and on types and sizes of gastrostomy tubes are available on appendix p 1. The study was funded for 38 months and stopped on Jan 31, 2013, at which point all patients had undergone data collection for the primary outcome. Nine patients did not undergo formal 3-month assessments and 93 patients did not undergo 12-month assessments.

Results Between Nov 2, 2010, and Jan 31, 2014, 330 patients underwent gastrostomy and were included in the analysis for the primary outcome (figure 1). Table 1 shows their baseline characteristics. 163 (49%) patients underwent percutaneous endoscopic gastrostomy, 121 (37%) underwent radiologically inserted gastrostomy, 43 (13%) underwent per-oral image-guided gastrostomy, and three (1%) underwent surgical gastrostomy. Table 2 summarises the differences across the three gastrostomy groups. Data for criteria used for gastrostomy method selection, indication, and predicted benefits on influence of patients on timing of gastrostomy, and on types and sizes of gastrostomy tubes are available on appendix p 1. The study was funded for 38 months and stopped on Jan 31, 2013, at which point all patients had undergone data collection for the primary outcome. Nine patients did not undergo formal 3-month assessments and 93 patients did not undergo 12-month assessments. 12 (4%, 95% CI 2–6) of 330 patients died within the first 30 days after gastrostomy: five (3%, 1–7) of 163 after percutaneous endoscopic gastrostomy, four (3%, 1–8) of 121 after radiologically inserted gastrostomy, and three (7%, 2–19) of 43 after per-oral image-guided gastrostomy (p=0·46). We did not find evidence of a difference in 30-day mortality between the procedures after adjustment for case mix variables (age at onset, weight loss, functional decline rate, forced vital capacity, and site of onset) and treatment centre (appendix p 2).

%, 2–19) of 43 after per-oral image-guided gastrostomy (p=0·46). We did not find evidence of a difference in 30-day mortality between the procedures after adjustment for case mix variables (age at onset, weight loss, functional decline rate, forced vital capacity, and site of onset) and treatment centre (appendix p 2). Overall median survival after gastrostomy was 325 days (95% CI 289–361). Median survival time after percutaneous endoscopic gastrostomy was 341 days (25th IQR inderteminate–164), after radiologically inserted gastrostomy was 361 days (25th IQR inderteminate–171), and after per-oral image-guided gastrostomy was 201 days (326–116; figure 2). We noted some evidence of a difference in survival times (log-rank χ2 1·4 on 2 df, p=0·03) between the three gastrostomy insertion methods before any adjustment for case mix variables (age at onset, weight loss, functional decline rate, forced vital capacity, and site of onset) and treatment centre. However, after adjustment for treatment centre and case mix variables, we did not note any evidence of a difference in survival times between the three gastrostomy insertion methods (appendix p 2).

ables (age at onset, weight loss, functional decline rate, forced vital capacity, and site of onset) and treatment centre. However, after adjustment for treatment centre and case mix variables, we did not note any evidence of a difference in survival times between the three gastrostomy insertion methods (appendix p 2). Irrespective of method of gastrostomy, among the 12 patients who died within the first 30 days following the procedure, one patient (8%) had lost up to 10% of their bodyweight compared with that at diagnosis (2·6% loss), eight patients (67%) had lost more than 10% of their weight (mean 17·1% [SD 5·6]), one patient (8%) had gained weight (3·1% gain; χ2, n=252, p=0·031), and for two patients (17%) weight data were missing. Binary logistic regression analysis showed that the odds for 30-day mortality were 10·7 times higher (95% CI 1·3–87·0; p=0·027) for patients who had lost more than 10% of their weight from diagnosis compared with those who had lost 10% or less of weight.

2, p=0·031), and for two patients (17%) weight data were missing. Binary logistic regression analysis showed that the odds for 30-day mortality were 10·7 times higher (95% CI 1·3–87·0; p=0·027) for patients who had lost more than 10% of their weight from diagnosis compared with those who had lost 10% or less of weight. Cox proportional hazards regression was done to ascertain the effect of gastrostomy method on survival from the time of gastrostomy insertion, with adjustment for covariates that might also affect survival. Variables that were inserted into the regression model were gastrostomy insertion method (percutaneous endoscopic gastrostomy, radiologically inserted gastrostomy, and per-oral image-guided gastrostomy subgroups), forced vital capacity at the time of gastrostomy insertion, percentage of weight difference at gastrostomy compared with diagnosis weight, and three additional well established predictors of survival in patients with amyotrophic lateral sclerosis:19 age at the onset of amyotrophic lateral sclerosis, site of amyotrophic lateral sclerosis symptom onset (bulbar and limb subgroups), and monthly rate of decline of the revised amyotrophic lateral sclerosis functional rating scale (ALSFRS-R). The results showed that the hazard of death after gastrostomy insertion was significantly affected by two main factors: the age at onset of amyotrophic lateral sclerosis (hazard ratio [HR] 1·032, 95% CI 1·007–1·059; p=0·013) and the percentage of weight difference at gastrostomy compared with weight at diagnosis (HR 0·956, 0·930–0·983; p=0·001). The hazard of death was not affected by the gastrostomy insertion method. Figure 2 shows the survival functions for the subgroups of patients who underwent percutaneous endoscopic gastrostomy, radiologically inserted gastrostomy, or per-oral image-guided gastrostomy.

eight at diagnosis (HR 0·956, 0·930–0·983; p=0·001). The hazard of death was not affected by the gastrostomy insertion method. Figure 2 shows the survival functions for the subgroups of patients who underwent percutaneous endoscopic gastrostomy, radiologically inserted gastrostomy, or per-oral image-guided gastrostomy. A Cox proportional hazards regression model including the same variables showed that the hazard of death from the time of amyotrophic lateral sclerosis onset was significantly affected by the age at onset (HR 1·045 [95% CI 1·015–1·075]; p=0·003), the ALSFRS-R monthly decline rate (1·768 [1·541–2·028]; p=0·001), and the site of amyotrophic lateral sclerosis symptom onset (bulbar compared with limb subgroups, HR 2·082 [1·204–3·601]; p=0·009).

yotrophic lateral sclerosis onset was significantly affected by the age at onset (HR 1·045 [95% CI 1·015–1·075]; p=0·003), the ALSFRS-R monthly decline rate (1·768 [1·541–2·028]; p=0·001), and the site of amyotrophic lateral sclerosis symptom onset (bulbar compared with limb subgroups, HR 2·082 [1·204–3·601]; p=0·009). To further explore the effect of weight loss on survival after insertion, we did a Cox proportional hazards analysis with adjustment for covariates that might also affect survival. The regression model included as variables weight at the time of gastrostomy compared with weight at diagnosis (<10% weight loss and >10% weight loss subgroups), forced vital capacity at the time of gastrostomy insertion, age at the onset of amyotrophic lateral sclerosis, site of amyotrophic lateral sclerosis symptom onset (bulbar and limb subgroups), and ALSFRS-R monthly decline rate. Hazard of death after gastrostomy insertion was significantly affected by the age at onset (HR 1·035 [95% CI 1·008–1·063]; p=0·011) and the percentage of weight loss from diagnosis to gastrostomy (>10% weight loss subgroup compared with the <10% weight loss subgroup, 2·514 [1·490–4·243]; p=0·001). The median survival after gastrostomy for patients who had lost 10% or less of weight from diagnosis was 12 months (95% CI was indeterminate because survival was greater than 50% at the last timepoint in this subgroup) and for those who had lost more than 10% of their weight from diagnosis was 7·7 months (n=223, 95% CI 6·5–8·9; log-rank test p=0·001). Figure 3 shows the survival functions for the different subgroups of patients in terms of weight loss at gastrostomy compared with weight at the time of diagnosis.

int in this subgroup) and for those who had lost more than 10% of their weight from diagnosis was 7·7 months (n=223, 95% CI 6·5–8·9; log-rank test p=0·001). Figure 3 shows the survival functions for the different subgroups of patients in terms of weight loss at gastrostomy compared with weight at the time of diagnosis. Periprocedural complications did not differ significantly across the three gastrostomy insertion methods, apart from the higher perioperational distress experienced by percutaneous endoscopic gastrostomy patients (table 3). Table 4 summarises complications in the first 3 months after gastrostomy. Patients who received balloon-retention tubes (radiologically inserted gastrostomy) had a significantly higher rate of tube-related complications than did those who received bumper-retention tubes, including displacement (20 [31%] of 96 patients vs one [1%] of 154 patients; p=0·001), leakage (21 [22%] vs 16 [10%]; p=0·011), replacement (29 [30%] vs four [3%]; p=0·001), and repeated gastrostomy (14/96 [15%] vs one [1%]; 0·001; appendix p 1); percutaneous endoscopic gastrostomy and per-oral image-guided gastrostomy odds ratio data are given in appendix p 1).

nts vs one [1%] of 154 patients; p=0·001), leakage (21 [22%] vs 16 [10%]; p=0·011), replacement (29 [30%] vs four [3%]; p=0·001), and repeated gastrostomy (14/96 [15%] vs one [1%]; 0·001; appendix p 1); percutaneous endoscopic gastrostomy and per-oral image-guided gastrostomy odds ratio data are given in appendix p 1). Valid weight measurements at 3 months after gastrostomy were collected for 170 (53%) of 323 patients, owing to attrition and difficulty in obtaining weight measurements from wheelchair-bound patients. After gastrostomy insertion, 43 (25%) of 170 patients gained more than 1 kg compared with weight at gastrostomy, 43 (25%) had loss or gain of 1 kg or less compared with weight at gastrostomy, and 84 (49%) lost more than 1 kg compared to weight at gastrostomy. The method of gastrostomy insertion did not influence the bodyweight post procedure (χ2, n=170; p=0·082). In the 43 patients who gained weight, these gains were small (median weight gain compared with weight at gastrostomy 3·1 kg, IQR 1·8–6·5). Appendix p 6 shows the nutritional outcome for patients in terms of weight at 3 months compared with weight at diagnosis. Continuing weight loss at 3 months after gastrostomy was associated with poor survival (appendix p 2).

gains were small (median weight gain compared with weight at gastrostomy 3·1 kg, IQR 1·8–6·5). Appendix p 6 shows the nutritional outcome for patients in terms of weight at 3 months compared with weight at diagnosis. Continuing weight loss at 3 months after gastrostomy was associated with poor survival (appendix p 2). The differences between patient quality of life at baseline and 3 months after gastrostomy were not statistically significant (mean [SD] total MQOL score 6·3 [1·6] at baseline vs 6·4 [1·6] at 3 months; p=0·749; appendix p 2). However, the strain of caregiving activities had increased significantly for carers at 3 months after gastrostomy (mean [SD] total MCSI score 9·9 [6·4] at baseline vs 11·8 [6·5] at 3 months; p=0·001; appendix p 2). The results of post-hoc multiple imputation and propensity score analyses of the survival endpoints suggested that our findings for 30-day mortality and predictors of survival were robust to both missing data and gastrostomy method preferences in the 24 participating sites (appendix pp 2, 3).

The differences between patient quality of life at baseline and 3 months after gastrostomy were not statistically significant (mean [SD] total MQOL score 6·3 [1·6] at baseline vs 6·4 [1·6] at 3 months; p=0·749; appendix p 2). However, the strain of caregiving activities had increased significantly for carers at 3 months after gastrostomy (mean [SD] total MCSI score 9·9 [6·4] at baseline vs 11·8 [6·5] at 3 months; p=0·001; appendix p 2). The results of post-hoc multiple imputation and propensity score analyses of the survival endpoints suggested that our findings for 30-day mortality and predictors of survival were robust to both missing data and gastrostomy method preferences in the 24 participating sites (appendix pp 2, 3). Discussion In our study, 30-day mortality was similar for percutaneous endoscopic gastrostomy, radiologically inserted gastrostomy, or per-oral image-guided gastrostomy, indicating that the three methods were as safe as each other in relation to procedure risk. The results suggested that weight loss at gastrostomy could affect procedure outcome, although caution with interpretation is necessary because 30-day mortality was low in this cohort. Our data indicate that overall mortality after gastrostomy insertion is independent of the gastrostomy method and is driven by the patient age at the onset of amyotrophic lateral sclerosis and the percentage of weight loss from diagnosis to the timepoint of gastrostomy.

because 30-day mortality was low in this cohort. Our data indicate that overall mortality after gastrostomy insertion is independent of the gastrostomy method and is driven by the patient age at the onset of amyotrophic lateral sclerosis and the percentage of weight loss from diagnosis to the timepoint of gastrostomy. In terms of periprocedural complications, the three different methods of gastrostomy were similar apart from the increased rate of distress, related to procedure tolerance, experienced by patients who underwent percutaneous endoscopic gastrostomy. This finding can be explained by the nature of the percutaneous endoscopic gastrostomy procedure, during which the patient's throat is intubated with an endoscope and the gastrostomy tube is pulled through the mouth towards its placement site.20 Patients who underwent radiologically inserted gastrostomy had a significantly increased rate of gastrostomy tube-related complications. Perhaps this is not surprising, because radiologically inserted gastrostomy tubes are usually relatively narrow in diameter (10–14 Fr), have a balloon-retention system (balloons could burst or deflate, causing gastrostomy tubes to migrate or fall out), and are not as securely fixed as those inserted by percutaneous endoscopic gastrostomy or per-oral image-guided gastrostomy.

serted gastrostomy tubes are usually relatively narrow in diameter (10–14 Fr), have a balloon-retention system (balloons could burst or deflate, causing gastrostomy tubes to migrate or fall out), and are not as securely fixed as those inserted by percutaneous endoscopic gastrostomy or per-oral image-guided gastrostomy. In terms of the nutritional outcome for the patient, gastrostomy feeding prevented further weight loss in only about half of the patients. In the 43 (25%) patients who gained weight, these gains were small and of doubtful clinical benefit. Continuing weight loss at 3 months after gastrostomy was associated with poor survival. The nutritional data suggested that the greater the percentage of weight loss at the time of gastrostomy from diagnosis, the less likely it was for patients to recover this loss after gastrostomy. This finding was more evident for patients who at the time of gastrostomy had had more than 10% loss of their diagnosis weight; this subgroup of patients had also a significantly shorter survival compared with those who had lost up to 10% of their diagnosis weight. These results suggest that patients might benefit from early gastrostomy, before substantial weight loss that might not be reversible.

d more than 10% loss of their diagnosis weight; this subgroup of patients had also a significantly shorter survival compared with those who had lost up to 10% of their diagnosis weight. These results suggest that patients might benefit from early gastrostomy, before substantial weight loss that might not be reversible. The reasons for the fairly poor nutritional outcome that we noted need further investigation. Perhaps weight loss due to continued denervation-induced skeletal muscle atrophy is masking nutritional benefits,21,22 which could be related to the change in metabolic state. Patients with amyotrophic lateral sclerosis can present hypermetabolism, and the caloric requirements of patients after gastrostomy might have been underestimated such that their energy intake was lower than energy expenditure.23 A small phase 2 study showed a potential benefit in terms of survival and nutritional gains for patients fed high calorific diets through a percutaneous endoscopic gastrostomy tube.24 Therefore, further study and subsequent evidence-based guidance on nutritional management post gastrostomy tube insertion is needed. A further potential metabolic explanation for our findings is related to the concept of refractory cachexia. The body of a patient with cachexia (defined as weight loss of more than 5%) is recognised to undergo irreparable metabolic changes, making artificial nutritional support ineffective.25 This idea is well recognised in oncology but has not been explored in patients with amyotrophic lateral sclerosis.

efractory cachexia. The body of a patient with cachexia (defined as weight loss of more than 5%) is recognised to undergo irreparable metabolic changes, making artificial nutritional support ineffective.25 This idea is well recognised in oncology but has not been explored in patients with amyotrophic lateral sclerosis. The effect of gastrostomy on the quality of life of patients in our study seemed to be neutral. Conversely, the strain of caregiving activities increased significantly after gastrostomy, although this was independent of insertion method. However, consequences of amyotrophic lateral sclerosis including increasing motor disability and dependency might contribute to caregiver strain. These results highlight the importance of provision of information and support from health-care professionals to carers, as well as to patients, before and after gastrostomy.

However, consequences of amyotrophic lateral sclerosis including increasing motor disability and dependency might contribute to caregiver strain. These results highlight the importance of provision of information and support from health-care professionals to carers, as well as to patients, before and after gastrostomy. This study has limitations. This study was not a randomised controlled trial and the assignment of patients to a specific gastrostomy method was not done at random, but based on practical and clinical considerations. Therefore we can make associations but we are limited in the ability to draw conclusions with regard to the direct effects of gastrostomy on survival and nutritional outcome compared with not having had a gastrostomy. Another limitation is that, of 484 patients who had been referred for a gastrostomy in the 24 participating centres, we recruited 345 patients (participation rate 71%). Patient refusal and several logistical issues hindered full recruitment. Unfortunately, we could not obtain meaningful information for the potential participants who were not recruited to our study because we did not have the consent of these patients to do so, and we could not compare their characteristics with those of patients in this study. Practical difficulties in obtaining weight measurements at 3 months after gastrostomy introduced another limitation. The prospective element of this study allowed us to follow up a large number of patients after gastrostomy insertion and to consistently collect data related to the predetermined primary and secondary outcomes. A major strength of this study is that our sample is representative of the wider amyotrophic lateral sclerosis population: the baseline characteristics of the patients who took part are very similar to those of other reported cohorts of patients with amyotrophic lateral sclerosis.19,26

d primary and secondary outcomes. A major strength of this study is that our sample is representative of the wider amyotrophic lateral sclerosis population: the baseline characteristics of the patients who took part are very similar to those of other reported cohorts of patients with amyotrophic lateral sclerosis.19,26 We noted significantly worse respiratory impairment in the per-oral image-guided gastrostomy group. Despite this, 30-day mortality was similar to the other groups. This observation would suggest that percutaneous endoscopic gastrostomy might be the optimum method of gastrostomy when respiratory function is largely unimpaired and per-oral image-guided gastrostomy when respiratory function is significantly compromised. Both percutaneous endoscopic gastrostomy and per-oral image-guided gastrostomy seemed to offer easier post-insertion tube management than radiologically inserted gastrostomy; ease of management is crucial, especially in very frail patients who undergo gastrostomy late, when they are more likely to feel the burden of other consequences of amyotrophic lateral sclerosis, such as respiratory problems and the loss of mobility and speech.

tube management than radiologically inserted gastrostomy; ease of management is crucial, especially in very frail patients who undergo gastrostomy late, when they are more likely to feel the burden of other consequences of amyotrophic lateral sclerosis, such as respiratory problems and the loss of mobility and speech. Our study showed that delay might lead to diminishing gains, especially for patients who at the time of gastrostomy have experienced excessive weight loss from their diagnosis weight. From a safety and efficacy perspective, the current guidelines of 10% weight loss might not be ideal and perhaps a better threshold would be to recommend gastrostomy insertion at a threshold similar to the one for cachexia—ie, at roughly 5% weight loss. Another recently suggested approach is to consider gastrostomy based on the ability of an individual to meet their total daily energy requirements.27 Delay of gastrostomy until after weight loss of more than 10% might convey minimal clinically meaningful benefit. However, some patients will not wish to undergo early gastrostomy. For such patients, gastrostomy will still have a role alleviating the difficulties caused by advanced dysphagia—eg, to allow administration of drugs and hydration—but in view of the possible diminishing nutritional benefits of delayed gastrostomy, other options of palliative support should also be considered. Correspondence to: Dr Christopher J McDermott, Sheffield Institute for Translational Neuroscience, The University of Sheffield, 385A Glossop Road, Sheffield, S10 2HQ, UK c.j.mcdermott@sheffield.ac.uk

Our study showed that delay might lead to diminishing gains, especially for patients who at the time of gastrostomy have experienced excessive weight loss from their diagnosis weight. From a safety and efficacy perspective, the current guidelines of 10% weight loss might not be ideal and perhaps a better threshold would be to recommend gastrostomy insertion at a threshold similar to the one for cachexia—ie, at roughly 5% weight loss. Another recently suggested approach is to consider gastrostomy based on the ability of an individual to meet their total daily energy requirements.27 Delay of gastrostomy until after weight loss of more than 10% might convey minimal clinically meaningful benefit. However, some patients will not wish to undergo early gastrostomy. For such patients, gastrostomy will still have a role alleviating the difficulties caused by advanced dysphagia—eg, to allow administration of drugs and hydration—but in view of the possible diminishing nutritional benefits of delayed gastrostomy, other options of palliative support should also be considered. Correspondence to: Dr Christopher J McDermott, Sheffield Institute for Translational Neuroscience, The University of Sheffield, 385A Glossop Road, Sheffield, S10 2HQ, UK c.j.mcdermott@sheffield.ac.uk Supplementary Material Supplementary appendix

Our study showed that delay might lead to diminishing gains, especially for patients who at the time of gastrostomy have experienced excessive weight loss from their diagnosis weight. From a safety and efficacy perspective, the current guidelines of 10% weight loss might not be ideal and perhaps a better threshold would be to recommend gastrostomy insertion at a threshold similar to the one for cachexia—ie, at roughly 5% weight loss. Another recently suggested approach is to consider gastrostomy based on the ability of an individual to meet their total daily energy requirements.27 Delay of gastrostomy until after weight loss of more than 10% might convey minimal clinically meaningful benefit. However, some patients will not wish to undergo early gastrostomy. For such patients, gastrostomy will still have a role alleviating the difficulties caused by advanced dysphagia—eg, to allow administration of drugs and hydration—but in view of the possible diminishing nutritional benefits of delayed gastrostomy, other options of palliative support should also be considered. Correspondence to: Dr Christopher J McDermott, Sheffield Institute for Translational Neuroscience, The University of Sheffield, 385A Glossop Road, Sheffield, S10 2HQ, UK c.j.mcdermott@sheffield.ac.uk Supplementary Material Supplementary appendix Acknowledgments ProGas was funded by the Motor Neurone Disease Association of England, Wales, and Northern Ireland and the Sheffield Institute for Translational Neuroscience. We are very grateful to the patients and carers who participated in this study and to our funders for making this research possible.

Supplementary Material Supplementary appendix Acknowledgments ProGas was funded by the Motor Neurone Disease Association of England, Wales, and Northern Ireland and the Sheffield Institute for Translational Neuroscience. We are very grateful to the patients and carers who participated in this study and to our funders for making this research possible. Contributors CJM was the chief investigator and study manager, helped develop the protocol and all study material, assessed patient eligibility, helped to recruit participants, participated in data collection at the principal site, (the Sheffield MND Care and Research Centre for Motor Neurone Disorders) advised on the conduct of the study, participated in data analysis, helped to interpret the results, and revised the manuscript. PJS was the principal investigator at the lead site (the Sheffield MND Care and Research Centre for Motor Neurone Disorders), helped to develop the protocol and all study material, assessed patient eligibility, helped to recruit participants, participated in data collection at the principal site, advised on the conduct of the study, participated in data analysis, helped to interpret the results, and revised the manuscript. TS was a coinvestigator and the study coordinator, helped develop the protocol and all study material, helped to recruit participants, participated in data collection at the principal site, facilitated recruitment and data collection from all other sites, created and updated the study database, conducted the data analysis, helped to interpret the results, and drafted and revised the manuscript. SJW was the statistician who advised on statistical matters, helped to conduct the data analysis, and revised the manuscript. AA-C, SC, FC, DD, CD, PE, MF, CG, GG, HH, COH, MJ, TM, AM, KM, RO, AP, AR, MR, KT, MRT, TW, CY were the principal investigators in the other ProGas sites, helped to develop the protocol, assess patient eligibility, recruit participants, collect data at their sites, advise on interpreting the results, and revise the manuscript.

, FC, DD, CD, PE, MF, CG, GG, HH, COH, MJ, TM, AM, KM, RO, AP, AR, MR, KT, MRT, TW, CY were the principal investigators in the other ProGas sites, helped to develop the protocol, assess patient eligibility, recruit participants, collect data at their sites, advise on interpreting the results, and revise the manuscript. ProGas Writing Committee Christopher J McDermott, Pamela J Shaw, Theocharis Stavroulakis, Stephen J Walters, Ammar Al-Chalabi, Siddharthan Chandran, Francesca Crawley, David Dick, Colette Donaghy, Penelope Eames, Mark Fish, Carol Gent, George Gorrie, Hisham Hamdalla, C Oliver Hanemann, Michael Johnson, Tahir Majeed, Andrea Malaspina, Karen Morrison, Richard Orrell, Ashwin Pinto, Aleksandar Radunovic, Mark Roberts, Kevin Talbot, Martin R Turner, Timothy Williams, Carolyn Young. Declaration of interests CJM, PJS, AA-C, KT, and MRT are supported by the EU Joint Programme—Neurodegenerative Disease Research (JPND), UK Medical Research Council, and Economic and Social Research Council. PJS is supported as a National Institute for Health Research (NIHR) Senior Investigator. AA-C receives salary support from the NIHR Dementia Biomedical Research Unit at South London and Maudsley NHS Foundation Trust and King's College London and from the European Community's Health Seventh Framework Programme (FP7/2007–2013; grant agreement number 259867). The other authors declare no competing interests. Figure 1 Study profile Figure 2 Survival functions for patients who underwent PEG, RIG, or PIG

Declaration of interests CJM, PJS, AA-C, KT, and MRT are supported by the EU Joint Programme—Neurodegenerative Disease Research (JPND), UK Medical Research Council, and Economic and Social Research Council. PJS is supported as a National Institute for Health Research (NIHR) Senior Investigator. AA-C receives salary support from the NIHR Dementia Biomedical Research Unit at South London and Maudsley NHS Foundation Trust and King's College London and from the European Community's Health Seventh Framework Programme (FP7/2007–2013; grant agreement number 259867). The other authors declare no competing interests. Figure 1 Study profile Figure 2 Survival functions for patients who underwent PEG, RIG, or PIG Subsequent Cox proportional hazards analysis suggested that the method of gastrostomy insertion was not significantly associated with survival. PEG=percutaneous endoscopic gastrostomy. RIG=radiologically inserted gastrostomy. PIG=per-oral image-guided gastrostomy. Figure 3 Survival functions according to weight loss Table 1 Baseline demographic and clinical characteristics

Subsequent Cox proportional hazards analysis suggested that the method of gastrostomy insertion was not significantly associated with survival. PEG=percutaneous endoscopic gastrostomy. RIG=radiologically inserted gastrostomy. PIG=per-oral image-guided gastrostomy. Figure 3 Survival functions according to weight loss Table 1 Baseline demographic and clinical characteristics Baseline value in all patients (n=330) Age (years) 64·4 (11·7), n=315 Sex Women 150/330 (45%) Men 180/330 (55%) Forced vital capacity (%) 62% (22·6), n=258 % Weight loss from diagnosis to baseline (kg) 8·6 (9·8), n=252 ALSFRS-R score 28 (8·5), n=307 Body-mass index (kg/m2) 23·3 (4·4), n=274 Monthly ALSFRS-R decline 2·2% (1·7%), n=290 Disease duration from diagnosis (months) 16·7 (5·8–14·9), n=309 Non-invasive ventilation routine users 81/323 (25%) Site of disease onset Limb 152/324 (47%) Bulbar 165/324 (51%) Both limb and bulbar 6/324 (2%) Respiratory 1/324 (0·3%) Data are mean (SD), median (IQR), or n/N (%). ALSFRS-R=amyotrophic lateral sclerosis functional rating scale revised. Table 2 Baseline differences of patients who underwent PEG, RIG, or PIG

Baseline value in all patients (n=330) Age (years) 64·4 (11·7), n=315 Sex Women 150/330 (45%) Men 180/330 (55%) Forced vital capacity (%) 62% (22·6), n=258 % Weight loss from diagnosis to baseline (kg) 8·6 (9·8), n=252 ALSFRS-R score 28 (8·5), n=307 Body-mass index (kg/m2) 23·3 (4·4), n=274 Monthly ALSFRS-R decline 2·2% (1·7%), n=290 Disease duration from diagnosis (months) 16·7 (5·8–14·9), n=309 Non-invasive ventilation routine users 81/323 (25%) Site of disease onset Limb 152/324 (47%) Bulbar 165/324 (51%) Both limb and bulbar 6/324 (2%) Respiratory 1/324 (0·3%) Data are mean (SD), median (IQR), or n/N (%). ALSFRS-R=amyotrophic lateral sclerosis functional rating scale revised. Table 2 Baseline differences of patients who underwent PEG, RIG, or PIG PEG RIG PIG Statistic (n) p value Age (years) 64·2 (11·7), n=157 63·6 (9·8), n=114 67·2 (12·6), n=41 F (312) 0·200 Sex χ2 (327) 0·683 Women 73/163 (45%) 59/121 (49%) 18/43 (42%) Men 90/163 (55%) 62/121 (51%) 25/43 (58%) FVC (%) 65·4 (22·2), n=136 59 (23·1), n=87 52 (19·7), n=33 F (256) 0·004 % Weight loss from diagnosis to baseline (kg) 7·1 (8·5), n=117 8·7 (9·9), n=98 13 (12·3), n=35 F (250) 0·008 ALSFRS-R score 29·1 (8·2), n=152 27·7 (8·8), n=114 24·7 (7·9), n=39 F (305) 0·014 Body-mass index (kg/m2) 23·7 (4), n=135 23·4 (5·1), n=102 21·8 (3), n=34 F (271) 0·091 Monthly ALSFRS-R decline 2·1% (1·5), n=144 2·4% (2·1), n=105 2·1% (1·2), n=39 F (288) 0·302 NIV routine users 29/162 (18%) 23/118 (19%) 28/42 (67%) χ2 (322) 0·001 Site of disease onset χ2 (321) 0·369 Limb 74/161 (46%) 54/117 (46%) 23/43 (53%) Bulbar 86/161 (53%) 59/117 (50%) 18/43 (42%) Both limb and bulbar 1/161 (1%) 3/117 (3%) 2/43 (5%) Respiratory 0/161 1/117 (1%) 0/43 Data are mean (SD) or n/N (%). PEG=percutaneous endoscopic gastrostomy. RIG=radiologically inserted gastrostomy. PIG=per-oral image-guided gastrostomy. F=one-way ANOVA F test. FVC=forced vital capacity. ALSFRS-R=amyotrophic lateral sclerosis functional rating scale revised. NIV=non-invasive ventilation.

1/117 (1%) 0/43 Data are mean (SD) or n/N (%). PEG=percutaneous endoscopic gastrostomy. RIG=radiologically inserted gastrostomy. PIG=per-oral image-guided gastrostomy. F=one-way ANOVA F test. FVC=forced vital capacity. ALSFRS-R=amyotrophic lateral sclerosis functional rating scale revised. NIV=non-invasive ventilation. Table 3 Periprocedural complications PEG RIG PIG Total p value Overall complication rate 41/169 (24%) 20/121 (17%) 8/42 (19%) 69/332 (21%) 0·266 Difficult procedure 26/168 (16%) 13/119 (11%) 7/42 (17%) 46/329 (14%) 0·475 Failed attempt 11/171 (6%) 7/125 (6%) 3/45 (7%) 21/341 (6%) 0·947 O2desaturation 6/166 (4%) 2/117 (2%) 3/42 (7%) 11/325 (3%) 0·241 Patient distress 26/166 (16%) 4/117 (3%) 2/42 (5%) 32/325 (10%) 0·002 Respiratory arrest 0/166 0/117 0/42 0/325 NA Laryngeal spasm 2/166 (1%) 1/117 (1%) 0/42 3/325 (1%) 0·763 Haemorrhage 0/166 3/117 (3%) 0/42 3/325 (1%) 0·068 Numbers are patients who experienced each event (n)/total patients in each group (N). The periprocedural period is the time during the gastrostomy procedure. PEG=percutaneous endoscopic gastrostomy. RIG=radiologically inserted gastrostomy. PIG=per-oral image-guided gastrostomy. NA=not applicable. Table 4 Postprocedural complications

PEG RIG PIG Total p value Overall complication rate 41/169 (24%) 20/121 (17%) 8/42 (19%) 69/332 (21%) 0·266 Difficult procedure 26/168 (16%) 13/119 (11%) 7/42 (17%) 46/329 (14%) 0·475 Failed attempt 11/171 (6%) 7/125 (6%) 3/45 (7%) 21/341 (6%) 0·947 O2desaturation 6/166 (4%) 2/117 (2%) 3/42 (7%) 11/325 (3%) 0·241 Patient distress 26/166 (16%) 4/117 (3%) 2/42 (5%) 32/325 (10%) 0·002 Respiratory arrest 0/166 0/117 0/42 0/325 NA Laryngeal spasm 2/166 (1%) 1/117 (1%) 0/42 3/325 (1%) 0·763 Haemorrhage 0/166 3/117 (3%) 0/42 3/325 (1%) 0·068 Numbers are patients who experienced each event (n)/total patients in each group (N). The periprocedural period is the time during the gastrostomy procedure. PEG=percutaneous endoscopic gastrostomy. RIG=radiologically inserted gastrostomy. PIG=per-oral image-guided gastrostomy. NA=not applicable. Table 4 Postprocedural complications PEG RIG PIG Total p value Infection 20/129 (16%) 21/96 (22%) 3/25 (12%) 44/250 (18%) 0·745 Granulation tissue 15/129 (12%) 19/96 (20%) 3/25 (12%) 37/250 (15%) 0·214 Pain 25/129 (19%) 34/96 (35%) 10/25 (40%) 69/250 (28%) 0·010 Anxiety 10/129 (8%) 24/96 (25%) 1/25 (4%) 35/250 (14%) 0·001 Nausea 12/129 (9%) 10/96 (10%) 2/25 (8%) 24/250 (10%) 0·923 Diarrhoea 6/129 (5%) 10/96 (10%) 3/25 (12%) 19/250 (8%) 0·185 Pneumonia 4/129 (3%) 4/96 (4%) 4/25 (16%) 12/250 (5%) 0·021 Constipation 16/129 (13%) 22/96 (24%) 9/25 (36%) 47/250 (19%) 0·010 Fatigue 15/129 (12%) 23/96 (24%) 4/25 (16%) 42/250 (17%) 0·050 Numbers are patients who experienced each event (n)/total patients in each group (N). The postprocedural period is the first 3 months after the completion of the gastrostomy procedure. PEG=percutaneous endoscopic gastrostomy. RIG=radiologically inserted gastrostomy. PIG=per-oral image-guided gastrostomy.

Introduction Treatment of neonatal encephalopathy with moderate hypothermia is now standard care in several countries.1 Cooling to 33–34°C after birth asphyxia increases survival without impairments in childhood by about 15%, but roughly 25% of treated infants with moderate or severe encephalopathy die and 20% of survivors develop sensorimotor and cognitive impairments. Additional therapies might further improve outcomes in these infants. Several neuroprotectants are effective in experimental studies; the challenge is to find the most promising candidate treatment to take forward to clinical trials. In a comparative review of potential neuroprotectants,2 inhaled xenon was rated highly but, because of the need for specialist equipment and training, there were concerns about cost and ease of administration.

l studies; the challenge is to find the most promising candidate treatment to take forward to clinical trials. In a comparative review of potential neuroprotectants,2 inhaled xenon was rated highly but, because of the need for specialist equipment and training, there were concerns about cost and ease of administration. Xenon is a monoatomic gas that rapidly crosses the blood–brain barrier. It is an approved inhalational anaesthetic at a minimum alveolar concentration of 60–70% in adults and is not associated with adverse cardiovascular effects, or anaesthetic-associated neurotoxic effects.3 Broad interest in xenon as a potential neuroprotectant is based on strong experimental evidence, but the drug is difficult to use in clinical practice. Xenon might provide neuroprotection after asphyxia by different mechanisms. It is an inhibitor of NMDA glutamate receptors and so could reduce neuronal injury caused by excessive glutamate concentrations and lessen seizures, and it reduces apoptosis by activation of anti-apoptotic factors.4, 5, 6 Xenon reduced cerebral injury in models of hypoxic ischaemic injury in different animal species and the neuroprotective effect was stronger when xenon was used in combination with cooling.7 The combination of 20% xenon with cooling to 35°C provided synergistic neuroprotection in both in-vitro and in-vivo models, with improvement in function for 30 days, but neither intervention alone was effective.8 Research in context Evidence before this study

Xenon is a monoatomic gas that rapidly crosses the blood–brain barrier. It is an approved inhalational anaesthetic at a minimum alveolar concentration of 60–70% in adults and is not associated with adverse cardiovascular effects, or anaesthetic-associated neurotoxic effects.3 Broad interest in xenon as a potential neuroprotectant is based on strong experimental evidence, but the drug is difficult to use in clinical practice. Xenon might provide neuroprotection after asphyxia by different mechanisms. It is an inhibitor of NMDA glutamate receptors and so could reduce neuronal injury caused by excessive glutamate concentrations and lessen seizures, and it reduces apoptosis by activation of anti-apoptotic factors.4, 5, 6 Xenon reduced cerebral injury in models of hypoxic ischaemic injury in different animal species and the neuroprotective effect was stronger when xenon was used in combination with cooling.7 The combination of 20% xenon with cooling to 35°C provided synergistic neuroprotection in both in-vitro and in-vivo models, with improvement in function for 30 days, but neither intervention alone was effective.8 Research in context Evidence before this study We searched PubMed with the terms “xenon neuroprotection” and “xenon hypothermia” for all articles published in English until July 23, 2015. Our search returned several preclinical studies, in which the neuroprotective effects of xenon after asphyxia were shown; these effects were enhanced when xenon was used in combination with hypothermia. Investigators of clinical studies had reported the feasibility of treatment with xenon in combination with hypothermia for neuroprotection in neonates after birth asphyxia and in adults after cardiac arrest, but we found no reports of neuroprotective effects associated with this therapy.

n combination with hypothermia. Investigators of clinical studies had reported the feasibility of treatment with xenon in combination with hypothermia for neuroprotection in neonates after birth asphyxia and in adults after cardiac arrest, but we found no reports of neuroprotective effects associated with this therapy. Added value of this study Our trial is the first randomised clinical study of the neuroprotective effects of xenon in combination with hypothermia after birth asphyxia. The treatment regimen that we used is generally applicable in high-resource settings, and we assessed it with qualified cerebral magnetic resonance endpoints. Our proof-of-concept study showed that, in the complex situation of neonatal care, delayed intervention with xenon beyond 6 h after birth does not have additional neuroprotective effects compared with induction of hypothermia alone after birth asphyxia. Implications of all available evidence Strong experimental evidence supports the use of xenon as a neuroprotectant, but treatment with 30% xenon for 24 h begun more than 6 h after birth combined with early hypothermia is unlikely to improve clinical outcomes compared with hypothermia alone after birth asphyxia. Qualified magnetic resonance biomarkers offer the potential to speed up the assessment of promising neuroprotective treatments before a large pragmatic trial, would substantially reduce opportunity costs, and could lead to redirection of future research.

al outcomes compared with hypothermia alone after birth asphyxia. Qualified magnetic resonance biomarkers offer the potential to speed up the assessment of promising neuroprotective treatments before a large pragmatic trial, would substantially reduce opportunity costs, and could lead to redirection of future research. In adults with stroke, many neuroprotectants that were effective in preclinical studies were not associated with any benefits in large randomised trials.9 In view of the neuroprotective effect of cooling after birth asphyxia, a study of about 750 infants would be needed to detect a further 10% improvement in neurological outcomes with additional or modified therapy.10 The substantial financial and opportunity costs associated with large pragmatic clinical trials that yield negative results can be avoided by first assessing candidate treatments in small proof-of-concept trials, in which qualified biomarkers and surrogate endpoints are used to test efficacy in the clinical context. Treatments that show promise at this stage are candidates for large definitive trials with clinical endpoints.

egative results can be avoided by first assessing candidate treatments in small proof-of-concept trials, in which qualified biomarkers and surrogate endpoints are used to test efficacy in the clinical context. Treatments that show promise at this stage are candidates for large definitive trials with clinical endpoints. In neonates after hypoxic ischaemic injury, the ratio of cerebral lactate to N-acetyl aspartate (assessed with magnetic resonance spectroscopy [MRS]) and fractional anisotropy (FA), a measure of tissue integrity in white matter tracts measured by diffusion tensor MRI, have been used in work in animals to assess potential neuroprotectants and can be used to predict subsequent neurological outcomes after birth asphyxia, including in infants treated with moderate hypothermia.11, 12, 13, 14 To assess whether the combination of cooling with inhaled xenon—administered at a concentration and within a timeframe suitable for general clinical application—could further improve neurological outcomes after birth asphyxia and neonatal encephalopathy, we compared the effects of combined therapy with cooling alone on the lactate to N-acetyl aspartate ratio in the thalamus and FA in white matter tracts within 15 days of birth.

meframe suitable for general clinical application—could further improve neurological outcomes after birth asphyxia and neonatal encephalopathy, we compared the effects of combined therapy with cooling alone on the lactate to N-acetyl aspartate ratio in the thalamus and FA in white matter tracts within 15 days of birth. Methods Study design and participants Total Body hypothermia plus Xenon (TOBY-Xe) was a proof-of-concept, pragmatic, open-label, parallel-group randomised controlled trial at four UK neonatal intensive-care units in London (University College Hospital, St Thomas' Hospital, Queen Charlotte and Chelsea Hospital) and Liverpool (Liverpool Women's Hospital). The National Perinatal Epidemiology Unit (University of Oxford, Oxford, UK), was the coordinating centre for the trial, and managed the study.

r UK neonatal intensive-care units in London (University College Hospital, St Thomas' Hospital, Queen Charlotte and Chelsea Hospital) and Liverpool (Liverpool Women's Hospital). The National Perinatal Epidemiology Unit (University of Oxford, Oxford, UK), was the coordinating centre for the trial, and managed the study. Infants were eligible if their gestational age was 36–43 weeks and they had at least one of the following: Apgar score of 5 or less 10 min after birth; continued need for resuscitation, including endotracheal or mask ventilation, 10 min after birth; or acidosis (defined as pH <7 or base deficit >15 mmol/L, or both, in umbilical cord blood or any blood sample) within 1 h of birth. Furthermore, eligible infants showed signs of moderate to severe encephalopathy, consisting of altered state of consciousness (reduced or absent response to stimulation), hypotonia or severe hypotonia, and abnormal primitive reflexes (weak or absent suck or Moro response), and had moderately or severely abnormal background activity for at least 30 min or seizures as shown by amplitude-integrated EEG (aEEG).15 We excluded infants if cooling was started after age 6 h or if they were older than 12 h at randomisation. We also excluded infants with an oxygen requirement of greater than 60%, those who needed nitric oxide inhalation or ventilation with a high-frequency oscillator, those who needed extracorporeal membrane oxygen, and those with major congenital abnormalities.

ter age 6 h or if they were older than 12 h at randomisation. We also excluded infants with an oxygen requirement of greater than 60%, those who needed nitric oxide inhalation or ventilation with a high-frequency oscillator, those who needed extracorporeal membrane oxygen, and those with major congenital abnormalities. Infants were recruited and assessed on admission by study personnel and cared for in participating centres. Infants born at hospitals that refer patients to the participating neonatal intensive-care units were also eligible for inclusion. Transport teams provided written information about the study for the parents of infants from referring hospitals, and study personnel assessed the infants on admission to the participating centre. The trial was approved by the UK National Research Ethics Service (approval number 10/H0707/33). Parents provided written informed consent; if neither parent was available, consent was first obtained by telephone and then written consent was obtained at the earliest opportunity. Consent was reaffirmed within 24 h of receiving written consent.

y the UK National Research Ethics Service (approval number 10/H0707/33). Parents provided written informed consent; if neither parent was available, consent was first obtained by telephone and then written consent was obtained at the earliest opportunity. Consent was reaffirmed within 24 h of receiving written consent. Randomisation and masking Eligible infants were randomly assigned (1:1) to cooling plus inhaled xenon or cooling only. Assignment to a treatment group was overseen by the National Perinatal Epidemiology Unit, and was done through a secure web-based system with a computer-generated randomisation sequence, with telephone back-up. Minimisation was used to ensure balance of treatment assignment among infants with moderate or severe grades of abnormality on aEEG and within each participating centre. Masking of investigators and parents to allocation was not practical because of the need for a special ventilator to administer xenon, and thus the trial was open label. However, investigators who assessed the primary outcome measures—ie, NT who assessed MRI data and AB who assessed MRS data—were masked to treatment allocation.

f investigators and parents to allocation was not practical because of the need for a special ventilator to administer xenon, and thus the trial was open label. However, investigators who assessed the primary outcome measures—ie, NT who assessed MRI data and AB who assessed MRS data—were masked to treatment allocation. Procedures We used servo-controlled equipment to cool all infants to a target rectal temperature of 33·5°C for 72 h starting within 6 h of birth. If cooling equipment was not available at the referring hospital, passive cooling was commenced and active cooling was started by the transport team and continued during transport to the treatment centre. Infants in the inhaled xenon group also received 30% xenon (Lenoxe, Air Liquide, Paris, France) through an uncuffed endotracheal tube connected to a recirculating device developed for the trial (SLE, Croydon, UK).16 The system provided automated control of xenon, air, and oxygen mixture and continuous monitoring of xenon, oxygen, and carbon dioxide concentrations in inhaled gas. Xenon was commenced immediately after randomisation and continued for 24 h. After xenon administration ended, the infant was ventilated with a standard ventilator according to the unit's practice.

ined diagnostic criteria or outcome events,8 and in which several different statistical methods were used to calculate the risk and predictors of ICH. Findings from these studies have left uncertainty about the magnitude of the risk of ICH and its predictors. Additionally, there remains an absence of prediction models. Therefore, we sought to address these uncertainties by undertaking a systematic review and meta-analysis of individual patient data from cohort studies of people with CCMs with similar designs and outcome definitions.12, 13 By using consistent methods of analysis, we aimed to estimate the risks of first ICH or FND during follow-up and to identify predictors of these outcomes.

enon, air, and oxygen mixture and continuous monitoring of xenon, oxygen, and carbon dioxide concentrations in inhaled gas. Xenon was commenced immediately after randomisation and continued for 24 h. After xenon administration ended, the infant was ventilated with a standard ventilator according to the unit's practice. All MRS and MRI studies were done with 3·0 Tesla systems (Philips Healthcare, Best, Netherlands) at each centre. The trial research physicist (GF) undertook rigorous standardisation and a quality-control programme with phantoms and repeated scanning. Comparability test objects were transported from site to site during the project. An adult volunteer was imaged at each site periodically to provide direct comparison data. Images were obtained according to a standard protocol that included T1-weighted and T2-weighted diffusion tensor MRI with 32 non-collinear directions, MRS from a single voxel on the left thalamus, and motion tolerant T1 and T2 structural scans. We used a study-specific SmartExam card (Philips Healthcare) to aid the planning of the various sequences (appendix). Total examination time for the study protocol was around 1 h. The SENSE 8 channel head coil (Philips Healthcare) was used for all infants.

left thalamus, and motion tolerant T1 and T2 structural scans. We used a study-specific SmartExam card (Philips Healthcare) to aid the planning of the various sequences (appendix). Total examination time for the study protocol was around 1 h. The SENSE 8 channel head coil (Philips Healthcare) was used for all infants. Outcomes The primary outcomes were reduction in lactate to N-acetyl aspartate ratio in the thalamus on MRS or preserved FA in the posterior limb of the internal capsule on diffusion tensor MRI, as assessed by tract-based spatial statistics, an automated observer-independent method of aligning FA images from several patients to allow group-wise comparisons of diffusion tensor imaging data free from partial volume effects.17, 18

preserved FA in the posterior limb of the internal capsule on diffusion tensor MRI, as assessed by tract-based spatial statistics, an automated observer-independent method of aligning FA images from several patients to allow group-wise comparisons of diffusion tensor imaging data free from partial volume effects.17, 18 Secondary outcomes assessed before discharge from hospital were maximum Thompson hypoxic ischaemic encephalopathy score (range 0–22, with higher scores corresponding to worse encephalopathy);19 neurological examination at discharge from treatment centre (which were done by experienced nominated physicians at each treatment centre);20 occurrence of seizures; intracranial haemorrhage; persistent hypotension; pulmonary haemorrhage; pulmonary hypertension; prolonged blood coagulation time (activated partial thromboplastin time >41 s or international normalised ratio >3); thrombocytopenia (platelet count <150 × 109 per L); major venous thrombosis; cardiac arrhythmia (heart rate <80 beats per min); culture-proven late-onset sepsis; necrotising enterocolitis; pneumonia; pulmonary air leak; anuria or urine output of less than 0·5 mL/kg/h for 24 h; age at which full oral feeding was achieved; and duration of hospital stay. We also measured the grade of abnormalities on visual analysis of MRI (scored 0–11, with higher scores corresponding to worse abnormalities).21 We did not compare aEEG results between groups after randomisation as initially planned, because during the study we noted that xenon treatment suppressed the reading, hindering a comparative analysis.6

of abnormalities on visual analysis of MRI (scored 0–11, with higher scores corresponding to worse abnormalities).21 We did not compare aEEG results between groups after randomisation as initially planned, because during the study we noted that xenon treatment suppressed the reading, hindering a comparative analysis.6 Non-serious adverse events and reactions were reported on the data collection forms, but adverse events commonly associated with neonatal encephalopathy were not recorded. Serious adverse events that we monitored were deaths, hypertension (mean blood pressure >85 mmHg), hypotension (mean blood pressure <25 mmHg), cardiac arrhythmia (severe bradycardia [heart rate <60 beats per min] or ventricular arrhythmia), and inability to achieve adequate ventilation despite appropriate adjustment of ventilator settings.

s that we monitored were deaths, hypertension (mean blood pressure >85 mmHg), hypotension (mean blood pressure <25 mmHg), cardiac arrhythmia (severe bradycardia [heart rate <60 beats per min] or ventricular arrhythmia), and inability to achieve adequate ventilation despite appropriate adjustment of ventilator settings. Statistical analysis The National Perinatal Epidemiology Unit had data entry and management functions, provided an OpenClinica clinical database system, and did most of the analyses (it was masked to treatment allocation). Sample size was estimated primarily for detection of differences in the geometric mean lactate to N-acetyl aspartate ratio between groups because this value was greater than the number needed to detect a difference in FA. Based on assumptions from previous data, and allowing for a mortality rate of 20% before 15 days, a study of 138 infants would have 80% power to detect a geometric mean ratio of lactate to N-acetyl aspartate of 0·6, with a coefficient of variation of 1·2.14 A geometric mean ratio of 1 suggests no difference in mean values between the two groups, whereas a ratio of less than 1 favours the intervention group.

5 days, a study of 138 infants would have 80% power to detect a geometric mean ratio of lactate to N-acetyl aspartate of 0·6, with a coefficient of variation of 1·2.14 A geometric mean ratio of 1 suggests no difference in mean values between the two groups, whereas a ratio of less than 1 favours the intervention group. For changes in FA detected with tract-based spatial statistics, power was estimated by computational modelling and previous data, which suggested that with 80% power and a two-sided 5% significance level, a 10% change in FA would be detected in a study of 60 infants (higher FA shows less tissue damage).11, 22 Analysis of changes in FA induced by neuroprotective hypothermia showed that a substantial clinical effect was associated with changes of 10–20%.11, 14 All infants for whom magnetic resonance data were available were analysed in the groups that they were randomly allocated to, irrespective of allocation or protocol deviation. Diffusion tensor MRI and MRS analyses were done masked to treatment allocation (data were anonymised and allocation group was not included). We could not adjust for the minimisation factors used during randomisation because of small numbers.

were randomly allocated to, irrespective of allocation or protocol deviation. Diffusion tensor MRI and MRS analyses were done masked to treatment allocation (data were anonymised and allocation group was not included). We could not adjust for the minimisation factors used during randomisation because of small numbers. We used linear regression to analyse differences between the intervention and control groups in mean thalamic ln(lactate/N-acetyl aspartate). The difference between the natural logarithm of two ratios is equivalent to the ratio of geometric means—ie, the geometric mean ratio. A ln(x + 1) transformation was used because of the presence of zero values for the lactate to N-acetyl aspartate ratio in the data. We used Randomise (a program for non-parametric permutation inference for neuroimaging data; v2.9) to analyse data for FA, controlling for postmenstrual age, with the two-sample t test with nuisance variable options.23

used because of the presence of zero values for the lactate to N-acetyl aspartate ratio in the data. We used Randomise (a program for non-parametric permutation inference for neuroimaging data; v2.9) to analyse data for FA, controlling for postmenstrual age, with the two-sample t test with nuisance variable options.23 We did two sets of analysis for the two primary outcomes: the first included all infants who had MRI, and the second excluded those who subsequently died before discharge to account for a possible differential rate of scanning among these cases. We prespecified a subgroup analysis of lactate to N-acetyl aspartate by severity of abnormality of aEEG at randomisation. We also explored the effect of the time from birth to start of xenon therapy on the lactate to N-acetyl aspartate ratio and on FA, and the relation between these measures and neurological findings at discharge. A p value of 0·05 (two-sided 5% significance level) was deemed significant for the primary outcomes, and a p value of 0·01 (two-sided 1% significance level) was deemed significant for the exploratory analyses of secondary outcomes. We used Stata/SE (version 13.1) for all analyses.

nd neurological findings at discharge. A p value of 0·05 (two-sided 5% significance level) was deemed significant for the primary outcomes, and a p value of 0·01 (two-sided 1% significance level) was deemed significant for the exploratory analyses of secondary outcomes. We used Stata/SE (version 13.1) for all analyses. A data monitoring committee oversaw the study, and the members of the committee had no involvement in the day-to-day running of the trial. Decisions and recommendations made by the data monitoring committee were communicated to the trial steering committee in writing. The data monitoring committee met four times between September, 2010, and December, 2013, and received reports of analyses of safety data after incremental enrolments of around 25 infants. A study statistician, who attended the open part of the meeting only, provided the reports. Although the charter allowed interim assessment of efficacy, no interim analyses of efficacy were done: they were not requested by the data monitoring committee because the study did not include the pre-specified sample size. The committee decided to look only at safety outcomes. This trial is registered with ClinicalTrials.gov, number NCT00934700, and with ISRCTN, as ISRCTN08886155.

A data monitoring committee oversaw the study, and the members of the committee had no involvement in the day-to-day running of the trial. Decisions and recommendations made by the data monitoring committee were communicated to the trial steering committee in writing. The data monitoring committee met four times between September, 2010, and December, 2013, and received reports of analyses of safety data after incremental enrolments of around 25 infants. A study statistician, who attended the open part of the meeting only, provided the reports. Although the charter allowed interim assessment of efficacy, no interim analyses of efficacy were done: they were not requested by the data monitoring committee because the study did not include the pre-specified sample size. The committee decided to look only at safety outcomes. This trial is registered with ClinicalTrials.gov, number NCT00934700, and with ISRCTN, as ISRCTN08886155. Role of the funding source The funder of the study had no role in study design; data collection, analysis, or interpretation; or writing of the report. The study statisticians (OO and LL) had full access to all the data in the study, and provided data to the corresponding author after data analysis was completed. Summary data were provided to all the authors. The corresponding author had final responsibility for the decision to submit for publication.

g of the report. The study statisticians (OO and LL) had full access to all the data in the study, and provided data to the corresponding author after data analysis was completed. Summary data were provided to all the authors. The corresponding author had final responsibility for the decision to submit for publication. Results The study was done from Jan 31, 2012, to Sept 30, 2014. We screened 220 infants for eligibility, 92 of whom were enrolled up to completion of the enrolment period (figure 1). Failure to recruit the target sample size of 138 was primarily due to the closing of recruitment at the Liverpool participating centre after one infant had been enrolled because of incompatible configuration of scanner magnet gradient coils. 46 infants were randomly assigned to the cooling only group, and 46 to the xenon group (figure 1). Baseline clinical characteristics of the infants who were assigned to the two groups were broadly similar (table 1). All infants allocated to the cooling plus xenon group received xenon. Ventilation with xenon commenced a median of 10·0 h (IQR 8·2–11·2, range 4·0–12·6) after birth and continued for a median of 24 h (IQR 24–24). The mean concentration of inhaled xenon was 32·2% (SD 6·9). Ventilation with xenon was started within 6 h of birth in seven (15%) of 46 infants and after 12 h in five (11%) infants (range 12·1–12·6 h), and was discontinued in two (4%) infants before 24 h because of increasing oxygen requirements due to persistent pulmonary hypertension. Median xenon leakage was 12 mL/min (IQR 10–15).

ation with xenon was started within 6 h of birth in seven (15%) of 46 infants and after 12 h in five (11%) infants (range 12·1–12·6 h), and was discontinued in two (4%) infants before 24 h because of increasing oxygen requirements due to persistent pulmonary hypertension. Median xenon leakage was 12 mL/min (IQR 10–15). Cerebral magnetic resonance scans were done in 37 (80%) of 46 infants a mean of 5·8 days (SD 2·0) after birth in the cooling group and 41 (89%) of 46 infants at 6·0 days (2·1) after birth in the cooling plus xenon group. Lactate to N-acetyl aspartate ratio in the thalamus and FA values in the posterior limb of the internal capsule were similar in the two groups. The thalamic geometric mean ratio of lactate to N-acetyl aspartate was 1·09 (95% CI 0·90 to 1·32) and mean difference in FA was −0·01 (−0·03 to 0·02); exclusion of deaths from the analysis did not significantly affect results (table 2). Two adverse events were reported during the study, both in the cooling plus xenon group. Subcutaneous fat necrosis, which is associated with cooling therapy, was noted in one, and transient desaturation during the MRI (done after cessation of cooling and xenon) in another. No serious adverse events occurred, but nine (20%) infants in the cooling group and 11 (24%) in the cooling plus xenon group died (relative risk 1·22, 99% CI 0·44–3·41). Neither event rates of adverse outcomes and other clinical measures examined before discharge from hospital (table 3) nor the distribution of MRI scores between groups (table 4) differed significantly.

infants in the cooling group and 11 (24%) in the cooling plus xenon group died (relative risk 1·22, 99% CI 0·44–3·41). Neither event rates of adverse outcomes and other clinical measures examined before discharge from hospital (table 3) nor the distribution of MRI scores between groups (table 4) differed significantly. Lactate to N-acetyl aspartate ratio results did not differ significantly according to severity of abnormality of the aEEG at randomisation (geometric mean ratio 1·02 [95% CI 0·97–1·09] in the moderately abnormal aEEG group vs 1·09 [0·87–1·36] in the severely abnormal aEEG group; pinteraction=0·80). No significant relations were noted between time from birth to start of xenon therapy and the magnetic resonance measures (Spearman's correlation −0·14 for FA and −0·05 for the lactate to N-acetyl aspartate ratio). When we compared infants with normal or mildly abnormal results on neurological examination at discharge and those with moderately or severely abnormal results, the difference in means was 0·52 (95% CI 0·46–0·60; p<0·0001) for the lactate to N-acetyl aspartate ratio and 0·03 (0·01–0·06; p=0·02) for FA in the posterior limb of the internal capsule.

normal or mildly abnormal results on neurological examination at discharge and those with moderately or severely abnormal results, the difference in means was 0·52 (95% CI 0·46–0·60; p<0·0001) for the lactate to N-acetyl aspartate ratio and 0·03 (0·01–0·06; p=0·02) for FA in the posterior limb of the internal capsule. Discussion Our results showed that when xenon was used in a real-world context—ie, when it is given only at specialist centres and not during transport in a population of referred infants—qualified biomarkers of brain damage were not significantly affected and there was no treatment benefit. The use of magnetic resonance biomarkers to assess potential treatments rapidly and at low cost might be applicable to other neuroprotective therapies.

es and not during transport in a population of referred infants—qualified biomarkers of brain damage were not significantly affected and there was no treatment benefit. The use of magnetic resonance biomarkers to assess potential treatments rapidly and at low cost might be applicable to other neuroprotective therapies. We planned to enrol 138 infants to acquire primary outcome data for 111 infants, but only had data for 78 infants for the lactate to N-acetyl aspartate ratio and for 73 infants for the FA analyses. The study was powered primarily for lactate to N-acetyl aspartate ratio, but we used data from a previous study of successful neuroprotection with hypothermia to predict a 10–20% increase in FA because of treatment, for which a study size of 60 infants would be sufficient to detect a clinically significant difference of 10%.11, 14 This estimate was supported by the results of an in-silico modelling study22 in which the effect of changing FA values on a tract-based spatial statistics study was simulated and the number of voxels showing a significant difference by FA change was estimated; the results showed that a study size of 60 infants would be sufficient to detect clinically important differences in FA between the study groups. Thus, our study was adequately powered to detect changes in FA.22 The model was validated with infant data, showing that the model predicts real-world data accurately. Our trial is underpowered for the lactate to N-acetyl aspartate ratio, but provides a reliable estimate of lack of biological effect through tract-based spatial statistics, and that all outcome measures are concordant is relevant.

s validated with infant data, showing that the model predicts real-world data accurately. Our trial is underpowered for the lactate to N-acetyl aspartate ratio, but provides a reliable estimate of lack of biological effect through tract-based spatial statistics, and that all outcome measures are concordant is relevant. The duration of the enrolment period was 32 months, 2 months longer than was planned in the protocol, and we did not seek to extend that period largely because the sample size needed to detect a significant change in FA had been reached, and our goal was a rapid analysis of the suitability of the intervention for a large pragmatic trial. Although the study size was smaller than initially planned, judging by our results there is only a remote possibility that outcomes would materially change if we had enrolled the planned number. We initially planned to include three participating centres in the trial, but later sought a fourth centre to ensure that recruitment would be to target. However, during the quality-assurance check done after the first baby was recruited at the fourth centre, the MRI scanner gradient set-up was shown to differ from that of the other scanners resulting in a potential discrepancy of 10% in the data. We therefore closed recruitment at that centre, which reduced the number of infants who could be recruited. This issue shows the complexity of establishing magnetic resonance biomarkers across several sites.

-up was shown to differ from that of the other scanners resulting in a potential discrepancy of 10% in the data. We therefore closed recruitment at that centre, which reduced the number of infants who could be recruited. This issue shows the complexity of establishing magnetic resonance biomarkers across several sites. Follow-up of the study cohort is ongoing according to our clinical practice and because of the novelty of the intervention. Xenon's lack of efficacy despite promising experimental studies in animals can be explained in several ways. The timing, dose, and duration of treatment with inhaled xenon might have been suboptimum. Our regimen was based on clinical and safety factors: higher doses had not previously been given for such prolonged periods and could not be delivered to infants with substantial oxygen requirements, and earlier intervention would not be feasible for most infants born outside treatment centres. In a feasibility study in newborn infants given up to 50% xenon for 3–18 h, xenon was begun a median of 11 h (range 5–18) after birth;24, 25 earlier treatment might be possible if xenon can be delivered during transport to a specialist centre, but even then substantial delays are probable.

rn outside treatment centres. In a feasibility study in newborn infants given up to 50% xenon for 3–18 h, xenon was begun a median of 11 h (range 5–18) after birth;24, 25 earlier treatment might be possible if xenon can be delivered during transport to a specialist centre, but even then substantial delays are probable. We also based the treatment regimen on experimental studies,7, 8 the results of which suggested both that the therapeutic window for neural rescue was extended with cooling and that the combination of cooling with xenon has synergistic or at least additive neuroprotective efficacy and thus subanaesthetic doses could be effective. In addition to neuroprotective effects through inhibition of NMDA receptors, which have an important role in the early phase of reperfusion injury, xenon reduces apoptotic cell death, which occurs in the later phase of reperfusion injury. Thus, the hypothesis that delayed treatment with xenon in combination with early hypothermia might have neuroprotective effects is plausible. However, the effects of delayed treatment with xenon are variable in work in animals and no studies have been done of hypothermia augmented with xenon starting 6 h after insult.26, 27 We did not find a relation between timing of xenon inhalation in the time range used in this study and the magnetic resonance biomarkers. Since only seven (15%) of 46 infants in the xenon group received xenon by 6 h, we cannot exclude the possibility that starting xenon within 6 h of birth might be beneficial.

t.26, 27 We did not find a relation between timing of xenon inhalation in the time range used in this study and the magnetic resonance biomarkers. Since only seven (15%) of 46 infants in the xenon group received xenon by 6 h, we cannot exclude the possibility that starting xenon within 6 h of birth might be beneficial. Experimental studies of cooling for neuroprotection suggest that treatment for 72 h is needed when initiation is delayed.28 Early clinical studies of MRS in neonates also showed that the secondary reperfusion phase of injury after asphyxia lasts about 72 h, so our 24 h treatment might have been too short.29 Further evidence that the optimum duration of treatment could be longer than 24 h was provided by our previously reported finding of a transient recurrence of seizures after discontinuation of xenon.6 However, much shorter treatment durations are neuroprotective in experimental studies, although in all cases the delay to treatment was much less than 12 h.

ation of treatment could be longer than 24 h was provided by our previously reported finding of a transient recurrence of seizures after discontinuation of xenon.6 However, much shorter treatment durations are neuroprotective in experimental studies, although in all cases the delay to treatment was much less than 12 h. Perhaps participants in our study were too severely asphyxiated and thus had little prospect of benefit from any intervention after birth. We used inclusion criteria modified from those used in cooling trials, which necessitated the presence of the entire main neurological criteria for selection of participants, and that could account for the high rate of severe abnormalities in the aEEG and the high median score for hypoxic ischaemic encephalopathy at trial entry. However, mortality in our cohort was similar to that in previous trials of cooling (although in our trial, death was recorded only until discharge from hospital). The similar death rate to that in the cooling trials despite evidence of worse encephalopathy in our cohort might suggest that early use of routine hypothermia led to better outomes in this trial than in those in which consent had to be obtained before initiation of cooling.

ecorded only until discharge from hospital). The similar death rate to that in the cooling trials despite evidence of worse encephalopathy in our cohort might suggest that early use of routine hypothermia led to better outomes in this trial than in those in which consent had to be obtained before initiation of cooling. Another possible explanation of our negative results is that the chosen biomarkers are insufficiently sensitive. However, a raised lactate to N-acetyl aspartate ratio was the best predictor of subsequent neurodevelopmental outcome in a meta-analysis13 and is a sensitive indicator of subtle effects after birth asphyxia.13, 30 Furthermore, changes in FA have been used to identify treatment effects in small groups of infants with asphyxia; these changes correlated closely with subsequent outcome.11, 14 The significant association in our study between these markers and early neurological assessment provides further support for use of these biomarkers. The similar rates of abnormalities in both groups on visual assessment of the MRIs were also consistent with the main findings of the study.

th subsequent outcome.11, 14 The significant association in our study between these markers and early neurological assessment provides further support for use of these biomarkers. The similar rates of abnormalities in both groups on visual assessment of the MRIs were also consistent with the main findings of the study. Magnetic resonance biomarkers have great potential for use in the early development of neuroprotectants before undertaking large trials with clinical outcomes. A major challenge is validation of the markers across several magnetic resonance scanners to enable multisite studies. Using a standardised magnetic resonance scanning protocol, we acquired primary outcome data from 85% of participants, showing that cerebral magnetic resonance biomarkers such as lactate to N-acetyl aspartate ratio and FA are useful for rapid, preliminary assessment of potential neuroprotectants and planning of larger definitive trials. Supplementary Material Supplementary appendix

Magnetic resonance biomarkers have great potential for use in the early development of neuroprotectants before undertaking large trials with clinical outcomes. A major challenge is validation of the markers across several magnetic resonance scanners to enable multisite studies. Using a standardised magnetic resonance scanning protocol, we acquired primary outcome data from 85% of participants, showing that cerebral magnetic resonance biomarkers such as lactate to N-acetyl aspartate ratio and FA are useful for rapid, preliminary assessment of potential neuroprotectants and planning of larger definitive trials. Supplementary Material Supplementary appendix Acknowledgments The study was funded by the UK Medical Research Council (G0701714/1), and the sponsor was Imperial College London, London, UK. Air Products and Chemicals (Allentown, PA, USA) supported experimental work leading up to the trial and the purchase of the xenon gas. SLE (Croydon, UK) provided the xenon delivery device and the London Medicine for Children Research Network (London, UK) provided additional research nurse support for the study. We thank the Biological Research Centres at King's College London and Guy's and St Thomas' NHS Foundation Trust, Imperial College London and Imperial College Healthcare NHS Trust, and the National Perinatal Epidemiology Unit Clinical Trials Unit at the University of Oxford for their support; Stuart Faulkner (University College London) for his contribution to the development of the xenon delivery system and, together with Nazakat Merchant (King's College London) and Latha Srinivasan (Imperial College London), for training the study investigators in use of the xenon delivery device; Isabelle Viac (Guy's and St Thomas' NHS Foundation Trust) and Mary Dinan (University College London Hospitals NHS Foundation Trust) for training staff at participating centres and collecting and checking study data; the London Neonatal Transport Service for assessing referrals and transporting eligible infants to participating centres; medical and nursing staff at each participating centre for helping with assessment of infants, obtaining consent, and data collection; Serena Counsell (King's College London) for advising on diffusion tensor MRI and tract-based spatial statistics; and Mark Turner, who was the study investigator at the Liverpool Women's Hospital, Liverpool, UK.

at each participating centre for helping with assessment of infants, obtaining consent, and data collection; Serena Counsell (King's College London) for advising on diffusion tensor MRI and tract-based spatial statistics; and Mark Turner, who was the study investigator at the Liverpool Women's Hospital, Liverpool, UK. Contributors DA, EC, ADE, NPF, JH, EJ, BK, MM, and NJR contributed to study design. JG and BS administered the study. AB, GC-E, AD, ADE, GF, NT, BK, and DA contributed to data collection. EJ, LL, and OO provided statistical advice, and DA, AB, LL, OO, and NT did the data analysis. All authors contributed to the interpretation of the data and the writing of the Article. Data monitoring committee Michael Weindling, Liverpool Women's Hospital, Liverpool, UK (chair); Alan Fenton, Royal Victoria Infirmary, Newcastle, UK; Rafael Perera, University of Oxford, Oxford, UK (statistician). Trial steering committee Henry Halliday (chair); David Sweet, Royal Maternity Hospital, Belfast, UK; Eleri Adams, John Radcliffe Hospital, Oxford, UK; Topun Austin, Addenbrooke's Hospital, Cambridge, UK; Joanna Jenkinson, Medical Research Council, London, UK.

Data monitoring committee Michael Weindling, Liverpool Women's Hospital, Liverpool, UK (chair); Alan Fenton, Royal Victoria Infirmary, Newcastle, UK; Rafael Perera, University of Oxford, Oxford, UK (statistician). Trial steering committee Henry Halliday (chair); David Sweet, Royal Maternity Hospital, Belfast, UK; Eleri Adams, John Radcliffe Hospital, Oxford, UK; Topun Austin, Addenbrooke's Hospital, Cambridge, UK; Joanna Jenkinson, Medical Research Council, London, UK. Declaration of interests NPF and MM are directors and own equity in the spin-out company Neuroprotexeon, and one of the aims of this company is to develop the use of xenon as a neuroprotectant. NJR reports grants from Air Products and the UK Medical Research Council during the conduct of the study; grants from Chiesi and Air Liquides, outside the submitted work; and has a patent intellectual property related to melatonin neuroprotection licensed. The other authors declare no competing interests. Figure Trial profile *Could not give consent. †When parents were unmarried, only the mother of the infant could provide consent. Table 1 Baseline clinical characteristics in the intention-to-treat population

Declaration of interests NPF and MM are directors and own equity in the spin-out company Neuroprotexeon, and one of the aims of this company is to develop the use of xenon as a neuroprotectant. NJR reports grants from Air Products and the UK Medical Research Council during the conduct of the study; grants from Chiesi and Air Liquides, outside the submitted work; and has a patent intellectual property related to melatonin neuroprotection licensed. The other authors declare no competing interests. Figure Trial profile *Could not give consent. †When parents were unmarried, only the mother of the infant could provide consent. Table 1 Baseline clinical characteristics in the intention-to-treat population Cooling only (n=46) Cooling plus xenon (n=46) Treatment hospital (n) University College London 15 15 St Thomas′ 17 17 Queen Charlotte and Chelsea 14 13 Liverpool Women's 0 1 Birth in treatment centre 15 (33%) 16 (35%) Male sex 21 (46%) 26 (57%) Birthweight (g), mean (SD) 3213 (448) 3392 (685) Gestation at delivery (weeks), mean (SD) 39·8 (1·3) 39·8 (1·7) Apgar at 10 min, median (IQR) 5 (4 to 7) 5 (3 to 6) Median cord or first blood pH (IQR) 6·9 (6·7 to 7·0) 6·9 (6·8 to 7·1) Mean cord or first blood pH (SD) 6·9 (0·2) 6·9 (0·2) Base excess (mmol/L), median (IQR) –19·7 (–23·7 to −14·0) –17·7 (–22 to −13·5) Thompson hypoxic ischaemic encephalopathy score* at trial entry 0–10 2 (4%) 5 (11%) 11–14 30 (65%) 21 (46%) 15–22 14 (30%) 20 (43%) Median (IQR) 14 (12 to 15) 14 (12 to 16) Abnormality on amplitude-integrated EEG Moderate 7 (15%) 6 (13%) Severe 39 (85%) 40 (87%) Age cooling commenced, n/N (%) <4 h 41/44 (93%) 41/44 (93%) 4–6 h 3/44 (7%) 3/44 (7%) Median (IQR) 0·3 (0·0 to 0·8) 0·2 (0·0 to 1·5) Head circumference at admission to neonatal unit (cm), mean (SD) 34·4 (1·5) 34·5 (1·8) Data are n (%), unless otherwise indicated.

ed EEG Moderate 7 (15%) 6 (13%) Severe 39 (85%) 40 (87%) Age cooling commenced, n/N (%) <4 h 41/44 (93%) 41/44 (93%) 4–6 h 3/44 (7%) 3/44 (7%) Median (IQR) 0·3 (0·0 to 0·8) 0·2 (0·0 to 1·5) Head circumference at admission to neonatal unit (cm), mean (SD) 34·4 (1·5) 34·5 (1·8) Data are n (%), unless otherwise indicated. * Score ranges from 0 to 22, with higher scores corresponding to more severe encephalopathy. Table 2 Analysis of primary outcomes Cooling only Cooling plus xenon Geometric mean ratio (95% CI) Mean difference (95%CI) Infants with MRI scans Lactate to N-acetyl aspartate ratio 1·09 (0·90 to 1·32) .. n 37 41 Arithmetic mean (SD) 0·47 (0·94) 0·68 (1·12) Coefficient of variation* 2·19 1·68 Geometric mean 0·34 0·47 Fractional anisotropy .. –0·01 (–0·03 to 0·02) n 35 38 Mean (SD) 0·41 (0·01) 0·40 (0·01) Infants with MRI scans surviving to discharge Lactate to N-acetyl aspartate ratio 0·98 (0·85 to 1·12) .. n 34 33 Arithmetic mean (SD) 0·32 (0·42) 0·34 [0·77] Coefficient of variation* 1·41 1·30 Geometric mean 0·28 0·25 Fractional anisotropy .. –0·01 (–0·01 to 0·01) n 33 30 Mean (SD) 0·40 (0·05) 0·40 (0·05) Geometric mean ratios were calculated after log (x + 1) transformation. Fractional anisotropy data were extracted from a mask of the posterior limb of the internal capsule via tract-based spatial statistics. * Coefficient of variation=√(exp(var)–1), where var is the variance on the log scale. Table 3 Analysis of secondary outcomes

Cooling only Cooling plus xenon Geometric mean ratio (95% CI) Mean difference (95%CI) Infants with MRI scans Lactate to N-acetyl aspartate ratio 1·09 (0·90 to 1·32) .. n 37 41 Arithmetic mean (SD) 0·47 (0·94) 0·68 (1·12) Coefficient of variation* 2·19 1·68 Geometric mean 0·34 0·47 Fractional anisotropy .. –0·01 (–0·03 to 0·02) n 35 38 Mean (SD) 0·41 (0·01) 0·40 (0·01) Infants with MRI scans surviving to discharge Lactate to N-acetyl aspartate ratio 0·98 (0·85 to 1·12) .. n 34 33 Arithmetic mean (SD) 0·32 (0·42) 0·34 [0·77] Coefficient of variation* 1·41 1·30 Geometric mean 0·28 0·25 Fractional anisotropy .. –0·01 (–0·01 to 0·01) n 33 30 Mean (SD) 0·40 (0·05) 0·40 (0·05) Geometric mean ratios were calculated after log (x + 1) transformation. Fractional anisotropy data were extracted from a mask of the posterior limb of the internal capsule via tract-based spatial statistics. * Coefficient of variation=√(exp(var)–1), where var is the variance on the log scale. Table 3 Analysis of secondary outcomes Cooling only (n=46) Cooling plus xenon (n=46) Relative risk (99% CI) Death before discharge 9 (20%) 11 (24%) 1·22 (0·44 to 3·41) Maximum Thompson hypoxic ischaemic encephalopathy score in first week of life 0–10 0 (0%) 1 (2%) 1·22 (0·82 to 1·82) 11–14 19 (41%) 12 (26%) 15–22 27 (59%) 33 (72%) Median (IQR) 16 (13 to 19) 15 (14 to 18) Neurological assessment at discharge* 0·66 (0·17 to 2·51) Normal or mildly abnormal 29 (78%) 30 (86%) Moderately abnormal 7 (19%) 3 (9%) Very abnormal 1 (3%) 2 (6%) Persistent hypotension 29 (63%) 31 (67%) 1·06 (0·72 to 1·58) Cardiac arrhythmia (heart rate <80 beats per min) 4 (9%) 2 (4%) 0·50 (0·06 to 4·36) Thrombocytopenia (platelet count <150 × 109 per L) 20 (43%) 18 (39%) 0·90 (0·55 to 1·47) Prolonged blood coagulation time (activated partial thromboplastin time >41 s or international normalised ratio >3) 32 (70%) 36 (78%) 1·13 (0·82 to 1·55) Major venous thrombosis 1 (2%) 0 (0%) .. Anuria or urine output <0·5 mL/kg/h for >24 h, n/N (%) 3/20 (15%) 6/23 (26%) 2·00 (0·38 to 10·5) Culture-proven late-onset sepsis 0 (0%) 2 (4%) .. Necrotising enterocolitis 0 (0%) 0 (0%) .. Pneumonia 1 (2%) 1 (2%) 1·00 (0·03 to 36·71) Pulmonary air leak 0 (0%) 3 (7%) .. Pulmonary haemorrhage 3 (7%) 1 (2%) 0·33 (0·02 to 6·21) Persistent pulmonary hypertension 3 (7%) 3 (7%) 1·00 (0·13 to 7·64) Intracranial haemorrhage 3 (7%) 4 (9%) 1·33 (0·20 to 8·85) Seizures 36 (78%) 36 (78%) 1·00 (0·75 to 1·33) Median age (IQR) full oral feeding achieved (days)† 9 (7 to 11) 9 (7 to 12) .. Did not achieve full oral feeding by discharge† 6 (17%) 4 (12%) 0·73 (0·16 to 3·40) Median hospital stay (IQR) to discharge* (days) 14 (10 to 17) 12 (9 to 22) –1 (–5 to 4)‡ Data are n (%) unless otherwise specified. Hypotension was defined as a mean blood pressure of less than 40 mmHg. Seizures included both clinical and subclinical events, and were identified by amplitude-integrated EEG.

·16 to 3·40) Median hospital stay (IQR) to discharge* (days) 14 (10 to 17) 12 (9 to 22) –1 (–5 to 4)‡ Data are n (%) unless otherwise specified. Hypotension was defined as a mean blood pressure of less than 40 mmHg. Seizures included both clinical and subclinical events, and were identified by amplitude-integrated EEG. * Calculated only in infants alive at discharge. † Data available for 36 infants in the cooling group and 33 in the xenon group. Median (IQR) based only on those who achieved full oral feeding by discharge. ‡ These data are median difference (95% CI). Table 4 Visual analysis of MRI by score (secondary outcome)

·16 to 3·40) Median hospital stay (IQR) to discharge* (days) 14 (10 to 17) 12 (9 to 22) –1 (–5 to 4)‡ Data are n (%) unless otherwise specified. Hypotension was defined as a mean blood pressure of less than 40 mmHg. Seizures included both clinical and subclinical events, and were identified by amplitude-integrated EEG. * Calculated only in infants alive at discharge. † Data available for 36 infants in the cooling group and 33 in the xenon group. Median (IQR) based only on those who achieved full oral feeding by discharge. ‡ These data are median difference (95% CI). Table 4 Visual analysis of MRI by score (secondary outcome) Cooling only (n=39) Cooling plus xenon (n=44) Relative risk (99% CI) Mean difference in scores (99% CI) Posterior limb of internal capsule Score 0 18 21 0·07 (–0·44 to 0·57) Score 1 11 8 0·97 (0·57 to 1·65) 0·07 (–0·44 to 0·57) Score 2 10 15 0·97 (0·57 to 1·65) 0·07 (–0·44 to 0·57) Basal ganglia and thalamus Score 0 6 14 –0·05 (–0·71 to 0·60) Score 1 9 3 –0·05 (–0·71 to 0·60) Score 2 16 13 1·00 (0·64 to 1·56) –0·05 (–0·71 to 0·60) Score 3 8 14 1·00 (0·64 to 1·56) –0·05 (–0·71 to 0·60) White matter Score 0 16 14 0·33 (–0·35 to 1·00) Score 1 7 8 0·33 (–0·35 to 1·00) Score 2 11 10 1·22 (0·65 to 2·29) 0·33 (–0·35 to 1·00) Score 3 5 12 1·22 (0·65 to 2·29) 0·33 (–0·35 to 1·00) Cortex Score 0 29 30 0·33 (–0·33 to 0·99) Score 1 4 2 0·33 (–0·33 to 0·99) Score 2 3 2 1·77 (0·56 to 5·64) 0·33 (–0·33 to 0·99) Score 3 3 10 1·77 (0·56 to 5·64) 0·33 (–0·33 to 0·99) Data are n. Relative risk is calculated for the moderate and severe changes groups combined, so only one relative risk and 99% CI is listed for each brain site. For the posterior limb of the internal capsule scores, 0=normal, 1=equivocal (reduced or asymmetrical signal intensity), and 2=loss (reversed or abnormal signal intensity bilaterally on T1-weighted or T2-weighted sequences, or both). For basal ganglia and thalamic scores, 0=normal, 1=mild (focal abnormal signal intensity), 2=moderate (multifocal abnormal signal intensity), and 3=severe (widespread abnormal signal intensity). For white matter scores, 0=normal, 1=mild (exaggerated long T1 and long T2 in periventricular white matter only), 2=moderate (long T1 and long T2 extending out to subcortical white matter or focal punctate lesions or focal area of infarction, or any combination thereof), and 3=severe (widespread abnormalities including overt infarction, haemorrhage, and long T1 and long T2). Cortical involvement was scored as the presence of abnormal signal intensity, usually decreased T1 or cortical highlighting (ie, increased signal intensity in the cortex). For cortical scores, 0=normal, 1=mild (one or two sites involved), 2=moderate (three sites involved), and 3=severe (more than three sites involved).

Cortical involvement was scored as the presence of abnormal signal intensity, usually decreased T1 or cortical highlighting (ie, increased signal intensity in the cortex). For cortical scores, 0=normal, 1=mild (one or two sites involved), 2=moderate (three sites involved), and 3=severe (more than three sites involved). The sites included the central sulcus, interhemispheric fissure, and the insula. All the scans were assessed and graded by NT, who was masked to intervention.

Introduction Cerebral cavernous malformations (CCMs) are the second commonest incidental vascular finding—after aneurysms1—on brain MRI, with a prevalence of one in 625 neurologically asymptomatic people.2, 3 Because brain MRI is needed for diagnosis of CCMs without pathological examination,4 the number of people in whom CCM has been detected has risen since the advent of MRI.5, 6 CCMs can be asymptomatic or can cause epileptic seizures,7 stroke due to symptomatic intracranial haemorrhage (ICH),8 or new focal neurological deficit (FND) without evidence on brain imaging of recent haemorrhage.8 As MRI use increases over time,9 so too does the need for information about the magnitude and predictors of the risk of ICH or FND from CCMs. These risks can inform decisions about whether to treat CCMs with neurosurgical excision or stereotactic radiosurgery, although the use of the latter remains controversial and neither treatment has been assessed in a randomised trial.10 Some cohort studies have estimated the risk of ICH from untreated CCMs. However, findings from a recent systematic review11 showed that these studies were mostly retrospective hospital-based series, with sample sizes not exceeding 139 people and short durations of follow-up, without clearly defined diagnostic criteria or outcome events,8 and in which several different statistical methods were used to calculate the risk and predictors of ICH. Findings from these studies have left uncertainty about the magnitude of the risk of ICH and its predictors. Additionally, there remains an absence of prediction models.

ertaking a systematic review and meta-analysis of individual patient data from cohort studies of people with CCMs with similar designs and outcome definitions.12, 13 By using consistent methods of analysis, we aimed to estimate the risks of first ICH or FND during follow-up and to identify predictors of these outcomes. Methods Study design We undertook this study according to a protocol finalised on May 9, 2012, which was approved by the Cerebral Cavernous Malformations Individual Patient Data Meta-analysis Collaborators. Two authors with training in undertaking systematic reviews (MAH and RA-SS) used electronic search strategies (appendix p 2)11 to search Ovid MEDLINE and Embase from inception until April 30, 2015, for published cohort studies; they screened the bibliographies of studies for other potentially eligible cohorts, established their eligibility, and resolved any disagreements by discussion. Cohorts were eligible for inclusion regardless of language of publication if they included people aged 16 years or older—a common age cutoff for transition to adult services14—at the time of a definite diagnosis of CCM confirmed by brain MRI, and if they included at least symptomatic ICH due to CCM and death as outcomes after an inception point of first diagnosis of CCM but before first CCM treatment with neurosurgical excision or stereotactic radiosurgery. After sending a copy of the protocol and invitation to collaborate to the corresponding authors of the reports that described cohorts that were eligible for inclusion, followed by one reminder, we included cohorts from studies for which the study investigators confirmed their eligibility and provided patient-level data on baseline covariates, outcomes, and CCM treatment. We were unable to use aggregate data from cohorts for which individual patient data were unavailable because time-to-event analyses require patient-level data.

from studies for which the study investigators confirmed their eligibility and provided patient-level data on baseline covariates, outcomes, and CCM treatment. We were unable to use aggregate data from cohorts for which individual patient data were unavailable because time-to-event analyses require patient-level data. Research ethics committees or other entities overseeing the use of patients' data approved the collaborating cohorts. Cohorts shared only anonymised data, so neither individual consent nor specific approval for this individual patient data meta-analysis were required.15 Data collection Collaborating cohorts provided patient-level data at baseline (sex, mode of symptomatic presentation leading to diagnosis of CCMs, age at CCM diagnosis, date of CCM diagnosis, CCM multiplicity, and primary CCM location) and during follow-up (all outcome types and dates, all treatment types and dates, and date of last follow-up) for time-to-event analyses. We excluded people who had been first diagnosed when younger than 16 years or who had already received treatment for CCMs.

CCM diagnosis, CCM multiplicity, and primary CCM location) and during follow-up (all outcome types and dates, all treatment types and dates, and date of last follow-up) for time-to-event analyses. We excluded people who had been first diagnosed when younger than 16 years or who had already received treatment for CCMs. Investigators in each cohort distinguished two clinical events attributable to CCMs, at presentation and during follow-up, where possible, according to the Angioma Alliance definitions.8 ICH was defined as acute or subacute onset of symptoms of haemorrhage with recent extralesional or intralesional haemorrhage confirmed by investigations (CT or MRI, or pathological examination at autopsy). FND was defined as new or worsened neurological deficit referable to the CCM anatomical location with or without timely investigation to rule out evidence of recent haemorrhage.8 We checked data from each cohort for internal consistency against published reports of the cohort and resolved any queries with the relevant collaborators. For missing data, we contacted collaborators to request the missing values. We assessed the risk of bias of each cohort according to an eight-item instrument published by the Cochrane Methods Bias group.16

rnal consistency against published reports of the cohort and resolved any queries with the relevant collaborators. For missing data, we contacted collaborators to request the missing values. We assessed the risk of bias of each cohort according to an eight-item instrument published by the Cochrane Methods Bias group.16 Outcomes The primary outcome was first symptomatic ICH due to CCM. The secondary outcome was a composite of first symptomatic ICH or FND due to CCM. Time-to-event analyses started at diagnosis of CCM and terminated at the earliest occurrence of ICH only for the analysis of the primary outcome. For the secondary outcome, these analyses were terminated at the earliest occurrence of ICH or FND. If an outcome did not occur, we censored analyses at the earliest occurrence of CCM treatment, death unrelated to CCM, last available follow-up, or 5 years after CCM diagnosis.

ly for the analysis of the primary outcome. For the secondary outcome, these analyses were terminated at the earliest occurrence of ICH or FND. If an outcome did not occur, we censored analyses at the earliest occurrence of CCM treatment, death unrelated to CCM, last available follow-up, or 5 years after CCM diagnosis. Statistical analysis We agreed a detailed statistical analysis plan with collaborators in Oct 25, 2013, before data analysis began. We categorised mode of presentation as one of four mutually exclusive categories: ICH, FND, epileptic seizure if the seizure was neither symptomatic of a concomitant ICH nor more likely to be due to another cause, or incidental if a person was asymptomatic or if their symptoms (eg, headache) could not be ascribed to the underlying CCM. We attributed one CCM location to people harbouring more than one CCM by using the location of the symptomatic CCM; when a person presented asymptomatically with more than one CCM, brainstem CCM location took precedence since this location seemed to be a predictor of ICH from a systematic review of aggregate data from existing studies.11 We calculated the relative risk of having a CCM located in the brainstem for ICH or FND versus other presentations. We used survival analysis to estimate the 5-year risk of symptomatic ICH attributable to CCMs. The inception point was the earliest date of definite diagnosis of CCMs by radiographic or pathological investigation.

Statistical analysis We agreed a detailed statistical analysis plan with collaborators in Oct 25, 2013, before data analysis began. We categorised mode of presentation as one of four mutually exclusive categories: ICH, FND, epileptic seizure if the seizure was neither symptomatic of a concomitant ICH nor more likely to be due to another cause, or incidental if a person was asymptomatic or if their symptoms (eg, headache) could not be ascribed to the underlying CCM. We attributed one CCM location to people harbouring more than one CCM by using the location of the symptomatic CCM; when a person presented asymptomatically with more than one CCM, brainstem CCM location took precedence since this location seemed to be a predictor of ICH from a systematic review of aggregate data from existing studies.11 We calculated the relative risk of having a CCM located in the brainstem for ICH or FND versus other presentations. We used survival analysis to estimate the 5-year risk of symptomatic ICH attributable to CCMs. The inception point was the earliest date of definite diagnosis of CCMs by radiographic or pathological investigation. We prespecified five potential predictors for investigation of their association with outcome on the basis of their clinical relevance, likelihood of being associated with outcome,11, 17, 18, 19, 20 reliability, accuracy of measurement, completeness, and availability at the time of diagnosis.21 Core predictors were mode of presentation (ICH or FND vs other) and CCM location (brainstem vs other). We dichotomised mode of clinical presentation because previous ICH seemed to be a predictor from a systematic review of aggregate data from existing studies, and included FND because in certain circumstances FND can suggest undetected ICH.8 Putative predictors were sex (female vs male), CCM multiplicity (more than one vs one), and increasing age at diagnosis (which we treated as a continuous variable in association analyses).22 We did separate univariable analyses of each cohort and the pooled data. We used Cox regression to calculate the unadjusted hazard ratios (HRs) for each predictor. We used log-minus-log plots to check that the proportional-hazards assumption was met before undertaking Cox proportional-hazards multivariable regression. In the statistical analysis plan, we specified that in multivariable adjusted analyses, we would enter the core predictors into the model first; then, to ascertain whether any of the remaining three putative predictors added significant information over and above the core predictors, we entered each putative predictor into the model, provided the conventional rule that there should be at least ten outcome events per predictor was fulfilled.21, 23 We followed this clinically driven approach because we were undertaking a two-stage individual patient data meta-analysis and we were aware that the number of events during follow-up in each individual cohort would be insufficient to permit a full data-driven selection of variables to be included in the model.

d.21, 23 We followed this clinically driven approach because we were undertaking a two-stage individual patient data meta-analysis and we were aware that the number of events during follow-up in each individual cohort would be insufficient to permit a full data-driven selection of variables to be included in the model. We did two-stage meta-analyses of the univariable associations of each of the five predictors with outcome, in which we derived study-specific unadjusted HRs and combined them using a random-effects model to generate a weighted unadjusted pooled HR.24, 25 We did meta-analyses for each of the three putative predictors, both unadjusted and adjusted for the core predictors. We assessed heterogeneity between studies using the I2 index of inconsistency, which measures the proportion of total variation in study estimates that is due to heterogeneity.26, 27

ed pooled HR.24, 25 We did meta-analyses for each of the three putative predictors, both unadjusted and adjusted for the core predictors. We assessed heterogeneity between studies using the I2 index of inconsistency, which measures the proportion of total variation in study estimates that is due to heterogeneity.26, 27 We planned to build prognostic models of the estimated 5-year risk of ICH after diagnosis on the basis of the findings of the multivariable analyses and meta-analyses.23, 28, 29 Because we had five potential predictors and outcome events tend to be infrequent, we used the entire dataset to develop our models, rather than split it into a derivation and a test set. For internal validation, we used bootstrapping to derive 95% CIs for the multivariable Cox regression analyses, by creating 10 000 random samples of the same size as the study cohort using sampling with replacement. We used Kaplan-Meier plots to assess the separation achieved by prognostic models, and life tables to estimate annual hazard rates of experiencing an ICH within 5 years of diagnosis of CCMs. We did descriptive and survival analyses using IBM SPSS Statistics 19 and 21, and used Stata IC12 for the individual patient data meta-analysis.

We planned to build prognostic models of the estimated 5-year risk of ICH after diagnosis on the basis of the findings of the multivariable analyses and meta-analyses.23, 28, 29 Because we had five potential predictors and outcome events tend to be infrequent, we used the entire dataset to develop our models, rather than split it into a derivation and a test set. For internal validation, we used bootstrapping to derive 95% CIs for the multivariable Cox regression analyses, by creating 10 000 random samples of the same size as the study cohort using sampling with replacement. We used Kaplan-Meier plots to assess the separation achieved by prognostic models, and life tables to estimate annual hazard rates of experiencing an ICH within 5 years of diagnosis of CCMs. We did descriptive and survival analyses using IBM SPSS Statistics 19 and 21, and used Stata IC12 for the individual patient data meta-analysis. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or in the decision to submit for publication. All authors had full access to all the data in the study, but only MAH and GDM analysed the data; the corresponding author had final responsibility for the decision to submit for publication.

esign, data collection, data analysis, data interpretation, writing of the report, or in the decision to submit for publication. All authors had full access to all the data in the study, but only MAH and GDM analysed the data; the corresponding author had final responsibility for the decision to submit for publication. Results From 22 publications of potentially eligible cohorts that included 2957 people (figure 1; appendix p 3), six research groups provided published data on seven cohorts involving 1620 people for inclusion in this meta-analysis (appendix p 4).11, 17, 18, 19, 20, 30 For missing data, in every instance the investigators of the original study were able to check their database and provide the missing values for all covariates. Of the 16 research groups who did not join the collaboration, whose studies involved 1337 people (45% of the published data), three groups agreed to collaborate but did not share data, two groups no longer had access to their data, one group did not have clinical data available, and ten groups did not respond to our invitations (figure 1; appendix pp 19–20).

in the collaboration, whose studies involved 1337 people (45% of the published data), three groups agreed to collaborate but did not share data, two groups no longer had access to their data, one group did not have clinical data available, and ten groups did not respond to our invitations (figure 1; appendix pp 19–20). Five cohorts were from tertiary referral centres (Toronto Western Hospital, ON, Canada, 1987–2007, n=345; Mayo Clinic, Rochester, MN, USA, 1984–98, n=267; Hôpital Lariboisière, Paris, France, 1994–2011, n=81; Seoul National University Hospital, Seoul, South Korea, 1998–2010, n=326; and Beijing Tiantan Hospital, Beijing, China, 1985–2012, n=306) and one was from the Scottish population (1999–2003, n=135). Data from a seventh cohort (Scottish population, 2006–10, n=160) and 63% (217 of 345 adults) of the Toronto cohort were previously unpublished. The Beijing cohort was restricted to adults with a brainstem CCM. Follow-up was prospective in the population-based cohorts11 and one of the hospital-based cohorts,20 both retrospective and prospective in another hospital-based cohort,17 and retrospective in the remainder (appendix p 3).18, 19, 30 We did not identify problems (eg, incompatibility with different study inception points or differences in CCM, ICH, or FND definitions) when we checked the individual patient data. We used the date of symptom onset leading to diagnosis of CCM or the date of first medical assessment for 27 people without a date of diagnosis of CCMs in one cohort.17 All seven cohorts recorded FND as a mode of clinical presentation, but four cohorts did not record FND during follow-up.18, 19, 20, 30 Risk of bias in the seven cohorts was low (appendix p 5).

to diagnosis of CCM or the date of first medical assessment for 27 people without a date of diagnosis of CCMs in one cohort.17 All seven cohorts recorded FND as a mode of clinical presentation, but four cohorts did not record FND during follow-up.18, 19, 20, 30 Risk of bias in the seven cohorts was low (appendix p 5). The median age at diagnosis was 45 years (range 16–91), 867 (54%) of people were women, 822 (51%) presented with ICH or FND, 282 (17%) had multiple CCMs, and 575 (35%) had CCMs located in the brainstem (table 1). People who had presented with ICH or FND were more likely to have a brainstem CCM than people who presented with a seizure or incidentally (prevalence ratio 6·0, 95% CI 4·8–7·5). The primary outcome event was recorded in all seven cohorts; follow-up ended at the occurrence of the first ICH outcome event (n=204) or censoring (owing to CCM treatment [n=193], death unrelated to ICH or FND [n=46], end of follow-up before 5 years [n=596], or end of follow-up at 5 years [n=581]). Total follow-up was 5197 person-years (median 3·5 years per person, IQR 1·6–5·0, 70% completeness31 of all potential follow-up). Only the Scottish and Toronto cohorts recorded both ICH and FND during follow-up (in 640 [40%] of 1620 included people), and so we restricted analysis of the secondary composite outcome event of ICH or FND to these cohorts.

years (median 3·5 years per person, IQR 1·6–5·0, 70% completeness31 of all potential follow-up). Only the Scottish and Toronto cohorts recorded both ICH and FND during follow-up (in 640 [40%] of 1620 included people), and so we restricted analysis of the secondary composite outcome event of ICH or FND to these cohorts. 204 of 1620 people experienced ICH within 5 years of diagnosis of CCMs (Kaplan-Meier estimated 5-year risk 15·8%, 95% CI 13·7–17·9), four of which were fatal (1-month case fatality rate after ICH due to CCM; 2·0% [95% CI 0·1–3·9]). The estimated risk of first ICH within 5 years of diagnosis of CCM in the pooled dataset was higher for people presenting with ICH or FND versus other modes of presentation (26·4% [95% CI 23·1–29·7] vs 4·3% [2·5–6·1]; pooled unadjusted HR 5·6 [95% CI 3·2–9·7]) and for people with a primary CCM location in the brainstem versus another location (27·7% [95% CI 23·6–31·8] vs 8·2% [6·2–10·2]; pooled unadjusted HR 4·4, [95% CI 2·3–8·6]); these findings were similar in individual cohorts (figure 2; appendix pp 6–7). We found no evidence—even after multivariable adjustment for the two core predictors—that age, sex, or CCM multiplicity affected the risk of ICH, and these findings were generally consistent between cohorts (appendix pp 8–10).

CI 2·3–8·6]); these findings were similar in individual cohorts (figure 2; appendix pp 6–7). We found no evidence—even after multivariable adjustment for the two core predictors—that age, sex, or CCM multiplicity affected the risk of ICH, and these findings were generally consistent between cohorts (appendix pp 8–10). We found no evidence of publication bias in a funnel plot of the HR for the primary outcome by mode of presentation (appendix p 11). We assessed risk of publication bias by inspecting a funnel plot of the HR against the standard error of log (HR). As a sensitivity analysis, we repeated the meta-analyses of the five cohorts with proportionally fewer outcome events. Although the Chinese cohort, which consisted entirely of adults with brainstem CCMs, contributed 44% of ICHs during follow-up (90 of 204), a sensitivity analysis excluding the Chinese and South Korean cohorts—both of which contained proportionally more people with ICH during follow-up than the other five cohorts (South Korea 52 of 204)—produced similar results (data not shown).

lts with brainstem CCMs, contributed 44% of ICHs during follow-up (90 of 204), a sensitivity analysis excluding the Chinese and South Korean cohorts—both of which contained proportionally more people with ICH during follow-up than the other five cohorts (South Korea 52 of 204)—produced similar results (data not shown). The findings of the univariable and multivariable survival analyses led us to create two prognostic models, using the two core predictors—mode of clinical presentation and CCM location—in multivariable Cox regression analyses. By dichotomising each of these predictors, we identified four prognostic subgroups with significant differences in their risks of outcomes according to stratified Kaplan-Meier plots (figure 3; log-rank p<0·0001 for both proportion progressing to ICH and proportion progressing to ICH or FND) or according to Cox regression of 5-year event rates and HRs (table 2), after checking proportional hazards assumptions were met (appendix p 12). People with CCMs outside the brainstem who had not presented with ICH or FND were in the majority and had the lowest risk, whereas people with brainstem CCMs presenting with ICH or FND were the highest-risk group (figure 3; table 2; appendix p 13). Estimates of the annual risk of ICH during each year of follow-up decreased from 6·2% (95% CI 4·9–7·4) in the first year of follow-up to 2·0% (0·9–3·0) in the fifth year overall (p<0·0001; appendix p 14). This reduction was evident only in people presenting with ICH or FND (appendix p 15).

f level (table 1). Patients in the Charcot-Marie-Tooth disease 1A group had significantly higher all-muscle fat fraction and T2 and lower MTR than matched controls at calf level but not at thigh level, whereas their all-muscle CSA was lower than for the controls at calf level but not at thigh level (table 1, figure 3). Repeat assessments (table 2) were done after a mean of 12·4 months (SD 1·0). In the inclusion body myositis group, the MRC lower limb score, inclusion body myositis functional rating scale (IBMFRS), and overall myometric knee extension strength deteriorated significantly during follow-up. In the Charcot-Marie-Tooth disease 1A group, none of the clinical measures changed significantly. Although myometric ankle dorsiflexion strength showed a significant increase (p=0·049) compared with baseline, this was not significant (p=0·24) compared with the change in matched controls. Quantitative MRI values for individual muscles are given in the appendix. Strong correlations were noted between the three quantitative MRI measures (fat fraction, MTR, and T2) of individual muscles (all r>0·89, p<0·0001, appendix).

The findings of the univariable and multivariable survival analyses led us to create two prognostic models, using the two core predictors—mode of clinical presentation and CCM location—in multivariable Cox regression analyses. By dichotomising each of these predictors, we identified four prognostic subgroups with significant differences in their risks of outcomes according to stratified Kaplan-Meier plots (figure 3; log-rank p<0·0001 for both proportion progressing to ICH and proportion progressing to ICH or FND) or according to Cox regression of 5-year event rates and HRs (table 2), after checking proportional hazards assumptions were met (appendix p 12). People with CCMs outside the brainstem who had not presented with ICH or FND were in the majority and had the lowest risk, whereas people with brainstem CCMs presenting with ICH or FND were the highest-risk group (figure 3; table 2; appendix p 13). Estimates of the annual risk of ICH during each year of follow-up decreased from 6·2% (95% CI 4·9–7·4) in the first year of follow-up to 2·0% (0·9–3·0) in the fifth year overall (p<0·0001; appendix p 14). This reduction was evident only in people presenting with ICH or FND (appendix p 15). In the subgroup of 640 people with data recorded for the composite secondary outcome of ICH or FND, 36 people had an ICH and 52 had an FND during follow-up. In this subgroup, the magnitude, direction, and consistency of associations of the core and putative predictors with the secondary outcome were similar to the primary analysis (table 2; figure 2; appendix pp 8–10 and 16–17), but the event rate was higher (17·0%, 95% CI 13·6–20·3).

CH and 52 had an FND during follow-up. In this subgroup, the magnitude, direction, and consistency of associations of the core and putative predictors with the secondary outcome were similar to the primary analysis (table 2; figure 2; appendix pp 8–10 and 16–17), but the event rate was higher (17·0%, 95% CI 13·6–20·3). Discussion To the best of our knowledge, this is the largest analysis of the clinical course of untreated CCMs so far, in which we found that brainstem CCM location and CCM presentation with ICH or FND were independently associated with the occurrence of ICH after diagnosis of CCMs, whereas age, sex, and CCM multiplicity did not contribute any additional prognostic information. The risk of ICH during 5 years of follow-up differed significantly according to the possession of either, or both, of these risk factors; annual ICH incidence decreased over time in people initially presenting with ICH or FND.

reas age, sex, and CCM multiplicity did not contribute any additional prognostic information. The risk of ICH during 5 years of follow-up differed significantly according to the possession of either, or both, of these risk factors; annual ICH incidence decreased over time in people initially presenting with ICH or FND. This study has the following strengths. The sample size of the pooled cohort was large and contained unpublished data in addition to data from people who were already described in published studies. There was 100% completeness for data on baseline covariates (appendix p 4). We ensured that outcome event definitions and statistical methods of survival analysis were consistent across the collaborating cohorts. The duration and completeness of follow-up enabled us to construct 5-year survival curves with more precise estimations of associations and event incidences than previously possible. The risk of bias in the participating cohorts was low, and the findings were consistent in retrospective hospital-based cohorts and prospective population-based cohorts (figure 2). These strengths in design are reinforced by robustness of the results when compared with findings from individual cohorts for which event rates of variable magnitude were estimated and for which findings of associations between outcome and our core and putative predictors have been inconsistent.11

d cohorts (figure 2). These strengths in design are reinforced by robustness of the results when compared with findings from individual cohorts for which event rates of variable magnitude were estimated and for which findings of associations between outcome and our core and putative predictors have been inconsistent.11 This study has some limitations. Not all of the eligible published cohorts had retained data to share in this collaborative analysis. Outcome data were available on both ICH and FND for only 40% of the included people, but these data were sufficient to quantify event rates with reasonable precision and show associations between predictors and outcomes that were consistent with the primary analysis. Our statistical analysis assumed that censoring was non-informative, but even if informative censoring did occur (ie, because of treatment of CCMs that were in the brainstem or that presented with ICH or FND),32 it happened infrequently. We set a lower limit of 16 years for age at diagnosis, to focus this study on patients who were referred to neurologists and neurosurgeons who care for adults (ie, age at least 16 years), resulting in the exclusion of data on 46 (3%) of the 1666 people available in the collaborating cohorts; however, the proportion of events in people aged 16–18 years was comparable with the rest of the cohort (data not shown).

e referred to neurologists and neurosurgeons who care for adults (ie, age at least 16 years), resulting in the exclusion of data on 46 (3%) of the 1666 people available in the collaborating cohorts; however, the proportion of events in people aged 16–18 years was comparable with the rest of the cohort (data not shown). The main implication of our findings for clinical practice is that people with CCMs can be stratified into four groups to predict the 5-year risk of ICH. These risks can inform decisions about CCM treatment, by indirect comparison with estimates of the effects of treatment. Future research is needed to externally validate our prognostic model and establish whether other factors (eg, genotype, pregnancy, and statin or antithrombotic drug use) are independently associated with ICH in addition to the two core predictors; even larger sample sizes will be needed for this research. Furthermore, long-term risks remain to be quantified for a disorder that is often diagnosed in young people, but has only been well recognised by MRI since the 1980s.4 Supplementary Material Supplementary appendix

The main implication of our findings for clinical practice is that people with CCMs can be stratified into four groups to predict the 5-year risk of ICH. These risks can inform decisions about CCM treatment, by indirect comparison with estimates of the effects of treatment. Future research is needed to externally validate our prognostic model and establish whether other factors (eg, genotype, pregnancy, and statin or antithrombotic drug use) are independently associated with ICH in addition to the two core predictors; even larger sample sizes will be needed for this research. Furthermore, long-term risks remain to be quantified for a disorder that is often diagnosed in young people, but has only been well recognised by MRI since the 1980s.4 Supplementary Material Supplementary appendix Acknowledgments This study was supported by the Medical Research Council (G84/5176, G108/613, G1002605) and the Edinburgh Hub for Trials Methodology Research (G0800803); the Chief Scientist Office of the Scottish Government (K/MRS/50/C2704 and CZB/4/35); and the Stroke Association (TSA04/01). We thank Rosemary Anderson, Aidan Hutchison, and all the people in the Scottish Audit of Intracranial Vascular Malformations; and Gail Nixon, clinical nurse coordinator, and Alex Kostynskyy, clinical research associate, of the Toronto Brain Vascular Malformation study group.

/35); and the Stroke Association (TSA04/01). We thank Rosemary Anderson, Aidan Hutchison, and all the people in the Scottish Audit of Intracranial Vascular Malformations; and Gail Nixon, clinical nurse coordinator, and Alex Kostynskyy, clinical research associate, of the Toronto Brain Vascular Malformation study group. Contributors MAH, GDM, and RA-SS designed the study. RA-SS did the primary search of the published work; MAH did a similar search and found no additional articles. MAH, KDF, I-CS, CS, JPJ, DL, SSM, PW, RA, W-SC, CWO, ZW, J-TZ, JEK, KtB, RW, RDB, and RA-SS collected the data. MAH checked, analysed, and interpreted the data according to a protocol and statistical analysis plan developed and approved by the authors involved in the first iteration of this meta-analysis (MAH, KDF, I-CS, CS, SSM, TJC, RA, KtB, RW, RDB, GDM, and RA-SS). MAH and RA-SS drafted the manuscript and generated the figures, and all coauthors reviewed the final version.

ccording to a protocol and statistical analysis plan developed and approved by the authors involved in the first iteration of this meta-analysis (MAH, KDF, I-CS, CS, SSM, TJC, RA, KtB, RW, RDB, GDM, and RA-SS). MAH and RA-SS drafted the manuscript and generated the figures, and all coauthors reviewed the final version. Cerebral Cavernous Malformations Individual Patient Data Meta-analysis Collaborators Scottish Audit of Intracranial Vascular Malformations steering committee R Al-Shahi Salman (NHS Lothian, Edinburgh), S Baird (NHS National Services Scotland, Edinburgh), J J Bhattacharya (NHS Greater Glasgow and Clyde, Glasgow), C E Counsell (NHS Grampian, Aberdeen), E J St George (NHS Greater Glasgow and Clyde, Glasgow), P M White (Newcastle Hospitals NHS Foundation Trust, Newcastle upon Tyne), V Ritchie (Fauldhouse Health Centre, Fauldhouse until retirement in 2013) R C Roberts (Dundee University, until retirement in 2013), R J Sellar (Edinburgh University until retirement in 2013), and C P Warlow (Edinburgh University, until retirement in 2008). Mayo Clinic Intracerebral Cavernous Malformation study group (Rochester, MN, USA) K D Flemming, R D Brown Jr, M J Link, T J Christianson. Toronto Brain Vascular Malformation study group (Toronto Western Hospital, Toronto, ON, Canada) K ter Brugge, R Willinsky, R Agid, M Tymianski, M C Wallace, C Stapf. Hôpital Lariboisière Cerebral Cavernous Malformation study group (Paris, France): D Hervé, F Riant, H M Schneble.

Mayo Clinic Intracerebral Cavernous Malformation study group (Rochester, MN, USA) K D Flemming, R D Brown Jr, M J Link, T J Christianson. Toronto Brain Vascular Malformation study group (Toronto Western Hospital, Toronto, ON, Canada) K ter Brugge, R Willinsky, R Agid, M Tymianski, M C Wallace, C Stapf. Hôpital Lariboisière Cerebral Cavernous Malformation study group (Paris, France): D Hervé, F Riant, H M Schneble. Seoul National University Hospital Cerebral Cavernous Malformation study group (Seoul, South Korea) J P Jeon, J E Kim, Y S Chung, S Oh, J H Ahn, W-S Cho, Y-J Son, J S Bang, H-S Kang, C-H Sohn, C W Oh. Beijing Tiantan Hospital Cerebral Cavernous Malformation study group (Beijing, China) D Li, S-Y Hao, G-J Jia, Z Wu, L-W Zhang, J-T Zhang. Declaration of interests MAH reports grants from the Edinburgh Hub for Trials Methodology Research. GDM reports grants from the Medical Research Council, the Edinburgh Hub for Trials Methodology Research, the Chief Scientist Office of the Scottish Government, and the Stroke Association. KDF, I-CS, CS, JPJ, DL, SSM, PW, TJC, RA, W-SC, CWO, ZW, J-TZ, JEK, KtB, RW, RDB, and RA-SS declare no competing interests. Figure 1 Study flowchart CCM=cerebral cavernous malformation. *See appendix pp 18–19. †See appendix pp 19–20 for references. ‡One eligible study provided data from two time periods, which are included as two separate cohorts. §See appendix p 20. Figure 2 Forest plots of associations between mode of presentation and cerebral cavernous malformation location with primary and secondary outcomes

CCM=cerebral cavernous malformation. *See appendix pp 18–19. †See appendix pp 19–20 for references. ‡One eligible study provided data from two time periods, which are included as two separate cohorts. §See appendix p 20. Figure 2 Forest plots of associations between mode of presentation and cerebral cavernous malformation location with primary and secondary outcomes Plots show cohort-level and pooled estimates of associations between ICH or FND at presentation (A and C) or brainstem CCM location (B and D) and outcome during 5 years of follow-up. The area of each shaded box is proportional to the weight of the cohort it represents. CCM=cerebral cavernous malformation. FND=focal neurological deficit. HR=hazard ratio. ICH=intracranial haemorrhage. Figure 3 Kaplan-Meier plots of progression to intracranial haemorrhage or to intracranial haemorrhage or focal neurological deficit Plots show the proportion of people progressing to ICH (A) or ICH or FND (B) during follow-up, stratified by ICH or FND presentation from brainstem CCMs, ICH or FND presentation from non-brainstem CCMs, other presentation from brainstem CCMs, and other presentation from non-brainstem CCMs. CCM=cerebral cavernous malformation. FND=focal neurological deficit. HR=hazard ratio. ICH=intracranial haemorrhage. Table 1 Baseline and follow-up characteristics

Plots show the proportion of people progressing to ICH (A) or ICH or FND (B) during follow-up, stratified by ICH or FND presentation from brainstem CCMs, ICH or FND presentation from non-brainstem CCMs, other presentation from brainstem CCMs, and other presentation from non-brainstem CCMs. CCM=cerebral cavernous malformation. FND=focal neurological deficit. HR=hazard ratio. ICH=intracranial haemorrhage. Table 1 Baseline and follow-up characteristics Mode of presentation leading to CCM diagnosis Incidental (n=461) Seizure (n=337) ICH (n=576) FND (n=246) Total (n=1620) Age at diagnosis (years) 51 (37–62) 42 (30–57) 41 (32–51) 47 (36–60) 45 (33–58) Sex Female 259 (56%) 160 (47%) 310 (54%) 138 (56%) 867 (54%) Male 202 (44%) 177 (53%) 266 (46%) 108 (44%) 753 (46%) More than one CCM 77 (17%) 70 (21%) 90 (16%) 45 (18%) 282 (17%) Primary CCM location Lobar 300 (65%) 289 (86%) 154 (27%) 69 (28%) 812 (50%) Deep 46 (10%) 18 (5%) 41 (7%) 24 (10%) 129 (8%) Cerebellum 50 (11%) 15 (4%) 22 (4%) 17 (7%) 104 (6%) Brainstem 65 (14%) 15 (4%) 359 (62%) 136 (55%) 575 (35%) CCM management Surgery or stereotactic radiosurgery 28 (6%) 77 (23%) 172 (30%) 35 (14%) 312 (19%) Conservative management 433 (94%) 260 (77%) 404 (70%) 211 (86%) 1308 (81%) First outcome event during untreated follow-up* ICH 12 (3%) 12 (4%) 151 (26%) 29 (12%) 204 (13%) FND 10 (2%) 3 (1%) 18 (3%) 24 (10%) 55 (3%) None 439 (95%) 322 (96%) 407 (71%) 193 (78%) 1361 (84%) Censored follow-up (years) 3·9 (2·0–5·0) 3·6 (1·5–5·0) 3·0 (1·1–5·0) 4·2 (2·1–5·0) 3·5 (1·6–5·0) Data are median (IQR) or number (%). Some percentages do not add up to 100 because of rounding. CCM=cerebral cavernous malformation. FND=non-haemorrhagic focal neurological deficit. ICH=intracranial haemorrhage.

1361 (84%) Censored follow-up (years) 3·9 (2·0–5·0) 3·6 (1·5–5·0) 3·0 (1·1–5·0) 4·2 (2·1–5·0) 3·5 (1·6–5·0) Data are median (IQR) or number (%). Some percentages do not add up to 100 because of rounding. CCM=cerebral cavernous malformation. FND=non-haemorrhagic focal neurological deficit. ICH=intracranial haemorrhage. * 1620 people contributed data on the occurrence of ICH outcomes. 640 people contributed data on the occurrence of the composite outcome of ICH or FND. Table 2 Hazard ratios and estimated 5-year risks of outcome events for core predictors in prognostic models Number of people (%) Number of outcome events during 5-year follow-up Hazard ratio (95% CI)* Estimated 5-year risk (95% CI) Primary outcome: ICH (n=1620) ICH or FND presentation, brainstem CCM location 495 (31%) 135 10·2 (5·0–23·9) 30·8% (26·3–35·2) ICH or FND presentation, other CCM location 327 (20%) 45 5·6 (3·7–9·4) 18·4% (13·3–23·5) Other presentation, brainstem CCM location 80 (5%) 4 1·8 (1·3–2·6) 8·0% (0·1–15·9) Other presentation, other CCM location 718 (44%) 20 Reference 3·8% (2·1–5·5) Secondary outcome: ICH or FND (n=640) ICH or FND presentation, brainstem CCM location 113 (18%) 48 16·3 (5·8–53·7) 50·7% (40·1–61·4) ICH or FND presentation, other CCM location 141 (22%) 24 5·1 (2·9–10·0) 22·4% (14·2–30·6) Other presentation, brainstem CCM location 31 (5%) 5 3·2 (2·0–5·4) 22·9% (3·7–42·2) Other presentation, other CCM location 355 (55%) 11 Reference 3·7% (1·5–5·9) CCM=cerebral cavernous malformation. FND=non-haemorrhagic focal neurological deficit. ICH=intracranial haemorrhage.

1 (22%) 24 5·1 (2·9–10·0) 22·4% (14·2–30·6) Other presentation, brainstem CCM location 31 (5%) 5 3·2 (2·0–5·4) 22·9% (3·7–42·2) Other presentation, other CCM location 355 (55%) 11 Reference 3·7% (1·5–5·9) CCM=cerebral cavernous malformation. FND=non-haemorrhagic focal neurological deficit. ICH=intracranial haemorrhage. * Bootstrapped 95% CIs: 10 000 samples with replacement.

Introduction In experimental trials, objective, responsive, and valid outcome measures are needed to test treatment efficacy. Neuromuscular disorders are common,1, 2 disabling, muscle wasting disorders, largely without proven therapy. The increasing identification of tractable targets that suggest new therapies drives a concomitant need for effective trial outcome measures.3 Responsive outcome measurement in neuromuscular disorders is challenging: longitudinal progression is typically slow, so disease progression might be masked by age-related changes4 or measurement variation, and new therapies are more likely to halt or slow progression than reverse established tissue damage. This limits outcome measure responsiveness, expressed as the standardised response mean (SRM);5 SRM is a key determinant of study power, with an inverse square relation to required sample size for a stated statistical power by Lehr's formula (panel 1).6 Outcome measure validity in terms of correlation against measures held to be valid markers of patient function or experience, such as relevant neuromuscular disorder functional rating scales, must also be established.

e square relation to required sample size for a stated statistical power by Lehr's formula (panel 1).6 Outcome measure validity in terms of correlation against measures held to be valid markers of patient function or experience, such as relevant neuromuscular disorder functional rating scales, must also be established. The shortcomings of established outcome measures were exemplified in the UK–Italian ascorbic acid trial6 in patients with Charcot-Marie-Tooth disease 1A. The trial showed no treatment benefit, and the placebo group natural history data showed only low primary outcome measure responsiveness (Charcot-Marie-Tooth disease Neuropathy Score SRM=0·19). The study therefore might have been underpowered, with the negative result suggesting type II error. Increasing group sizes to compensate for low responsiveness is challenging in rare diseases, and a pressing need exists to identify and validate adequately responsive outcome measures for neuromuscular disorder clinical trials.7

therefore might have been underpowered, with the negative result suggesting type II error. Increasing group sizes to compensate for low responsiveness is challenging in rare diseases, and a pressing need exists to identify and validate adequately responsive outcome measures for neuromuscular disorder clinical trials.7 Acute or early pathological muscle changes, whether due to denervation, dystrophy, or inflammation, typically involve tissue water changes, whereas chronic disease is characterised by intramuscular fat accumulation. Intramuscular fat accumulation is the final common pathological pathway for many different primary genetic and acquired neuromuscular disorders. MRI Dixon fat-water imaging,8 which quantifies tissue fat content on a 0–100% fat-fraction scale, has been applied to skeletal muscle in neuromuscular disorders.9, 10, 11, 12, 13 Muscle MRI transverse magnetisation relaxation time (T2) and magnetisation transfer ratio (MTR) are sensitive to changes in muscle water distribution14, 15 and lipid content. Potential demonstrated in these studies to quantify both early or acute and chronic pathology non-invasively and objectively suggests these markers might be ideal candidate outcome measures for neuromuscular disorders. We previously showed their reliability and normative age dependencies and sex dependencies in healthy volunteers.16 However, before application in trials, their validity and responsiveness as disease markers in specific neuromuscular disorders must be established.

candidate outcome measures for neuromuscular disorders. We previously showed their reliability and normative age dependencies and sex dependencies in healthy volunteers.16 However, before application in trials, their validity and responsiveness as disease markers in specific neuromuscular disorders must be established. We aimed to test the hypothesis that these indices are valid and responsive markers of disease progression using clinical, myometric, and MRI assessments at baseline and 12 months. We studied two deeply phenotyped cohorts with Charcot-Marie-Tooth disease 1A and inclusion body myositis—representing neurogenic and myopathic, and genetic and acquired, conditions with variable disease progression rates—and age-matched and sex-matched controls. We aimed to examine the validity of MRI-quantified intramuscular fat accumulation as an outcome measure by direct correlation with clinical and myometric measures and establish the sensitivity of MRI measures to early muscle water changes before substantial intramuscular fat accumulation.

ched and sex-matched controls. We aimed to examine the validity of MRI-quantified intramuscular fat accumulation as an outcome measure by direct correlation with clinical and myometric measures and establish the sensitivity of MRI measures to early muscle water changes before substantial intramuscular fat accumulation. Methods Study design and participants We did a prospective longitudinal observational study of patients attending the inherited neuropathy or muscle clinics at the Medical Research Council (MRC) Centre for Neuromuscular Diseases at the National Hospital for Neurology and Neurosurgery, London, UK (figure 1). Inclusion criteria were genetic confirmation of the chromosome 17p11·2 duplication for patients with Charcot-Marie-Tooth disease 1A, and classification as pathologically or clinically definite by MRC criteria17 for patients with inclusion body myositis, and being aged 17 years or older (as the study was based in an adult hospital). Exclusion criteria were concomitant diseases and safety-related MRI contraindications (appendix). Patients meeting these inclusion criteria and attending the clinics were invited to participate. Healthy participants were enrolled as controls in two overlapping subgroups comprising 20 participants each, one matched for age, sex, weight, and body-mass index distribution to the Charcot-Marie-Tooth disease 1A group and one matched for the same variables to the inclusion body myositis group. Controls were recruited in the first instance from spouses, friends, and relatives of the patients, who underwent assessments on the same day as the respective patient (relatives of patients with Charcot-Marie-Tooth disease 1A were included only if genetic testing was negative). Because patients with inclusion body myositis group are older men (usually older than 45 years),18 the controls recruited were not well matched to those patients, and the youngest six individuals in the inclusion body myositis control subgroup were replaced by six healthy individuals recruited from hospital and research staff whose ages fell within the specific demographics of the inclusion body myositis group. Clinical assessments and myometry were done at the MRC Centre for Neuromuscular Diseases, UCL Institute of Neurology, UK. MRI was done at the National Hospital for Neurology and Neurosurgery, London, UK.

m hospital and research staff whose ages fell within the specific demographics of the inclusion body myositis group. Clinical assessments and myometry were done at the MRC Centre for Neuromuscular Diseases, UCL Institute of Neurology, UK. MRI was done at the National Hospital for Neurology and Neurosurgery, London, UK. Cross-sectional MRI data from all of the control participants, combined with data from other participants scanned using the same MRI protocol, have been reported previously.16 The study was approved by the local ethics committee and all participants provided written informed consent at enrolment. Procedures The following assessments were done in all patients: medical history, neurological examination, bedside strength examination using MRC grading (appendix), and SF-36 quality of life questionnaire.19 Patients with Charcot-Marie-Tooth disease 1A were assessed with the Charcot-Marie-Tooth examination score (version 2; CMTES),20 a 28-point score based on signs and symptoms where 28 is very severely affected. Patients with inclusion body myositis were assessed with the inclusion body myositis functional rating scale (IBMFRS), a 40-point score where 0 is most functionally impaired.21 Patients and controls underwent detailed lower limb myometry on a HUMAC NORM dynamometer (CSMi, MA, USA; see appendix for full assessment protocol).

nts with inclusion body myositis were assessed with the inclusion body myositis functional rating scale (IBMFRS), a 40-point score where 0 is most functionally impaired.21 Patients and controls underwent detailed lower limb myometry on a HUMAC NORM dynamometer (CSMi, MA, USA; see appendix for full assessment protocol). Participants were examined lying feet-first and supine in a 3T MRI scanner (TIM Trio, Siemens, Erlangen, Germany). The quantitative MRC Centre MRI protocol was developed and included fat fraction, transverse relaxation time (T2), and magnetisation transfer ratio (MTR) measurement. These parameters were analysed during the 12-month follow-up, by measuring correlation with functionally relevant clinical measures, and for T2 and MTR, sensitivity in muscles with fat fraction less than the 95th percentile of the control group. The following measurements were done: 3-point-Dixon fat-fraction quantification resulting in fat-fraction maps expressed as percentage fat (0–100%),8 non-fat-suppressed T2 measurement by dual-contrast turbo-spin-echo imaging, and MTR imaging requiring two 3D-FLASH images with and without a magnetisation transfer pre-pulse with radiofrequency field (B1) non-uniformity correction resulting in MTR maps with values expressed by convention in percentage units (0–100 pu).22 Sequence details and variables are listed in the appendix; the total acquisition time per participant per visit was about 35 min.

ithout a magnetisation transfer pre-pulse with radiofrequency field (B1) non-uniformity correction resulting in MTR maps with values expressed by convention in percentage units (0–100 pu).22 Sequence details and variables are listed in the appendix; the total acquisition time per participant per visit was about 35 min. Statistical analysis The number of patients recruited from each of the disease groups was based pragmatically on the number of patients known to have the disease of interest and eligible for enrolment. 20 patients in each disease group were deemed sufficient to identify trends in MRI variables and provide data useful to inform statistical power estimates in future studies. A single observer (a radiologist with 4 years of specialist experience in neuromuscular imaging) masked to all clinical details including diagnosis defined whole muscle and small regions of interest (ROIs) for each participant on single slices (appendix) for all muscles at mid-thigh and mid-calf levels from an unprocessed Dixon acquisition (echo time=3·45 ms) using ITK-SNAP software (figure 2).23

muscular imaging) masked to all clinical details including diagnosis defined whole muscle and small regions of interest (ROIs) for each participant on single slices (appendix) for all muscles at mid-thigh and mid-calf levels from an unprocessed Dixon acquisition (echo time=3·45 ms) using ITK-SNAP software (figure 2).23 The small ROIs were transferred to the co-registered fat fraction, T2, and MTR variable maps. For each muscle, mean T2, mean fat fraction, and mean MTR were recorded from the small ROIs, and mean fat fraction and muscle cross-sectional area (CSA) were recorded from the whole muscle ROIs. A systematic bias was noted in T2 measurements after a routine scanner software upgrade necessitating a minor correction to post-upgrade values (appendix). For each participant, individual muscle values were combined into a summary measure calculated for each variable for all muscles (left and right limb) at thigh level and at calf level, and for relevant functional muscle groups (quadriceps, hamstrings, anterior tibial compartment, and triceps surae) for left and right limbs separately. Summary measures were calculated for small ROIs as the mean of the individual muscle values, and for whole muscle ROIs as the mean of individual muscle values weighted by cross-sectional area. To assess early pathological changes in patients' muscles for which intramuscular fat accumulation lay within the normal healthy range, additional analyses included only muscles with mean fat fraction less than the 95th percentile of the healthy control data for thigh and calf muscles separately. As a measure of the functional muscle CSA, the metric remaining muscle area was calculated with the following equation, where CSA and fat fraction refer to the whole muscle ROI values:

ed only muscles with mean fat fraction less than the 95th percentile of the healthy control data for thigh and calf muscles separately. As a measure of the functional muscle CSA, the metric remaining muscle area was calculated with the following equation, where CSA and fat fraction refer to the whole muscle ROI values: Remaining muscle area=CSA×(100-fat fraction)100% Longitudinal changes were quantified on a muscle-by-muscle, variable-by-variable basis, and combined in the same way as for cross-sectional data to create separate all-muscle summary variables at thigh level and calf level. Statistical analyses were done with IBM-SPSS Statistics (version 20). At baseline, cross-sectional differences between each patient group and their respective matched controls were assessed for each measure using two-tailed t tests. Correlations between measures at baseline were assessed with Pearson or Spearman correlation as appropriate. Missing data were excluded from the analysis. Differences between baseline and follow-up values for each measure for each patient group were assessed with paired t tests, and two-tailed t tests compared absolute change over 12 months between the patient and respective matched-control groups. Multivariate linear regressions were done to assess the dependence of T2 and MTR upon disease group membership in muscles without substantial intramuscular fat accumulation (defined as fat fraction less than the control group 95th percentile), while adjusting for the influence of residual fat fraction as a covariate.

. Multivariate linear regressions were done to assess the dependence of T2 and MTR upon disease group membership in muscles without substantial intramuscular fat accumulation (defined as fat fraction less than the control group 95th percentile), while adjusting for the influence of residual fat fraction as a covariate. Role of the funding source The study funders had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author and all coauthors had full access to all the data in the study and had final responsibility for the decision to submit for publication.

. Multivariate linear regressions were done to assess the dependence of T2 and MTR upon disease group membership in muscles without substantial intramuscular fat accumulation (defined as fat fraction less than the control group 95th percentile), while adjusting for the influence of residual fat fraction as a covariate. Role of the funding source The study funders had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author and all coauthors had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between Jan 19, 2010, and July 7, 2011, we recruited 20 patients with Charcot-Marie-Tooth disease 1A, 20 patients with inclusion body myositis, and 29 healthy controls allocated to one or both of the 20-participant control subgroups (figure 1). Age, sex, height, and weight were similar between the two patient groups and their respective matched controls (table 1). In both patient groups, baseline myometric muscle strength was significantly reduced in all muscle groups compared with their matched controls (table 1, appendix). At baseline, patients with inclusion body myositis had significantly higher all-muscle fat fraction and T2, and lower all-muscle MTR, at both thigh level and calf level than matched controls (table 1, figure 3). The all-muscle CSA in the inclusion body myositis group was lower than that of matched controls at thigh level but not at calf level (table 1). Patients in the Charcot-Marie-Tooth disease 1A group had significantly higher all-muscle fat fraction and T2 and lower MTR than matched controls at calf level but not at thigh level, whereas their all-muscle CSA was lower than for the controls at calf level but not at thigh level (table 1, figure 3).

Repeat assessments (table 2) were done after a mean of 12·4 months (SD 1·0). In the inclusion body myositis group, the MRC lower limb score, inclusion body myositis functional rating scale (IBMFRS), and overall myometric knee extension strength deteriorated significantly during follow-up. In the Charcot-Marie-Tooth disease 1A group, none of the clinical measures changed significantly. Although myometric ankle dorsiflexion strength showed a significant increase (p=0·049) compared with baseline, this was not significant (p=0·24) compared with the change in matched controls. Quantitative MRI values for individual muscles are given in the appendix. Strong correlations were noted between the three quantitative MRI measures (fat fraction, MTR, and T2) of individual muscles (all r>0·89, p<0·0001, appendix). In patients with inclusion body myositis, at 12-month follow-up all-muscle thigh-level and calf-level fat fraction (figure 4), for both small and whole muscle ROIs, and calf muscle T2 had increased significantly from baseline, whereas calf muscle MTR significantly decreased (table 2). All-muscle CSA significantly decreased compared with baseline (p=0·01) at calf level but not at thigh level (p=0·08). In patients with Charcot-Marie-Tooth disease 1A, calf level all-muscle fat fraction increased significantly during the 12-month follow-up (figure 4), whereas T2, MTR, and muscle CSA did not change significantly (table 3), nor did any MRI measures at thigh level. In the control groups, no significant 12-month changes in fat fraction, T2, or MTR were reported at either level.

vel all-muscle fat fraction increased significantly during the 12-month follow-up (figure 4), whereas T2, MTR, and muscle CSA did not change significantly (table 3), nor did any MRI measures at thigh level. In the control groups, no significant 12-month changes in fat fraction, T2, or MTR were reported at either level. For the 29 control participants, the fat-fraction 95th percentiles were 4·8% for thigh muscles and 4·7% for calf-level muscles, defining the upper thresholds for intramuscular fat accumulation in these healthy controls. On regression analyses done separately for the inclusion body myositis and Charcot-Marie-Tooth disease 1A groups, combined with their respective controls, including only data from all individual patient and control muscles with fat fraction lower than these thresholds, T2 and MTR remained significantly dependent on fat fraction (appendix). However, the dependencies for control versus disease status were also significant in each case (appendix), with increased T2 (thigh 4·0 ms [SE 0·5], calf 3·5 ms [0·6]) and reduced MTR (thigh −1·5 pu [0·2], calf −1·1 pu [0·2]) independent of fat fraction in this regression model, in patients with inclusion body myositis versus matched controls. Smaller but significant fat-fraction-adjusted dependencies were also reported in patients with Charcot-Marie-Tooth disease 1A versus matched-controls, especially at calf level (T2 thigh 1·0 ms [0·3], calf 2·0 ms [0·3]; MTR thigh −0·3 pu [0·1], calf −0·7 pu [0·1]). The distribution of T2 and MTR values in muscles without substantial intramuscular fat accumulation in patients with inclusion body myositis is shown (figure 3).

ie-Tooth disease 1A versus matched-controls, especially at calf level (T2 thigh 1·0 ms [0·3], calf 2·0 ms [0·3]; MTR thigh −0·3 pu [0·1], calf −0·7 pu [0·1]). The distribution of T2 and MTR values in muscles without substantial intramuscular fat accumulation in patients with inclusion body myositis is shown (figure 3). In all groups, for all movements assessed, muscle strength correlated strongly with the corresponding muscle group total CSA (appendix). In regions where patients' fat fraction was significantly increased (inclusion body myositis thigh and calf level, Charcot-Marie-Tooth 1A calf level), whole muscle fat-fraction correlated negatively with muscle strength (appendix). The combined variable remaining muscle area correlated more strongly with strength than with either CSA or fat fraction separately (figure 3, appendix).

sed (inclusion body myositis thigh and calf level, Charcot-Marie-Tooth 1A calf level), whole muscle fat-fraction correlated negatively with muscle strength (appendix). The combined variable remaining muscle area correlated more strongly with strength than with either CSA or fat fraction separately (figure 3, appendix). Significant correlations were also noted between summary clinical measures and all-muscle fat fraction measurements. In patients with inclusion body myositis, all-muscle thigh-level fat fraction correlated with disease duration (ρ=0·50, p=0·03, appendix), IBMFRS (ρ=–0·53, p=0·02), IBMFRS lower limb components (ρ=–0·64, p=0·002, figure 3), total MRC lower limb score (ρ=–0·60, p=0·005), and the physical function domain of the Short-Form 36 Quality of Life Score (SF36-PF; ρ=–0·60, p=0·007). Overall SF36 (ρ=–0·11, p=0·66) and age (ρ=0·14, p=0·55) were not significantly correlated with fat fraction. In patients with Charcot-Marie-Tooth 1A disease, all-muscle calf-level fat fraction correlated with disease duration (ρ=0·89, p<0·0001, appendix), age (ρ=0·84, p<0·0001), CMTES (ρ=0·63, p=0·003, appendix), lower limb motor component of the CMTES (ρ=0·77, p<0·0001), and reduced total MRC lower limb score (ρ=–0·76, p<0·0001). The correlation between all-muscle calf fat fraction and total SF36 score was not significant (ρ=–0·34, p=0·18); however, the correlation with the SF36-PF was significant (ρ=–0·63, p=0·007).

ppendix), lower limb motor component of the CMTES (ρ=0·77, p<0·0001), and reduced total MRC lower limb score (ρ=–0·76, p<0·0001). The correlation between all-muscle calf fat fraction and total SF36 score was not significant (ρ=–0·34, p=0·18); however, the correlation with the SF36-PF was significant (ρ=–0·63, p=0·007). Correlations between the change in clinical variables and the change in MRI variables were not significant in the Charcot-Marie-Tooth group. However, in patients with inclusion body myositis, the change in myometric strength of knee extension correlated with change in several MRI variables, including quadriceps remaining muscle area (right r=0·66, p=0·005, left r=0·81, p=0·0001; figure 4). The highest SRMs, greater than 1, were for whole muscle fat fraction at thigh and calf level, and T2 at calf level, whereas no clinical variables had an SRM greater than 1 (table 2). In the CMT group, significant 12-month change occurred only in whole muscle fat fraction at calf level, with an SRM of 0·83 (table 3).

Correlations between the change in clinical variables and the change in MRI variables were not significant in the Charcot-Marie-Tooth group. However, in patients with inclusion body myositis, the change in myometric strength of knee extension correlated with change in several MRI variables, including quadriceps remaining muscle area (right r=0·66, p=0·005, left r=0·81, p=0·0001; figure 4). The highest SRMs, greater than 1, were for whole muscle fat fraction at thigh and calf level, and T2 at calf level, whereas no clinical variables had an SRM greater than 1 (table 2). In the CMT group, significant 12-month change occurred only in whole muscle fat fraction at calf level, with an SRM of 0·83 (table 3). Discussion In this study, quantitative MRI measures changed significantly during the 12-month follow-up in calf muscles of patients with Charcot-Marie-Tooth disease 1A and both thigh and calf muscles in patients with inclusion body myositis. MRI-measured fat fraction showed greater responsiveness (higher SRM) than clinical or myometric measures. Fat fraction correlated with strength and function. Even after adjustment for fat fraction, T2 was increased and MTR was reduced in muscles without substantial intramuscular fat accumulation in both patient groups compared with controls, suggesting sensitivity to early and potentially reversible changes in muscle water distribution. MRI therefore provides responsive outcome measures with validity suitable for application in future clinical trials of neuromuscular disorders.

l intramuscular fat accumulation in both patient groups compared with controls, suggesting sensitivity to early and potentially reversible changes in muscle water distribution. MRI therefore provides responsive outcome measures with validity suitable for application in future clinical trials of neuromuscular disorders. The development of responsive outcome measures has proven especially difficult so far in neuromuscular disorders such as Charcot-Marie-Tooth 1A. The CMTNS was selected at the 136th European Neuromuscular Centre workshop24 as the most appropriate primary outcome measure for Charcot-Marie-Tooth 1A trials and was used in the ascorbic acid trials in adults.6, 25, 26 The SRM of the CMTNS and other outcome measures in the largest of these trials6 together with 5-year natural history data4 are shown in the appendix. On average, a 0·3 points per year increase in CMTNS is reported, resulting in minimal responsiveness over 2 years and small responsiveness over 5 years. Neurophysiology showed either no significant change from baseline6 or change no greater than seen in controls4 so seems unsuitable as an outcome measure. Ankle dorsiflexion showed mild responsiveness during 2 years using a custom built frame but no responsiveness during 5 years with hand-held myometry and in our study an apparent improvement in Charcot-Marie-Tooth 1A and controls was probably due to a learning effect. By marked contrast, in this study, calf-level MRI-quantified fat fraction showed large responsiveness (SRM=0·83) with a highly significant (p=0·002) increase in fat fraction at 12 months. This finding has important implications for future trial design. For a hypothetical Charcot-Marie-Tooth 1A treatment trial powered to detect a 50% reduction in disease progression during a 1-year period with 80% power at p<0·05 significance, the number of patients needed in active and placebo groups27 would be roughly 93 with calf muscle MRI-determined fat fraction as the primary outcome measure, as opposed to 7700 patients for the equivalent statistical power with CMTNS as the outcome. MRI-quantified calf muscle fat fraction is therefore the most responsive outcome measure proposed so far in Charcot-Marie-Tooth 1A.

would be roughly 93 with calf muscle MRI-determined fat fraction as the primary outcome measure, as opposed to 7700 patients for the equivalent statistical power with CMTNS as the outcome. MRI-quantified calf muscle fat fraction is therefore the most responsive outcome measure proposed so far in Charcot-Marie-Tooth 1A. Progression in inclusion body myositis is somewhat faster than in Charcot-Marie-Tooth 1A, with a 10-year median interval between symptom onset and significant disability.28 By contrast with the Charcot-Marie-Tooth 1A group, in the inclusion body myositis group, significant change during the 12-month follow-up was detected in IBMFRS, knee extension strength, MRC scores, myometry-measured knee extension, and many MRI measures. However, only MRI measure SRMs exceeded 1. Furthermore, quadriceps remaining muscle area correlated with change in knee extension strength (figure 4), showing for the first time a direct link between MRI-detected change and functional deficit. Thus, although MRI detected significant fat fraction progression during 12 months in both disease groups, in the more slowly progressive Charcot-Marie-Tooth 1A, calf-level MRI-measured fat fraction was the only measure to change significantly. In the more progressive inclusion body myositis, significant changes in several MRI, and some clinical and myometric, measures showed that the MRI indices were more responsive.

oups, in the more slowly progressive Charcot-Marie-Tooth 1A, calf-level MRI-measured fat fraction was the only measure to change significantly. In the more progressive inclusion body myositis, significant changes in several MRI, and some clinical and myometric, measures showed that the MRI indices were more responsive. Additionally to responsiveness, outcome measures should show validity through correlation to relevant patient function. We showed strong clinical-MRI correlations for both overall participant and individual muscle measures. In both Charcot-Marie-Tooth 1A and inclusion body myositis, strong correlations exist between overall MRI measures and quality of life indices, functional or composite scales (IBMFRS and CMTES), and bedside strength examination, consistent with previous research.9, 13, 29, 30 At the individual muscle level, the intuitively expected correlation between CSA and strength in healthy controls was shown here, similar to a previous study.31 In both sets of patients, negative correlation between strength and fat fraction was noted, also shown previously in myotonic dystrophy for ankle dorsiflexion9 and in this study for both ankle and knee movements in both diseases. Muscle CSA and fat infiltration, when combined as remaining muscle area, correlated most strongly with strength. This might be superior to simple T1 and T2 measurements, which failed to show a correlation with strength in a study of patients with poliomyelitis.32 Thus, MRI provides indices of chronic muscle pathology that are highly correlated to muscle strength, but independent of participant effort or operator involvement, which lead to poor test-retest and interobserver reliability33 of direct muscle strength measures. MRI, therefore, provides a valid, reliable16 surrogate measure of muscle strength.

of chronic muscle pathology that are highly correlated to muscle strength, but independent of participant effort or operator involvement, which lead to poor test-retest and interobserver reliability33 of direct muscle strength measures. MRI, therefore, provides a valid, reliable16 surrogate measure of muscle strength. Because the T2 of fat greatly exceeds that of muscle water, and vice versa for MTR, T2 and MTR are affected by both fat fraction (exemplified by the strong inter-MRI variable correlations) and potentially independent early tissue water distribution changes. For this reason, T2 obtained in this manner is referred to as total T2 to distinguish from water-specific T2 obtained with methods eliminating the fat signal.34 Our data from muscles with fat fraction below the 95th percentile of the control range suggest that adjustment for T2 and MTR dependence on residual fat fraction shows significant differences between patient and control groups independent of fat fraction. These differences, greater for the inclusion body myositis group but also significant in Charcot-Marie-Tooth 1A, might represent early changes in muscle water distribution occurring before significant intramuscular fat accumulation. These changes might be reversible with therapy,35 and muscle T2 and MTR might thus provide useful biomarkers in clinical trials focused on early or active disease.

gnificant in Charcot-Marie-Tooth 1A, might represent early changes in muscle water distribution occurring before significant intramuscular fat accumulation. These changes might be reversible with therapy,35 and muscle T2 and MTR might thus provide useful biomarkers in clinical trials focused on early or active disease. This study had several limitations. We analysed data from only single slices of thigh and calf blocks, with only small ROI for T2 and MTR sequences. This might have contributed to variation if muscle pathology was anatomically non-homogeneous. To ensure consistency in the volumes of tissue assessed longitudinally, slice positions were defined by measured distance from bony landmarks, a more reliable method than surface anatomy-based slice positioning,36 and follow-up ROIs were drawn with direct reference to the baseline ROI. T2 estimation by sampling two turbo-spin echo images is potentially less accurate than the Carr-Purcell-Meiboom-Gill multiple-spin-echo method.37 However, our approach provides time-efficient, wide-coverage quantification of T2 change using a method that can be implemented on standard MRI systems without specialist modification. Outcome measure SRMs derived from observational studies are applicable to interventional studies only when the intervention will affect the outcome measurement. For example, if an intervention had an effect on muscle that improved strength without an effect on muscle size or quantitative MRI variables, a functionally important benefit might have been missed. This association will need to be established in a disease-specific and intervention-specific manner.

utcome measurement. For example, if an intervention had an effect on muscle that improved strength without an effect on muscle size or quantitative MRI variables, a functionally important benefit might have been missed. This association will need to be established in a disease-specific and intervention-specific manner. Chronic intramuscular fat accumulation is common to a wide range of neuromuscular disorders. Although the precise molecular events responsible for intramuscular fat accumulation are not fully understood, one mechanistic study38 suggests that a range of different primary genetic muscle diseases, and potentially denervation, stimulate muscle precursor stems cells to differentiate into adipose cells and fibroblasts. That intramuscular fat accumulation seems to be a common pathway in many genetic and acquired neuromuscular disorders, underlines its potential usefulness as an outcome measure across neuromuscular disorders. As an example, we have shown here that the same types of change can be quantified in these two different muscle wasting diseases: an acquired late-onset progressive proximal and distal myopathy (inclusion body myositis) and an inherited childhood-onset slowly progressive distal predominant neuropathy (Charcot-Marie-Tooth 1A).

xample, we have shown here that the same types of change can be quantified in these two different muscle wasting diseases: an acquired late-onset progressive proximal and distal myopathy (inclusion body myositis) and an inherited childhood-onset slowly progressive distal predominant neuropathy (Charcot-Marie-Tooth 1A). Although differences between diseases result in different distributions and degrees of muscle abnormalities, their presence, direction, and clinical correlations are consistent (panel 2), making these outcome measures potentially applicable across other neuromuscular disorders with lower limb weakness. Through selection of MRI variables targeted to disease and intervention, MRI outcome measures can be optimised to provide maximum responsiveness for a specific clinical trial. For example, in Charcot-Marie-Tooth 1A, in which the absence of significant MRI abnormalities at thigh level is a reflection of the length-dependent distribution of weakness, the optimum MRI protocol would include assessment of fat infiltration of calf muscles with a sequence such as 3-point Dixon. In inclusion body myositis, to assess whether an intervention reversed acute pathological processes, additional water-sensitive sequences such as T2 or MTR quantification should be included. In this study, T2 and MTR measurements showed similar responses, so the additional benefit of measuring both sequences remains unproven. For any specific neuromuscular disorders, understanding of basic disease mechanisms, disease distribution of the affected muscles, and treatment mechanisms will enable trial-specific selection of MRI outcome measures and appropriate anatomical imaging levels to provide optimum responsiveness.

equences remains unproven. For any specific neuromuscular disorders, understanding of basic disease mechanisms, disease distribution of the affected muscles, and treatment mechanisms will enable trial-specific selection of MRI outcome measures and appropriate anatomical imaging levels to provide optimum responsiveness. In these representative neuromuscular disorders, the comprehensive MRC Centre MRI protocol provides outcome measures closely correlated to strength, function, and disease severity at baseline and longitudinally. In Charcot-Marie-Tooth 1A, responsiveness far exceeded that of existing outcome measures. This finding might allow progress in the design of adequately powered clinical trials in this gradually progressive but debilitating disease. In inclusion body myositis, T2 and MTR showed early changes in muscles before significant intramuscular fat accumulation, providing potential measures of early disease before irreversible changes have occurred. Together, the methods provide objective, non-invasive, valid, responsive outcome measures in inclusion body myositis and Charcot-Marie-Tooth 1A. Supplementary Material Supplementary appendix

In these representative neuromuscular disorders, the comprehensive MRC Centre MRI protocol provides outcome measures closely correlated to strength, function, and disease severity at baseline and longitudinally. In Charcot-Marie-Tooth 1A, responsiveness far exceeded that of existing outcome measures. This finding might allow progress in the design of adequately powered clinical trials in this gradually progressive but debilitating disease. In inclusion body myositis, T2 and MTR showed early changes in muscles before significant intramuscular fat accumulation, providing potential measures of early disease before irreversible changes have occurred. Together, the methods provide objective, non-invasive, valid, responsive outcome measures in inclusion body myositis and Charcot-Marie-Tooth 1A. Supplementary Material Supplementary appendix Acknowledgments This work was supported by the MRC (grant numbers G0601943 2008–2013 and MR/K000608/1 2013–2018). We gratefully acknowledge the capital and research support of the NIHR University College London Hospitals Biomedical Research Centre (2008–2013 and 2013–2018). AF was supported by a grant from the Lorenzo-Piaggio Foundation Switzerland. PMM was supported by a National Institute of Health Research, Translational Research Collaboration Fellowship. JMM was in receipt of a Medical Research Council Centenary award. We thank patient organisations including CMT UK and the Myositis Support Group who made patients aware of this study. We thank all patients who took part in this study.

tional Institute of Health Research, Translational Research Collaboration Fellowship. JMM was in receipt of a Medical Research Council Centenary award. We thank patient organisations including CMT UK and the Myositis Support Group who made patients aware of this study. We thank all patients who took part in this study. Declaration of interests JMM and CDJS report grants from the Medical Research Council during the study. MMR received research support from the Medical Research Council, the National Institutes of Neurological Diseases and Stroke, Office of Rare Diseases (U54NS065712), and CMT UK. AF reports grants from Lorenzo Piaggio Foundation, Roche AG, and Bracco AG outside the submitted work. PMM reports grants from the National Institute for Health Research (NIHR) Rare Diseases Translational Research Collaboration (RD TRC) and from the NIHR University College London Hospitals (UCLH) Biomedical Research Centre, during the conduct of the study. TAY reports grants from MRC and NIHR University College London Hospitals Biomedical Research Centre during the conduct of the study and grants from Biogen Idec, GlaxoSmithKline, and Novartis, outside the submitted work. JST reports grants from MRC, and NIHR University College London Hospitals Biomedical Research Centre during the conduct of the study and grants from Siemens Healthcare and GlaxoSmithKline and Engineering and Physical Sciences Research Council outside the submitted work. MGH reports grants from MRC and NIHR University College London Hospitals Biomedical Research Centre during the conduct of the study.

l Research Centre during the conduct of the study and grants from Siemens Healthcare and GlaxoSmithKline and Engineering and Physical Sciences Research Council outside the submitted work. MGH reports grants from MRC and NIHR University College London Hospitals Biomedical Research Centre during the conduct of the study. Figure 1 Flow chart of patient assessments and dropout CMT1A=Charcot-Marie-Tooth 1A. IBM=inclusion body myositis. *11 controls were common to both disease control groups. †Three controls dropped out because they had undertaken baseline assessments accompanying patients who dropped out before repeat assessments. Figure 2 Regions of interest and sample axial images of left lower limb In each panel the upper row shows the mid-thigh level and the lower row shows the mid-calf level. (A) Unprocessed Dixon sequence (echo time=3·45 ms) in a healthy control (left), with overlaid whole muscle regions of interest (centre) and small regions of interest (right) for the same participant. (B) Fat-fraction map of a participant from each group. (C) Transverse relaxation time (T2) map of a participant from each group. (D) Magnetisation transfer ratio map of a participant from each group. RF=rectus femoris. VL=vastus lateralis. VM=vastus medialis. VI=vastus intermedius. Sa=sartorius. G=gracilis. AM=adductor magnus. SM=semimembranosus. ST=semitendinosus. BF=biceps femoris. TA=tibialis anterior group. MG=medial head of gastrocnemius. So=soleus. TP=tibialis posterior. PL=peroneus longus. LG=lateral head of gastrocnemius. pu=percentage units. Figure 3 Cross-sectional data

In each panel the upper row shows the mid-thigh level and the lower row shows the mid-calf level. (A) Unprocessed Dixon sequence (echo time=3·45 ms) in a healthy control (left), with overlaid whole muscle regions of interest (centre) and small regions of interest (right) for the same participant. (B) Fat-fraction map of a participant from each group. (C) Transverse relaxation time (T2) map of a participant from each group. (D) Magnetisation transfer ratio map of a participant from each group. RF=rectus femoris. VL=vastus lateralis. VM=vastus medialis. VI=vastus intermedius. Sa=sartorius. G=gracilis. AM=adductor magnus. SM=semimembranosus. ST=semitendinosus. BF=biceps femoris. TA=tibialis anterior group. MG=medial head of gastrocnemius. So=soleus. TP=tibialis posterior. PL=peroneus longus. LG=lateral head of gastrocnemius. pu=percentage units. Figure 3 Cross-sectional data (A) Overall fat fraction is significantly increased at both thigh and calf level in patients with inclusion body myositis and at calf level in patients with CMT1A compared with matched controls. Boxes represent median and IQR, whiskers show range, and filled circles are outliers. (B) Combination of all muscles without substantial intramuscular fat accumulation shows that muscle T2 is increased and MTR is reduced in muscles from patients with IBM, showing early pathological changes. Similar significant differences of lower magnitude were also identified in CMT1A. (C) Fat-fraction maps of the right thigh in a healthy control and a patient with IBM. In the patient, fatty infiltration of muscles is greatest in the quadriceps (red region of interest), which also has a reduced CSA. The mean fat fraction and CSA can be combined to calculate the composite MRI metric, RMA. (D) RMA of quadriceps muscle showed significant correlation with knee extension strength in patients with IBM, CMT1A, and controls. Equivalent graphs of other movements are shown in the appendix. (E) Strong correlations were observed between mean thigh fat fraction and IBMFRS-LL (r=–0·64) for patients with IBM. FF=fat fraction. IBM=inclusion body myositis. CMT1A=Charcot-Marie-Tooth 1A. T2=transverse relaxation time. MTR=magnetisation transfer ratio. CSA=cross-sectional area. RMA=remaining muscle area. IBMFRS-LL=inclusion body myositis functional rating score lower limb.

fraction and IBMFRS-LL (r=–0·64) for patients with IBM. FF=fat fraction. IBM=inclusion body myositis. CMT1A=Charcot-Marie-Tooth 1A. T2=transverse relaxation time. MTR=magnetisation transfer ratio. CSA=cross-sectional area. RMA=remaining muscle area. IBMFRS-LL=inclusion body myositis functional rating score lower limb. Figure 4 Longitudinal data Boxes represent median and IQR, whiskers show range and filled circles are outliers. (A) Fat-fraction maps of the right calf at baseline and after 1 year are shown for control participants and patients with CMT1A and IBM. Minimal change is seen in the mean overall fat fraction in the control, but an increase of 2·0% is shown in the patient with CMT1A and of 7·9% in the patient with IBM. (B) Group comparison against matched controls shows significant increases in overall mean fat fraction in patients with IBM at thigh and calf level and in patients with CMT1A at calf level. (C) Change in quadriceps RMA correlated with change in quadriceps strength over 12 months in patients with IBM for both left and right legs. No significant correlations were seen between 1-year changes in myometric and MRI measures in the CMT1A group. CMT1A=Charcot-Marie-Tooth 1A. IBM=inclusion body myositis. FF=fat fraction. RMA=remaining muscle area. Table 1 Baseline characteristics

Boxes represent median and IQR, whiskers show range and filled circles are outliers. (A) Fat-fraction maps of the right calf at baseline and after 1 year are shown for control participants and patients with CMT1A and IBM. Minimal change is seen in the mean overall fat fraction in the control, but an increase of 2·0% is shown in the patient with CMT1A and of 7·9% in the patient with IBM. (B) Group comparison against matched controls shows significant increases in overall mean fat fraction in patients with IBM at thigh and calf level and in patients with CMT1A at calf level. (C) Change in quadriceps RMA correlated with change in quadriceps strength over 12 months in patients with IBM for both left and right legs. No significant correlations were seen between 1-year changes in myometric and MRI measures in the CMT1A group. CMT1A=Charcot-Marie-Tooth 1A. IBM=inclusion body myositis. FF=fat fraction. RMA=remaining muscle area. Table 1 Baseline characteristics Charcot-Marie-Tooth 1A group Control group for Charcot-Marie-Tooth 1A disease p value Inclusion body myositis group Control group for inclusion body myositis p value Demographics Sex Male 11 11 1 16 12 0·17 Female 9 9 1 4 8 0·17 Age (years) 42·8 (13·9) 45·8 (14·2) 0·51 66·7 (8·9) 61·8 (10·3) 0·12 Height (cm) 167 (12) 171 (11) 0·34 175 (8) 171 (10) 0·28 Weight (kg) 70 (16) 75 (19) 0·44 84 (16) 77 (19) 0·22 Body-mass index 25·1 (4·6) 25·4 (4·9) 0·84 27·4 (4·0) 26·0 (4·6) 0·30 Clinical parameters Age of onset (years) 6·0 (4·4) NA NA 59·0 (8·5) NA NA Disease duration (years) 35·8 (17·5) NA NA 7·7 (3·1) NA NA CMTSS (0–12)* 3·1 (2·0) NA NA NA NA NA CMTES (0–28)* 8·0 (5·1) NA NA NA NA NA MRC-LL (0–110)* 95·4 (15·4) NA NA 93·4 (15·7) NA NA SF36 (0–100%) 73·9% (15·2) NA NA 61·5% (15·3) NA NA SF36-PF (0–100%) 65·3% (23·2) NA NA 39·5% (19·1) NA NA IBMFRS (0–40)* NA NA NA 27·6 (5·4) NA NA Myometric measures Knee extension (Nm) 93·6 (44·1) 134·9 (43·5) 0·005 26·8 (26·0) 119·9 (43·0) <0·0001 Knee flexion (Nm) 47·2 (20·1) 66·1 (20·4) 0·006 35·3 (19·9) 60·6 (20·1) <0·0001 Ankle plantarflexion (Nm) 26·0 (14·3) 62·0 (19·1) <0·0001 29·6 (15·7) 51·6 (18·4) <0·0001 Ankle dorsiflexion (Nm) 10·8 (7·5) 30·0 (9·8) <0·0001 13·2 (10·8) 28 (10·6) <0·0001 Thigh MRI Fat fraction (%) 3·7% (6·8) 1·7% (1·3) 0·23 26·6% (15·5) 2·0% (1·2) <0·0001 T2 (ms) 45·7 (9·1) 41·8 (3·1) 0·08 83·5 (17·7) 43·1 (2·5) <0·0001 MTR (pu) 30·8 (3·0) 32·0 (0·8) 0·09 22·9 (5·0) 31·6 (0·7) <0·0001 Fat fraction whole (%) 5·8% (8·5) 3·3% (1·8) 0·23 27·5% (15·7) 4·0% (1·6) <0·0001 CSA (cm2) 201 (53) 219 (56) 0·31 169 (46) 212 (51) 0·009 Calf MRI Fat fraction (%) 15·8% (25·5) 1·6% (1·0) 0·02 18·0% (13·2) 2·0% (0·9) <0·0001 T2 (ms) 59·7 (29·3) 40·2 (3·6) 0·005 71·1 (21·4) 41·9 (3·8) <0·0001 MTR (pu) 26·1 (9·0) 32·1 (1·0) 0·007 23·7 (5·7) 31·6 (0·9) <0·0001 Fat fraction whole (%) 15·5% (24·0) 2·7% (1·5) 0·03 19·2% (13·7) 3·5% (1·3) <0·0001 CSA (cm2) 100 (26) 123 (27) 0·01 112 (26) 120 (31) 0·39 Data are presented mean (SD), unless otherwise indicated. MRI values are all-muscle region-of-interest means at the respective anatomical levels.

·007 23·7 (5·7) 31·6 (0·9) <0·0001 Fat fraction whole (%) 15·5% (24·0) 2·7% (1·5) 0·03 19·2% (13·7) 3·5% (1·3) <0·0001 CSA (cm2) 100 (26) 123 (27) 0·01 112 (26) 120 (31) 0·39 Data are presented mean (SD), unless otherwise indicated. MRI values are all-muscle region-of-interest means at the respective anatomical levels. For sex, the p value is for the Pearson χ2 test and for all other variables, the p value is for the two-tailed test. NA=not applicable. CMTSS=Charcot-Marie-Tooth symptom score. CMTES=Charcot-Marie-Tooth examination score. MRC-LL=Medical Research Council lower limb score. SF36=Short-Form 36 Quality of Life Score. PF=physical function domain. IBMFRS=inclusion body myositis functional rating scale. MTR=magnetisation transfer ratio. CSA=cross-sectional area. pu=percentage units * Controls had no neuromuscular symptoms and were normal on neurological examination so all scored 0 on CMTES and CMTSS and attained the maximum score on IBMFRS (40) and MRC-LL (110). Table 2 Change in all-muscle MRI and clinical measures between baseline and 12-month follow-up in patients with inclusion body myositis

For sex, the p value is for the Pearson χ2 test and for all other variables, the p value is for the two-tailed test. NA=not applicable. CMTSS=Charcot-Marie-Tooth symptom score. CMTES=Charcot-Marie-Tooth examination score. MRC-LL=Medical Research Council lower limb score. SF36=Short-Form 36 Quality of Life Score. PF=physical function domain. IBMFRS=inclusion body myositis functional rating scale. MTR=magnetisation transfer ratio. CSA=cross-sectional area. pu=percentage units * Controls had no neuromuscular symptoms and were normal on neurological examination so all scored 0 on CMTES and CMTSS and attained the maximum score on IBMFRS (40) and MRC-LL (110). Table 2 Change in all-muscle MRI and clinical measures between baseline and 12-month follow-up in patients with inclusion body myositis Inclusion body myositis group (baseline vs 12-month follow-up) Control group for inclusion body myositis (baseline vs 12-month follow-up) Pairedttest p value (12-month follow-up vs baseline in patient group) Two-tailed p value (patient vs matched control over 12 month follow-up) Standardised response mean Overall clinical measures MRC-LL*† −3·4 (5·6; 0·6 to 6·2) NA 0·02 NA 0·63 SF36 (%) −1·1% (7·0; −2·7 to 4·8) NA 0·55 NA 0·16 SF36-PF (%) −4·1% (18·5; −13·9 to 5·8) NA 0·39 NA 0·22 IBMFRS*† −2·8 (2·9; 1·3 to 4·2) NA 0·0008 NA 0·97 Thigh myometry Knee extension (Nm) −6·0 (5·2; −8·4 to −3·6) −4·2 (11·4; −9·5 to 1·1) 0·002 0·55 −1·15 Knee flexion (Nm) −1·7 (4·3; −3·7 to 0·2) 2·9 (7·1; −0·4 to 6·2) 0·08 0·02 −0·40 Calf myometry Ankle plantarflexion (Nm) −0·7 (3·9; −2·5 to 1·1) 5·6 (14·6; −1·2 to 12·4) 0·42 0·09 −0·19 Ankle dorsiflexion (Nm) −0·4 (4·2; −2·4 to 1·5) 2·0; 5·3 (−0·4 to 4·4) 0·65 0·14 −0·10 Thigh MRI Fat fraction (%)† 3·3% (4·0; 1·4 to 5·2) 0·1% (0·4; −0·3 to 0·1) 0·005 0·001 0·83 T2 (ms) 2·6 (4·2; 0·5 to 4·7) 0·5 (1·6; −0·3 to 1·3) 0·03 0·07 0·62 MTR (pu) −0·9 (1·6; −1·7 to 0·0) 0·0 (0·5; −0·2 to 0·2) 0·06 0·06 −0·54 Fat-fraction whole (%)† 3·3% (3·2; 1·8 to 4·9) 0·2% (0·8; −0·2 to 0·6) 0·0007 0·0004 1·06 CSA (% change) −2·7% (7·9; −6·5 to 1·1) 0·2 (5·7; −2·5 to 2·9) 0·08 0·23 −0·34 Calf MRI Fat fraction (%)† 2·6% (2·7; 1·1 to 4·1) 0·0 (0·4; −0·2 to 0·2) 0·004 0·0007 0·97 T2 (ms)† 4·5 (3·7; 2·6 to 6·4) 0·0 (1·5; −0·7 to 0·7) 0·0005 <0·0001 1·21 MTR (pu)† −0·7 (0·7; −1·1 to −0·3) 0·2 (0·8; −0·2 to 0·6) 0·004 0·003 −0·99 Fat-fraction whole (%)† 2·6% (2·4; 1·3 to 4·0) 0·1% (0·4; −0·1 to 0·3) 0·002 0·0006 1·07 CSA (% change) −2·5% (3·9; −4·4 to −0·6) 0·1% (5·0; −2·3 to 2·5) 0·01 0·11 −0·63 Data are mean (SD; 95% CI), unless otherwise indicated. Fat-fraction whole and CSA are from whole muscle regions of interest, whereas fat fraction, T2, and MTR are from small regions of interest. NA=not applicable. MRC-LL=Medical Research Council lower limb score. SF36=Short-Form 36 Quality of Life Score. PF=physical function domain. IBMFRS=inclusion body myositis functional rating scale. MTR=magnetisation transfer ratio. CSA=cross-sectional area. pu=percentage units.

and MTR are from small regions of interest. NA=not applicable. MRC-LL=Medical Research Council lower limb score. SF36=Short-Form 36 Quality of Life Score. PF=physical function domain. IBMFRS=inclusion body myositis functional rating scale. MTR=magnetisation transfer ratio. CSA=cross-sectional area. pu=percentage units. * Controls had no neuromuscular symptoms and were normal on neurological examination so attained the maximum score on IBMFRS (40) and MRC-LL (110), and they did not have quality of life assessments. † Follow-up value is significantly different from baseline and change is significantly different from change in controls (both paired t test and two-tailed t test are significant). Table 3 Change in all-muscle MRI and clinical measures between baseline and 12-month follow-up in patients with Charcot-Marie-Tooth 1A disease

† Follow-up value is significantly different from baseline and change is significantly different from change in controls (both paired t test and two-tailed t test are significant). Table 3 Change in all-muscle MRI and clinical measures between baseline and 12-month follow-up in patients with Charcot-Marie-Tooth 1A disease Charcot-Marie-Tooth 1A group (baseline vs 12-month follow-up) Control group for Charcot-Marie-Tooth 1A disease Pairedttest p value (12-month follow-up vs baseline in patient group) Two-tailed p value (patient vs matched control over 12-month follow-up) Standardised response mean Overall clinical measures MRC-LL* −0·4 (3·8; −1·5 to 2·3) NA 0·65 NA −0·11 SF36 (%) −2·5 (15·2; −5·3 to 10·3) NA 0·51 NA −0·16 SF36-PF (%) −0·9 (12·3; −5·4 to 7·2) NA 0·77 NA −0·08 CMTES* −0·3 (1·3; −0·9 to 0·4) NA 0·37 NA −0·23 Thigh myometry Knee extension (Nm) 1·0 (8·1; −2·8 to 4·8) −5·2 (10·2; −9·9 to −0·4) 0·57 0·06 0·12 Knee flexion (Nm) 2·5 (6·8; −0·7 to 5·8) 1·7 (7·6; −1·7 to 5·2) 0·15 0·75 0·37 Calf myometry Ankle plantarflexion (Nm) 3·8 (7·6; 0·0 to 7·4) 2·1 (10·8; −2·9 to 7·1) 0·06 0·59 0·51 Ankle dorsiflexion (Nm) 2·1 (4·0; 0·2 to 4·0) 0·5 (3·8; −1·3 to 2·2) 0·05 0·24 0·51 Thigh MRI Fat fraction (%) 0·4 (1·0; −0·1 to 0·9) 0·0 (0·3; −0·1 to 0·1) 0·15 0·12 0·36 T2 (ms) 1·3 (1·5; 0·6 to 2·1) 0·6 (1·6; −0·2 to 1·4) 0·003 0·21 0·86 MTR (pu) 0·0 (0·7; −0·3 to 0·3) −0·1 (0·3; −0·3 to 0·1) 0·96 0·55 −0·01 Fat-fraction whole (%) 0·2 (0·8; −0·2 to 0·6) 0·2 (0·8; −0·2 to 0·6) 0·38 0·97 0·22 CSA (% change) −0·6 (5·3; −3·1 to 1·9) −0·7 (7·2; −4·0 to 2·6) 0·62 0·96 −0·12 Calf MRI Fat fraction (%) 1·1 (2·4; 0·2 to 2·2) 0·0 (0·4; −0·2 to 0·2) 0·07 0·07 0·46 T2 (ms) 1·4 (2·6; 0·0 to 2·6) 0·3 (1·1; −0·2 to 0·9) 0·05 0·13 0·54 MTR (pu) −0·2 (0·6; −0·5 to 0·1) 0·0 (0·7; −0·3 to 0·4) 0·30 0·28 −0·34 Fat-fraction whole (%)† 1·2 (1·5; 0·5 to 1·9) 0·2 (0·4; 0·0 to 0·4) 0·002 0·008 0·83 CSA (% change) 0·6 (6·6; −2·5 to 3·8) 0·0 (4·4; −2·0 to 2·0) 0·67 0·74 0·10 Data are mean (SD; 95% CI), unless otherwise indicated. Fat fraction whole and CSA are from whole muscle regions of interest, whereas fat fraction, T2, and MTR are from small regions of interest. NA=not applicable. MRC-LL=Medical Research Council lower limb score. SF36=Short-Form 36 Quality of Life Score. PF=physical function domain. CMTES=Charcot-Marie-Tooth 1A examination score. T2=transverse relaxation time. MTR=magnetisation transfer ratio. CSA=cross-sectional area.

d MTR are from small regions of interest. NA=not applicable. MRC-LL=Medical Research Council lower limb score. SF36=Short-Form 36 Quality of Life Score. PF=physical function domain. CMTES=Charcot-Marie-Tooth 1A examination score. T2=transverse relaxation time. MTR=magnetisation transfer ratio. CSA=cross-sectional area. * Controls had no neuromuscular symptoms and were normal on neurological examination so scored 0 on CMTES and attained the maximum score on MRC-LL (110), and they did not have quality of life tests. † Follow-up value is significantly different from baseline and is significantly different from change in controls (both paired t test and two-tailed t test are significant). Panel 1 Technical terms Statistical terms Responsiveness: the ability of a measure to change over a prespecified timeframe (internal responsiveness). External responsiveness represents the extent to which a change in a measure relates to corresponding change in a reference measure of clinical or health status. Standardised response mean (SRM): a measure of outcome measure responsiveness calculated from natural history or placebo data by dividing the mean change by the standard deviation. An SRM of less than 0·2 is thought to have minimal responsiveness and an SRM of more than 0·8 is deemed to have large responsiveness. In addition to having a high SRM, the measurement must also be expected to respond to the planned intervention. Lehr's formula: n=16(E·SRM)2

Standardised response mean (SRM): a measure of outcome measure responsiveness calculated from natural history or placebo data by dividing the mean change by the standard deviation. An SRM of less than 0·2 is thought to have minimal responsiveness and an SRM of more than 0·8 is deemed to have large responsiveness. In addition to having a high SRM, the measurement must also be expected to respond to the planned intervention. Lehr's formula: n=16(E·SRM)2 A formula to estimate the number of participants in each group for a study with 80% power to detect an intergroup difference with p<0·05 significance. It is based on SRM and effect size in terms of reduction in progression versus placebo (E). For example, if a clinically significant benefit is regarded as a 50% reduction in disease progression (E=0·5), use of an outcome measure with minimal responsiveness (SRM=0·2), would require 1600 participants in each study group, whereas an outcome measure with large responsiveness (SRM=0·8) would require 100 participants in each group. Validity: measured by comparing the results of a measurement to the results of a gold standard test. For an outcome measure, validity here refers to correspondence with clinical measures of patient function, considered along with mortality as possible “true outcome measures” in a clinical trial. MRI terms Fat fraction: the proportion of water (W) to fat (F) signal originating in the measurement volume reported as a percentage. Fat fraction=F(F+W)×100% 3-point-Dixon fat water separation: an MRI technique yielding separate fat-only and water-only images

Validity: measured by comparing the results of a measurement to the results of a gold standard test. For an outcome measure, validity here refers to correspondence with clinical measures of patient function, considered along with mortality as possible “true outcome measures” in a clinical trial. MRI terms Fat fraction: the proportion of water (W) to fat (F) signal originating in the measurement volume reported as a percentage. Fat fraction=F(F+W)×100% 3-point-Dixon fat water separation: an MRI technique yielding separate fat-only and water-only images Remaining muscle area: defined here as the cross-sectional area of muscle tissue within the muscle fascial boundary (CSA) not replaced by fat. Transverse relaxation time (T2): a key source of MRI contrast representing changes in molecular mobility, commonly interpreted in terms of tissue water or lipid distribution changes. Magnetisation transfer ratio (MTR): an MRI measure being the signal intensity ratio between a control image and one made sensitive to changes in the relative population of free water in tissue and the rate of magnetisation exchange with other proton pools. Panel 2 Research in context Systematic review

Transverse relaxation time (T2): a key source of MRI contrast representing changes in molecular mobility, commonly interpreted in terms of tissue water or lipid distribution changes. Magnetisation transfer ratio (MTR): an MRI measure being the signal intensity ratio between a control image and one made sensitive to changes in the relative population of free water in tissue and the rate of magnetisation exchange with other proton pools. Panel 2 Research in context Systematic review We searched PubMed for reports published before July 1, 2015, with the MeSH terms “magnetic resonance imaging” and “Charcot-Marie-Tooth disease”, “myositis, inclusion body”, or “neuromuscular diseases”, with no language restrictions. We restricted our search to reports in human participants. Five articles were identified reporting skeletal muscle MRI in inclusion body myositis, but none included quantitative MRI sequences and none included longitudinal data. 12 articles were identified reporting skeletal muscle MRI in Charcot-Marie-Tooth 1A. Only two of these14, 39 used quantitative MRI sequences but did not include longitudinal data. Of the ten studies reporting qualitative MRI in Charcot-Marie-Tooth 1A, one study included a longitudinal scan on a single patient,40 whereas one study reported qualitative MRI imaging in 14 patients at a 2 year interval but identified no progression.41 Taking into account all neuromuscular disorders, published quantitative MRI studies have quantified fat fraction,9, 10, 11, 12, 13 T2,9, 11, 29, 30, 42, 43 and magnetisation transfer ratio,14, 44 in disorders that are myopathic,9, 10, 11, 12, 13, 29, 30, 42, 43, 44 neuropathic,10, 14 inherited,9, 10, 11, 12, 13, 14, 29, 30, 42, 44 or acquired,14, 43 at thigh level10, 11, 13, 30, 42, 43 or calf level,9, 11, 14, 29, 42, 44 in one muscle9, 10, 29 or many muscles,11, 12, 13, 14, 30, 42, 43, 44 and with controls9, 11, 14, 29, 42, 43, 44 or without controls.10, 12, 13, 30 Few studies in patients with neuromuscular disorders have been published so far with longitudinal quantitative MRI data.13, 15, 45, 46 Observational studies have shown longitudinal change in muscle fat fraction in patients with limb girdle muscular dystrophy type 2I,13 and longitudinal change in muscle T2 in Duchenne muscular dystrophy.15 Studies in Duchenne have used MRI longitudinally after steroid initiation and showed a reduction in T245 or a variable effect on T2.46

s have shown longitudinal change in muscle fat fraction in patients with limb girdle muscular dystrophy type 2I,13 and longitudinal change in muscle T2 in Duchenne muscular dystrophy.15 Studies in Duchenne have used MRI longitudinally after steroid initiation and showed a reduction in T245 or a variable effect on T2.46 Interpretation Accurately and reliably monitoring disease progression is an important barrier to successful experimental clinical trials to test new therapies for muscle-wasting neuromuscular diseases. We report longitudinal data in patients with one of two representative neuromuscular disorders and healthy controls. We have systematically investigated quantitative MRI biomarkers and have shown they can reliably and accurately track disease progression. In patients with Charcot-Marie-Tooth 1A, calf muscle fat fraction showed high responsiveness, exceeding all clinical measures. Abnormalities in T2 and MTR, measures sensitive to muscle water accumulation, were seen in muscles before significant fat infiltration. We have also shown correlations with important measures of patient function including strength, functional scores, and quality of life measures. The biomarkers we selected detect and quantify water and fat accumulation within muscle, which both occur across a range of different muscle-wasting diseases. Therefore these quantitative MRI methods might be applicable to other muscle-wasting neuromuscular diseases.

Introduction Huntington's disease is a slowly progressive neurodegenerative disorder for which no proven disease-modifying treatments yet exist. Knowledge of its genetic cause, CAG repeat expansions in the HTT gene leading to the formation of mutant huntingtin (mHTT), has enabled focused study of the disease and the development of advanced therapeutics targeting known aspects of its pathobiology.1 Although extensive efforts have established well characterised clinical, cognitive, and neuroimaging biomarkers of progression,2, 3, 4, 5 few biochemical markers have been identified that enable direct assessment of relevant aspects of pathology.6, 7 No prognostic biomarkers for assessing neuronal damage, disease progression, or therapeutic response have been validated, which limits the ability to test novel therapeutics, especially in mutation carriers with premanifest Huntington's disease for whom treatment is most likely to result in long-term meaningful benefits. Accessible, reliable, biochemical markers would greatly facilitate the development of novel therapeutics for Huntington's disease.1

the ability to test novel therapeutics, especially in mutation carriers with premanifest Huntington's disease for whom treatment is most likely to result in long-term meaningful benefits. Accessible, reliable, biochemical markers would greatly facilitate the development of novel therapeutics for Huntington's disease.1 Many potential markers in CSF have been proposed, but only a few (eg, mHTT itself, microtubule-associated protein tau, and chitinase-3-like protein 1) have shown associations with clinical phenotype beyond established predictors, such as age and HTT CAG triplet repeat count.7, 8, 9, 10 None of the potential biomarkers has been studied longitudinally.7 Moreover, CSF is more difficult and expensive to obtain than other fluids, but no substance detectable in a highly accessible biofluid, such as blood, has robustly reflected Huntington's disease-related alterations due to CNS pathology, either in cross-sectional or longitudinal studies. We previously showed that mHTT concentrations in blood leucocytes were associated with clinical severity cross-sectionally,11 but that this was probably due to peripheral mHTT production rather than CNS pathology. This and other such markers of peripheral pathology,12 therefore, are poor candidates for trials involving direct CNS delivery of disease-modifying agents. Research in context Evidence before this study

Many potential markers in CSF have been proposed, but only a few (eg, mHTT itself, microtubule-associated protein tau, and chitinase-3-like protein 1) have shown associations with clinical phenotype beyond established predictors, such as age and HTT CAG triplet repeat count.7, 8, 9, 10 None of the potential biomarkers has been studied longitudinally.7 Moreover, CSF is more difficult and expensive to obtain than other fluids, but no substance detectable in a highly accessible biofluid, such as blood, has robustly reflected Huntington's disease-related alterations due to CNS pathology, either in cross-sectional or longitudinal studies. We previously showed that mHTT concentrations in blood leucocytes were associated with clinical severity cross-sectionally,11 but that this was probably due to peripheral mHTT production rather than CNS pathology. This and other such markers of peripheral pathology,12 therefore, are poor candidates for trials involving direct CNS delivery of disease-modifying agents. Research in context Evidence before this study We searched PubMed by use of the MeSH terms “([intermediate filaments] AND [nerve degeneration OR Huntington disease OR Alzheimer disease OR Parkinson disease OR Pick disease of the brain OR frontotemporal dementia OR amyotrophic lateral sclerosis OR supranuclear palsy, progressive] AND [blood OR plasma OR serum OR cerebrospinal fluid] AND [humans])” and their natural language variants for human studies of neurofilament light protein (NfL, also known as NF-L) in neurodegenerative disorders. Four small cross-sectional studies reported raised concentrations of NfL in the CSF of people with Huntington's disease, and four longitudinal studies had assessed blood NfL concentrations in other disorders. The largest longitudinal study of people with neurodegenerative disease measured NfL in 174 patients with progressive supranuclear palsy over 1 year. The largest study of neurodegenerative disease onset was done in 34 GRN mutation carriers with premanifest frontotemporal dementia, among whom only two had disease progression during follow-up. The association between concentrations of NfL in plasma has never been reported in people with Huntington's disease. We investigated whether NfL concentrations in plasma could serve as a potential prognostic marker of neurodegeneration in a genetically homogeneous cohort of HTT CAG expansion mutation carriers, including a substantial subgroup with premanifest disease, followed up over 3 years.

ted in people with Huntington's disease. We investigated whether NfL concentrations in plasma could serve as a potential prognostic marker of neurodegeneration in a genetically homogeneous cohort of HTT CAG expansion mutation carriers, including a substantial subgroup with premanifest disease, followed up over 3 years. Added value of this study Increased concentrations of NfL in plasma were seen throughout the course of Huntington's disease and, after controlling for age and CAG repeat count, were independently associated with cognitive and motor dysfunction and global and regional brain volume at any given timepoint. We also found a strong association between increased NfL concentration and CAG repeat count, which suggests a firm link between this factor and the genetic basis of Huntington's disease. In HTT mutation carriers with premanifest disease at baseline, increased concentration of NfL in plasma at baseline was associated with subsequent clinical onset beyond the known prognostic variables of age, CAG repeat count, and brain volume; such an association has not previously been found for other potential biofluid biomarkers. Associations were also seen with disease progression assessed by cognitive, functional, and brain atrophy measures. In an independent cohort of 37 individuals (23 HTT mutation carriers and 14 controls), we showed a strong correlation between raised NfL concentrations in plasma and matched CSF. Implications of all the available evidence

Increased concentrations of NfL in plasma were seen throughout the course of Huntington's disease and, after controlling for age and CAG repeat count, were independently associated with cognitive and motor dysfunction and global and regional brain volume at any given timepoint. We also found a strong association between increased NfL concentration and CAG repeat count, which suggests a firm link between this factor and the genetic basis of Huntington's disease. In HTT mutation carriers with premanifest disease at baseline, increased concentration of NfL in plasma at baseline was associated with subsequent clinical onset beyond the known prognostic variables of age, CAG repeat count, and brain volume; such an association has not previously been found for other potential biofluid biomarkers. Associations were also seen with disease progression assessed by cognitive, functional, and brain atrophy measures. In an independent cohort of 37 individuals (23 HTT mutation carriers and 14 controls), we showed a strong correlation between raised NfL concentrations in plasma and matched CSF. Implications of all the available evidence Measurement of NfL in plasma could be useful to assess the risk of Huntington's disease onset and progression beyond currently known prognostic factors. Our findings in individuals with premanifest Huntington's disease suggest that NfL has potential as a biomarker in the preclinical phases of other neurodegenerative diseases. The availability of accessible, reliable biochemical markers might also be useful in assessing novel therapeutics. Finally, this study affirms the benefits of systematically studying genetically homogeneous cohorts to assess the earliest changes in neurodegeneration.

clinical phases of other neurodegenerative diseases. The availability of accessible, reliable biochemical markers might also be useful in assessing novel therapeutics. Finally, this study affirms the benefits of systematically studying genetically homogeneous cohorts to assess the earliest changes in neurodegeneration. Neurofilament light protein (NfL, also known as NF-L) is the smallest of three subunits that make up neurofilaments, which are major components of the neuronal cytoskeleton. NfL is released from damaged neurons. Concentrations in CSF are increased in people with neurodegenerative diseases, including Alzheimer's disease,13 amyotrophic lateral sclerosis,13 and frontotemporal dementia.13, 14 Four small-scale cross-sectional studies found raised concentrations of NfL in the CSF of individuals with Huntington's disease,8, 15, 16, 17 and we have shown a close association between increased concentrations of NfL and mHTT in CSF.8

r's disease,13 amyotrophic lateral sclerosis,13 and frontotemporal dementia.13, 14 Four small-scale cross-sectional studies found raised concentrations of NfL in the CSF of individuals with Huntington's disease,8, 15, 16, 17 and we have shown a close association between increased concentrations of NfL and mHTT in CSF.8 NfL is detectable in blood plasma or serum. Cross-sectional studies have shown increased concentrations in blood in people with frontotemporal dementia,18 Alzheimer's disease, amyotrophic lateral sclerosis,19 and atypical parkinsonism,20 and in longitudinal studies of those with frontotemporal dementia,21, 22 amyotrophic lateral sclerosis,23 and progressive supranuclear palsy.24 A 1-year study followed up 147 patients with progressive supranuclear palsy, but these cohorts were genetically and pathologically heterogeneous and only a few included premanifest individuals, in small numbers.24 NfL concentrations in blood have not been reported in people with Huntington's disease.

ranuclear palsy.24 A 1-year study followed up 147 patients with progressive supranuclear palsy, but these cohorts were genetically and pathologically heterogeneous and only a few included premanifest individuals, in small numbers.24 NfL concentrations in blood have not been reported in people with Huntington's disease. We aimed to investigate whether NfL in plasma could act as a potential prognostic marker of neurodegeneration and disease progression for Huntington's disease. The study involved the 366 participants of the TRACK-HD cohort,5, 25 who had been assessed by standardised blood sampling, clinical testing, and MRI annually over 3 years. This cohort offers a unique resource for studying neurodegeneration in a genetically uniform disease cohort and for assessing changes at different stages of disease, including in the premanifest phase. We tested the hypotheses that NfL concentrations would be raised in individuals with Huntington's disease, would increase as disease progressed, and that concentrations would correlate with disease onset in HTT mutation carriers with premanifest Huntington's disease and with clinical progression in those with manifest disease, therefore acting as an indicator of neurodegeneration. We additionally investigated whether concentrations of NfL in CSF would correlate with those in plasma.

s would correlate with disease onset in HTT mutation carriers with premanifest Huntington's disease and with clinical progression in those with manifest disease, therefore acting as an indicator of neurodegeneration. We additionally investigated whether concentrations of NfL in CSF would correlate with those in plasma. Methods Study design and participants We did a retrospective study involving 366 participants enrolled in the TRACK-HD study at four international study sites in 2008. Participants were assessed annually with standardised 3-Tesla T1 volumetric MRI, clinical, cognitive, quantitative motor, and neuropsychiatric assessments, as previously described.25 The study protocol is available online. At enrolment, participants with HTT CAG expansion mutations were classified as having premanifest or manifest Huntington's disease based on the Unified Huntington's Disease Rating Scale (UHDRS) Total Motor Score (TMS). Participants with premanifest disease were further separated into two subgroups of early and late premanifest disease (termed preHD A and preHD B in some other TRACK-HD publications), with the group median for predicted number of years to onset of manifest disease (10·8) as the threshold. Using the UHDRS Total Functional Capacity (TFC) score, patients with manifest Huntington's disease were separated into subgroups with clinical stage 1 (TFC score >10) and stage 2 (TFC score 7–10) disease. Controls were healthy partners or siblings of HTT mutation carriers. All human studies are compliant with the Declaration of Helsinki and approved by local ethics committees, and all participants gave written informed consent.

ed into subgroups with clinical stage 1 (TFC score >10) and stage 2 (TFC score 7–10) disease. Controls were healthy partners or siblings of HTT mutation carriers. All human studies are compliant with the Declaration of Helsinki and approved by local ethics committees, and all participants gave written informed consent. Clinical and imaging assessments T1 volumetric MRI scans were subjected to rigorous quality control followed by analysis by operators unaware of participants' statuses at specialist image analysis sites with use of optimised, standardised techniques, as previously described.5 Briefly, cross-sectional putamen volumes were calculated by automated segmentation; whole-brain, caudate, lateral ventricles, and total intracranial volumes by semiautomated segmentation; and grey-matter and white-matter volumes by voxel-based morphometry. Longitudinal changes in whole-brain, ventricles, and caudate were calculated with the boundary shift integral technique, and changes in grey-matter and white-matter volume were assessed with voxel compression mapping within voxel-based morphometry segmentations.5 All cross-sectional imaging measures were calculated as a percentage of total intracranial volume. Cognitive function was assessed with the Symbol-Digit Modality Test (SDMT) and Stroop Word Reading (SWR) task, and clinical severity was assessed with the UHDRS TMS and TFC.25

ithin voxel-based morphometry segmentations.5 All cross-sectional imaging measures were calculated as a percentage of total intracranial volume. Cognitive function was assessed with the Symbol-Digit Modality Test (SDMT) and Stroop Word Reading (SWR) task, and clinical severity was assessed with the UHDRS TMS and TFC.25 NfL quantification in plasma At each TRACK-HD visit, blood was collected in BD Vacutainer tubes containing edetic acid (Franklin Lakes, NJ, USA.). Samples were processed onsite to isolate plasma, as previously described.26 Plasma samples were frozen, stored at −80°C, then shipped frozen for analysis by operators unaware of participants' disease statuses. NfL concentrations were quantified with an ultrasensitive single-molecule array method.18 All NfL values were within the linear ranges of the assays. CSF cohort A London-based independent cohort of 37 participants (14 controls, three HTT mutation carriers with premanifest disease [UHDRS diagnostic confidence scores <4], and 20 participants with manifest Huntington's disease) underwent CSF and plasma collection standardised for diet, time of day, clinical procedures, and processing.8 Blood was collected within 30 min of CSF in sodium heparin cell-preparation BD Vacutainer tubes and processed to isolate plasma. NfL concentrations in CSF were quantified with a commercial ELISA, used according to the manufacturer's protocol (UmanDiagnostics, Umeå, Sweden).8

f day, clinical procedures, and processing.8 Blood was collected within 30 min of CSF in sodium heparin cell-preparation BD Vacutainer tubes and processed to isolate plasma. NfL concentrations in CSF were quantified with a commercial ELISA, used according to the manufacturer's protocol (UmanDiagnostics, Umeå, Sweden).8 Statistical analysis Analysis of plasma NfL in TRACK-HD was done per a prespecified statistical plan that was designed with and done by a statistician experienced in analysing the TRACK-HD dataset (DL). Outcomes of interest (UHDRS TMS, UHDRS TFC, SWR, SDMT, and volumes of whole-brain, caudate, putamen, lateral ventricles, grey matter, and white matter) were chosen because they had the strongest effect sizes in TRACK-HD.5 Clinical and imaging data were paired to NfL concentrations in plasma at each study visit. Longitudinal changes were calculated over the longest available period and converted to annualised rates. NfL concentrations in plasma were non-normally distributed because of biologically plausible higher values. Natural log-transformation produced plausibly normal distributions and was used for all analyses. Longitudinal changes in UHDRS TMS were calculated by square-root transformation of cross-sectional values.

Statistical analysis Analysis of plasma NfL in TRACK-HD was done per a prespecified statistical plan that was designed with and done by a statistician experienced in analysing the TRACK-HD dataset (DL). Outcomes of interest (UHDRS TMS, UHDRS TFC, SWR, SDMT, and volumes of whole-brain, caudate, putamen, lateral ventricles, grey matter, and white matter) were chosen because they had the strongest effect sizes in TRACK-HD.5 Clinical and imaging data were paired to NfL concentrations in plasma at each study visit. Longitudinal changes were calculated over the longest available period and converted to annualised rates. NfL concentrations in plasma were non-normally distributed because of biologically plausible higher values. Natural log-transformation produced plausibly normal distributions and was used for all analyses. Longitudinal changes in UHDRS TMS were calculated by square-root transformation of cross-sectional values. ANOVA was used to compare groups at baseline. Other cross-sectional associations were assessed for individuals' baseline and follow-up measurements in linear models that included a random participant effect. Sex and study site were not significantly associated with NfL and, therefore, were not included as covariates. When controlling for the interacting effects of age and CAG repeat count (known prognostic factors for progression of Huntington's disease), we used a polynomial model that included linear and squared terms for variables and interactions between them, since in our observations of other Huntington's disease phenomena such higher order terms have often been significantly associated with progression.

prognostic factors for progression of Huntington's disease), we used a polynomial model that included linear and squared terms for variables and interactions between them, since in our observations of other Huntington's disease phenomena such higher order terms have often been significantly associated with progression. Longitudinal changes in NfL concentrations were assessed with correlated random intercept and slope models. UHDRS TFC changes are negligible in individuals with premanifest Huntington's disease and, therefore, longitudinal analyses of changes in this measure were restricted to individuals with manifest Huntington's disease. For simplicity, we describe longitudinal associations with Pearson's correlations between cross-sectional NfL concentrations and other outcomes, and between baseline NfL concentrations and the annualised rates of change in other outcomes, calculated from data available at the last available assessment. However, all p values are derived from analogous random effects repeated measurement models, which allow proper inference but lack unambiguously defined corresponding correlation statistics.

entrations and the annualised rates of change in other outcomes, calculated from data available at the last available assessment. However, all p values are derived from analogous random effects repeated measurement models, which allow proper inference but lack unambiguously defined corresponding correlation statistics. We used Cox proportional hazard survival modelling to calculate hazard ratios (HRs) and 95% CIs for the correlation between baseline NfL concentration and subsequent onset of Huntington's disease within 3 years in premanifest HTT mutation carriers. The small number of confirmed new diagnoses (n=18) precluded simultaneous inclusion of multiple covariates, but their pattern was consistent with the proportional odds assumption. Thus, other known risk factors previously identified in the TRACK-HD data5, 27 were controlled separately to assess non-redundancy of NfL concentrations as a risk factor. The data for the independent CSF cohort were analysed without log transformation. Participants with premanifest and manifest Huntington's disease were pooled to create one Huntington's disease group. We used Wilcoxon's rank sum test to compare groups, and Pearson's correlation coefficient and partial correlations with bootstrap estimates of SEs with 1000 replications to assess associations between variables.

nts with premanifest and manifest Huntington's disease were pooled to create one Huntington's disease group. We used Wilcoxon's rank sum test to compare groups, and Pearson's correlation coefficient and partial correlations with bootstrap estimates of SEs with 1000 replications to assess associations between variables. The threshold for statistical significance for all analyses was p<0·05. Where prominent outliers remained after normalisation, they were excluded and analyses were repeated to evaluate their influence. The analysis was carried out in SAS (v9.4, SAS Institute Inc, Cary, NC, USA) using SAS/STAT 14.1. Role of the funding source The funders of the study had no role in the design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Of 366 TRACK-HD participants, baseline and follow-up plasma samples were available from 298 (81%) who completed the 3-year TRACK-HD study (97 controls and 201 HTT mutation carriers—58 with early premanifest and 46 with late premanifest disease, and 66 with stage 1 and 31 with stage 2 Huntington's disease). 293 had paired plasma samples from the 3-year follow-up visit, four from the 2-year visit, and one from the 1-year visit. The demographic and clinical characteristics of participants at baseline are presented in the appendix.

with late premanifest disease, and 66 with stage 1 and 31 with stage 2 Huntington's disease). 293 had paired plasma samples from the 3-year follow-up visit, four from the 2-year visit, and one from the 1-year visit. The demographic and clinical characteristics of participants at baseline are presented in the appendix. Concentrations of NfL in plasma at baseline were 2·6 times higher before log transformation in HTT mutation carriers than in controls (mean 3·63 [SD 0·54] log pg/mL vs 2·68 [0·52] log pg/mL, p<0·00001). Baseline NfL concentrations in plasma were significantly higher in all disease stage subgroups among HTT mutation carriers than those in controls (all p<0·0001, figure 1 and appendix). Additionally, NfL concentrations differed significantly with increasing disease stage except stage 2 versus stage 1 Huntington's disease (figure 1 and appendix). The stage 2 manifest subgroup did, however, differ significantly from the early premanifest subgroup (mean difference 0·785 [SE 0·113], p<0·0001) and the late premanifest subgroup (0·348 [0·102], p=0·0017).Figure 1 Associations between NfL concentrations in plasma and disease stage, age, and CAG triplet repeat count

). The stage 2 manifest subgroup did, however, differ significantly from the early premanifest subgroup (mean difference 0·785 [SE 0·113], p<0·0001) and the late premanifest subgroup (0·348 [0·102], p=0·0017).Figure 1 Associations between NfL concentrations in plasma and disease stage, age, and CAG triplet repeat count (A) Baseline NfL concentrations in plasma, by disease stage. Boxes show first and third quartiles, the central band shows the median, and the whiskers show data within 1·5 IQR of the median. The dots represent outliers. Data were log transformed for comparisons. (B) Associations between NfL concentration in plasma, age, and CAG repeat count, modelled with a polynomial function of age, CAG repeat counts, their squares, and their interactions in 201 HTT mutation carriers and 97 controls. The lines show quadratic fit for all participants with a given CAG repeat count or all controls. Each increase in CAG repeat count was associated with higher and more steeply rising NfL concentrations in plasma. Predicted values are truncated at the vertical inflection point of the parabola. Datapoints for each individual CAG repeat count and for controls are provided in the appendix. HD=Huntington's disease mutation carriers. CAG=mutation carriers' CAG repeat counts.

igher and more steeply rising NfL concentrations in plasma. Predicted values are truncated at the vertical inflection point of the parabola. Datapoints for each individual CAG repeat count and for controls are provided in the appendix. HD=Huntington's disease mutation carriers. CAG=mutation carriers' CAG repeat counts. We found positive associations between NfL concentrations in plasma and age in controls and all Huntington's disease subgroups. In controls, the association was roughly linear (slope 0·02 log pg/mL per year [SE 0·0042], p<0·0001; figure 1B). In HTT mutation carriers there was a significant positive association between NfL concentration in plasma and the CAG-age product, which measures the extent of exposure to the effects of the CAG expansion (appendix),4 but the non-linear association with age and CAG triplet repeat count was best described by CAG-repeat-count-dependent quadratic functions of age (appendix). The association is clear when each CAG count is considered separately (figure 1B, appendix) and shows that for a given age, overall NfL concentrations in plasma increased with increasing CAG repeat count, and that the steepness of the slopes declined with increasing age. Thus, maximum predicted NfL concentrations at older ages became similar.

when each CAG count is considered separately (figure 1B, appendix) and shows that for a given age, overall NfL concentrations in plasma increased with increasing CAG repeat count, and that the steepness of the slopes declined with increasing age. Thus, maximum predicted NfL concentrations at older ages became similar. In HTT mutation carriers, NfL concentrations in plasma at baseline were negatively associated with cognitive score on the SDMT and SWR and with the MRI measures of brain volume for putamen, caudate, grey matter, and white matter (higher NfL values were associated with smaller brain volumes, figure 2). These associations remained significant after adjustment for the combined effects of age and CAG repeat count (r=–0·293, p<0·0001 for SDMT, r=–0·239, p=0·0042 for SWR; r=–0·286, p<0·0001 for putamen, r=–0·187, p=0·017 for caudate, r=–0·198, p=0·0004 for grey matter, and r=–0·121, p=0·048 for white matter; appendix). The exception was whole-brain volume, which showed a significant negative association before adjustment (r=–0·447, p<0·0001), but not after (r=–0·120, p=0·150). Significant positive associations were seen with UHDRS TMS (higher NfL values were associated with worse motor performance) and lateral ventricle volume (higher NfL values were associated with larger ventricles; figure 2). These associations persisted after adjustment for age and CAG count (r=0·246, p<0·0001 for TMS, r=0·260, p<0·0001 for ventricular volume; appendix).Figure 2 Associations between NfL concentrations in plasma at baseline and cross-sectional measures of cognitive function, motor impairment, and brain volume on MRI

ure 2). These associations persisted after adjustment for age and CAG count (r=0·246, p<0·0001 for TMS, r=0·260, p<0·0001 for ventricular volume; appendix).Figure 2 Associations between NfL concentrations in plasma at baseline and cross-sectional measures of cognitive function, motor impairment, and brain volume on MRI (A, B) Association with cognitive scores. (C) Association with motor function. (D–H) Associations with global and regional brain volumes, expressed as percentages of total intracranial volume. UHDRS=Unified Huntington's Disease Rating Scale. Concentrations of NfL in plasma increased significantly from baseline in individuals with premanifest Huntington's disease, by 0·060 log pg/mL per year (SE 0·012, p<0·0001), and in those with manifest Huntington's disease, by 0·026 log pg/mL per year (0·0129, p=0·0442). No change was seen in controls (0·018 log pg/mL per year [0·0128], p=0·171). The rate of increase was significantly greater in HTT mutation carriers with premanifest Huntington's disease than in controls (0·043 log pg/mL per year [0·018], p=0·0161) but did not differ significantly between those with premanifest and manifest Huntington's disease (0·034 log pg/mL per year [0·018], p=0·0547) or between those with manifest Huntington's disease and controls (0·009 log pg/mL per year [0·018], p=0·630). The greater rate of increase in premanifest Huntington's disease is consistent with the non-linear associations between NfL concentrations, age, and CAG repeat count.

0·034 log pg/mL per year [0·018], p=0·0547) or between those with manifest Huntington's disease and controls (0·009 log pg/mL per year [0·018], p=0·630). The greater rate of increase in premanifest Huntington's disease is consistent with the non-linear associations between NfL concentrations, age, and CAG repeat count. 18 (17%) of 104 HTT mutation carriers with premanifest disease at baseline were newly diagnosed as having manifest Huntington's disease during the TRACK-HD study. The association between baseline NfL concentration in plasma and subsequent disease onset was significant (HR 3·29 per 1·0 log pg/mL, 95% CI 1·48–7·34, p=0·0036; figure 3), and remained so after adjustment for age, CAG repeat count, and their interactions (3·03 per 1·0 log pg/mL, 1·07–8·60, p=0·0371) and for each baseline brain volume measure (appendix).27 A receiver operating characteristic curve for risk of diagnosis within 3 years showed that mean sensitivity and specificity were highest when NfL concentration in plasma was 3·61 log pg/mL at baseline (appendix), which is close to the median value among participants with premanifest Huntington's disease at baseline of 3·69 log pg/mL.Figure 3 Association between baseline NfL concentration in plasma and progression to manifest Huntington's disease in HTT mutation carriers who were premanifest at baseline

L at baseline (appendix), which is close to the median value among participants with premanifest Huntington's disease at baseline of 3·69 log pg/mL.Figure 3 Association between baseline NfL concentration in plasma and progression to manifest Huntington's disease in HTT mutation carriers who were premanifest at baseline (A) NfL concentration in plasma at baseline by disease progression status at 3 years. Boxes show first and third quartiles, the central band shows the median, and the whiskers show data within 1·5 IQR of the median. The dots represent outliers. (B) Kaplan-Meier plot showing longitudinal survival in the premanifest state among HTT mutation carriers with NfL concentrations in plasma greater or less than the median. The Cox proportional hazards model is the more sensitive of the two models presented here.

ithin 1·5 IQR of the median. The dots represent outliers. (B) Kaplan-Meier plot showing longitudinal survival in the premanifest state among HTT mutation carriers with NfL concentrations in plasma greater or less than the median. The Cox proportional hazards model is the more sensitive of the two models presented here. We found significant associations between NfL concentration in plasma at baseline and subsequent decline in cognition and total functional capacity (figure 4). After adjustment for age and CAG repeat count, NfL concentration remained an independent prognostic factor for decline in SDMT (r=–0·173, p<0·0001) but not SWR (r=–0·040, p=0·4057) or UHDRS TFC (r=–0·151, p=0·1107). No association was found with change in motor score (r=0·112, p=0·0592). We found positive associations with atrophy of the caudate, whole-brain, grey matter, and white matter and with ventricular expansion (figure 4). These associations remained significant after adjustment for age and CAG repeat count (r=0·199, p=0·0043 for caudate, r=0·320, p<0·0001 for whole-brain; r=0·242, p=0·019 for grey matter; r=0·327, p<0·0001 for white matter; and r=–0·323, p=0·0002 for ventricular expansion).Figure 4 Associations between baseline NfL concentration in plasma and longitudinal change in cognitive, motor, and functional decline and brain atrophy

for caudate, r=0·320, p<0·0001 for whole-brain; r=0·242, p=0·019 for grey matter; r=0·327, p<0·0001 for white matter; and r=–0·323, p=0·0002 for ventricular expansion).Figure 4 Associations between baseline NfL concentration in plasma and longitudinal change in cognitive, motor, and functional decline and brain atrophy (A, B) Associations with cognitive scores. (C) Association with functional capacity. (D–H) Associations with global and regional brain volumes, expressed as percentages of total intracranial volume. By convention, negative values for change in lateral ventricle volumes indicate ventricular expansion (ie, brain atrophy). SDMT=Symbol-Digit Modality Test Score. SWR=Stroop Word Reading score. UHDRS TFC=Unified Huntington's Disease Rating Scale Total Functional Capacity score. HD=Huntington's disease mutation carriers.

By convention, negative values for change in lateral ventricle volumes indicate ventricular expansion (ie, brain atrophy). SDMT=Symbol-Digit Modality Test Score. SWR=Stroop Word Reading score. UHDRS TFC=Unified Huntington's Disease Rating Scale Total Functional Capacity score. HD=Huntington's disease mutation carriers. Diagnostic status is an obvious prognostic indicator in Huntington's disease and, therefore, we additionally assessed whether NfL concentrations in plasma at baseline were associated with change in brain volumes after controlling for baseline diagnostic status, age, and CAG repeat count. We found independent associations with changes in whole-brain, grey-matter, and white-matter volumes and with ventricular expansion. Furthermore, the associations were independently significant in the premanifest and manifest Huntington's disease subgroups for whole-brain atrophy, grey-matter atrophy, and lateral ventricular expansion (appendix), but for all features where the difference between these subgroups was significant, the prognostic value was a stronger indicator in manifest than in premanifest Huntington's disease.

t and manifest Huntington's disease subgroups for whole-brain atrophy, grey-matter atrophy, and lateral ventricular expansion (appendix), but for all features where the difference between these subgroups was significant, the prognostic value was a stronger indicator in manifest than in premanifest Huntington's disease. The median concentration of NfL in CSF in the 37 participants in the independent CSF cohort was significantly higher in HTT mutation carriers than in controls (1871 pg/mL, IQR 1312–2461 vs 300 pg/mL, 234–368, figure 5), and in mutation carriers a positive association was seen with UHDRS TMS (r=0·4815, p=0·012). In matched plasma samples from 30 participants, the median NfL concentration was also significantly higher in HTT mutation carriers than in controls (31·7 pg/mL, IQR 24·9–50·6 vs 9·9 pg/mL, 8·4–13·7, figure 5), which in mutation carriers was also positively associated with UHDRS TMS (r=0·709, p<0·0001). In keeping with CNS origin, the median NfL concentration in CSF was 46·4 times higher than that in plasma. The ratio differed significantly between controls and HTT mutation carriers (30·1 vs 62·11, p<0·0001), but there was a positive association between concentrations in plasma and CSF (figure 5).Figure 5 NfL concentrations in paired CSF and plasma samples

NfL concentration in CSF was 46·4 times higher than that in plasma. The ratio differed significantly between controls and HTT mutation carriers (30·1 vs 62·11, p<0·0001), but there was a positive association between concentrations in plasma and CSF (figure 5).Figure 5 NfL concentrations in paired CSF and plasma samples Concentrations in CSF (A) and plasma (B) in HTT mutation carriers and controls. Boxes show first and third quartiles, the central band shows the median, and the whiskers show data within 1·5 IQR of the median. The dots represent outliers. (C) Correlation between NfL concentration in CSF and plasma. HD=Huntington's disease mutation carriers.

(A) and plasma (B) in HTT mutation carriers and controls. Boxes show first and third quartiles, the central band shows the median, and the whiskers show data within 1·5 IQR of the median. The dots represent outliers. (C) Correlation between NfL concentration in CSF and plasma. HD=Huntington's disease mutation carriers. Discussion In this genetically homogeneous cohort of HTT mutation carriers, uniquely large and well characterised among neurodegenerative disease cohorts, we found that NfL concentrations in plasma are increased compared with controls. Additionally, concentrations increased with advancing disease—most steeply in participants with premanifest Huntington's disease at baseline—and with increasing CAG triplet repeat counts. The concentration at a given timepoint reflected the degree of motor and cognitive impairment as well as global and regional brain volumes. NfL in plasma was a prognostic indicator for disease onset within 3 years in participants who were premanifest HTT mutation carriers at baseline, independently of previously known prognostic factors. Furthermore, it was indicative of the likely rate of worsening of cognition, functional ability, and brain atrophy beyond age and CAG repeat count. We are unaware of other substances in blood, or for that matter in CSF, that have shown similarly strong prognostic power longitudinally and across a broad range of clinical, functional, and imaging measures. Our findings suggest that NfL concentrations in plasma offer a rapid and accessible means of assessing and predicting neuronal damage in people with Huntington's disease.

tter in CSF, that have shown similarly strong prognostic power longitudinally and across a broad range of clinical, functional, and imaging measures. Our findings suggest that NfL concentrations in plasma offer a rapid and accessible means of assessing and predicting neuronal damage in people with Huntington's disease. The closer association between NfL concentrations in plasma and the rate of whole-brain atrophy than with the striatal atrophy rate suggests that this factor better reflects the global rate of neuronal damage. Whole-brain change is well described in all stages of Huntington's disease, although the striatum undergoes proportionately greater atrophy in the early stages.5 A blood marker of neuronal damage across the whole brain that would be expected to respond to amelioration of CNS pathology would be extremely helpful for therapeutic development. The striking association we found between NfL concentrations in plasma and HTT CAG repeat count establishes a genetic dose–response connection in Huntington's disease.

damage across the whole brain that would be expected to respond to amelioration of CNS pathology would be extremely helpful for therapeutic development. The striking association we found between NfL concentrations in plasma and HTT CAG repeat count establishes a genetic dose–response connection in Huntington's disease. This study allowed assessment of the longitudinal prognostic power of NfL concentrations in plasma. We were able to analyse in detail the associations with NfL from before disease onset and during manifest disease supported by rigorous clinical, cognitive, and neuroimaging data. Additionally, we were able to replicate the principal plasma findings in an independent cohort and confirm that concentrations in CSF are increased in individuals with HTT mutations, as has been previously reported.8, 9, 15, 16, 17 We further showed that NfL concentrations in plasma and CSF are closely correlated, which affirms the likely CNS origin of the NfL detected in both biofluids and suggests that NfL in blood could be a reliable estimator of concentrations in CSF.

iduals with HTT mutations, as has been previously reported.8, 9, 15, 16, 17 We further showed that NfL concentrations in plasma and CSF are closely correlated, which affirms the likely CNS origin of the NfL detected in both biofluids and suggests that NfL in blood could be a reliable estimator of concentrations in CSF. Our study is not without limitations. First, some of the cross-sectional and longitudinal correlations of NfL with existing outcome measures are slight, probably due to both biological and measurement variability. Accurate quantification of putaminal atrophy, for example, is particularly challenging. One potential advantage of measuring NfL is that repeated assessment is not needed to indicate the rate of change in the brain at a given timepoint. Thus, modest associations in a natural history study do not preclude interpretable changes in NfL concentrations in plasma in response to an intervention that ameliorates neuronal damage. Second, the analysis of the independent CSF cohort was not powered to compare the relative effect sizes of NfL concentrations in CSF and plasma and, therefore, we could not determine whether measurement in plasma is a sufficient alternative or whether there remains an additional value in quantification in CSF. Third, we do not yet have longitudinal data on NfL concentrations in CSF or predictive power of this measurement for Huntington's disease progression. Fourth, TRACK-HD did not include participants with advanced Huntington's disease, and further study is needed to understand the patterns of NfL concentrations across the whole disease spectrum. To address these issues and to enable head-to-head comparison of NfL with other proposed biochemical markers, we have recruited 80 participants in whom NfL concentrations in CSF will be measured longitudinally, supported by neuroimaging,28 and have launched a multisite CSF study, HDClarity (NCT02855476), that will include 600 participants with premanifest to advanced Huntington's disease and controls. Finally, we note that although NfL was a strong predictor of onset and progression overall in this study, its variability was too great to allow confident prediction in individuals. Moreover, the clinical relevance of any predicted changes cannot be inferred from this work.

o advanced Huntington's disease and controls. Finally, we note that although NfL was a strong predictor of onset and progression overall in this study, its variability was too great to allow confident prediction in individuals. Moreover, the clinical relevance of any predicted changes cannot be inferred from this work. Measurement of NfL concentrations in plasma yielded promising results as a prognostic blood biomarker of onset, progression, and neuronal damage in Huntington's disease. We suggest that this approach has a potential role, once validated to regulatory standards, in facilitating development of novel disease-modifying therapeutics and, possibly, guiding treatment decisions once such treatments become available. We recommend that quantification of NfL concentrations in plasma be included in future observational and therapeutic trials for Huntington's disease. Retrospective analysis in blood samples collected in previous trials might also be useful, to test for evidence that interventions had effects on neuronal damage, even if the clinical outcomes were negative. Supplementary Material Supplementary appendix

Measurement of NfL concentrations in plasma yielded promising results as a prognostic blood biomarker of onset, progression, and neuronal damage in Huntington's disease. We suggest that this approach has a potential role, once validated to regulatory standards, in facilitating development of novel disease-modifying therapeutics and, possibly, guiding treatment decisions once such treatments become available. We recommend that quantification of NfL concentrations in plasma be included in future observational and therapeutic trials for Huntington's disease. Retrospective analysis in blood samples collected in previous trials might also be useful, to test for evidence that interventions had effects on neuronal damage, even if the clinical outcomes were negative. Supplementary Material Supplementary appendix Acknowledgments Funding was provided by the Medical Research Council UK, GlaxoSmithKline, CHDI Foundation, Swedish Research Council, European Research Council, the Knut and Alice Wallenberg Foundation, and the Torsten Söderberg Foundation. This work was supported by the National Institute for Health Research UCL Hospitals Biomedical Research Centre and the Leonard Wolfson Experimental Neurology Centre, University College London. We thank Beth Borowsky for her assistance in planning and initiating this project. We acknowledge the Track-HD investigators who contributed data (appendix).

National Institute for Health Research UCL Hospitals Biomedical Research Centre and the Leonard Wolfson Experimental Neurology Centre, University College London. We thank Beth Borowsky for her assistance in planning and initiating this project. We acknowledge the Track-HD investigators who contributed data (appendix). Contributors LMB and EJW did the literature search. LMB, KB, AD, BRL, RACR, RIS, SJT, HZ, and EJW designed the study. Data were collected by LMB, AD, BRL, RACR, RIS, SJT, HZ, and EJW, analysed by LMB, FBR, RIS, and EJW, and interpreted by all authors. LMB, DL, and EJW created the figures. All authors were involved in the writing of the report. Declaration of interests EJW has been a participant on scientific advisory boards for Novartis, F Hoffmann-La Roche, Ionis, Shire, and Wave Life Sciences. SJT has been a participant on scientific advisory boards for F Hoffmann-La Roche Ltd, Ionis, Shire, and Teva Pharmaceuticals and received these honoraria through UCL Consultants Ltd, a wholly owned subsidiary of University College London, London, UK; University College London Hospitals NHS Foundation Trust receives funds as compensation for conducting clinical trials for Ionis Pharmaceuticals, Pfizer, and Teva Pharmaceuticals. KB has served at advisory boards or as a consultant for Alzheon, Eli Lilly, Fujirebio Europe, IBL International, and Roche Diagnostics. KB and HZ are co-founders of Brain Biomarker Solutions, Gothenburg, Sweden, part of the holding company GU Ventures at the University of Gothenburg.

Introduction Cognitive impairment in multiple sclerosis can occur from the earliest stages of the disease and its prevalence can exceed 80% in some studies of secondary progressive multiple sclerosis (SPMS).1, 2 The cognitive domains most frequently affected in multiple sclerosis are speed of information processing, attention, episodic memory, and executive function.1 The effect of cognitive impairment in multiple sclerosis on daily function can be substantial, and greater than the effect of physical disability on quality-of-life measures such as independence, social inclusion, and mood.3 In view of this effect, a 2013 international position paper highlighted development of effective interventions to treat cognitive impairment as a key goal in multiple sclerosis.4

ubstantial, and greater than the effect of physical disability on quality-of-life measures such as independence, social inclusion, and mood.3 In view of this effect, a 2013 international position paper highlighted development of effective interventions to treat cognitive impairment as a key goal in multiple sclerosis.4 So far, most studies of cognition in SPMS have been cross-sectional. Longitudinal studies have largely focused on other multiple sclerosis subtypes such as relapsing-remitting multiple sclerosis,5 and overall have differed with regard to the rate of progression of cognitive decline. Some studies have reported stability, whereas others have described declines in patient subgroups, in specific domains, or in global cognition.6, 7, 8 However, detailed understanding of the longitudinal pattern of cognitive decline in observational studies, specifically in SPMS, has been limited by small sample sizes, with this group typically investigated as part of larger cohorts that also included patients with different multiple sclerosis subtypes.8, 9 In a large study of patients with primary progressive multiple sclerosis (PPMS),7 baseline impairments of verbal memory, attention, verbal fluency, and spatial reasoning were identified, with cognitive decline occurring in a third of patients after 2 years. Research in context Evidence before this study

So far, most studies of cognition in SPMS have been cross-sectional. Longitudinal studies have largely focused on other multiple sclerosis subtypes such as relapsing-remitting multiple sclerosis,5 and overall have differed with regard to the rate of progression of cognitive decline. Some studies have reported stability, whereas others have described declines in patient subgroups, in specific domains, or in global cognition.6, 7, 8 However, detailed understanding of the longitudinal pattern of cognitive decline in observational studies, specifically in SPMS, has been limited by small sample sizes, with this group typically investigated as part of larger cohorts that also included patients with different multiple sclerosis subtypes.8, 9 In a large study of patients with primary progressive multiple sclerosis (PPMS),7 baseline impairments of verbal memory, attention, verbal fluency, and spatial reasoning were identified, with cognitive decline occurring in a third of patients after 2 years. Research in context Evidence before this study We searched MEDLINE (from 1948); Embase (from 1980); and PubMed, the Cochrane Database of Systematic Reviews, the Cochrane Central Register of Controlled Trials, DARE, the Health Technology Assessment database, and the TRIP database (no date restrictions) up to April 30, 2016, for studies with the key words “multiple sclerosis” AND “cognition” OR “neuropsychiatric features” OR “SF-36” OR more general MESH terms for quality of life: “Health Status Indicators”, “Quality of Life”, “Health Status”, “Outcome Assessment (Health Care)”, “Health Surveys”, and “Activities of Daily Living”. We included trials, observational cohort studies (longitudinal and cross-sectional), and systematic reviews. Additionally, we searched abstract books from the European Committee for Treatment and Research in Multiple Sclerosis meetings for the previous 10 years. The search yielded 174 observational cohorts of patients with progressive multiple sclerosis (alone or mixed populations), of which 26 were longitudinal studies of cognition or neuropsychiatric symptoms, or both. After excluding studies in which details of the multiple sclerosis phenotype or results of the cognitive or neuropsychiatric battery were not fully defined, 19 cohorts remained: 11 predominantly examined cognition and included two to 31 patients with secondary progressive multiple sclerosis (SPMS) whose follow-up ranged from 6 months to 17 years. Three Cochrane reviews have been published on pharmacological and neuropsychological treatments of cognition or memory in multiple sclerosis. In the review of pharmacological approaches, seven randomised controlled trials of those assessed by the Cochrane reviewers were suitable for inclusion. Donepezil, ginkgo biloba, memantine, and rivastigmine were examined with trial sizes of 43–126 patients; the maximum number of patients with SPMS was 39, treatment durations were up to 24 weeks, and all trials were judged to be negative. The original review of neuropsychological treatment was updated in 2014 with 20 randomised controlled trials deemed suitable.

vastigmine were examined with trial sizes of 43–126 patients; the maximum number of patients with SPMS was 39, treatment durations were up to 24 weeks, and all trials were judged to be negative. The original review of neuropsychological treatment was updated in 2014 with 20 randomised controlled trials deemed suitable. The individual trial sizes were 15–240 patients, the maximum number of patients with SPMS was 94 (where disease subtype was recorded), and follow-up ranged from 4 weeks to 1 year. There was insufficient evidence for efficacy, in part because of heterogeneity of the evidence base. In 2016, the Cochrane group added seven new trials to its review of techniques for memory rehabilitation, bringing the total number of assessed trials to 15, of which eight were also assessed in the neuropsychological review. Trial sizes were 19–240 patients, the maximum number of patients with SPMS was 94 (where disease subtype was recorded), and the follow-up ranged from 5 weeks to 8 months. Overall, the updated Cochrane review of memory rehabilitation concluded there was limited evidence to support such interventions, particularly with regards to objective memory testing, and that higher-quality studies were needed. Added value of this study

The individual trial sizes were 15–240 patients, the maximum number of patients with SPMS was 94 (where disease subtype was recorded), and follow-up ranged from 4 weeks to 1 year. There was insufficient evidence for efficacy, in part because of heterogeneity of the evidence base. In 2016, the Cochrane group added seven new trials to its review of techniques for memory rehabilitation, bringing the total number of assessed trials to 15, of which eight were also assessed in the neuropsychological review. Trial sizes were 19–240 patients, the maximum number of patients with SPMS was 94 (where disease subtype was recorded), and the follow-up ranged from 5 weeks to 8 months. Overall, the updated Cochrane review of memory rehabilitation concluded there was limited evidence to support such interventions, particularly with regards to objective memory testing, and that higher-quality studies were needed. Added value of this study This is, to our knowledge, the largest and longest detailed assessment of cognition and cognitive interventions in SPMS. This study was done in the context of a randomised controlled trial of simvastatin. Baseline assessment showed impairment on tests of frontal lobe function and several memory domains, which is largely consistent with previous findings. After 2 years, verbal and non-verbal memory had declined significantly. Depression worsened, although remained mild. We noted a positive effect from simvastatin on frontal lobe function, although we identified no specific association with frontal atrophy using MRI. We also identified a treatment effect for health-related quality of life (HRQoL—physical component).

d declined significantly. Depression worsened, although remained mild. We noted a positive effect from simvastatin on frontal lobe function, although we identified no specific association with frontal atrophy using MRI. We also identified a treatment effect for health-related quality of life (HRQoL—physical component). Implications of all the available evidence This study reinforces the fact that the domains of memory (verbal and non-verbal) and frontal lobe function (executive function) are preferentially affected in SPMS compared with other cognitive domains and should be targeted in interventional trials. We found a beneficial effect of simvastatin on the frontal cognitive domain, as well as on the physical component of HRQoL, adding to the substantive effect on whole brain atrophy (as well as clinician-reported and patient-reported outcome measures) noted in the original MS-STAT trial. We make the following recommendation for future trial design for cognitive treatments in multiple sclerosis: focus on single phenotypes (eg, SPMS), focus on specific cognitive domains with high baseline impairments (frontal lobe function, episodic memory, and attention and speed of information processing), include a minimum of 12–24 months of follow-up, and develop MRI interim outcomes beyond volumetry, such as other structural measures and functional connectivity models.

specific cognitive domains with high baseline impairments (frontal lobe function, episodic memory, and attention and speed of information processing), include a minimum of 12–24 months of follow-up, and develop MRI interim outcomes beyond volumetry, such as other structural measures and functional connectivity models. Various methods have been used to try to improve cognition in patients with multiple sclerosis, including disease-modifying drugs, acetylcholinesterase inhibitors, cognitive rehabilitation, and exercise, but have yielded variable results with no consistent evidence of benefit.4, 10 A Cochrane review of pharmacological treatment for memory impairment in patients with multiple sclerosis concluded with no evidence of any useful pharmacological approach (seven trials),11 although in another Cochrane review, some support was noted for various neuropsychological rehabilitation techniques (20 trials).12 Studies were hampered by mixed disease phenotypes, short durations, and differing outcome measures. A third Cochrane review13 concluded that there was limited evidence for memory rehabilition techniques and that more rigorous trial evidence was needed. The effects of exercise on cognition in multiple sclerosis has also been examined, with a pilot study in 42 patients showing some effect of high-intensity aerobic training on learning, memory, and attention.14

here was limited evidence for memory rehabilition techniques and that more rigorous trial evidence was needed. The effects of exercise on cognition in multiple sclerosis has also been examined, with a pilot study in 42 patients showing some effect of high-intensity aerobic training on learning, memory, and attention.14 As well as cognitive impairment, neuropsychiatric symptoms commonly occur in multiple sclerosis, with approximate prevalence of 20% for depression, 9% for anxiety, and 5% for bipolar disorder.15 More com-prehensive investigations have identified the occurrence of agitation, irritability, and apathy in multiple sclerosis.16 The inter-relationship and directionality between neuropsychiatric symptoms and cognitive impairment remains uncertain, with some reports suggesting an association, but others not.5 Compared with the general population, health-related quality of life (HRQoL) is reduced in patients with multiple sclerosis, particularly in those with progressive disease. In a 10-year study using the 36-Item Short Form Survey Instrument (SF-36) to assess HRQoL, decline related predominantly to problems affecting physical status.17 In the placebo arm of the IMPACT study18 of interferon beta-1a versus placebo in SPMS, the mean change over 2 years in the SF-36 physical component was −0·70 (SD 8·2) and in the psychological component was −1·6 (9·7).

rument (SF-36) to assess HRQoL, decline related predominantly to problems affecting physical status.17 In the placebo arm of the IMPACT study18 of interferon beta-1a versus placebo in SPMS, the mean change over 2 years in the SF-36 physical component was −0·70 (SD 8·2) and in the psychological component was −1·6 (9·7). MS-STAT19 was a phase 2 trial of high-dose simvastatin (80 mg) in patients with SPMS. A significant 43% reduction in the annualised rate of whole brain atrophy (the primary outcome) was noted, with additional positive effects on clinician-reported and patient-reported outcome measures, namely the Expanded Disability Status Scale (EDSS) and Multiple Sclerosis Impact Scale-29 (MSIS-29v2). The MS-STAT trial included a pre-planned secondary analysis of cognitive and neuropsychiatric outcome measures together with HRQoL, the results of which are presented here. These secondary analyses were designed to obtain detailed longitudinal imformation on cognitive impairment, neuropsychiatric symptoms and HRQoL, as well as to identify an interventional effect of simvastatin in patients with SPMS.

sychiatric outcome measures together with HRQoL, the results of which are presented here. These secondary analyses were designed to obtain detailed longitudinal imformation on cognitive impairment, neuropsychiatric symptoms and HRQoL, as well as to identify an interventional effect of simvastatin in patients with SPMS. Methods Study design and patients MS-STAT was a double-blind, parallel-group, randomised, placebo-controlled trial of simvastatin in patients with SPMS done at three neuroscience centres in the UK between Jan 28, 2008, and Nov 4, 2011. Details of MS-STAT, including the sample size calculation, have been published previously.19 In brief, the key study inclusion criteria included age 18–65 years, EDSS score 4·0–6·5, and fulfilment of the revised McDonald diagnostic criteria for multiple sclerosis with evidence of secondary progression over at least the preceding 2 years.20 No patients were on disease-modifying treatment. The protocol was approved by the institutional review board at each study site, and ethics approval was granted by the Berkshire Research Ethics Committee (reference 07/Q1602/73). Patients provided written informed consent before entering the study.

Methods Study design and patients MS-STAT was a double-blind, parallel-group, randomised, placebo-controlled trial of simvastatin in patients with SPMS done at three neuroscience centres in the UK between Jan 28, 2008, and Nov 4, 2011. Details of MS-STAT, including the sample size calculation, have been published previously.19 In brief, the key study inclusion criteria included age 18–65 years, EDSS score 4·0–6·5, and fulfilment of the revised McDonald diagnostic criteria for multiple sclerosis with evidence of secondary progression over at least the preceding 2 years.20 No patients were on disease-modifying treatment. The protocol was approved by the institutional review board at each study site, and ethics approval was granted by the Berkshire Research Ethics Committee (reference 07/Q1602/73). Patients provided written informed consent before entering the study. Randomisation and masking In MS-STAT, patients were randomly assigned (1:1) to receive simvastatin 80 mg daily or matching placebo for 24 months. Randomisation was by a centralised web-based service with minimisation on the following variables: age (<45 years and ≥45 years), sex (male and female), EDSS (4–5·5 and 6·0–6·5), centre, and assessing physician. Patients, treating physicians, and outcome assessors (including MRI scan analysts) were masked to treatment allocation.

on was by a centralised web-based service with minimisation on the following variables: age (<45 years and ≥45 years), sex (male and female), EDSS (4–5·5 and 6·0–6·5), centre, and assessing physician. Patients, treating physicians, and outcome assessors (including MRI scan analysts) were masked to treatment allocation. Procedures Patients received simvastatin 80 mg daily (two 40 mg tablets inside opaque hard gelatine capsules) or matching placebo (both groups received one tablet per day for the first month before having two per day from then on) for 24 months. Participants were considered compliant with treatment if they reported taking, on average, at least 90% of their drug at a dose of two tablets per day (80 mg simvastatin or matching placebo). Compliance was assessed by the self-reported proportion of capsules taken in the month before assessments at 6, 12, 18, and 24 months. Patients were assessed at baseline and months 1, 6, 12, and 24, with telephone follow-up at months 3 and 18. Cognitive and neuropsychiatric assessments were done at baseline, 12 months, and 24 months, by certified psychologists at three neuroscience centres in southeast England or at patients' homes if needed. HRQoL was assessed using the self-reported SF-36 (version 2), which was completed by patients at baseline, 12 months, and 24 months.

europsychiatric assessments were done at baseline, 12 months, and 24 months, by certified psychologists at three neuroscience centres in southeast England or at patients' homes if needed. HRQoL was assessed using the self-reported SF-36 (version 2), which was completed by patients at baseline, 12 months, and 24 months. Outcomes Analyses of cognitive and neuropsychiatric outcome measures together with HRQoL were prespecified as part of the MS-STAT trial. We used a neuropsychological battery designed specifically to cover a broad range of cognitive domains. The cognitive tests administered (panel) were the National Adult Reading Test (NART; one measure); Wechsler Abbreviated Scale of Intelligence (WASI; seven measures); Graded Naming Test (GNT; one measure); Birt Memory and Information Processing Battery (BMIPB; five measures); cube analysis task from the Visual Object and Space Perception battery (VOSP; one measure); Frontal Assessment Battery (FAB; one measure); and Paced Auditory Serial Addition Test (PASAT-3). The NART (premorbid intelligence quotient [IQ]) and the figure copying measure of the BMIPB (ability to copy figure when placed in front of patient, assessing visuoperceptual function) were done at first visit only as a baseline; thus, 15 longitudinal cognitive outcomes were assessed. Neuropsychiatric status was assessed using the Hamilton Depression Rating Scale (HAM-D) and the Neuropsychiatric Inventory Questionnaire (NPIQ), which has two subscales—severity and distress—representing a brief questionnaire form of the Neuropsychiatric Inventory. HRQoL was assessed using self-reported SF-36 (version 2) and standard scoring methods were used to convert SF-36 (version 2) scores into eight domains of HRQoL from which were derived a physical component summary (PCS), and a mental component summary.Panel Prespecified cognitive outcome measures • National Adult Reading Test—assessing premorbid IQ*

reported SF-36 (version 2) and standard scoring methods were used to convert SF-36 (version 2) scores into eight domains of HRQoL from which were derived a physical component summary (PCS), and a mental component summary.Panel Prespecified cognitive outcome measures • National Adult Reading Test—assessing premorbid IQ* • Wechsler Abbreviated Scale of Intelligence: • Overall IQ—assessing general intellectual function • Verbal IQ—assessing verbal IQ • Performance IQ—assessing non-verbal IQ • Vocabulary—assessing verbal intelligence • Similarities—assessing abstract verbal reasoning • Block design—assessing spatial perception and visuomotor skills • Matrix reasoning—assessing non-verbal abstract reasoning • Graded Naming Test—assessing semantic memory • Birt Memory and Information Processing Battery: • Story: immediate recall—assessing verbal episodic memory • Story: delayed recall—assessing verbal episodic memory • Figure copying: ability to copy figure when placed in front of participant*—assessing visuoperceptual functioning • Figure copying: immediate recall—assessing non-verbal episodic memory • Figure copying: delayed recall—assessing non-verbal episodic memory • Cube analysis task from the Visual Object and Space Perception battery—assessing spatial perception • Frontal Assessment Battery—assessing executive function (conceptual thinking, mental flexibility, motor programming, sensitivity to interference, inhibitory control, and environmental autonomy) • Paced Auditory Serial Addition Test—assessing speed of information processing, attention, and working memory

• Cube analysis task from the Visual Object and Space Perception battery—assessing spatial perception • Frontal Assessment Battery—assessing executive function (conceptual thinking, mental flexibility, motor programming, sensitivity to interference, inhibitory control, and environmental autonomy) • Paced Auditory Serial Addition Test—assessing speed of information processing, attention, and working memory We also did a post-hoc analysis of frontal lobe volumes (MRI scans were done in the trial at baseline, 12 months, and 25 months, as described previously19). The brain was initially parcellated using a multi-atlas propagation and fusion approach, as described by Cardoso and colleagues,21 which involves registration of atlas images with associated manual segmentations to each MS-STAT dataset. These propagated segmentations were then fused according to the local similarity between each atlas and the new images. We used the Hammers atlas as the multi-atlas propagation and fusion template database, resulting in 83 non-overlapping brain regions, 24 of which encompassed the frontal lobes. Because these regions were non-overlapping, total frontal lobe volumes were calculated from the summed volumes of these 24 frontal regions. Further details of methods are provided in the appendix (pp 1–2 and 4–7).

mplate database, resulting in 83 non-overlapping brain regions, 24 of which encompassed the frontal lobes. Because these regions were non-overlapping, total frontal lobe volumes were calculated from the summed volumes of these 24 frontal regions. Further details of methods are provided in the appendix (pp 1–2 and 4–7). Statistical analysis All analyses were by intention to treat (ie, included patients in the group to which they were randomly assigned, regardless of compliance with the study protocol). Cognitive scores and HRQoL scores were rescaled with reference to means and SDs of a healthy control group (appendix p 4) to create T scores. On the T score scale, healthy control scores have a mean of 50 and a SD of 10. The reference T scores were calculated using control means and SDs taken from test manuals (NART, WASI, and BMIPB) or reference papers from the published work (WASI and GNT). Age-specific control means were used for the BMIPB, for which age is known to be influential. The VOSP cube analysis and FAB scores are not normally distributed in healthy controls and thus unsuitable for conversion to T scores. Instead, these results are presented as raw scores, as are the measures of neuropsychiatric status and PASAT-3 scores. PASAT-3 scores were normally distributed, but have been previously reported as raw scores and are retained in this format in this Article to preserve consistency with the previous report.19 Impairment on cognitive scores and HRQoL scores at baseline was defined as a score of more than 1·5 SDs below the mean of the reference healthy control group (appendix p 4). On the T score scale this is a score less than 35. Neuropsychiatric symptoms were assessed in terms of prevalence and severity. Since these are symptoms indicative of disease and thus not normally present in healthy control individuals, severity was not investigated in terms of differences from normative values.

p 4). On the T score scale this is a score less than 35. Neuropsychiatric symptoms were assessed in terms of prevalence and severity. Since these are symptoms indicative of disease and thus not normally present in healthy control individuals, severity was not investigated in terms of differences from normative values. We used linear mixed models to examine how HRQoL, cognitive scores, neuropsychiatric scores, and frontal lobe volumes changed between baseline, 12 months, and 24 months, and to assess the difference in means between the placebo and simvastatin groups at 12 months and 24 months. For analysis of cognitive and neuropsychiatric scores, we used a mixed effect model to compare the mean of each outcome between the placebo and simvastatin groups at the visits at 12 months and 24 months, with adjustment for minimisation variables. In the model, measurements made at baseline, 12 months, and 24 months were classed as three correlated outcomes. Interactions between treatment group and visit were included to estimate the treatment effects, with that at baseline set as zero to reflect the fact that randomisation ensures that the true treatment effect at baseline is zero. We adjusted for minimisation variables by including interactions between visit and each variable. An unstructured residual variance–covariance matrix allowed for assessment of the correlation between repeated measures in the same patient. Data were included from all patients who had an outcome measured at one or more of the visits, providing an unbiased estimate of the treatment effect under the assumption that the model is correctly specified, and that data are missing at random. For each outcome, we estimated three separate fitted mean changes between baseline and 24 months from the relevant mixed effect model. First, we calculated the mean change for the cohort overall for a population with the same distribution of baseline values of the minimisation variables as the cohort as a whole, and 1:1 randomisation to treatment group. The other two mean changes were those for populations with the same distribution of minimisation variables as above, but with (1) all patients randomly assigned to placebo, and (2) all patients randomly assigned to simvastatin.

e minimisation variables as the cohort as a whole, and 1:1 randomisation to treatment group. The other two mean changes were those for populations with the same distribution of minimisation variables as above, but with (1) all patients randomly assigned to placebo, and (2) all patients randomly assigned to simvastatin. The mean rates of frontal lobe atrophy were compared between the two treatment groups using a linear mixed model for change per year. The model included an interaction between treatment group and time since baseline MRI as well as minimisation variables, MRI site, and their interactions with time. The treatment effect was set as zero at baseline, as described earlier. The model included a random intercept and random slope as correlated random effects, to allow for repeated measures.

between treatment group and time since baseline MRI as well as minimisation variables, MRI site, and their interactions with time. The treatment effect was set as zero at baseline, as described earlier. The model included a random intercept and random slope as correlated random effects, to allow for repeated measures. The assumption that residuals follow a homoscedastic normal distribution, which is required for valid parametric statistical inference, was not met for VOSP, FAB, NPIQ distress and NPIQ severity, or HRQoL domains of physical functioning, role limitations–physical, or role limitations–emotional. For these variables, inference was based on non-parametric, bias-corrected and accelerated bootstrap 95% CIs calculated from 2000 replications stratified by treatment group and clustered by patient. As a result of bootstrapping, p values cannot be provided. However, statistical significance (p<0·05) can be inferred by whether or not the 95% CI crosses zero. Since we class all the variables as of independent scientific interest, no formal statistical adjustments for multiple comparisons were made. However, the data must be interpreted with caution in view of the number of variables analysed. We used Stata (version 14.1) for all analyses. MS-STAT is registered with ClinicalTrials.gov, number NCT00647348.

s as of independent scientific interest, no formal statistical adjustments for multiple comparisons were made. However, the data must be interpreted with caution in view of the number of variables analysed. We used Stata (version 14.1) for all analyses. MS-STAT is registered with ClinicalTrials.gov, number NCT00647348. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit the manuscript. CF, JMN, JC, and RN had full access to all the data in the study; DC, SB, and DW had access to the cognitive data; and MJC and SO had access to the frontal lobe volumetry data. The corresponding author had final responsibility for the decision to submit for publication. Results 140 patients were recruited to the MS-STAT trial (figure 1). Patient demographics were similar between groups, with a mean age of 51·3 years (SD 6·9), multiple sclerosis duration of 21·2 years (8·6), and SPMS duration of 7·2 years (5·2; table 1). T scores of around 50 suggested that the premorbid IQ (derived from the NART) was in the normal range (table 1). At baseline, both study groups had similar scores on cognitive, neuropsychiatric, and HRQoL measures.Figure 1 Trial profile Table 1 Baseline characteristics of participants

Results 140 patients were recruited to the MS-STAT trial (figure 1). Patient demographics were similar between groups, with a mean age of 51·3 years (SD 6·9), multiple sclerosis duration of 21·2 years (8·6), and SPMS duration of 7·2 years (5·2; table 1). T scores of around 50 suggested that the premorbid IQ (derived from the NART) was in the normal range (table 1). At baseline, both study groups had similar scores on cognitive, neuropsychiatric, and HRQoL measures.Figure 1 Trial profile Table 1 Baseline characteristics of participants Combined Placebo Simvastatin N Mean (SD) N Mean (SD) N Mean (SD) Age (years) 140 51·3 (6·9) 70 51·1 (6·8) 70 51·5 (7·0) Multiple sclerosis duration (years) 140 21·2 (8·6) 70 20·3 (8·8) 70 22·1 (8·3) SPMS duration (years) 140 7·2 (5·2) 70 7·1 (4·8) 70 7·3 (5·6) Expanded Disability Status Scale 140 5·8 (0·8) 70 5·9 (0·8) 70 5·8 (0·8) Education (years) 132 13·5 (3·2) 66 13·4 (3·3) 66 13·7 (3·1) Premorbid verbal IQ (NART)* 135 52·7 (6·5) 68 51·7 (6·8) 67 53·7 (6·1) Premorbid performance IQ (NART)* 135 54·7 (4·6) 68 54·0 (4·8) 67 55·4 (4·4) Premorbid full-scale IQ(NART)* 135 53·9 (6·2) 68 53·0 (6·5) 67 54·9 (5·9) Cognitive scores Wechsler Abbreviated Scale of Intelligence* Overall IQ 130 53·1 (9·6) 66 53·4 (9·6) 64 52·8 (9·8) Verbal IQ 131 54·5 (9·5) 66 54·1 (9·7) 65 54·8 (9·3) Performance IQ 130 51·1 (9·5) 66 51·8 (9·1) 64 50·4 (10·0) Vocabulary 130 56·0 (10·2) 65 55·9 (9·3) 65 56·1 (11·1) Similarities 129 51·4 (9·3) 63 51·0 (9·6) 66 51·8 (9·0) Block design 129 48·2 (9·2) 64 48·6 (9·0) 65 47·7 (9·5) Matrix reasoning 129 53·4 (10·5) 65 54·6 (9·4) 64 52·2 (11·4) Graded Naming Test* 130 53·2 (12·2) 66 52·8 (12·0) 64 53·6 (12·6) Birt Memory and Information Processing Battery* Story immediate 131 43·2 (10·7) 66 42·7 (9·8) 65 43·7 (11·7) Story delay 131 43·7 (10·7) 66 43·1 (9·4) 65 44·3 (11·9) Figure copying 130 48·5 (11·5) 66 49·3 (12·1) 64 47·7 (10·8) Figure copying: immediate recall 130 41·0 (11·1) 66 40·5 (11·2) 64 41·5 (11·0) Figure copying: delayed recall 130 46·4 (9·2) 66 46·4 (9·4) 64 46·4 (9·1) VOSP cube analysis 132 9·2 (1·6) 67 9·3 (1·5) 65 9·1 (1·7) Frontal Assessment Battery 133 16·1 (2·4) 67 15·9 (2·5) 66 16·3 (2·4) Paced Auditory Serial Addition Test 134 34·2 (15·0) 67 33·7 (16·1) 67 34·8 (13·8) Neuropsychiatric scores Hamilton Depression Rating Scale 133 9·2 (6·2) 67 9·3 (6·1) 66 9·1 (6·3) NPIQ severity 112 4·5 (4·7) 55 4·9 (5·2) 57 4·1 (4·1) NPIQ distress 112 5·0 (5·5) 55 5·6 (6·1) 57 4·3 (4·7) Health-related quality of life* Mental component score 122 48·9 (11·1) 59 47·9 (12·6) 63 49·8 (9·5) Physical component score 122 33·7 (8·7) 59 32·8 (8·5) 63 34·5 (8·9) Physical functioning 130 27·1 (8·8) 62 25·6 (7·9) 68 28·4 (9·3) Role limitations physical 130 35·

NPIQ distress 112 5·0 (5·5) 55 5·6 (6·1) 57 4·3 (4·7) Health-related quality of life* Mental component score 122 48·9 (11·1) 59 47·9 (12·6) 63 49·8 (9·5) Physical component score 122 33·7 (8·7) 59 32·8 (8·5) 63 34·5 (8·9) Physical functioning 130 27·1 (8·8) 62 25·6 (7·9) 68 28·4 (9·3) Role limitations physical 130 35· 8 (9·9) 64 36·0 (10·2) 66 35·7 (9·7) Bodily pain 135 48·4 (11·4) 66 46·3 (11·6) 69 50·4 (10·9) General health 131 39·6 (10·7) 65 39·1 (9·7) 66 40·1 (11·6) Vitality 130 42·4 (9·0) 65 42·2 (9·5) 65 42·7 (8·4) Social functioning 134 39·2 (11·8) 66 38·0 (11·9) 68 40·3 (11·7) Role limitations emotional 133 44·3 (13·7) 66 43·8 (14·0) 67 44·8 (13·5) Mental health 131 47·3 (9·9) 64 45·9 (10·6) 67 48·6 (9·1) Premorbid IQ data were considered a baseline demographic characteristic, against which cognitive data could be compared. SPMS=secondary progressive multiple sclerosis. IQ=intelligence quotient. NART=National Adult Reading Test. VOSP=Visual Object and Space Perception. NPIQ=Neuropsychiatric Inventory Questionnaire. * T score (mean 50 [SD 10] in healthy reference population).

8 (9·9) 64 36·0 (10·2) 66 35·7 (9·7) Bodily pain 135 48·4 (11·4) 66 46·3 (11·6) 69 50·4 (10·9) General health 131 39·6 (10·7) 65 39·1 (9·7) 66 40·1 (11·6) Vitality 130 42·4 (9·0) 65 42·2 (9·5) 65 42·7 (8·4) Social functioning 134 39·2 (11·8) 66 38·0 (11·9) 68 40·3 (11·7) Role limitations emotional 133 44·3 (13·7) 66 43·8 (14·0) 67 44·8 (13·5) Mental health 131 47·3 (9·9) 64 45·9 (10·6) 67 48·6 (9·1) Premorbid IQ data were considered a baseline demographic characteristic, against which cognitive data could be compared. SPMS=secondary progressive multiple sclerosis. IQ=intelligence quotient. NART=National Adult Reading Test. VOSP=Visual Object and Space Perception. NPIQ=Neuropsychiatric Inventory Questionnaire. * T score (mean 50 [SD 10] in healthy reference population). The proportion of patients with cognitive impairment at baseline differed across cognitive domains (figure 2; appendix p 8). Three (2%) of 131 to ten (8%) of 129 patients showed impairment at baseline of general intellectual functioning in verbal, visuomotor, or abstract reasoning domains, as measured by the WASI. 14 (11%) of 130 were impaired in semantic memory (GNT). Between 13 (10%) of 130 and 43 (33%) of 130 patients were impaired on tests of immediate and delayed verbal and non-verbal episodic memory recall (BMIPB story and figure scores), with 15 (11%) of 132 impaired on higher visual processing using the VOSP cube analysis. 60 (45%) of 133 patients exhibited impairment in executive function assessed by the FAB, and 62 (46%) of 134 were impaired on the PASAT-3.Figure 2 Percentage of patients impaired in each measure at baseline in both groups combined

, with 15 (11%) of 132 impaired on higher visual processing using the VOSP cube analysis. 60 (45%) of 133 patients exhibited impairment in executive function assessed by the FAB, and 62 (46%) of 134 were impaired on the PASAT-3.Figure 2 Percentage of patients impaired in each measure at baseline in both groups combined Numerators and denominators used to calculate these percentages are provided in the appendix (p 8). WASI=Wechsler Abbreviated Scale of Intelligence. IQ=intelligence quotient. BMIPB=Birt Memory and Information Processing Battery. VOSP=Visual Object and Space Perception. HRQoL=health-related quality of life. MCS=mental component score. PCS=physical component score. HAM-D scores suggested that 76 (57%) of 133 patients had depression: 47 (35%) mild, 21 (16%) moderate, and eight (6%) severe or very severe. Although the NPIQ does not subdivide neuropsychiatric symptoms into mild, moderate, or severe categories, NPIQ severity (degree of symptoms in the patient) and distress (effect on caregiver) scores of 4·5 (SD 4·7) and 5·0 (5·5) were above mean scores of 1·5 and 1·6 for cognitively intact older adults, but below the mean scores of 7·9 and 9·4 associated with dementia disorders (appendix p1, p4).

erate, or severe categories, NPIQ severity (degree of symptoms in the patient) and distress (effect on caregiver) scores of 4·5 (SD 4·7) and 5·0 (5·5) were above mean scores of 1·5 and 1·6 for cognitively intact older adults, but below the mean scores of 7·9 and 9·4 associated with dementia disorders (appendix p1, p4). For HRQoL, 15 (12%) of 122 patients had impairment in the MCS and 68 (56%) in the PCS at baseline (figure 2; appendix p 8). Of the eight HRQoL domains, the highest proportion with impairment was for physical functioning (109 [84%] of 130), followed by role limitations–physical (67 [52%]). 18 (13%) of 135 patients had impairment on bodily pain and 17 (13%) of 135 had impairment on mental health.

the PCS at baseline (figure 2; appendix p 8). Of the eight HRQoL domains, the highest proportion with impairment was for physical functioning (109 [84%] of 130), followed by role limitations–physical (67 [52%]). 18 (13%) of 135 patients had impairment on bodily pain and 17 (13%) of 135 had impairment on mental health. At 24 months, frontal lobe function (FAB scores) was significantly better with simvastatin treatment than with placebo (difference 1·2 points, 95% CI 0·2 to 2·3; table 2). FAB score increased from baseline in the simvastatin group (change 0·3 points, 95% CI −0·4 to 0·9), whereas it decreased in the placebo group (change −0·9, −1·9 to −0·1; table 2). There was no significant difference between the placebo and simvastatin groups for any other cognitive or neuropsychiatric outcome (table 2), but we noted weak evidence of improvement in the simvastatin group compared with placebo at 24 months in the WASI block T score (difference 2·1, 95% CI −0·1 to 4·3; p=0·064) and PASAT-3 score (3·9 points, −0·3 to 8·1; p=0·070). The appendix (p 9) shows the individual patient changes in the FAB score.Table 2 Changes in cognitive and neuropsychiatric scores, health-related quality of life, and frontal lobe volume between baseline and 24 months

re (difference 2·1, 95% CI −0·1 to 4·3; p=0·064) and PASAT-3 score (3·9 points, −0·3 to 8·1; p=0·070). The appendix (p 9) shows the individual patient changes in the FAB score.Table 2 Changes in cognitive and neuropsychiatric scores, health-related quality of life, and frontal lobe volume between baseline and 24 months Combined Placebo Simvastatin Treatment effect (95% CI) p value N Change (95% CI) p value N Change (95% CI) N Change (95% CI) Cognitive scores Wechsler Abbreviated Scale of Intelligence* Overall IQ 135 −0·2 (−1·4 to 0·9) 0·67 68 −0·3 (−1·9 to 1·3) 67 −0·2 (−1·7 to 1·3) 0·1 (−2·0 to 2·2) 0·92 Verbal IQ 136 −1·4 (−2·8 to 0·0) 0·058 68 −1·2 (−3·2 to 0·7) 68 −1·5 (−3·4 to 0·3) −0·3 (−2·8 to 2·2) 0·82 Performance IQ 135 0·8 (−0·2 to 1·8) 0·13 68 0·3 (−1·1 to 1·8) 67 1·2 (−0·1 to 2·6) 0·9 (−1·0 to 2·9) 0·35 Vocabulary 137 −1·2 (−3·0 to 0·5) 0·17 68 −1·3 (−3·7 to 1·1) 69 −1·1 (−3·4 to 1·2) 0·2 (−2·9 to 3·3) 0·90 Similarities 136 −0·8 (−2·3 to 0·7) 0·30 67 −0·6 (−2·7 to 1·5) 69 −1·0 (−2·9 to 1·0) −0·3 (−3·1 to 2·4) 0·80 Block design 136 0·6 (−0·5 to 1·7) 0·29 68 −0·4 (−2·0 to 1·2) 68 1·6 (0·1 to 3·2) 2·1 (−0·1 to 4·3) 0·064 Matrix reasoning 136 0·9 (−0·8 to 2·6) 0·30 68 1·4 (−1·0 to 3·8) 68 0·4 (−1·9 to 2·7) −1·0 (−4·2 to 2·2) 0·55 Graded Naming Test* 135 −0·5 (−1·7 to 0·7) 0·44 68 −0·3 (−2·0 to 1·4) 67 −0·7 (−2·3 to 0·9) −0·4 (−2·7 to 1·9) 0·75 Birt Memory and Information Processing Battery* Story: immediate recall 136 −2·5 (−4·5 to −0·4) 0·018 68 −3·0 (−6·0 to −0·1) 68 −2·0 (−4·7 to 0·8) 1·1 (−2·9 to 5·0) 0·60 Story: delayed recall 136 −5·7 (−7·8 to −3·6) <0·0001 68 −6·5 (−9·5 to −3·4) 68 −5·0 (−7·8 to −2·1) 1·5 (−2·6 to 5·7) 0·47 Figure copying: immediate recall 135 −1·8 (−4·1 to 0·5) 0·12 68 −1·4 (−4·8 to 1·9) 67 −2·2 (−5·3 to 0·8) −0·8 (−5·2 to 3·6) 0·72 Figure copying: delayed recall 135 −6·8 (−8·7 to −4·8) <0·0001 68 −7·0 (−9·9 to −4·2) 67 −6·5 (−9·1 to −3·9) 0·5 (−3·4 to 4·4) 0·79 VOSP cube analysis† 137 0·0 (−0·3 to 0·3) .. 69 0·0 (−0·4 to 0·3) 68 0·0 (−0·4 to 0·4) 0·0 (−0·5 to 0·5) .. Frontal Assessment Battery† 138 −0·3 (−0·9 to 0·3) .. 69 −0·9 (−1·9 to −0·1) 69 0·3 (−0·4 to 0·9) 1·2 (0·2 to 2·3) .. Paced Auditory Serial Addition Test 140 2·0 (−0·2 to 4·2) 0·074 70 0·0 (−3·1 to 3·2) 70 3·9 (1·0 to 6·8) 3·9 (−0·3 to 8·1) 0·070 Neuropsychiatric measures Hamilton Depression Rating Scale 138 2·8 (1·5 to 4·0) <0·0001 69 3·3 (1·6 to 5·0) 69 2·3 (0·7 to 3·8) −1·0 (−3·2 to 1·2) 0·37 NPIQ severity† 127 0·5 (−0·7 to 1·7) ..

Paced Auditory Serial Addition Test 140 2·0 (−0·2 to 4·2) 0·074 70 0·0 (−3·1 to 3·2) 70 3·9 (1·0 to 6·8) 3·9 (−0·3 to 8·1) 0·070 Neuropsychiatric measures Hamilton Depression Rating Scale 138 2·8 (1·5 to 4·0) <0·0001 69 3·3 (1·6 to 5·0) 69 2·3 (0·7 to 3·8) −1·0 (−3·2 to 1·2) 0·37 NPIQ severity† 127 0·5 (−0·7 to 1·7) .. 64 0·7 (−0·7 to 2·6) 63 0·2 (−1·4 to 1·6) −0·6 (−2·7 to 1·6) .. NPIQ distress† 127 1·0 (−0·4 to 2·5) .. 64 1·6 (−0·5 to 3·9) 63 0·3 (−1·3 to 1·9) −1·3 (−3·9 to 1·1) .. Health-related quality of life* Mental component score 136 −0·8 (−3·0 to 1·4) 0·48 67 −0·7 (−3·8 to 2·3) 69 −0·8 (−3·7 to 2·0) −0·1 (−4·0 to 3·8) 0·96 Physical component score 136 −0·6 (−1·9 to 0·7) 0·36 67 −1·9 (−3·6 to −0·1) 69 0·7 (−1·0 to 2·3) 2·5 (0·3 to 4·8) 0·028 Physical functioning† 138 −2·5 (−3·8 to −1·4) .. 68 −2·3 (−4·0 to −0·7) 70 −2·8 (−4·4 to −1·4) −0·5 (−2·8 to 1·5) .. Role limitations–physical† 137 0·8 (−1·2 to 2·7) .. 68 −0·5 (−2·9 to 1·9) 69 2·1 (−0·6 to 4·8) 2·6 (−0·7 to 6·0) .. Bodily pain 139 −1·0 (−2·6 to 0·7) 0·24 69 −1·7 (−4·0 to 0·7) 70 −0·3 (−2·5 to 1·9) 1·4 (−1·7 to 4·4) 0·39 General health 139 −0·1 (−1·6 to 1·4) 0·89 69 0·1 (−2·1 to 2·3) 70 −0·3 (−2·3 to 1·7) −0·4 (−3·3 to 2·4) 0·77 Vitality 137 0·2 (−1·4 to 1·8) 0·82 68 0·3 (−1·9 to 2·6) 69 0·0 (−2·1 to 2·2) −0·3 (−3·2 to 2·7) 0·85 Social functioning 139 −1·9 (−4·2 to 0·5) 0·12 69 −0·8 (−4·0 to 2·4) 70 −2·9 (−5·9 to 0·1) −2·1 (−6·2 to 2·0) 0·31 Role limitations–emotional† 139 −1·4 (−4·1 to 1·3) .. 69 −2·3 (−6·0 to 1·3) 70 −0·5 (−4·2 to 2·9) 1·8 (−2·9 to 6·3) .. Mental health 138 −1·3 (−3·2 to 0·6) 0·19 68 −0·5 (−3·1 to 2·1) 70 −2·0 (−4·5 to 0·4) −1·5 (−4·9 to 1·9) 0·38 MRI Frontal lobe atrophy (mL/year)‡ 140 −1·0 (−1·3 to −0·6) <0·0001 70 −1·0 (−1·5 to −0·4) 70 −0·9 (−1·4 to −0·4) 0·0 (−0·7 to 0·7) 0·97 Imaging was done between baseline and 25 months. IQ=intelligence quotient. VOSP=Visual Object and Space Perception. NPIQ=Neuropsychiatric Inventory Questionnaire.

5 (−4·9 to 1·9) 0·38 MRI Frontal lobe atrophy (mL/year)‡ 140 −1·0 (−1·3 to −0·6) <0·0001 70 −1·0 (−1·5 to −0·4) 70 −0·9 (−1·4 to −0·4) 0·0 (−0·7 to 0·7) 0·97 Imaging was done between baseline and 25 months. IQ=intelligence quotient. VOSP=Visual Object and Space Perception. NPIQ=Neuropsychiatric Inventory Questionnaire. * T score. † As a result of bootstrapping, p values cannot be provided; however, significance (p<0·05) can be inferred by the 95% CI not crossing zero. ‡ Post-hoc analysis. In terms of HRQoL, there was a significant treatment effect on PCS (difference 2·5 points, 95% CI 0·3 to 4·8; p=0·028), which corresponded to a mean increase in PCS of 0·7 (95% CI −1·0 to 2·3) in the simvastatin group and a decrease of 1·9 (0·1 to 3·6) in the placebo group. The appendix (pp 9, 10) shows the individual patient changes in PCS and MCS. There was no evidence of differences between the placebo and simvastatin groups for MCS or any of the eight individual HRQoL domains (table 2).

% CI −1·0 to 2·3) in the simvastatin group and a decrease of 1·9 (0·1 to 3·6) in the placebo group. The appendix (pp 9, 10) shows the individual patient changes in PCS and MCS. There was no evidence of differences between the placebo and simvastatin groups for MCS or any of the eight individual HRQoL domains (table 2). Figure 3 shows mean changes in cognitive, neuropsychiatric, and HRQoL outcomes between baseline and 24 months. General intellectual functioning as measured by WASI did not change significantly across the duration of the study. There was no significant change in naming and verbal semantic memory (GNT). For the cohort as a whole, we noted the greatest decline on measures of delayed episodic memory recall (BMIPB), with the mean T-score worsening by 5·7 points (95% CI −7·8 to −3·6; p<0·0001) for delayed verbal recall and by 6·8 points (–8·7 to −4·8; p<0·0001) for non-verbal recall. Immediate verbal episodic recall showed a mean decline of 2·5 points (–4·5 to −0·4; p=0·018), but the 1·8 point drop in immediate non-verbal episodic recall was not significant (–4·1 to 0·5; p=0·12). There was no significant change in spatial perception (VOSP). There was an increase of 2·8 points (1·5 to 4·0) on the HAM-D (p<0·0001), although 64 (68%) of 94 patients' scores at 24 months still suggested either no depression or mild depression. There was no significant change in mean NPIQ severity or distress scores. Neither the MCS nor PCS HRQoL changed significantly between baseline and 24 months for the cohort as a whole. However, as mentioned earlier, the mean PCS decreased in the placebo group, but did not change substantially in the simvastatin group. Of the individual HRQoL domains, there was a 2·5 point decline in mean physical functioning score, whereas the other domains did not change significantly between baseline and 24 months.Figure 3 Change in cognitive scores, neuropsychiatric scores, and health-related quality of life between baseline and 24 months in both groups combined

l HRQoL domains, there was a 2·5 point decline in mean physical functioning score, whereas the other domains did not change significantly between baseline and 24 months.Figure 3 Change in cognitive scores, neuropsychiatric scores, and health-related quality of life between baseline and 24 months in both groups combined Bars are 95% CIs. In all tests, except HAM-D and NPIQ, a positive change suggests improvement. WASI=Wechsler Abbreviated Scale of Intelligence. IQ=intelligence quotient. BMIPB=Birt Memory and Information Processing Battery. VOSP=Visual Object and Space Perception. HAM-D=Hamilton Depression Rating Scale. NPIQ=Neuropsychiatric Inventory Questionnaire. HRQoL=health-related quality of life. MCS=mental component score. PCS=physical component score. *T score. In view of the effect on frontal lobe function, we did a post-hoc analysis using frontal lobe volumetry, which was judged to be the most appropriate technique for our stated region of interest. We noted no significant difference in the rate of change in frontal lobe volume between the simvastatin and placebo groups (table 2). There was no evidence of an association between change in FAB score between baseline and 24 months and change in frontal lobe volume between baseline and 25 months, either in the placebo (Spearman rank correlation 0·17; p=0·29; n=40) or simvastatin (–0·19; p=0·21; n=45) group.

simvastatin and placebo groups (table 2). There was no evidence of an association between change in FAB score between baseline and 24 months and change in frontal lobe volume between baseline and 25 months, either in the placebo (Spearman rank correlation 0·17; p=0·29; n=40) or simvastatin (–0·19; p=0·21; n=45) group. Discussion This is, to our knowledge, the largest reported cohort of patients with SPMS to have undergone detailed longitudinal assessment of cognition, neuropsychiatric status, and HRQoL. At baseline, the most prominent cognitive deficits were in attention and speed of information processing, frontal lobe function, verbal and non-verbal recall, and working memory. This profile is similar to that reported previously in other cross-sectional SPMS groups.2, 22, 23, 24 In terms of the effect of simvastatin on cognition, over the 24-month trial, changes in the overall cognitive profile were similar for both simvastatin and placebo groups, with no effect of simvastatin on most cognitive domains. However, there was evidence of a positive treatment effect on frontal lobe function, assessed using the FAB. Overall cognitive decline was greatest in verbal and non-verbal episodic memory recall, without significant changes in general intellectual function, naming, or higher visual processing. Although the much smaller cohorts of previous longitudinal SPMS studies (generally 20% of this study) precludes like-for-like comparison, changes over time were noted in episodic memory, learning, attention, speed of information, and visual processing.6, 8, 9

intellectual function, naming, or higher visual processing. Although the much smaller cohorts of previous longitudinal SPMS studies (generally 20% of this study) precludes like-for-like comparison, changes over time were noted in episodic memory, learning, attention, speed of information, and visual processing.6, 8, 9 In terms of neuropsychiatric outcome measures, over 24 months, depression increased significantly, as determined by the HAM-D score, and there was a non-significant increase in NPIQ severity and distress scores, although no treatment effect was noted on any neuropsychiatric outcome measure.

intellectual function, naming, or higher visual processing. Although the much smaller cohorts of previous longitudinal SPMS studies (generally 20% of this study) precludes like-for-like comparison, changes over time were noted in episodic memory, learning, attention, speed of information, and visual processing.6, 8, 9 In terms of neuropsychiatric outcome measures, over 24 months, depression increased significantly, as determined by the HAM-D score, and there was a non-significant increase in NPIQ severity and distress scores, although no treatment effect was noted on any neuropsychiatric outcome measure. The potential treatment effect on frontal lobe function warrants further discussion. Several studies have described impaired frontal lobe function in multiple sclerosis,22, 23, 24, 25, 26 and a recent, large cross-sectional study,25 including 74 patients with SPMS, revealed a heavy burden of executive dysfunction in all progressive multiple sclerosis subtypes. Although Ruano and colleagues25 showed patients with PPMS more frequently had executive dysfunction compared with patients with other subtypes of multiple sclerosis, others have shown frontal lobe capabilities to be affected to a greater extent in SPMS than in PPMS.22, 23, 24 In one study of patients with relapsing-remitting multiple sclerosis, impairment of frontal executive function was noted in the context of otherwise intact cognitive function,26 suggesting that this cognitive domain might be one of the earliest affected in multiple sclerosis. However, the reason for the apparently selective effect of simvastatin on frontal lobe function is unclear. There is no obvious pharmacological reason based on the current understanding of the mode of action of simvastatin that would result in a preferential improvement in frontal lobe function. Therefore, this finding might be due in part to the study population, in particular the level of impairment at baseline, with the FAB being one of the tests in which the greatest proportion of patients were affected (about 45%). As such, any treatment effect on other cognitive tests might have been more subtle, because proportionally fewer patients were impaired, and this study did not have sufficient power to detect these effects.

line, with the FAB being one of the tests in which the greatest proportion of patients were affected (about 45%). As such, any treatment effect on other cognitive tests might have been more subtle, because proportionally fewer patients were impaired, and this study did not have sufficient power to detect these effects. Various methods have been used in previous studies to assess frontal lobe function in multiple sclerosis, including bespoke frontal lobe test batteries,24 or as part of tests within global batteries such as the MindStreams Global Assessment Battery23 or Brief Repeatable Battery,7, 9 in some cases augmented by additional frontal lobe assessments.25 The FAB was chosen for this study in view of its ability to probe differing aspects of frontal lobe function, for which it has been used widely in the study of patients with frontotemporal dementia and other neurodegenerative disorders affecting the frontal lobes. Furthermore, the reproducibility and ease of administration of the FAB confers advantages in terms of application to large multiple sclerosis cohorts. The FAB has previously been used in patients with multiple sclerosis, principally in studies focused on assessment of quality of life and coping strategies, but also as part of an executive function battery.2, 27 To our knowledge, this is the first study to use the FAB as an independently reported cognitive outcome measure within a longitudinal interventional study, and the study findings show the importance of including a comprehensive assessment of frontal lobe function in future multiple sclerosis interventional studies, either as an individual outcome measure or in addition to current batteries such as the Brief International Cognitive Assessment for MS.5

interventional study, and the study findings show the importance of including a comprehensive assessment of frontal lobe function in future multiple sclerosis interventional studies, either as an individual outcome measure or in addition to current batteries such as the Brief International Cognitive Assessment for MS.5 By contrast with the treatment effect on whole brain atrophy rates reported previously,19 we noted no significant effect on rates of frontal lobe atrophy, and there was no significant correlation between FAB scores and rates of frontal lobe atrophy. Several potential explanations exist for this apparent dissociation. Other imaging measures might be superior predictors of executive function than atrophy.5, 28 Corpus callosum atrophy might outperform other imaging markers of cognitive dysfunction, such as grey or white matter fraction, and tracked over 17 years, was associated with cognitive dysfunction in multiple sclerosis subtypes including SPMS;29 other studies have used resting-state functional MRI to identify an association between cognitive impairment and altered functional connectivity.5 More recently, techniques such as thalamic volume and activation28, 30 and cortical lesions visible via high-field MRI31 have shown associations with cognition.

ding SPMS;29 other studies have used resting-state functional MRI to identify an association between cognitive impairment and altered functional connectivity.5 More recently, techniques such as thalamic volume and activation28, 30 and cortical lesions visible via high-field MRI31 have shown associations with cognition. We also showed evidence that simvastatin treatment had a positive effect on physical HRQoL, as measured using the SF-36 (version 2) PCS. This finding is consistent with the positive treatment effect on the EDSS and MSIS-29v2 (especially the physical subscale) previously reported.19 The absence of effect on the mental component scale of the HRQoL is consistent with the absence of effect on the neuropsychiatric outcome measures and MSIS-29v2 psychological subscale.

ng is consistent with the positive treatment effect on the EDSS and MSIS-29v2 (especially the physical subscale) previously reported.19 The absence of effect on the mental component scale of the HRQoL is consistent with the absence of effect on the neuropsychiatric outcome measures and MSIS-29v2 psychological subscale. This study has several limitations. First, in view of the need for comprehensive assessment of cognition and neuropsychiatric status, this study involved analysis of data from 15 cognitive outcomes and three neuropsychiatric outcomes. Caution is therefore warranted in the interpretation of the evidence of a positive treatment effect on the measure of frontal lobe function and confirmation of this finding in independent studies is needed. The fluctuating nature of disease activity in all patients with multiple sclerosis, and its susceptibility to environmental factors such as heat and concomitant illness, are confounders in the study of cognitive function in multiple sclerosis.32 Such fluctuations might lead to variability in test performance, which would reduce the power of the study to detect an effect of treatment on cognition. Although patients were screened for possible concurrent acute medical disorders at each study visit, the possibility of variability in test performance as a result of external factors cannot be entirely excluded. The effect of patient dropout on data analysis also needs to be considered. However, this effect was offset in this study by the use of a statistical model that maximised data inclusion by incorporating all available cognitive assessments for all patients, thus ensuring that the estimated treatment effect was unbiased if data were missing at random.

n data analysis also needs to be considered. However, this effect was offset in this study by the use of a statistical model that maximised data inclusion by incorporating all available cognitive assessments for all patients, thus ensuring that the estimated treatment effect was unbiased if data were missing at random. In conclusion, we found evidence of a positive effect of treatment with high-dose simvastatin on frontal lobe function and a physical quality-of-life measure, adding to our previous findings of a treatment effect on the annualised rate of whole brain atrophy.19 Supplementary Material Supplementary appendix

n data analysis also needs to be considered. However, this effect was offset in this study by the use of a statistical model that maximised data inclusion by incorporating all available cognitive assessments for all patients, thus ensuring that the estimated treatment effect was unbiased if data were missing at random. In conclusion, we found evidence of a positive effect of treatment with high-dose simvastatin on frontal lobe function and a physical quality-of-life measure, adding to our previous findings of a treatment effect on the annualised rate of whole brain atrophy.19 Supplementary Material Supplementary appendix Acknowledgments This study was funded by The Moulton Foundation (charity number 1109891), Berkeley Foundation (268369), MSTC (1113598), the Rosetrees Trust (298582), a personal contribution from A W Pidgley CBE, and the UK National Institute for Health Research (NIHR) University College London Hospitals Biomedical Research Centre and University College London. DC is funded by the Cambridge UK NIHR Biomedical Research Centre. SB holds an NIHR Academic Clinical Fellowship. SO receives funding from the EPSRC (EP/H046410/1, EP/J020990/1, EP/K005278), the MRC (MR/J01107X/1), the EU-FP7 (FP7-ICT-2011-9-601055), and NIHR UCLH BRC (BW.mn.BRC10269). JC has received funding from the NIHR University College London Hospitals Biomedical Research Centre and University College London. We thank Rebecca Cooper, Katie Meadmore, and Veronica Bradley for their contribution to the cognitive and neuropsychiatric testing of the MS-STAT patients. We thank all those who worked on the original MS-STAT study and the patients who participated in the MS-STAT trial.

l Research Centre and University College London. We thank Rebecca Cooper, Katie Meadmore, and Veronica Bradley for their contribution to the cognitive and neuropsychiatric testing of the MS-STAT patients. We thank all those who worked on the original MS-STAT study and the patients who participated in the MS-STAT trial. Contributors JC devised the overall MS-STAT study and DC conceived the cognitive component of MS-STAT. JC, DC, CF, and RN designed the study. JC, DC, SB, and RN did the literature search. JC, DC, SB, DW, and RN collected the data. SB, JN, and CF analysed data. MJC and SO developed the parcellation algorithms. MJC did the MRI quality control and analysis under the supervision of SO. All authors wrote the manuscript. Declaration of interests JC and RN are trustees of the Multiple Sclerosis Trials Collaboration (MSTC). JMN and DW have received grant support from the MSTC. All other authors declare no competing interests. * Assessed at baseline only; therefore, 15 cognitive outcomes were reported longitudinally. IQ=intelligence quotient.

Stroke was the leading cause of age-standardised DALY rates in 18 of 21 GBD regions, while migraine and Alzheimer's disease and other dementias were ranked among the top three causes in all regions except Oceania, south Asia, and the four sub-Saharan African regions, where epilepsy or meningitis ranked higher. Communicable neurological conditions ranked low in high-income regions and central Europe. Epilepsy and medication overuse headache ranked fourth, fifth, or sixth in almost all regions (figure 4).Figure 4 Ranking of age-standardised DALY rates for all neurological disorders by GBD region in 2015 Data are for both sexes. DALYs=disability-adjusted life-years. A more detailed breakdown of geographical variation is shown as maps of age-standardised prevalence, death rates, and DALY rates for each neurological disorder (appendix pp 156–69). The prevalence of stroke was highest in eastern Europe, central Asia, Oceania, Indonesia, Myanmar, and sub-Saharan African countries. High death rates in Mongolia, Afghanistan, and Central African Republic were responsible for these countries also having the highest age-standardised DALY rates for stroke. Prevalence, death, and DALY rates for Alzheimer's and other dementias were highest in North America and north Africa and the Middle East, while the lowest rates were estimated for sub-Saharan Africa, southern Latin America, and Australasia. Prevalence and DALY rates for Parkinson's disease were highest in high-income regions and lowest in sub-Saharan Africa and eastern Europe. The strong positive relationship with latitude was apparent in the distribution of prevalence, death, and DALY rates for multiple sclerosis. High death rates due to epilepsy were the cause of high DALY rates in sub-Saharan Africa. Prevalence of epilepsy was highest in central Latin America, Chile, north Africa and the Middle East, and Bangladesh. Highest rates of motor neuron disease occurred in high-income regions.

h, and DALY rates for multiple sclerosis. High death rates due to epilepsy were the cause of high DALY rates in sub-Saharan Africa. Prevalence of epilepsy was highest in central Latin America, Chile, north Africa and the Middle East, and Bangladesh. Highest rates of motor neuron disease occurred in high-income regions. The prevalence and DALY rates for the different types of headache varied by a factor of three to four between countries with the highest rates in high-income regions, north Africa and the Middle East, and tropical Latin America, while lowest rates were seen in sub-Saharan Africa and east Asia. Medication overuse headache was particularly common in eastern Europe and Iran. The highest death and DALY rates for brain and nervous system cancers were in central Europe and north Africa and the Middle East. Meningitis caused most disease burden in sub-Saharan Africa, especially in the meningitis belt across the Sahara. The largest burden from tetanus was estimated for a few countries in east Africa, particularly South Sudan and Somalia, which have low vaccination coverage owing to national conflict. India had the largest burden from encephalitis.

Introduction In 2006, WHO emphasised the importance of neurological disorders (a group that at the time included epilepsy, Alzheimer's disease and other dementias, Parkinson's disease, multiple sclerosis, migraine, stroke, poliomyelitis, tetanus, meningitis, and Japanese encephalitis) for public health,1 and estimated that these disorders accounted for 6·3% of the global disability-adjusted life-years (DALYs). However, the WHO report was largely based on the Global Burden of Disease Study 2000 (GBD 2000)2 and did not reflect understanding of the comparative global epidemiology of neurological disorders. Increasing life expectancy and population growth worldwide in recent years mean that more people are now reaching ages where neurological disorders are most prevalent. Additionally, there have been changes in the treatment of some neurological disorders (such as the introduction of acute stroke units, thrombolysis, and thrombectomy), as well as changes in the risk factors that affect the burden of neurological disorders in the world. Research in context Evidence before this study The Global Burden of Disease Study 2015 (GBD 2015) produced comprehensive and comparable estimates of the burden of various disorders (prevalence, deaths, and disability-adjusted life years [DALYs]) for 195 countries and territories from 1990 to 2015. However, aggregated estimates in the category of neurological disorders did not include data for stroke, brain and nervous system cancer, or communicable neurological disorders, which are all of great relevance to neurologists. Added value of this study

The Global Burden of Disease Study 2015 (GBD 2015) produced comprehensive and comparable estimates of the burden of various disorders (prevalence, deaths, and disability-adjusted life years [DALYs]) for 195 countries and territories from 1990 to 2015. However, aggregated estimates in the category of neurological disorders did not include data for stroke, brain and nervous system cancer, or communicable neurological disorders, which are all of great relevance to neurologists. Added value of this study For GBD 2015, the GBD 2013 search strategy was replicated to take into account all relevant epidemiological studies published between 2013 and 2015. In this systematic analysis, we produced estimates of the burden of neurological disorders (stroke, Alzheimer's disease and other dementias, Parkinson's disease, epilepsy, multiple sclerosis, migraine and tension-type headache, medication overuse headache, meningitis, tetanus, encephalitis, brain and nervous system cancers, motor neuron disease, and a residual category of other neurological disorders), separately and combined. For the first time, we included stroke, brain and nervous system cancer, and communicable neurological disorders. We quantified the global disease burden (as measured by prevalence, mortality, DALYs, years of life lost, and years lived with disability) and explored variation in the burden by neurological disorder type, age, sex, and overall country development level as measured by the Socio-demographic Index (SDI). The study showed that neurological disorders ranked as the leading cause of DALYs in 2015 and the second-leading cause of death. The most prevalent neurological disorders were headaches (tension-type headache, medication overuse headache, and migraine) and Alzheimer's disease and other dementias.

mographic Index (SDI). The study showed that neurological disorders ranked as the leading cause of DALYs in 2015 and the second-leading cause of death. The most prevalent neurological disorders were headaches (tension-type headache, medication overuse headache, and migraine) and Alzheimer's disease and other dementias. Implications of all the available evidence Over the past 25 years, the burden of neurological disorders has increased substantially, and our expanded list of neurological disorders makes this the leading cause group of disability, and the second-leading cause group of mortality worldwide. Stroke remains the largest contributor to this burden globally. For all neurological disorders combined, there was a noticeable (by almost five times) gradual decrease in age-standardised DALY rates with increasing SDI, but these rates had large geographical variations. These findings could help to guide health systems and research activities to reduce the burden of neurological disorders. To improve health-care planning and health outcomes of people with neurological disorders, we must understand not only the number and distribution of people with these disorders between countries, but also how these disorders affect population health (in terms of both mortality and disability) compared with other diseases and injuries.

-care planning and health outcomes of people with neurological disorders, we must understand not only the number and distribution of people with these disorders between countries, but also how these disorders affect population health (in terms of both mortality and disability) compared with other diseases and injuries. The Global Burden of Disease Study 2015 (GBD 2015) GBD 20153, 4 produced comprehensive and comparable estimates of the burden (prevalence, deaths, and DALYs) of 315 diseases and injuries for 195 countries and territories from 1990 to 2015. However, aggregated estimates in the category of neurological disorders did not include stroke, brain and nervous system cancer, or communicable neurological disorders, which are of great relevance to neurologists. Ischaemic and haemorrhagic stroke were classified as cardiovascular diseases; brain and nervous system cancer as malignant neoplasms; and meningitis, encephalitis, and tetanus as communicable diseases. In this systematic analysis, we quantify the disease burden (as measured by prevalence, mortality, DALYs, and years lived with disability [YLDs]) due to neurological disorders and its relationship with development level as measured by the Socio-demographic Index (SDI; an indicator based on lagged-distributed income per person, average educational attainment among individuals older than 15 years, and total fertility rate).

and years lived with disability [YLDs]) due to neurological disorders and its relationship with development level as measured by the Socio-demographic Index (SDI; an indicator based on lagged-distributed income per person, average educational attainment among individuals older than 15 years, and total fertility rate). Methods Categorisation of disorders In an expanded category of neurological disorders we included stroke, Alzheimer's disease and other dementias, Parkinson's disease, epilepsy, multiple sclerosis, migraine, tension-type headache, medication overuse headache, meningitis, tetanus, encephalitis, brain and nervous system cancer, motor neuron disease, and a residual category of other neurological disorders that included diseases such as muscular dystrophy and Huntington's disease (appendix p 112).

multiple sclerosis, migraine, tension-type headache, medication overuse headache, meningitis, tetanus, encephalitis, brain and nervous system cancer, motor neuron disease, and a residual category of other neurological disorders that included diseases such as muscular dystrophy and Huntington's disease (appendix p 112). Non-fatal estimates GBD 2015 non-fatal burden estimates were based on a systematic review of the literature to obtain all available epidemiological data on prevalence, incidence, risk of mortality, and severity. In the appendix we provide for each neurological disorder analysed the following: the list of International Classification of Diseases (ICD) codes used for mapping neurological causes of death (pp 118–21); a list of GBD sequelae, health states, health state lay descriptions, and disability weights for neurological disorders (pp 122–39); the total number of site-years by neurological cause and source type (p140); and the data representativeness index for each neurological disorder, the percentage of GBD 2015 geographies with any data by cause pertaining to the period before 2005, 2005–15, and all years of data (p 141). Reference case definitions were based on ICD-9 or ICD-10 criteria with the addition of Diagnostic and Statistical Manual of Mental Disorders (DSM)-III and DSM-IV criteria for dementia and the International Classification of Headache Disorders criteria for headaches.5, 6 Sources of information used to estimate the burden of neurological disorders are on the Global Health Data Exchange website.

addition of Diagnostic and Statistical Manual of Mental Disorders (DSM)-III and DSM-IV criteria for dementia and the International Classification of Headache Disorders criteria for headaches.5, 6 Sources of information used to estimate the burden of neurological disorders are on the Global Health Data Exchange website. Detailed GBD methods for calculating non-fatal estimates have been reported elsewhere.4 The epidemiological data were analysed with DisMod-MR 2.1,7 a Bayesian meta-regression tool that adjusts datapoints for variations in study methods between data sources and enforces consistency between data for different parameters, such as incidence and prevalence. For each neurological disorder, we defined a parsimonious set of sequelae that best described different aspects of the disabling consequences. Each non-fatal sequela was estimated separately. We calculated the YLDs caused by the residual category of other neurological disorders indirectly using a ratio of YLDs to years of life lost (YLLs). We calculated the ratio of YLDs to YLLs for Alzheimer's disease and other dementias, Parkinson's disease, multiple sclerosis, and motor neuron disease, and multiplied this ratio by the YLL estimates for other neurological disorders. Further details of the non-fatal estimates of each of the included neurological disorders are in the appendix (pp 5–113).

YLLs for Alzheimer's disease and other dementias, Parkinson's disease, multiple sclerosis, and motor neuron disease, and multiplied this ratio by the YLL estimates for other neurological disorders. Further details of the non-fatal estimates of each of the included neurological disorders are in the appendix (pp 5–113). Disability weights for a set of 235 health states covering all sequelae of disease and injury estimated in GBD 2015 were estimated by pair-wise comparison methods presenting pairs of lay health state descriptions to respondents in surveys done in the general population in nine countries, and an open web-based survey.8 The sequelae of the neurological disorders included in this analysis each map to a unique health state with a corresponding disability weight (appendix pp 122–39). YLDs were computed by multiplying the prevalence of each sequela by a disability weight and aggregating estimates for all sequelae for a disease. We categorised countries by their overall development status level as determined by the SDI, classifying them in high, high–medium, medium, medium–low, and low SDI quintiles on the basis of values across the 1980–2015 period (appendix pp 142–46), as described in detail elsewhere.9 The average expected relationships between DALY rates and death rates from neurological disorders (individually and as a group) and SDI over the entire study period (1990–2015) across all geographies in males and females were estimated using spline regression. We also categorised countries into 21 GBD regions (appendix p 155).

age expected relationships between DALY rates and death rates from neurological disorders (individually and as a group) and SDI over the entire study period (1990–2015) across all geographies in males and females were estimated using spline regression. We also categorised countries into 21 GBD regions (appendix p 155). Causes of death GBD 2015 uses a database of 14 236 site-years of vital registration, verbal autopsy (a method of determining cause of death in countries that have no functional vital registration system), and maternal and child death surveillance data, covering the period from 1980 to 2015.9 Trained interviewers administer a structured questionnaire to relatives of a deceased person about their symptoms preceding death. The underlying cause of death is inferred by computer algorithms or physician review of the autopsy interview. We estimated deaths for all neurological disorders apart from headaches, to which no deaths are assigned as the underlying cause. For each neurological cause except dementia, we used the GBD Cause of Death Ensemble model (CODEm) strategy.10, 11 CODEm applies mixed effects or spatiotemporal Gaussian process regression models to mortality rates or cause fractions in varying combinations with predictive covariates. The ensemble of models with best credentials on out-of-sample predictive validity tests was selected for each cause of death. YLLs were calculated by multiplying the number of deaths at each age by the standard life expectancy at that age.12 Results from CODEm for each disease were scaled to fit all-cause mortality estimates derived from demographic sources by location, age, year, and sex.

validity tests was selected for each cause of death. YLLs were calculated by multiplying the number of deaths at each age by the standard life expectancy at that age.12 Results from CODEm for each disease were scaled to fit all-cause mortality estimates derived from demographic sources by location, age, year, and sex. We decided to use a natural history model for dementia because of a large inconsistency between the data for prevalence and mortality. For instance, in the USA, the rates of death from dementia increased five times between 1990 and 2014, whereas the available prevalence and incidence data showed no significant changes over the same period. Large increases in death rates assigned to dementia have also occurred in some other countries with high-quality vital registration systems. Furthermore, in GBD 2015, the prevalence of dementia varied among 187 countries by a factor of three, whereas dementia death rates varied by more than 20 times.13 The likely explanation was a change in coding practices between countries and within countries over time. To correct for this source of measurement bias, we assessed the most recent data from 23 high-income countries with high-quality vital registration systems and the highest ratio of registered dementia death rates to prevalent cases. This ratio is equivalent to the excess rate of mortality in cases of dementia. We derived a pooled estimate by age and sex using a linear regression of the log of these rates. We added these estimates as data in DisMod-MR 2.1 to derive estimates of cause-specific mortality rates that were consistent with prevalence data and the pooled estimate of excess mortality from the 23 countries that in their most recent year of vital registration were most willing to code a death to dementia as the underlying cause.10

timates as data in DisMod-MR 2.1 to derive estimates of cause-specific mortality rates that were consistent with prevalence data and the pooled estimate of excess mortality from the 23 countries that in their most recent year of vital registration were most willing to code a death to dementia as the underlying cause.10 Compilation of results DALYs were computed as the sum of YLLs and YLDs for each country, age, sex, and year with 95% uncertainty intervals (UIs) based on the 25th and 975th values of the ordered 1000 draws. Unless explicitly mentioned otherwise, all rates were age-standardised using the GBD standard.10 This study is compliant with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER; appendix pp 114–17).14, 15 Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding authors had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Compilation of results DALYs were computed as the sum of YLLs and YLDs for each country, age, sex, and year with 95% uncertainty intervals (UIs) based on the 25th and 975th values of the ordered 1000 draws. Unless explicitly mentioned otherwise, all rates were age-standardised using the GBD standard.10 This study is compliant with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER; appendix pp 114–17).14, 15 Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding authors had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Global burden In 2015, the neurological disorders included in this analysis caused 250·692 (95% UI 229·080 to 274·654) million DALYs, comprising 10·2% of global DALYs, and 9·399 (9·095 to 9·714) million deaths, comprising 16·8% of global deaths. We were unable to report the combined prevalence. Classifying stroke under neurological disorders rather than cardiovascular diseases made neurological disorders the largest cause group of DALYs and the second-largest behind cardiovascular diseases in terms of global deaths. Among the neurological disorders, stroke, migraine, meningitis, Alzheimer's disease and other dementias, and epilepsy each caused more than 10 million DALYs. The most prevalent neurological disorders were the various types of headache and Alzheimer's disease and other dementias (table 1).Table 1 Global absolute numbers and age-standardised rates per 100 000 people for neurological disorders and percentage changes between 1990 and 2015

each caused more than 10 million DALYs. The most prevalent neurological disorders were the various types of headache and Alzheimer's disease and other dementias (table 1).Table 1 Global absolute numbers and age-standardised rates per 100 000 people for neurological disorders and percentage changes between 1990 and 2015 All-age numbers (thousands) Age-standardised rate (per 100 000) 2015 Change from 1990 to 2015 2015 Change from 1990 to 2015 Tetanus DALYs 3510 (3002 to 4503) −86·3% (−88·1 to −83·3) 47 (40 to 61) −87·4% (−89·0 to −84·9) Deaths 57 (48 to 80) −83·4% (−85·5 to −78·8) 1 (1 to 1) −85·7% (−87·5 to −82·4) Prevalence 209 (205 to 215) 10·6% (5·1 to 15·2) 3 (3 to 3) −15·4% (−19·6 to −12·3) Meningitis DALYs 25 395 (21 653 to 30 649) −31·9% (−42·6 to −10·1) 342 (292 to 413) −39·8% (−48·8 to −22·0) Deaths 379 (323 to 445) −25·2% (−35·6 to −4·2) 5 (5 to 6) −38·5% (−45·9 to −23·7) Prevalence 8734 (8321 to 9107) 27·9% (24·3 to 31·7) 120 (114 to 125) −10·6% (−13·2 to −8·0) Encephalitis DALYs 8453 (7669 to 9412) −14·0% (−25·8 to −0·1) 115 (104 to 128) −29·0% (−38·4 to −18·7) Deaths 150 (138 to 167) −3·4% (−15·4 to 10·9) 2 (2 to 2) −29·6% (−38·1 to −20·1) Prevalence 4316 (3146 to 5876) 13·5% (9·3 to 19·2) 59 (43 to 80) −23·7% (−26·9 to −19·2) Stroke DALYs 118 627 (114 862 to 122 627) 21·7% (17·8 to 25·7) 1777 (1721 to 1835) −32·3% (−34·4 to −30·0) Deaths 6326 (6175 to 6493) 36·4% (32·4 to 40·8) 101 (99 to 104) −30·0% (−32·0 to −27·7) Prevalence 42 431 (42 068 to 42 767) 59·2% (58·4 to 59·9) 627 (621 to 631) −9·8% (−10·3 to −9·4) Alzheimer's disease and other dementias DALYs 23 779 (20 118 to 27 886) 98·4% (95·0 to 101·9) 396 (334 to 464) −5·7% (−7·2 to −4·2) Deaths 1908 (1587 to 2229) 114·9% (111·1 to 119·6) 32 (27 to 38) −3·4% (−4·8 to −1·4) Prevalence 45 956 (40 179 to 52 656) 111·7% (109·1 to 114·2) 761 (663 to 876) 2·4% (1·7 to 3·2) Parkinson's disease DALYs 2059 (1832 to 2321) 111·2% (102·4 to 118·1) 33 (30 to 37) 10·8% (6·5 to 14·3) Deaths 117 (114 to 121) 149·8% (135·0 to 161·4) 2 (2 to 2) 22·6% (15·7 to 28·4) Prevalence 6193 (5726 to 6777) 117·8% (113·2 to 122·8) 98 (90 to 107) 15·7% (13·3 to 18·3) Epilepsy DALYs 12 418 (10 438 to 14 479) 2·5% (−5·7 to 11·2) 168 (141 to 195) −22·5% (−28·2 to −16·8) Deaths 125 (119 to 131) 18·9% (6·4 to 32·1) 2 (2 to 2) −15·6% (−23·0 to −8·0) Prevalence 23 415 (21 550 to 25 419) 39·2% (33·4 to 45·2) 320 (295 to 347) 1·9% (−2·1 to 6·1) Multiple sclerosis DALYs 1234 (1033 to 1437) 42·4% (31·8 to 57·3) 17 (14 to 20) −16·0% (−22·1 to −7·6) Deat

al African Republic, Guinea-Bissau, Kiribati, and Somalia. The highest death rates (more than 280 per 100 000 people) were estimated for Afghanistan and the Central African Republic (figure 3).Figure 3 Age-standardised rates of (A) DALYs and (B) deaths per 100 000 people from all neurological disorders combined in 2015 Data are for both sexes. DALYs=disability-adjusted life-years. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. TLS=Timor-Leste. Marshall Isl=Marshall Islands. Sol Isl=Solomon Islands. FSM=Federated States of Micronesia. Stroke was the leading cause of age-standardised DALY rates in 18 of 21 GBD regions, while migraine and Alzheimer's disease and other dementias were ranked among the top three causes in all regions except Oceania, south Asia, and the four sub-Saharan African regions, where epilepsy or meningitis ranked higher. Communicable neurological conditions ranked low in high-income regions and central Europe. Epilepsy and medication overuse headache ranked fourth, fifth, or sixth in almost all regions (figure 4).Figure 4 Ranking of age-standardised DALY rates for all neurological disorders by GBD region in 2015 Data are for both sexes. DALYs=disability-adjusted life-years.

168 (141 to 195) −22·5% (−28·2 to −16·8) Deaths 125 (119 to 131) 18·9% (6·4 to 32·1) 2 (2 to 2) −15·6% (−23·0 to −8·0) Prevalence 23 415 (21 550 to 25 419) 39·2% (33·4 to 45·2) 320 (295 to 347) 1·9% (−2·1 to 6·1) Multiple sclerosis DALYs 1234 (1033 to 1437) 42·4% (31·8 to 57·3) 17 (14 to 20) −16·0% (−22·1 to −7·6) Deat hs 19 (17 to 20) 39·9% (24·9 to 61·6) 0 (0 to 0) −21·4% (−29·1 to −10·8) Prevalence 2012 (1866 to 2167) 59·0% (55·5 to 62·6) 28 (26 to 30) −4·6% (−6·7 to −2·5) Migraine DALYs 32 899 (20 295 to 48 945) 49·5% (46·8 to 52·3) 439 (271 to 654) 0·5% (−0·5 to 1·5) Deaths ·· ·· ·· ·· Prevalence 958 789 (872 109 to 1 055 631) 49·6% (46·9 to 52·3) 12 799 (11 651 to 14 065) 0·1% (−0·9 to 1·1) Tension-type headache DALYs 2261 (1055 to 4193) 49·2% (46·0 to 52·5) 30 (14 to 56) 0·2% (−1·0 to 1·4) Deaths ·· ·· ·· ·· Prevalence 1 505 892 (1 337 310 to 1 681 575) 49·2% (46·1 to 52·5) 20 121 (17 924 to 22 442) 0·0 (−1·1 to 1·2) Medication overuse headache DALYs 9165 (6089 to 13081) 57·7% (52·5 to 63·3) 124 (82 to 177) −0·9% (−3·9 to 2·2) Deaths ·· ·· ·· ·· Prevalence 58 455 (50 835 to 67 364) 57·8% (52·6 to 63·3) 790 (690 to 907) −1·2% (−4·2 to 1·8) Motor neuron disease DALYs 910 (872 to 959) 56·1% (34·6 to 69·6) 13 (13 to 14) −0·9% (−12·4 to 4·0) Deaths 35 (34 to 37) 97·3% (73·9 to 103·6) 1 (1 to 1) 12·7% (0·5 to 15·9) Prevalence 202 (190 to 216) 72·4% (69·2 to 75·2) 3 (3 to 3) 3·1% (1·5 to 4·7) Brain and nervous system cancer DALYs 7624 (6975 to 8219) 37·5% (16·9 to 54·1) 106 (97 to 114) −9·0% (−21·1 to 0·8) Deaths 229 (210 to 245) 65·8% (42·2 to 78·2) 3 (3 to 4) −0·4% (−13·5 to 6·4) Prevalence 1205 (1102 to 1323) 60·2% (31·7 to 87·2) 17 (16 to 19) 8·9% (−7·2 to 24·0) Other neurological disorders* DALYs 2360 (2217 to 2565) 31·7% (19·1 to 40·5) 34 (31 to 36) −9·4% (−15·8 to −5·7) Deaths 54 (53 to 58) 53·1% (42·9 to 58·4) 1 (1 to 1) −5·2% (−10·5 to −1·9) All neurological disorders DALYs 250 692 (229 080 to 274 654) 7·4%† 3639 (3342 to 3964) −29·7%† Deaths 9399 (9095 to 9714) 36·7%† 150 (145 to 155) −26·1%† Data are n or % with 95% uncertainty intervals (UIs).

34 (31 to 36) −9·4% (−15·8 to −5·7) Deaths 54 (53 to 58) 53·1% (42·9 to 58·4) 1 (1 to 1) −5·2% (−10·5 to −1·9) All neurological disorders DALYs 250 692 (229 080 to 274 654) 7·4%† 3639 (3342 to 3964) −29·7%† Deaths 9399 (9095 to 9714) 36·7%† 150 (145 to 155) −26·1%† Data are n or % with 95% uncertainty intervals (UIs). Prevalence is an aggregate of all sequelae for a condition. DALYs=disability-adjusted life-years. * Includes muscular dystrophy, Huntington's disease, and other less common neurological disorders (appendix p 112). † UIs were not calculated for aggregated estimates of percentage change as this was not built into the Global Burden of Disease Study 2015 computational machinery and would have required a large computational effort competing with demands of completing the Global Burden of Disease Study 2016.

* Includes muscular dystrophy, Huntington's disease, and other less common neurological disorders (appendix p 112). † UIs were not calculated for aggregated estimates of percentage change as this was not built into the Global Burden of Disease Study 2015 computational machinery and would have required a large computational effort competing with demands of completing the Global Burden of Disease Study 2016. Global burden by sex and age Substantial sex differences in age-standardised death, DALY, and prevalence rates existed globally in 2015. The only neurological disorders with less than 10% difference between males and females were Alzheimer's disease and other dementias in terms of death and DALY rates, and meningitis and epilepsy in terms of prevalence rate. Apart from multiple sclerosis, Alzheimer's disease and other dementias, and the three headache types, DALY and prevalence rates were higher in males than females. All these more substantial sex differences were significant apart from those for meningitis, tetanus, and encephalitis death rates and the meningitis and encephalitis DALY rates (appendix pp 153–54). The age pattern varied between the various neurological disorders. The bulk of burden due to communicable neurological disorders occurred in young individuals, particularly the 0–5 years age group. Epilepsy caused the most burden in children and young adults. Headaches peaked at ages 25–49 years, whereas the burden of other neurological disorders increased with age (figure 1).Figure 1 Global DALYs by age and neurological disorder in 2015

sorders occurred in young individuals, particularly the 0–5 years age group. Epilepsy caused the most burden in children and young adults. Headaches peaked at ages 25–49 years, whereas the burden of other neurological disorders increased with age (figure 1).Figure 1 Global DALYs by age and neurological disorder in 2015 DALYs=disability-adjusted life-years. Proportional contribution to the combined burden Stroke accounted for the largest proportion of total DALYs (47·3%) and deaths (67·3%) among all neurological disorders analysed (figure 2). Migraine, meningitis, and Alzheimer's disease and other dementias were the second, third, and fourth largest contributors of DALYs. The proportional contributions of the other neurological disorders analysed were less substantial and varied from 0·4% (for motor neuron disease) to 5·0% (epilepsy). The second largest contributor to deaths from neurological disorders was Alzheimer's disease and other dementias. The proportional contribution of deaths from other neurological disorders varied from 0·2% for (multiple sclerosis) to 4·0% (for meningitis; figure 2).Figure 2 Contribution of various neurological disorders to the overall burden from neurological disorders in 2015 Estimates are for (A) disability-adjusted life-years and (B) deaths.

Proportional contribution to the combined burden Stroke accounted for the largest proportion of total DALYs (47·3%) and deaths (67·3%) among all neurological disorders analysed (figure 2). Migraine, meningitis, and Alzheimer's disease and other dementias were the second, third, and fourth largest contributors of DALYs. The proportional contributions of the other neurological disorders analysed were less substantial and varied from 0·4% (for motor neuron disease) to 5·0% (epilepsy). The second largest contributor to deaths from neurological disorders was Alzheimer's disease and other dementias. The proportional contribution of deaths from other neurological disorders varied from 0·2% for (multiple sclerosis) to 4·0% (for meningitis; figure 2).Figure 2 Contribution of various neurological disorders to the overall burden from neurological disorders in 2015 Estimates are for (A) disability-adjusted life-years and (B) deaths. Geographical variations in the burden In 2015, the lowest age-standardised DALY rates (less than 3000 per 100 000 people) and death rates (less than 100 per 100 000 people) from neurological disorders were estimated for high-income regions and Latin America, whereas the highest DALY rates (more than 7000 per 100 000 people) were estimated for Afghanistan, Central African Republic, Guinea-Bissau, Kiribati, and Somalia. The highest death rates (more than 280 per 100 000 people) were estimated for Afghanistan and the Central African Republic (figure 3).Figure 3 Age-standardised rates of (A) DALYs and (B) deaths per 100 000 people from all neurological disorders combined in 2015

sease burden in sub-Saharan Africa, especially in the meningitis belt across the Sahara. The largest burden from tetanus was estimated for a few countries in east Africa, particularly South Sudan and Somalia, which have low vaccination coverage owing to national conflict. India had the largest burden from encephalitis. Changes in the burden from 1990 to 2015 Time trends from 1990 to 2015 in prevalence, deaths, and DALYs due to neurological disorders varied across the disorders studied (table 1). Despite a significant drop in age-standardised stroke rates, population increase and ageing combined led to increases in absolute numbers of the three indicators. There was a large increase in the absolute numbers of DALYs, deaths, and prevalent cases of Alzheimer's disease and other dementias even though the changes in age-standardised rates were small. For Parkinson's disease, the numbers of DALYs, deaths, and prevalent cases increased substantially while increases in age-standardised rates were more modest. There was a significant decrease in epilepsy death and DALY rates while the age-standardised prevalence rate remained stable. DALYs, deaths, and prevalent cases of multiple sclerosis increased despite decreasing age-standardised rates. The absolute numbers of DALYs and prevalent cases of motor neuron disease, brain cancer, and headaches increased substantially while there was little change in age-standardised rates. The largest decreases were observed in communicable neurological disorders (tetanus, meningitis, and encephalitis). Although the rates of DALYs and deaths for all neurological disorders combined decreased by more than a quarter from 1990 to 2015, the absolute number of DALYs and deaths from all neurological disorders combined over that period increased by 7·4% (from 233·4 million to 250·7 million) and 36·7% (from 6·9 million to 9·4 million), respectively.

ALYs and deaths for all neurological disorders combined decreased by more than a quarter from 1990 to 2015, the absolute number of DALYs and deaths from all neurological disorders combined over that period increased by 7·4% (from 233·4 million to 250·7 million) and 36·7% (from 6·9 million to 9·4 million), respectively. The decrease in DALY rates from 1990 to 2015 was estimated for most countries. The only countries with an increase in age-standardised DALY rates were the Philippines, Mongolia, Lesotho, Swaziland, and Zimbabwe, largely because these countries had an increase in DALYs from stroke over the 25-year period. The age-standardised DALY rates from neurological disorders dropped by more than half in Ethiopia, Equatorial Guinea, and South Korea (table 2).Table 2 Global absolute number of DALYs and age-standardised rates per 100 000 people for all neurological disorders combined and percentage changes between 1990 and 2015 by country

ge-standardised DALY rates from neurological disorders dropped by more than half in Ethiopia, Equatorial Guinea, and South Korea (table 2).Table 2 Global absolute number of DALYs and age-standardised rates per 100 000 people for all neurological disorders combined and percentage changes between 1990 and 2015 by country All-age DALYs (thousands) Age-standardised rate of DALYs (per 100 000) 1990 2015 Change from 1990 to 2015* 1990 2015 Change from 1990 to 2015* Afghanistan 1535 (1057 to 2053) 2071 (1667 to 2553) 34·9% 12 255 (9773 to 14 594) 10 669 (8798 to 12 830) −12·9% Albania 108 (97 to 120) 121 (108 to 135) 12·2% 4395 (4033 to 4804) 3691 (3287 to 4101) −16·0% Algeria 822 (707 to 961) 1107 (950 to 1267) 34·7% 4608 (4104 to 5162) 3462 (3021 to 3915) −24·9% Andorra 1 (1 to 2) 2 (2 to 3) 62·8% 2146 (1791 to 2536) 1766 (1435 to 2132) −17·7% Angola 871 (523 to 1284) 1067 (647 to 1956) 22·4% 8642 (4399 to 14 190) 6273 (3302 to 13 343) −27·4% Antigua and Barbuda 2 (2 to 2) 2 (2 to 3) 8·8% 4170 (3783 to 4612) 2866 (2516 to 3259) −31·3% Argentina 997 (908 to 1099) 977 (853 to 1118) −2·0% 3310 (3029 to 3644) 2152 (1873 to 2470) −35·0% Armenia 108 (96 to 120) 108 (97 to 121) 0 3878 (3519 to 4253) 3101 (2782 to 3487) −20·0% Australia 448 (392 to 513) 572 (490 to 667) 27·6% 2529 (2209 to 2897) 1877 (1571 to 2237) −25·8% Austria 300 (270 to 336) 263 (227 to 303) −12·5% 3022 (2661 to 3434) 2046 (1701 to 2431) −32·3% Azerbaijan 240 (219 to 266) 309 (274 to 350) 28·7% 4940 (4576 to 5348) 3619 (3242 to 4046) −26·7% Bahrain 10 (8 to 12) 21 (17 to 26) 110·8% 3231 (2800 to 3697) 2222 (1856 to 2627) −31·2% Bangladesh 6545 (5715 to 7383) 6416 (5551 to 7253) −2·0% 7004 (6347 to 7682) 5344 (4667 to 5998) −23·7% Barbados 10 (9 to 11) 10 (9 to 11) 3·1% 3768 (3404 to 4168) 2760 (2411 to 3123) −26·7% Belarus 486 (444 to 529) 519 (474 to 572) 6·8% 4414 (4021 to 4831) 3906 (3536 to 4328) −11·5% Belgium 377 (341 to 420) 374 (328 to 426) −0·6% 2954 (2624 to 3338) 2181 (1847 to 2562) −26·2% Belize 5 (4 to 5) 7 (6 to 9) 52·4% 3647 (3275 to 4045) 3055 (2681 to 3453) −16·2% Benin 330 (227 to 449) 431 (307 to 606) 30·7% 5984 (4947 to 7160) 5443 (3617 to 8060) −9·0% Bhutan 34 (25 to 44) 22 (18 to 26) −34·8% 6145 (5110 to 7428) 3525 (2930 to 4167) −42·6% Bolivia 251 (217 to 299) 258 (219 to 302) 2·7% 4444 (3949 to 5024) 2941 (2514 to 3412) −33·8% Bosnia and Herzegovina 165 (148 to 183) 151 (135 to 169) −8·8% 4323 (3931 to 4737) 2927 (2592 to 3313) −32·3% Botswana 33 (19 to 62) 61 (31 to 157) 84·3% 4495 (2344 to 87

) −34·8% 6145 (5110 to 7428) 3525 (2930 to 4167) −42·6% Bolivia 251 (217 to 299) 258 (219 to 302) 2·7% 4444 (3949 to 5024) 2941 (2514 to 3412) −33·8% Bosnia and Herzegovina 165 (148 to 183) 151 (135 to 169) −8·8% 4323 (3931 to 4737) 2927 (2592 to 3313) −32·3% Botswana 33 (19 to 62) 61 (31 to 157) 84·3% 4495 (2344 to 87 76) 4312 (2204 to 10667) −4·1% Brazil 4947 (4480 to 5482) 5874 (5149 to 6669) 18·7% 4988 (4619 to 5392) 3100 (2753 to 3480) −37·9% Brunei 5 (4 to 6) 8 (7 to 9) 57·6% 3414 (3080 to 3798) 2477 (2190 to 2796) −27·4% Bulgaria 570 (536 to 607) 458 (424 to 492) −19·7% 5401 (5059 to 5782) 3870 (3531 to 4237) −28·4% Burkina Faso 647 (482 to 880) 750 (552 to 1010) 16·0% 5990 (4957 to 7341) 4633 (3380 to 6265) −22·6% Burundi 453 (337 to 573) 470 (341 to 644) 3·7% 10 962 (7640 to 14 124) 5627 (3938 to 7954) −48·7% Cambodia 536 (440 to 674) 468 (404 to 531) −12·7% 7223 (6409 to 8176) 4512 (3938 to 5061) −37·5% Cameroon 595 (458 to 785) 865 (622 to 1200) 45·4% 5384 (4559 to 6359) 5148 (3530 to 7428) −4·4% Canada 724 (628 to 834) 984 (850 to 1130) 35·8% 2473 (2148 to 2844) 1999 (1685 to 2350) −19·1% Cape Verde 12 (10 to 14) 14 (12 to 16) 13·3% 4867 (4433 to 5327) 3577 (3086 to 4219) −26·5% Central African Republic 221 (173 to 269) 305 (197 to 439) 38·1% 9138 (7601 to 10697) 8838 (5307 to 13 092) −3·3% Chad 474 (349 to 634) 787 (555 to 1108) 66·2% 6416 (5233 to 7779) 6128 (4089 to 8876) −4·5% Chile 359 (318 to 406) 438 (382 to 501) 22·1% 3420 (3088 to 3784) 2223 (1926 to 2551) −35·0% China 45 845 (42 895 to 48 907) 49 486 (46 074 to 53 228) 7·9% 5529 (5208 to 5880) 3410 (3181 to 3658) −38·3% Colombia 826 (729 to 933) 949 (806 to 1112) 15·0% 3361 (3033 to 3710) 2244 (1928 to 2580) −33·2% Comoros 24 (18 to 31) 23 (17 to 30) −5·0% 7384 (5104 to 10 202) 4378 (3200 to 5834) −40·7% Congo (Brazzaville) 117 (92 to 146) 166 (117 to 225) 41·4% 7457 (6021 to 9064) 5566 (3793 to 7795) −25·4% Costa Rica 59 (50 to 68) 93 (78 to 109) 58·6% 2594 (2272 to 2954) 1996 (1684 to 2335) −23·1% Cote d'Ivoire 594 (475 to 735) 848 (614 to 1130) 42·9% 5778 (4923 to 6704) 5377 (3750 to 7587) −6·9% Croatia 234 (218 to 253) 192 (175 to 210) −18·0% 4309 (3986 to 4667) 2871 (2554 to 3217) −33·4% Cuba 324 (291 to 362) 379 (337 to 425) 16·7% 3295 (2976 to 3645) 2583 (2276 to 2944) −21·6% Cyprus 20 (18 to 22) 21 (18 to 25) 8·1% 2962 (2627 to 3321) 1922 (1593 to 2284) −35·1% Czech Republic 519 (483 to 561) 355 (315 to 400) −31·7% 4376 (4039 to 4748) 2311 (2007 to 2657) −47·2% Democratic Republic of the Congo 173

% Cuba 324 (291 to 362) 379 (337 to 425) 16·7% 3295 (2976 to 3645) 2583 (2276 to 2944) −21·6% Cyprus 20 (18 to 22) 21 (18 to 25) 8·1% 2962 (2627 to 3321) 1922 (1593 to 2284) −35·1% Czech Republic 519 (483 to 561) 355 (315 to 400) −31·7% 4376 (4039 to 4748) 2311 (2007 to 2657) −47·2% Democratic Republic of the Congo 173 0 (1316 to 2183) 2976 (2197 to 3938) 72·0% 5934 (4390 to 7735) 5521 (3927 to 7317) −7·0% Denmark 204 (184 to 226) 184 (162 to 207) −9·9% 3032 (2707 to 3425) 2286 (1967 to 2644) −24·6% Djibouti 27 (20 to 35) 34 (22 to 52) 26·7% 5845 (4301 to 7885) 5047 (3271 to 7902) −13·7% Dominica 2 (2 to 2) 2 (2 to 2) 7·1% 3479 (3103 to 3902) 3022 (2645 to 3436) −13·1% Dominican Republic 198 (174 to 222) 265 (231 to 301) 33·6% 3657 (3289 to 4039) 3038 (2691 to 3408) −16·9% Ecuador 246 (216 to 275) 326 (279 to 380) 32·8% 3266 (2917 to 3623) 2382 (2063 to 2734) −27·1% Egypt 2393 (2172 to 2607) 2696 (2395 to 3068) 12·6% 5267 (4828 to 5746) 3963 (3581 to 4494) −24·8% El Salvador 139 (124 to 156) 118 (99 to 139) −15·0% 3516 (3166 to 3892) 2105 (1788 to 2461) −40·1% Equatorial Guinea 30 (17 to 47) 29 (18 to 55) −3·5% 9305 (4456 to 15829) 4487 (2798 to 9079) −51·8% Eritrea 219 (161 to 281) 213 (146 to 314) −3·0% 9041 (7575 to 10 688) 6369 (3831 to 9981) −29·6% Estonia 86 (79 to 92) 50 (44 to 56) −42·1% 4813 (4451 to 5213) 2464 (2130 to 2836) −48·8% Ethiopia 4344 (2923 to 5858) 3387 (2546 to 4404) −22·0% 9284 (7605 to 10967) 4637 (3343 to 6232) −50·1% Federated States of Micronesia 3 (3 to 4) 3 (2 to 4) −15·6% 5884 (4146 to 7785) 4216 (3098 to 6093) −28·4% Fiji 24 (21 to 27) 28 (24 to 32) 15·7% 5429 (4749 to 6247) 3882 (3380 to 4409) −28·5% Finland 197 (178 to 219) 198 (175 to 224) 0·2% 3287 (2936 to 3678) 2309 (1980 to 2701) −29·8% France 1743 (1546 to 1966) 1901 (1658 to 2170) 9·1% 2493 (2169 to 2845) 1948 (1642 to 2305) −21·8% Gabon 44 (37 to 53) 53 (41 to 73) 20·4% 5763 (4886 to 6694) 4289 (3185 to 5984) −25·6% Georgia 289 (265 to 312) 222 (203 to 247) −23·0% 5268 (4847 to 5707) 4236 (3825 to 4740) −19·6% Germany 3115 (2799 to 3455) 2748 (2400 to 3123) −11·8% 2973 (2627 to 3343) 2041 (1711 to 2401) −31·4% Ghana 687 (549 to 853) 941 (644 to 1342) 36·9% 6309 (4812 to 8167) 5071 (3277 to 7649) −19·6% Greece 430 (394 to 470) 447 (403 to 496) 3·9% 3366 (3046 to 3735) 2365 (2058 to 2734) −29·7% Greenland 2 (2 to 2) 1 (1 to 2) −20·7% 4929 (4464 to 5370) 3214 (2843 to 3616) −34·8% Grenada 4 (3 to 4) 3 (3 to 4) −10·4% 4941 (4534 to 5365) 3714 (3350 to 4139) −24·8% Guatemala 191 (166 to 218

167) 5071 (3277 to 7649) −19·6% Greece 430 (394 to 470) 447 (403 to 496) 3·9% 3366 (3046 to 3735) 2365 (2058 to 2734) −29·7% Greenland 2 (2 to 2) 1 (1 to 2) −20·7% 4929 (4464 to 5370) 3214 (2843 to 3616) −34·8% Grenada 4 (3 to 4) 3 (3 to 4) −10·4% 4941 (4534 to 5365) 3714 (3350 to 4139) −24·8% Guatemala 191 (166 to 218 ) 304 (257 to 355) 59·6% 2874 (2531 to 3251) 2495 (2136 to 2865) −13·2% Guinea 558 (399 to 753) 550 (413 to 722) −1·3% 7288 (5950 to 8810) 5719 (4215 to 7617) −21·5% Guinea–Bissau 87 (57 to 130) 121 (71 to 218) 40·0% 7587 (4479 to 13 190) 7463 (3972 to 15 833) −1·6% Guyana 32 (29 to 35) 27 (24 to 31) −14·2% 7121 (6584 to 7631) 4805 (4309 to 5343) −32·5% Haiti 515 (441 to 624) 437 (358 to 525) −15·0% 8584 (7695 to 9626) 5696 (4751 to 6682) −33·6% Honduras 224 (198 to 247) 210 (179 to 243) −6·4% 4773 (4368 to 5208) 3530 (3022 to 4104) −26·0% Hungary 585 (545 to 626) 382 (343 to 424) −34·7% 4697 (4357 to 5065) 2567 (2250 to 2906) −45·3% Iceland 7 (6 to 8) 8 (7 to 9) 16·4% 2607 (2286 to 2973) 1976 (1653 to 2345) −24·2% India 49 938 (45 989 to 54 561) 45 738 (40 949 to 51 065) −8·4% 6590 (6135 to 7129) 4274 (3881 to 4706) −35·2% Indonesia 7770 (6393 to 9192) 9914 (8424 to 11268) 27·6% 5930 (5238 to 6639) 5115 (4375 to 5773) −13·7% Iran 1618 (1379 to 1855) 2033 (1664 to 2395) 25·7% 4596 (4042 to 5149) 3411 (2871 to 3950) −25·8% Iraq 640 (544 to 757) 1045 (875 to 1257) 63·2% 5733 (4939 to 6610) 4989 (4150 to 5944) −13·0% Ireland 101 (89 to 116) 111 (94 to 130) 9·3% 2811 (2446 to 3212) 2039 (1703 to 2412) −27·4% Israel 109 (96 to 125) 169 (143 to 199) 55·2% 2693 (2370 to 3069) 1968 (1642 to 2340) −26·9% Italy 2182 (1936 to 2447) 2378 (2058 to 2712) 9·0% 2991 (2611 to 3410) 2211 (1832 to 2637) −26·1% Jamaica 81 (73 to 89) 89 (79 to 101) 10·1% 4074 (3706 to 4489) 3249 (2873 to 3671) −20·2% Japan 3650 (3267 to 4087) 4489 (3999 to 5046) 23·0% 2587 (2302 to 2914) 1838 (1571 to 2163) −29·0% Jordan 76 (64 to 88) 129 (106 to 155) 69·9% 4125 (3579 to 4701) 2683 (2278 to 3104) −35·0% Kazakhstan 657 (605 to 720) 691 (623 to 765) 5·1% 5161 (4805 to 5562) 4518 (4102 to 4954) −12·4% Kenya 1143 (953 to 1509) 1427 (1222 to 1699) 24·9% 5323 (4608 to 6314) 4160 (3648 to 4684) −21·9% Kiribati 5 (4 to 6) 6 (5 to 7) 19·1% 8788 (7712 to 9853) 7420 (6390 to 8508) −15·6% Kuwait 30 (24 to 37) 59 (45 to 74) 94·5% 2447 (2078 to 2867) 2372 (1982 to 2802) −3·1% Kyrgyzstan 189 (174 to 205) 211 (190 to 234) 11·4% 6102 (5662 to 6542) 5062 (4661 to 5540) −17·1% Laos 327 (255 to 425) 221 (186 to 263) −3

4) −21·9% Kiribati 5 (4 to 6) 6 (5 to 7) 19·1% 8788 (7712 to 9853) 7420 (6390 to 8508) −15·6% Kuwait 30 (24 to 37) 59 (45 to 74) 94·5% 2447 (2078 to 2867) 2372 (1982 to 2802) −3·1% Kyrgyzstan 189 (174 to 205) 211 (190 to 234) 11·4% 6102 (5662 to 6542) 5062 (4661 to 5540) −17·1% Laos 327 (255 to 425) 221 (186 to 263) −3 2·4% 8564 (7132 to 10059) 4735 (4058 to 5497) −44·7% Latvia 160 (150 to 171) 112 (103 to 122) −29·9% 5136 (4784 to 5534) 3308 (2963 to 3682) −35·6% Lebanon 79 (67 to 91) 129 (104 to 155) 64·0% 3729 (3182 to 4297) 2421 (1969 to 2877) −35·1% Lesotho 45 (34 to 56) 79 (52 to 115) 74·2% 4296 (3155 to 5366) 6037 (3841 to 8956) 40·5% Liberia 160 (115 to 224) 154 (115 to 212) −4·0% 6601 (5212 to 8374) 4590 (3327 to 6043) −30·5% Libya 115 (99 to 131) 151 (128 to 176) 32·1% 3690 (3248 to 4146) 3324 (2880 to 3808) −9·9% Lithuania 142 (129 to 156) 132 (119 to 145) −6·8% 3550 (3227 to 3926) 2843 (2522 to 3193) −19·9% Luxembourg 16 (15 to 18) 15 (13 to 18) −4·8% 3480 (3113 to 3879) 2137 (1789 to 2551) −38·6% Macedonia 91 (84 to 99) 102 (93 to 112) 12·1% 5554 (5179 to 5963) 4160 (3765 to 4547) −25·1% Madagascar 648 (525 to 783) 925 (662 to 1286) 42·7% 7517 (6553 to 8475) 6186 (4189 to 8857) −17·7% Malawi 661 (472 to 950) 604 (452 to 810) −8·6% 6512 (5245 to 8155) 4212 (3071 to 5691) −35·3% Malaysia 437 (389 to 484) 689 (592 to 789) 57·8% 4020 (3696 to 4359) 2925 (2555 to 3305) −27·3% Maldives 5 (4 to 6) 5 (4 to 6) −0·2% 3712 (3308 to 4173) 2011 (1700 to 2359) −45·8% Mali 666 (498 to 885) 768 (570 to 1047) 15·4% 7318 (6215 to 8660) 5185 (3867 to 7169) −29·2% Malta 10 (9 to 11) 12 (10 to 13) 17·4% 2781 (2445 to 3159) 2007 (1686 to 2373) −27·8% Marshall Islands 2 (1 to 2) 2 (2 to 2) 28·2% 5588 (4953 to 6257) 4358 (3764 to 5009) −22·0% Mauritania 84 (63 to 112) 108 (81 to 140) 28·3% 5274 (4369 to 6368) 3664 (2624 to 4925) −30·5% Mauritius 34 (31 to 37) 36 (32 to 40) 5·8% 4922 (4593 to 5253) 2738 (2461 to 3056) −44·4% Mexico 1731 (1495 to 2002) 2414 (2048 to 2838) 39·5% 2772 (2441 to 3145) 2232 (1930 to 2577) −19·5% Moldova 201 (185 to 218) 165 (150 to 182) −18·0% 5098 (4714 to 5483) 3592 (3263 to 3976) −29·5% Mongolia 96 (75 to 120) 136 (123 to 151) 40·9% 5738 (5020 to 6651) 6468 (5944 to 7197) 12·7% Montenegro 28 (26 to 30) 32 (29 to 34) 12·6% 4875 (4486 to 5281) 3933 (3574 to 4319) −19·3% Morocco 962 (790 to 1143) 970 (796 to 1177) 0·8% 4956 (4362 to 5627) 3407 (2815 to 4100) −31·3% Mozambique 875 (690 to 1116) 1062 (746 to 1475) 21·4% 6921 (5890 to 8264) 5756 (3711 to 8423) −16

to 6651) 6468 (5944 to 7197) 12·7% Montenegro 28 (26 to 30) 32 (29 to 34) 12·6% 4875 (4486 to 5281) 3933 (3574 to 4319) −19·3% Morocco 962 (790 to 1143) 970 (796 to 1177) 0·8% 4956 (4362 to 5627) 3407 (2815 to 4100) −31·3% Mozambique 875 (690 to 1116) 1062 (746 to 1475) 21·4% 6921 (5890 to 8264) 5756 (3711 to 8423) −16 ·8% Myanmar 1912 (1439 to 2452) 2010 (1457 to 2662) 5·1% 6857 (4981 to 9046) 4991 (3657 to 6496) −27·2% Namibia 38 (31 to 44) 52 (38 to 72) 37·2% 4578 (3800 to 5351) 3426 (2455 to 4914) −25·2% Nepal 1702 (1339 to 2128) 904 (744 to 1066) −46·9% 7684 (6469 to 8998) 3994 (3309 to 4713) −48·0% Netherlands 476 (423 to 536) 508 (444 to 577) 6·9% 2742 (2418 to 3106) 2131 (1821 to 2484) −22·3% New Zealand 92 (81 to 103) 108 (94 to 124) 17·4% 2625 (2331 to 2959) 1897 (1620 to 2215) −27·8% Nicaragua 94 (82 to 106) 109 (92 to 129) 16·7% 3076 (2723 to 3465) 2312 (1980 to 2655) −24·8% Niger 1259 (801 to 2018) 1267 (883 to 1848) 0·7% 9958 (7256 to 14427) 5869 (4344 to 7786) −41·1% Nigeria 5744 (4238 to 7818) 5694 (4390 to 7992) −0·9% 5747 (4458 to 7261) 3618 (2864 to 4930) −37·1% North Korea 703 (558 to 876) 1178 (973 to 1412) 67·7% 4955 (3984 to 6118) 4679 (3888 to 5567) −5·6% Norway 161 (146 to 178) 141 (124 to 160) −12·3% 2864 (2565 to 3209) 2017 (1723 to 2343) −29·6% Oman 35 (29 to 43) 71 (57 to 88) 99·7% 3029 (2471 to 3652) 2516 (2140 to 2905) −16·9% Pakistan 6419 (5315 to 7793) 7283 (6255 to 8567) 13·5% 6131 (5327 to 7083) 5005 (4380 to 5716) −18·4% Palestine 51 (43 to 61) 96 (80 to 115) 88·6% 4369 (3718 to 5160) 3722 (3109 to 4323) −14·8% Panama 57 (50 to 65) 86 (73 to 100) 51·0% 3230 (2889 to 3597) 2423 (2080 to 2798) −25·0% Papua New Guinea 205 (154 to 269) 308 (223 to 438) 49·9% 7826 (5720 to 10186) 6275 (4550 to 8789) −19·8% Paraguay 109 (97 to 123) 167 (145 to 194) 52·7% 3882 (3513 to 4288) 3298 (2898 to 3786) −15·1% Peru 610 (533 to 689) 572 (471 to 686) −6·2% 3511 (3123 to 3921) 2066 (1721 to 2442) −41·2% Philippines 1617 (1439 to 1781) 2883 (2621 to 3173) 78·3% 3885 (3591 to 4190) 3968 (3642 to 4326) 2·1% Poland 1442 (1315 to 1580) 1351 (1206 to 1503) −6·3% 3743 (3420 to 4100) 2541 (2227 to 2876) −32·1% Portugal 541 (504 to 581) 401 (357 to 450) −25·8% 4501 (4160 to 4877) 2370 (2052 to 2748) −47·3% Qatar 9 (7 to 10) 32 (24 to 41) 270·0% 2970 (2543 to 3480) 2220 (1826 to 2683) −25·3% Romania 1224 (1139 to 1313) 1124 (1044 to 1208) −8·2% 5039 (4687 to 5412) 3817 (3491 to 4175) −24·3% Russia 8313 (7738 to 8910) 8507 (7857 to 9176) 2·3% 5380 (5009 to 5772) 4220 (38

25·8% 4501 (4160 to 4877) 2370 (2052 to 2748) −47·3% Qatar 9 (7 to 10) 32 (24 to 41) 270·0% 2970 (2543 to 3480) 2220 (1826 to 2683) −25·3% Romania 1224 (1139 to 1313) 1124 (1044 to 1208) −8·2% 5039 (4687 to 5412) 3817 (3491 to 4175) −24·3% Russia 8313 (7738 to 8910) 8507 (7857 to 9176) 2·3% 5380 (5009 to 5772) 4220 (38 51 to 4600) −21·6% Rwanda 460 (360 to 585) 378 (288 to 494) −17·9% 8493 (7054 to 10073) 4485 (3236 to 6286) −47·2% Saint Lucia 5 (5 to 6) 6 (5 to 7) 10·3% 4773 (4383 to 5192) 3237 (2879 to 3632) −32·2% St Vincent and the Grenadines 3 (3 to 4) 3 (3 to 4) 0·1% 4181 (3761 to 4616) 3483 (3122 to 3851) −16·7% Samoa 4 (4 to 5) 4 (3 to 5) −4·0% 4534 (3770 to 5412) 3109 (2547 to 3657) −31·4% São Tomé and Príncipe 5 (4 to 5) 5 (4 to 8) 19·0% 4788 (4119 to 5510) 4610 (3123 to 6711) −3·7% Saudi Arabia 317 (263 to 379) 555 (444 to 686) 75·1% 3332 (2859 to 3808) 2705 (2302 to 3158) −18·8% Senegal 397 (326 to 494) 541 (383 to 748) 36·0% 5637 (4846 to 6563) 5226 (3419 to 7715) −7·3% Serbia 488 (449 to 529) 476 (440 to 515) −2·5% 5146 (4755 to 5572) 3739 (3408 to 4110) −27·3% Seychelles 2 (2 to 2) 2 (2 to 3) 9·6% 4015 (3678 to 4382) 2636 (2311 to 3005) −34·4% Sierra Leone 304 (220 to 453) 308 (228 to 413) 1·3% 6568 (4982 to 8508) 6096 (4473 to 8013) −7·2% Singapore 61 (54 to 69) 76 (65 to 88) 24·8% 2864 (2613 to 3145) 1612 (1367 to 1896) −43·7% Slovakia 200 (182 to 220) 176 (156 to 199) −12·0% 3830 (3487 to 4200) 2535 (2225 to 2903) −33·8% Slovenia 76 (69 to 84) 64 (57 to 73) −15·9% 3553 (3228 to 3903) 2057 (1758 to 2390) −42·1% Solomon Islands 13 (10 to 17) 21 (14 to 29) 57·3% 7670 (5517 to 10124) 6077 (4252 to 8631) −20·8% Somalia 603 (401 to 896) 682 (428 to 1121) 13·1% 9690 (5059 to 16652) 7720 (3831 to 14872) −20·3% South Africa 1110 (1004 to 1226) 1458 (1291 to 1636) 31·4% 4340 (3955 to 4753) 3549 (3176 to 3953) −18·2% South Korea 1501 (1382 to 1640) 1371 (1216 to 1555) −8·7% 5331 (5028 to 5669) 2185 (1926 to 2496) −59·0% South Sudan 508 (321 to 771) 691 (423 to 1222) 36·0% 8134 (4561 to 14144) 6548 (3545 to 13257) −19·5% Spain 1387 (1249 to 1548) 1477 (1276 to 1697) 6·5% 2998 (2671 to 3385) 2006 (1685 to 2378) −33·1% Sri Lanka 389 (344 to 439) 469 (389 to 566) 20·5% 2748 (2476 to 3062) 2384 (1994 to 2865) −13·3% Sudan 984 (790 to 1194) 1347 (1095 to 1647) 36·9% 6558 (5444 to 7867) 5257 (4240 to 6467) −19·8% Suriname 14 (13 to 15) 18 (16 to 20) 30·6% 4520 (4132 to 4920) 3913 (3462 to 4393) −13·4% Swaziland 23 (18 to 29) 40 (25 to 61) 75·5% 4819 (3679 to 6099) 5399 (3167 t

66) 20·5% 2748 (2476 to 3062) 2384 (1994 to 2865) −13·3% Sudan 984 (790 to 1194) 1347 (1095 to 1647) 36·9% 6558 (5444 to 7867) 5257 (4240 to 6467) −19·8% Suriname 14 (13 to 15) 18 (16 to 20) 30·6% 4520 (4132 to 4920) 3913 (3462 to 4393) −13·4% Swaziland 23 (18 to 29) 40 (25 to 61) 75·5% 4819 (3679 to 6099) 5399 (3167 t o 8410) 12·0% Sweden 300 (271 to 333) 286 (254 to 326) −4·4% 2458 (2176 to 2803) 1896 (1622 to 2227) −22·9% Switzerland 222 (197 to 249) 227 (196 to 262) 2·2% 2554 (2236 to 2913) 1875 (1572 to 2224) −26·6% Syria 472 (398 to 543) 454 (390 to 523) −3·8% 6029 (5346 to 6815) 3733 (3253 to 4249) −38·1% Taiwan (Province of China) 509 (463 to 559) 533 (461 to 613) 4·8% 3439 (3176 to 3733) 1831 (1590 to 2109) −46·8% Tajikistan 189 (171 to 208) 237 (207 to 269) 25·1% 5248 (4808 to 5675) 4394 (3942 to 4849) −16·3% Tanzania 1186 (970 to 1421) 1740 (1290 to 2368) 46·7% 5600 (4737 to 6441) 4364 (3016 to 6521) −22·1% Thailand 1482 (1332 to 1652) 2033 (1751 to 2339) 37·2% 3723 (3397 to 4099) 2679 (2313 to 3084) −28·0% The Bahamas 7 (6 to 8) 11 (9 to 12) 48·5% 4025 (3668 to 4457) 2858 (2501 to 3232) −29·0% The Gambia 40 (31 to 51) 56 (43 to 71) 38·5% 5025 (3647 to 6938) 4166 (3164 to 5411) −17·1% Timor–Leste 38 (24 to 52) 30 (24 to 38) −19·1% 6251 (4945 to 7614) 3643 (2976 to 4343) −41·7% Togo 171 (142 to 215) 235 (174 to 308) 37·3% 5474 (4660 to 6395) 4902 (3506 to 6685) −10·4% Tonga 3 (2 to 3) 3 (2 to 3) 0·3% 3983 (3479 to 4540) 3252 (2775 to 3750) −18·3% Trinidad and Tobago 40 (36 to 44) 42 (38 to 48) 5·6% 4459 (4103 to 4845) 3114 (2760 to 3496) −30·2% Tunisia 259 (227 to 294) 317 (265 to 373) 22·5% 4332 (3877 to 4801) 3088 (2577 to 3633) −28·7% Turkey 2225 (1854 to 2660) 1970 (1693 to 2271) −11·5% 4564 (3926 to 5239) 2778 (2408 to 3176) −39·1% Turkmenistan 128 (117 to 141) 188 (171 to 207) 46·8% 5312 (4926 to 5798) 4786 (4427 to 5184) −9·9% Uganda 1254 (962 to 1677) 1558 (1143 to 2088) 24·3% 7413 (5838 to 8909) 5386 (3665 to 7530) −27·3% Ukraine 2900 (2698 to 3116) 2476 (2267 to 2671) −14·6% 4853 (4494 to 5242) 3676 (3327 to 4045) −24·3% United Arab Emirates 46 (37 to 57) 212 (168 to 263) 359·0% 4712 (3949 to 5507) 3647 (3072 to 4316) −22·6% UK 2274 (2057 to 2522) 2018 (1777 to 2307) −11·3% 3065 (2722 to 3466) 2207 (1875 to 2600) −28·0% England 1865 (1686 to 2070) 1651 (1453 to 1886) −11·5% 3001 (2668 to 3390) 2149 (1825 to 2530) −28·4% Northern Ireland 55 (49 to 61) 55 (48 to 63) 0·8% 3025 (2675 to 3415) 2261 (1933 to 2631) −25·3% Scotland 235 (210 to 263) 203 (177 to 235)

522) 2018 (1777 to 2307) −11·3% 3065 (2722 to 3466) 2207 (1875 to 2600) −28·0% England 1865 (1686 to 2070) 1651 (1453 to 1886) −11·5% 3001 (2668 to 3390) 2149 (1825 to 2530) −28·4% Northern Ireland 55 (49 to 61) 55 (48 to 63) 0·8% 3025 (2675 to 3415) 2261 (1933 to 2631) −25·3% Scotland 235 (210 to 263) 203 (177 to 235) −13·3% 3679 (3236 to 4186) 2708 (2270 to 3232) −26·4% Wales 120 (109 to 132) 109 (96 to 123) −8·9% 3081 (2746 to 3452) 2297 (1977 to 2690) −25·4% USA 6812 (6031 to 7722) 8776 (7741 to 9939) 28·8% 2447 (2154 to 2790) 2109 (1831 to 2432) −13·8% Uruguay 116 (106 to 127) 110 (99 to 122) −4·8% 3503 (3196 to 3849) 2509 (2232 to 2840) −28·4% Uzbekistan 680 (609 to 750) 919 (817 to 1028) 35·1% 5064 (4627 to 5483) 4126 (3747 to 4541) −18·5% Vanuatu 6 (5 to 8) 10 (7 to 14) 64·7% 7084 (5199 to 9626) 5797 (4304 to 7946) −18·2% Venezuela 435 (382 to 498) 649 (551 to 755) 49·2% 3378 (3051 to 3753) 2556 (2208 to 2935) −24·3% Vietnam 2310 (1974 to 2691) 3179 (2588 to 3745) 37·6% 4775 (4164 to 5511) 3862 (3134 to 4523) −19·1% Yemen 615 (436 to 791) 851 (629 to 1175) 38·3% 6756 (4862 to 9201) 5734 (4079 to 8182) −15·1% Zambia 557 (425 to 738) 636 (493 to 823) 14·3% 6691 (5443 to 7948) 5755 (4199 to 7711) −14·0% Zimbabwe 282 (237 to 336) 451 (339 to 595) 60·0% 3629 (2993 to 4329) 4009 (2809 to 5567) 10·5% Data are n or % with 95% uncertainty intervals (UI). For more details on disorder, country, year, age, and sex, see the Global Burden of Disase compare website https://vizhub.healthdata.org/gbd-compare/. DALYs=disability-adjusted life-years.

−13·3% 3679 (3236 to 4186) 2708 (2270 to 3232) −26·4% Wales 120 (109 to 132) 109 (96 to 123) −8·9% 3081 (2746 to 3452) 2297 (1977 to 2690) −25·4% USA 6812 (6031 to 7722) 8776 (7741 to 9939) 28·8% 2447 (2154 to 2790) 2109 (1831 to 2432) −13·8% Uruguay 116 (106 to 127) 110 (99 to 122) −4·8% 3503 (3196 to 3849) 2509 (2232 to 2840) −28·4% Uzbekistan 680 (609 to 750) 919 (817 to 1028) 35·1% 5064 (4627 to 5483) 4126 (3747 to 4541) −18·5% Vanuatu 6 (5 to 8) 10 (7 to 14) 64·7% 7084 (5199 to 9626) 5797 (4304 to 7946) −18·2% Venezuela 435 (382 to 498) 649 (551 to 755) 49·2% 3378 (3051 to 3753) 2556 (2208 to 2935) −24·3% Vietnam 2310 (1974 to 2691) 3179 (2588 to 3745) 37·6% 4775 (4164 to 5511) 3862 (3134 to 4523) −19·1% Yemen 615 (436 to 791) 851 (629 to 1175) 38·3% 6756 (4862 to 9201) 5734 (4079 to 8182) −15·1% Zambia 557 (425 to 738) 636 (493 to 823) 14·3% 6691 (5443 to 7948) 5755 (4199 to 7711) −14·0% Zimbabwe 282 (237 to 336) 451 (339 to 595) 60·0% 3629 (2993 to 4329) 4009 (2809 to 5567) 10·5% Data are n or % with 95% uncertainty intervals (UI). For more details on disorder, country, year, age, and sex, see the Global Burden of Disase compare website https://vizhub.healthdata.org/gbd-compare/. DALYs=disability-adjusted life-years. * UIs cannot be calculated for aggregated estimates of percentage changes.

−13·3% 3679 (3236 to 4186) 2708 (2270 to 3232) −26·4% Wales 120 (109 to 132) 109 (96 to 123) −8·9% 3081 (2746 to 3452) 2297 (1977 to 2690) −25·4% USA 6812 (6031 to 7722) 8776 (7741 to 9939) 28·8% 2447 (2154 to 2790) 2109 (1831 to 2432) −13·8% Uruguay 116 (106 to 127) 110 (99 to 122) −4·8% 3503 (3196 to 3849) 2509 (2232 to 2840) −28·4% Uzbekistan 680 (609 to 750) 919 (817 to 1028) 35·1% 5064 (4627 to 5483) 4126 (3747 to 4541) −18·5% Vanuatu 6 (5 to 8) 10 (7 to 14) 64·7% 7084 (5199 to 9626) 5797 (4304 to 7946) −18·2% Venezuela 435 (382 to 498) 649 (551 to 755) 49·2% 3378 (3051 to 3753) 2556 (2208 to 2935) −24·3% Vietnam 2310 (1974 to 2691) 3179 (2588 to 3745) 37·6% 4775 (4164 to 5511) 3862 (3134 to 4523) −19·1% Yemen 615 (436 to 791) 851 (629 to 1175) 38·3% 6756 (4862 to 9201) 5734 (4079 to 8182) −15·1% Zambia 557 (425 to 738) 636 (493 to 823) 14·3% 6691 (5443 to 7948) 5755 (4199 to 7711) −14·0% Zimbabwe 282 (237 to 336) 451 (339 to 595) 60·0% 3629 (2993 to 4329) 4009 (2809 to 5567) 10·5% Data are n or % with 95% uncertainty intervals (UI). For more details on disorder, country, year, age, and sex, see the Global Burden of Disase compare website https://vizhub.healthdata.org/gbd-compare/. DALYs=disability-adjusted life-years. * UIs cannot be calculated for aggregated estimates of percentage changes. Global burden by SDI The patterns of disease varied along the development spectrum as measured by the SDI (figure 5). Age-standardised DALY rates of communicable neurological disorders were the largest cause of DALYs at low levels of SDI. Stroke rates increased from low to middle levels of SDI and then decreased to their lowest values at the highest level of SDI. The headaches showed little change in rates with SDI. Rates of epilepsy gradually decreased with rising SDI (figure 5A). However, the all-age rates of combined neurological disorders increased from middle range to highest SDI values, particularly in females (figure 5B). Most of the changes in DALY rates of neurological disorders with development were driven by changes in YLLs. By comparison, the burden due to YLDs showed less variation over the range of SDI (figure 5C).Figure 5 Expected relationship between the Socio-demographic Index and DALY rates for neurological disorders per 100 000 people between 1990 and 2015

s of neurological disorders with development were driven by changes in YLLs. By comparison, the burden due to YLDs showed less variation over the range of SDI (figure 5C).Figure 5 Expected relationship between the Socio-demographic Index and DALY rates for neurological disorders per 100 000 people between 1990 and 2015 Age-standardised disability-adjusted life-years (DALYs) per 100 000 people (A) by sex; (B) all-age DALY rate per 100 000 by sex; and (C) age-standardised rate per 100 000 by years of life lost (YLLs) and years lived with disability (YLDs).

s of neurological disorders with development were driven by changes in YLLs. By comparison, the burden due to YLDs showed less variation over the range of SDI (figure 5C).Figure 5 Expected relationship between the Socio-demographic Index and DALY rates for neurological disorders per 100 000 people between 1990 and 2015 Age-standardised disability-adjusted life-years (DALYs) per 100 000 people (A) by sex; (B) all-age DALY rate per 100 000 by sex; and (C) age-standardised rate per 100 000 by years of life lost (YLLs) and years lived with disability (YLDs). On the basis of SDI, Oceania, east and southeast Asia, eastern and central Europe, and central Asia had higher than expected age-standardised DALY rates for stroke for males over the entire estimation period. In females, the stroke DALY rates followed the same pattern but crossed the expected line in a few years of estimation. In females, the stroke DALY rates were initially higher than expected on the basis of SDI but fell to below expected levels during the study period. Latin America, apart from tropical Latin America in the 1990s, eastern and western sub-Saharan Africa, western Europe, Australasia, and high-income North America had lower DALY rates than expected based on SDI (appendix p 170). North Africa and the Middle East had higher than expected DALY rates for Alzheimer's disease and other dementias for males and females. Females in central sub-Saharan Africa, high-income North America, and tropical Latin America had DALY rates close to expected for their SDI values, while all other regions had much lower than expected rates (appendix p 171). Migraine DALY rates in males in south Asia were furthest above the expected line based on SDI. In females, migraine DALY rates in tropical Latin America, south Asia, Australasia, western Europe, and high-income North America were higher than expected. East Asia had much lower DALY rates of migraine than expected in males and females (appendix p 172). Epilepsy DALY rates were higher than expected in eastern, southern, and central sub-Saharan Africa and central Asia. Epilepsy rates were a bit lower than expected in Oceania and east Asia. The DALY rates for meningitis were higher than expected based on SDI in western sub-Saharan Africa. All other regions closely followed the expected pattern (appendix p 173). Plots of regional age-standardised DALY rates and SDI for the remaining neurological disorders are in the appendix (pp 174–83).

nia and east Asia. The DALY rates for meningitis were higher than expected based on SDI in western sub-Saharan Africa. All other regions closely followed the expected pattern (appendix p 173). Plots of regional age-standardised DALY rates and SDI for the remaining neurological disorders are in the appendix (pp 174–83). Discussion Neurological disorders including stroke, communicable neurological diseases, and brain cancer accounted for 10·2% of global DALYs and 16·8% of all deaths in 2015. DALYs from all neurological disorders combined exceeded those from all injuries (249·8 million), cardiovascular disease (228·9 million, excluding stroke), cancer (209·4 million), and mental and substance use disorders (162·4 million).3 Our study provides a comprehensive assessment of the extent, patterns, and trends of DALYs for the combined neurological disorders at global, regional, and national levels for 195 countries and territories with important implications for health policy, including priority-setting and financing of health services.

ion).3 Our study provides a comprehensive assessment of the extent, patterns, and trends of DALYs for the combined neurological disorders at global, regional, and national levels for 195 countries and territories with important implications for health policy, including priority-setting and financing of health services. Despite an overall decrease in the age-standardised rates of DALYs, death, and YLDs of all neurological disorders combined between 1990 and 2015, the number of people dying from and affected by these disorders has increased substantially, contributing to higher health loss across the lifespan. Parkinson's disease was the only neurological disorder with increasing age-standardised rates of deaths, prevalence, and DALYs between 1990 and 2015. Although the burden of communicable neurological disorders has significantly decreased over this period, the burden of non-communicable neurological disorders has significantly increased. This finding is consistent with the overall global burden shift from communicable to non-communicable disorders.16 In terms of absolute number of people affected by neurological disorders, most of the increase in the burden was associated with ageing of the population and population growth.9 Increasing incidence of stroke in low-income and middle-income countries,17 increasing prevalence of multiple sclerosis,18, 19 increasing incidence of epilepsy in elderly people,20 increasing prevalence of tension-type headache,21 and increasing incidence of brain tumours in elderly people22 have been reported elsewhere. Findings from other studies have reported that it is difficult to assess trends in prevalence and incidence of Parkinson's disease because of changes in case definitions over time.23, 24 The evidence on secular trends in prevalence of migraine is mixed.21, 25

in tumours in elderly people22 have been reported elsewhere. Findings from other studies have reported that it is difficult to assess trends in prevalence and incidence of Parkinson's disease because of changes in case definitions over time.23, 24 The evidence on secular trends in prevalence of migraine is mixed.21, 25 The conclusions from a systematic review26 of similar studies over time to examine trends in prevalence, incidence, and mortality for people with dementia were that there is some evidence that the incidence of dementia might be declining in high-income countries, but evidence on trends in the prevalence of dementia is inconsistent. Our study showed a modest increase in the prevalence of Alzheimer's disease and other dementias in high-income North America, high-income Asia Pacific, east Asia, south Asia, the Caribbean, and southern sub-Saharan Africa, and a modest decrease elsewhere. The small increase in the global age-standardised prevalence of Alzheimer's disease and other dementias is in contrast with a significant reduction from 1990 to 2015 in the prevalence of stroke, despite stroke sharing risks with, and contributing to, vascular dementia. For stroke, blood pressure control and smoking cessation have been important contributors to the reduction of its incidence in high-income and middle-income countries.27, 28, 29 Diverging trends of stroke and dementia incidence rates have been reported elsewhere.30 However, the most striking change has been the more than doubling of people in the world who die or are disabled from Alzheimer's disease and other dementias over the past 25 years. As ageing of the global population continues, our findings have important major health-service implications for the care of patients and adequate support for affected families.

been the more than doubling of people in the world who die or are disabled from Alzheimer's disease and other dementias over the past 25 years. As ageing of the global population continues, our findings have important major health-service implications for the care of patients and adequate support for affected families. Vaccinations have contributed to the favourable trends in the DALY rates of tetanus, meningitis, and encephalitis.31, 32, 33, 34, 35, 36, 37 Our estimate of 23·4 million cases of active epilepsy in 2015 is lower than the 32·7 million cases estimated in a meta-analysis of 65 prevalence studies, although that study did not specify a year of estimate.38 Similar to findings from this meta-analysis, we noted large geographical variations in the prevalence of epilepsy, with significantly greater rates in low-income and middle-income countries. The high prevalence of epilepsy in low-income and middle-income countries can partly be accounted for by the greater number of cases with communicable causes in these countries.38, 39 Our 2015 prevalence estimates are within the ranges reported for Parkinson's disease (51–177 per 100 000 people),40 motor neuron disease (1·9–3·9 per 100 000 people),41, 42 tension-type headache (21–27%),43, 44 medication overuse headache (0·5%–7·2%),45 and migraine (mean estimate 10%, range 1–25),46 but lower than that reported for multiple sclerosis (65–74 per 100 000 people),47 dementias (4·2–8·0% in people aged 60 years or older),48 and stroke (0·4–2·1% in low-income and middle-income countries).49 Differences in study populations (eg, age range and countries included) might account for some of the observed differences in prevalence rates.

for multiple sclerosis (65–74 per 100 000 people),47 dementias (4·2–8·0% in people aged 60 years or older),48 and stroke (0·4–2·1% in low-income and middle-income countries).49 Differences in study populations (eg, age range and countries included) might account for some of the observed differences in prevalence rates. Between countries, age-standardised rates of DALYs and deaths from neurological disorders as a group varied up to six times, with the highest rates in low-income to middle-income countries. These geographical patterns of the burden and distribution of individual neurological disorders are important for global and regional health-care planning and might inform further research to examine possible causes of the diseases. For example, the clear latitudinal gradient we noted in the prevalence of multiple sclerosis (about two times higher prevalence in countries at highest latitudes compared with those on the equator) corresponds to that observed in other studies50, 51 and is suggestive of the role of environmental factors (eg, vitamin D deficiency and infection).52, 53

al gradient we noted in the prevalence of multiple sclerosis (about two times higher prevalence in countries at highest latitudes compared with those on the equator) corresponds to that observed in other studies50, 51 and is suggestive of the role of environmental factors (eg, vitamin D deficiency and infection).52, 53 Our findings of large geographical variations in the burden of stroke, dementias, Parkinson's disease, epilepsy, migraine, medication overuse headache, motor neuron disease, and brain and nervous system cancers concur with previous observations.17, 24, 26, 38, 45, 46, 48, 54, 55, 56, 57, 58 In non-communicable neurological disorders, the largest (greater than 20 times) variations in age-standardised DALYs were observed for stroke, motor neuron disease, and multiple sclerosis; in communicable neurological disorders, geographical differences ranged from 100 times for encephalitis to 10 000 times for tetanus. The greater geographical variation of communicable disorders was related to their overwhelming predominance in countries at low levels of socio-demographic development and emphasised the need for better prevention (including vaccination and sanitation measures), as well as better case management in these regions.

anus. The greater geographical variation of communicable disorders was related to their overwhelming predominance in countries at low levels of socio-demographic development and emphasised the need for better prevention (including vaccination and sanitation measures), as well as better case management in these regions. We noted significant sex differences in the burden of many neurological disorders analysed (higher prevalence of tetanus, stroke, Parkinson's disease, and motor neuron disease, but lower prevalence of multiple sclerosis and various types of headaches in males). The greater prevalence of stroke,59 Parkinson's disease,24 epilepsy,60 and motor neuron disease61 in males has also been reported by other studies. A greater prevalence in females has been reported elsewhere for migraine62 and multiple sclerosis.50 Our finding of 22% higher age-standardised prevalence of Alzheimer's disease and other dementias in women is in accordance with a finding from a meta-analysis of consistently higher estimates among 160 studies, although it was reported not to be significant.48 The higher age-standardised rate might be partly due to having a top age category of 80 years and older. As dementia is so highly prevalent at oldest ages, greater age detail in estimation of dementia prevalence might reduce the observed sex difference.

ong 160 studies, although it was reported not to be significant.48 The higher age-standardised rate might be partly due to having a top age category of 80 years and older. As dementia is so highly prevalent at oldest ages, greater age detail in estimation of dementia prevalence might reduce the observed sex difference. Our study findings have important health service implications. The large and increasing numbers of patients with neurological disorders necessitate careful planning by governments and other health-care providers to ensure adequate funding and staff for their treatment and rehabilitation services. However, a recent WHO–World Federation of Neurology survey63 of services and other resources for neurological disorders in 109 countries (90% of the world population) showed that there are large inequalities in access to neurological care across different populations, in particular for those living in low-income to middle-income countries. With a global shortage of neurologists, neurosurgeons, and rehabilitation professionals, improving neurological care will require innovative strategies within existing health systems. A good example of such innovative strategies is the use of nurses, nurse practitioners, and physician assistants trained in stroke care to care for patients with acute stroke and transient ischaemic attack in stroke units. A recently suggested implementation cycle for combating cardiovascular disease in low-income to middle-income countries64 provides a good template for similar interventional strategies for reducing the burden of neurological disorders in such countries. Although improving care and rehabilitation of people with neurological disorders is important for improving outcomes, there also are effective primary prevention strategies for communicable neurological disorders and stroke. However, proven effective preventive strategies are often underutilised.65, 66 More quality epidemiological research on risk factors, incidence, prevalence, and outcomes of neurological disorders in various countries is required to guide better prevention and management of these disorders, and our findings could help to prioritise such efforts. The sex, age, and regional and national differences and trends in the burden of neurological disorders necessitate the development, implementation, and prioritisation of treatments and preventive interventions that are specific for sex, age, and population to reduce the burden from these disorders.

ritise such efforts. The sex, age, and regional and national differences and trends in the burden of neurological disorders necessitate the development, implementation, and prioritisation of treatments and preventive interventions that are specific for sex, age, and population to reduce the burden from these disorders. Although ours was the most up-to-date overview of the global burden of major neurological disorders, this study was not free from some limitations in addition to overall GBD limitations.3, 4 First, we assumed that the excess mortality in Alzheimer's disease and other dementias implied by prevalence and mortality rates in countries that were most willing to code deaths to dementia in their vital registrations would apply to all other countries and periods. Although we realised that excess mortality was unlikely to be generalisable over location and time, we chose to make this assumption to address the much larger change we observed in deaths certified as Alzheimer's disease and other dementias between countries and over time than seen in prevalence or incidence studies. Second, estimates of cause-specific death rates in most low-income and middle-income countries rely on verbal autopsy data rather than physician-certified death records. Verbal autopsy instruments can identify deaths due to some neurological disorders, including stroke, meningitis, tetanus, and epilepsy, but are unable to capture other neurological disorders.67 Third, there are also many gaps in data availability by world region, and many low-income and middle-income regions do not have any epidemiological data. Epilepsy is the only neurological disorder with data sources for all 21 GBD regions. Fourth, heterogeneity in study methods and case definitions complicates the non-fatal estimation. Although we endeavoured to adjust for such methodological differences, this relied on generalising adjustment factors from few studies. Fifth, some categories of neurological disease were not included in this analysis because of an inability to aggregate cause-level and sequela-level data. For this reason, we were unable to include secondary epilepsy, the long-term neurological consequences of neonatal disorders, or traumatic brain injury and spinal cord injury, which are estimated in GBD as sequelae of injuries such as falls or road injuries.

an inability to aggregate cause-level and sequela-level data. For this reason, we were unable to include secondary epilepsy, the long-term neurological consequences of neonatal disorders, or traumatic brain injury and spinal cord injury, which are estimated in GBD as sequelae of injuries such as falls or road injuries. Adding currently missed neurological disorders would increase the health significance of the neurological disorders for public health systems and should be possible in a future GBD iteration. Sixth, sparse data on the severity of these neurological diseases did not allow us to differentiate severity by location or over time, with the exception of epilepsy (appendix pp 5–113). Relying on few studies, often from high-income countries only, meant we were unable to quantify any treatment effects on severity.

n. Sixth, sparse data on the severity of these neurological diseases did not allow us to differentiate severity by location or over time, with the exception of epilepsy (appendix pp 5–113). Relying on few studies, often from high-income countries only, meant we were unable to quantify any treatment effects on severity. In conclusion, we have shown that neurological disorders are a large cause of disability and death worldwide. Globally, the burden of neurological disorders has increased substantially over the past 25 years because of population ageing, despite substantial decreases in mortality rates from stroke and communicable neurological disorders. Because low-income and middle-income countries still have a long way to go through the demographic transition of reductions in child mortality and population ageing, the number of patients who will need neurological care will continue to grow in the coming decades. It is important that policy makers and health-care providers are aware of these past trends to be able to provide adequate services for the growing numbers of patients with neurological disorders. Correspondence to: Prof Valery L Feigin, National Institute for Stroke and Applied Neurosciences, School of Public Health and Psychosocial Studies, Faculty of Health and Environmental Studies, Auckland University of Technology, North Shore Campus, Auckland 1142, New Zealand valery.feigin@aut.ac.nz and Prof Theo Vos, Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA 98121, USA tvos@uw.edu

Correspondence to: Prof Valery L Feigin, National Institute for Stroke and Applied Neurosciences, School of Public Health and Psychosocial Studies, Faculty of Health and Environmental Studies, Auckland University of Technology, North Shore Campus, Auckland 1142, New Zealand valery.feigin@aut.ac.nz and Prof Theo Vos, Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA 98121, USA tvos@uw.edu For more on the Global Health Data Exchange see http://ghdx.healthdata.org/gbd-2015/data-input-sources Supplementary Material Supplementary appendix Acknowledgments This research was supported by the Bill & Melinda Gates Foundation and the Health Research Council of New Zealand (VLF, RVK, PGP). VLF was also partly funded by the Brain Research New Zealand Centre of Research Excellence and Ageing Well programme of the National Science Challenge, Ministry of Business, Innovation and Employment of New Zealand. The views expressed in this article are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute, National Institutes of Health, or the US Department of Health and Human Services.

l Science Challenge, Ministry of Business, Innovation and Employment of New Zealand. The views expressed in this article are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute, National Institutes of Health, or the US Department of Health and Human Services. GBD 2015 Neurological Disorders Collaborator Group Valery L Feigin, Amanuel Alemu Abajobir, Kalkidan Hassen Abate, Foad Abd-Allah, Abdishakur M Abdulle, Semaw Ferede Abera, Gebre Yitayih Abyu, Muktar Beshir Ahmed, Amani Nidhal Aichour, Ibtihel Aichour, Miloud Taki Eddine Aichour, Rufus Olusola Akinyemi, Samer Alabed, Rajaa Al-Raddadi, Nelson Alvis-Guzman, Azmeraw T. Amare, Hossein Ansari, Palwasha Anwari, Johan Ärnlöv, Hamid Asayesh, Solomon Weldegebreal Asgedom, Tesfay Mehari Atey, Leticia Avila-Burgos, Euripide FrinelG. Arthur Avokpaho, Mahmood Reza Azarpazhooh, Aleksandra Barac, Miguel Barboza, Suzanne L Barker-Collo, Till Bärnighausen, Neeraj Bedi, Ettore Beghi, Derrick A Bennett, Isabela M Bensenor, Adugnaw Berhane, Balem Demtsu Betsu, Soumyadeep Bhaumik, Sait Mentes Birlik, Stan Biryukov, Dube Jara Boneya, Lemma NegesaBulto Bulto, Hélène Carabin, Daniel Casey, Carlos A.

h, Aleksandra Barac, Miguel Barboza, Suzanne L Barker-Collo, Till Bärnighausen, Neeraj Bedi, Ettore Beghi, Derrick A Bennett, Isabela M Bensenor, Adugnaw Berhane, Balem Demtsu Betsu, Soumyadeep Bhaumik, Sait Mentes Birlik, Stan Biryukov, Dube Jara Boneya, Lemma NegesaBulto Bulto, Hélène Carabin, Daniel Casey, Carlos A. Castañeda-Orjuela, Ferrán Catalá-López, Honglei Chen, Abdulaal A Chitheer, Rajiv Chowdhury, Hanne Christensen, Lalit Dandona, Rakhi Dandona, Gabrielle A deVeber, Samath D Dharmaratne, Huyen Phuc Do, Klara Dokova, E Ray Dorsey, Richard G Ellenbogen, Sharareh Eskandarieh, Maryam S Farvid, Seyed-Mohammad Fereshtehnejad, Florian Fischer, Kyle J Foreman, Johanna M Geleijnse, Richard F Gillum, Giorgia Giussani, Ellen M Goldberg, Philimon N Gona, Alessandra Carvalho Goulart, Harish Chander Gugnani, Rahul Gupta, Rajeev Gupta, Vladimir Hachinski, Randah Ribhi Hamadeh, Mitiku Hambisa, Graeme J Hankey, Habtamu Abera Hareri, Rasmus Havmoeller, Simon I Hay, Pouria Heydarpour, Peter J Hotez, Mihajlo (Michael) B Jakovljevic, Mehdi Javanbakht, Panniyammakal Jeemon, Jost B Jonas, Yogeshwar Kalkonde, Amit Kandel, André Karch, Amir Kasaeian, Anshul Kastor, Peter Njenga Keiyoro, Yousef Saleh Khader, Ibrahim A Khalil, Ejaz Ahmad Khan, Young-Ho Khang, Abdullah TawfihAbdullah Khoja, Jagdish Khubchandani, Chanda Kulkarni, Daniel Kim, Yun Jin Kim, Mika Kivimaki, Yoshihiro Kokubo, Soewarta Kosen, Michael Kravchenko, Rita Vanmala Krishnamurthi, Barthelemy Kuate Defo, G Anil Kumar, Rashmi Kumar, Hmwe H Kyu, Anders Larsson, Pablo M Lavados, Yongmei Li, Xiaofeng Liang, Misgan Legesse Liben, Warren D Lo, Giancarlo Logroscino, Paulo A Lotufo, Clement T Loy, Mark T Mackay, Hassan Magdy Abd El Razek, Mohammed Magdy Abd El Razek, Azeem Majeed, Reza Malekzadeh, Treh Manhertz, Lorenzo G Mantovani, João Massano, Mohsen Mazidi, Colm McAlinden, Suresh Mehata, Man Mohan Mehndiratta, Ziad A Memish, Walter Mendoza, Mubarek Abera Mengistie, George A Mensah, Atte Meretoja, Haftay Berhane Mezgebe, Ted R Miller, Shiva Raj Mishra, Norlinah Mohamed Ibrahim, Alireza Mohammadi, Kedir Endris Mohammed, Shafiu Mohammed, Ali H Mokdad, Maziar Moradi-Lakeh, Ilais Moreno Velasquez, Kamarul Imran Musa, Mohsen Naghavi, Josephine Wanjiku Ngunjiri, Cuong Tat Nguyen, Grant Nguyen, Quyen Le Nguyen, Trang Huyen Nguyen, Emma Nichols, Dina NurAnggraini Ningrum, Vuong Minh Nong, Bo Norrving, Jean Jacques N Noubiap, Felix Akpojene Ogbo, Mayowa O Owolabi, Jeyaraj D Pandian, Priyakumari Ganesh Parmar, David M Pereira, Max Petzold,

sen Naghavi, Josephine Wanjiku Ngunjiri, Cuong Tat Nguyen, Grant Nguyen, Quyen Le Nguyen, Trang Huyen Nguyen, Emma Nichols, Dina NurAnggraini Ningrum, Vuong Minh Nong, Bo Norrving, Jean Jacques N Noubiap, Felix Akpojene Ogbo, Mayowa O Owolabi, Jeyaraj D Pandian, Priyakumari Ganesh Parmar, David M Pereira, Max Petzold, Michael Robert Phillips, Michael A Piradov, Richie G.

sen Naghavi, Josephine Wanjiku Ngunjiri, Cuong Tat Nguyen, Grant Nguyen, Quyen Le Nguyen, Trang Huyen Nguyen, Emma Nichols, Dina NurAnggraini Ningrum, Vuong Minh Nong, Bo Norrving, Jean Jacques N Noubiap, Felix Akpojene Ogbo, Mayowa O Owolabi, Jeyaraj D Pandian, Priyakumari Ganesh Parmar, David M Pereira, Max Petzold, Michael Robert Phillips, Michael A Piradov, Richie G. Poulton, Farshad Pourmalek, Mostafa Qorbani, Anwar Rafay, Mahfuzar Rahman, Mohammad HifzUr Rahman, Rajesh Kumar Rai, Sasa Rajsic, Annemarei Ranta, Salman Rawaf, Andre M.N. Renzaho, Mohammad Sadegh Rezai, Gregory A Roth, Gholamreza Roshandel, Enrico Rubagotti, Perminder Sachdev, Saeid Safiri, Ramesh Sahathevan, Mohammad Ali Sahraian, Abdallah M. Samy, Paula Santalucia, Itamar S Santos, Benn Sartorius, Maheswar Satpathy, Monika Sawhney, Mete I Saylan, Sadaf G Sepanlou, Masood Ali Shaikh, Raad Shakir, Morteza Shamsizadeh, Kevin N Sheth, Mika Shigematsu, Haitham Shoman, Diego AugustoSantos Silva, Mari Smith, Eugene Sobngwi, Luciano A Sposato, Jeffrey D Stanaway, Dan J Stein, Timothy J Steiner, Lars Jacob Stovner, Rizwan Suliankatchi Abdulkader, Cassandra EI Szoeke, Rafael Tabarés-Seisdedos, David Tanne, Alice M Theadom, Amanda G Thrift, David L Tirschwell, Roman Topor-Madry, Bach Xuan Tran, Thomas Truelsen, Kald Beshir Tuem, Kingsley Nnanna Ukwaja, Olalekan A Uthman, Yuri Y Varakin, Tommi Vasankari, Narayanaswamy Venketasubramanian, Vasiliy Victorovich Vlassov, Fiseha Wadilo, Tolassa Wakayo, Mitchell T Wallin, Elisabete Weiderpass, Ronny Westerman, Tissa Wijeratne, Charles Shey Wiysonge, Minyahil Alebachew Woldu, Charles D A Wolfe, Denis Xavier, Gelin Xu, Yuichiro Yano, Hassen Hamid Yimam, Naohiro Yonemoto, Chuanhua Yu, Zoubida Zaidi, Maysaa El Sayed Zaki, Joseph R Zunt, Christopher J L Murray, Theo Vos.

, Mitchell T Wallin, Elisabete Weiderpass, Ronny Westerman, Tissa Wijeratne, Charles Shey Wiysonge, Minyahil Alebachew Woldu, Charles D A Wolfe, Denis Xavier, Gelin Xu, Yuichiro Yano, Hassen Hamid Yimam, Naohiro Yonemoto, Chuanhua Yu, Zoubida Zaidi, Maysaa El Sayed Zaki, Joseph R Zunt, Christopher J L Murray, Theo Vos. Affiliations National Institute for Stroke and Applied Neurosciences (Prof V L Feigin PhD), Auckland University of Technology, Auckland, New Zealand (R V Krishnamurthi PhD, A M Theadom PhD); School of Public Health (A A Abajobir MPH), University of Queensland, Brisbane, QLD, Australia (S R Mishra MPH); Department of Epidemiology, College of Health Sciences (M B Ahmed MPH), Jimma University, Jimma, Ethiopia (K H Abate MS, M A Mengistie MS, T Wakayo MS); Department of Neurology, Cairo University, Cairo, Egypt (Prof F Abd-Allah MD); New York University Abu Dhabi, Abu Dhabi, United Arab Emirates (A M Abdulle PhD); School of Public Health, College of Health Sciences (S F Abera MSc, K E Mohammed MPH), Mekelle University, Mekelle, Ethiopia (Prof G Y Abyu MS, S W Asgedom MS, T M Atey MS, B D Betsu MS, H B Mezgebe MS, K B Tuem MS, M A Woldu MS); Food Security and Institute for Biological Chemistry and Nutrition, University of Hohenheim, Stuttgart, Germany (S F Abera MSc); University Ferhat Abbas of Setif, Setif, Algeria (A N Aichour BS); National Institute of Nursing Education, Setif, Algeria (I Aichour MS); High National School of Veterinary Medicine, Algiers, Algeria (M T Aichour MD); University of Ibadan, Ibadan, Nigeria (R O Akinyemi PhD); Newcastle University, Newcastle upon Tyne, UK (R O Akinyemi PhD); University of Sheffield, Sheffield, UK (S Alabed MS); Joint Program of Family and Community Medicine, Jeddah, Saudi Arabia (R Al-Raddadi PhD); Universidad de Cartagena, Cartagena de Indias, Colombia (Prof N Alvis-Guzman PhD); School of Medicine, University of Adelaide, Adelaide, SA, Australia (A T Amare MPH); College of Medicine and Health Sciences, Bahir Dar University, Bahir Dar, Ethiopia (A T Amare); Health Promotion Research Center, Department of Epidemiology and Biostatistics, Zahedan University of Medical Sciences, Zahedan, Iran (H Ansari PhD); self employed, Kabul, Afghanistan (P Anwari MS); Department of Neurobiology, Care Sciences and Society, Division of Family Medicine and Primary Care (Prof J Ärnlöv PhD, S Fereshtehnejad PhD), Department of Medical Epidemiology and Biostatistics (E Weiderpass PhD), Karolinska Institutet, Stockholm, Sweden (R Havmoeller Ph

lf employed, Kabul, Afghanistan (P Anwari MS); Department of Neurobiology, Care Sciences and Society, Division of Family Medicine and Primary Care (Prof J Ärnlöv PhD, S Fereshtehnejad PhD), Department of Medical Epidemiology and Biostatistics (E Weiderpass PhD), Karolinska Institutet, Stockholm, Sweden (R Havmoeller Ph D); School of Health and Social Studies, Dalarna University, Falun, Sweden (Prof J Ärnlöv); Department of Medical Emergency, School of Paramedic, Qom University of Medical Sciences, Qom, Iran (H Asayesh MS); National Institute of Public Health, Cuernavaca, Mexico (L Avila-Burgos PhD); Institut de Recherche Clinique du Bénin (IRCB), Cotonou, Benin (E F G A Avokpaho MPH); Laboratoire d'Etudes et de Recherche-Action en Santé (LERAS Afrique), Parakou, Benin (E F G A Avokpaho); Ghaem Hospital, Mashhad University of Medical Sciences, Mashhad, Iran (M R Azarpazhooh MD); Faculty of Medicine, University of Belgrade, Belgrade, Serbia (A Barac PhD); Hospital Dr Rafael A Calderón Guardia, CCSS, San José, Costa Rica (M Barboza MD); Universidad de Costa Rica, San Pedro, Costa Rica (M Barboza); School of Psychology, University of Auckland, Auckland, New Zealand (S L Barker-Collo PhD); Department of Global Health and Population, Harvard T H Chan School of Public Health (Prof T Bärnighausen MD), Department of Nutrition, Harvard T H Chan School of Public Health (M S Farvid PhD), Harvard University, Boston, MA, USA; Africa Health Research Institute, Mtubatuba, South Africa (Prof T Bärnighausen); Institute of Public Health, Heidelberg University, Heidelberg, Germany (Prof T Bärnighausen, S Mohammed PhD); College of Public Health and Tropical Medicine, Jazan, Saudi Arabia (N Bedi MD); IRCCS–Istituto di Ricerche Farmacologiche Mario Negri, Milan, Italy (E Beghi MD, G Giussani BiolD); Nuffield Department of Population Health (D A Bennett PhD), Oxford Big Data Institute, Li Ka Shing Centre for Health Information and Discovery (Prof S I Hay DSc), University of Oxford, Oxford, UK; Center for Clinical and Epidemiological Research Center, Hospital Universitario (A C Goulart PhD), Internal Medicine Department (Prof I S Santos PhD), University of São Paulo, São Paulo, Brazil (I M Bensenor PhD, Prof P A Lotufo DrPH); College of Health Sciences, Debre Berhan University, Debre Berhan, Ethiopia (A Berhane PhD); Centre for Control of Chronic Conditions (P Jeemon PhD), Public Health Foundation of India, Gurugram, India (S Bhaumik MSc, (Prof L Dandona MD, Prof R Dandona PhD, G A Kumar PhD); GBS-CID

M Bensenor PhD, Prof P A Lotufo DrPH); College of Health Sciences, Debre Berhan University, Debre Berhan, Ethiopia (A Berhane PhD); Centre for Control of Chronic Conditions (P Jeemon PhD), Public Health Foundation of India, Gurugram, India (S Bhaumik MSc, (Prof L Dandona MD, Prof R Dandona PhD, G A Kumar PhD); GBS-CID P International Foundation, Menemen, Turkey (S M Birlik BS); Institute for Health Metrics and Evaluation (S Biryukov BS, D Casey MPH, Prof L Dandona, Prof R Dandona, K J Foreman PhD, E M Goldberg BS, Prof S I Hay DSc, I A Khalil MD, H H Kyu PhD, T Manhertz BA, Prof A H Mokdad PhD, Prof M Naghavi PhD, G Nguyen MPH, E Nichols BA, M Smith MPA, Prof C J L Murray DPhil, G A Roth MD, J D Stanaway PhD, Prof T Vos PhD), Harborview/UW Medicine (R G Ellenbogen MD), Center for Health Trends and Forecasts, Institute for Health Metrics and Evaluation (Prof M B Jakovljevic PhD), University of Washington, Seattle, WA, USA (D L Tirschwell MD, J R Zunt MD); Department of Public Health, Debre Markos University, Debre Markos, Ethiopia (D J Boneya MPH); College of Health and Medical Sciences (M Hambisa MPH), Haramaya University, Harar, Ethiopia (L N B Bulto MS); Department of Biostatistics and Epidemiology, University of Oklahoma Health Sciences Center, Oklahoma City, OK, USA (H Carabin PhD); Colombian National Health Observatory, Instituto Nacional de Salud, Bogota, Colombia (C A Castañeda-Orjuela MSc); Epidemiology and Public Health Evaluation Group, Public Health Department, Universidad Nacional de Colombia, Bogota, Colombia (C A Castañeda-Orjuela); Department of Medicine, University of Valencia, INCLIVA Health Research Institute and CIBERSAM, Valencia, Spain (F Catalá-López PhD, Prof R Tabarés-Seisdedos PhD); Clinical Epidemiology Program, Ottawa Hospital Research Institute, Ottawa, ON, Canada (F Catalá-López); Michigan State University, East Lansing, MI, USA (H Chen PhD); Ministry of Health, Baghdad, Iraq (A A Chitheer MD); Department of Public Health and Primary Care, University of Cambridge, Cambridge, UK (R Chowdhury PhD); Bispebjerg University Hospital, Copenhagen, Denmark (Prof H Christensen DMSCi); The Hospital for Sick Children, University of Toronto, Toronto, ON, Canada (G A deVeber MD); Department of Community Medicine, Faculty of Medicine, University of Peradeniya, Peradeniya, Sri Lanka (S D Dharmaratne MD); Institute for Global Health Innovations, Duy Tan University, Da Nang, Vietnam (H P Do MSc, C T Nguyen MSc, Q L Nguyen MD, T H Nguyen MSc, V M Nong MSc); Departme

nto, ON, Canada (G A deVeber MD); Department of Community Medicine, Faculty of Medicine, University of Peradeniya, Peradeniya, Sri Lanka (S D Dharmaratne MD); Institute for Global Health Innovations, Duy Tan University, Da Nang, Vietnam (H P Do MSc, C T Nguyen MSc, Q L Nguyen MD, T H Nguyen MSc, V M Nong MSc); Departme nt of Social Medicine, Faculty of Public Health, Medical University–Varna, Varna, Bulgaria (K Dokova PhD); University of Rochester Medical Center, Rochester, NY, USA (E R Dorsey MD); Multiple Sclerosis Research Center, Tehran, Iran (S Eskandarieh PhD); Harvard/MGH Center on Genomics, Vulnerable Populations, and Health Disparities, Mongan Institute for Health Policy, Massachusetts General Hospital, Boston, MA, USA (M S Farvid PhD); School of Public Health, Bielefeld University, Bielefeld, Germany (F Fischer PhD); Department of Primary Care and Public Health (Prof A Majeed MD), Division of Brain Sciences (Prof T J Steiner PhD), Imperial College London, London, UK (K J Foreman, Prof S Rawaf MD, Prof R Shakir MD, H Shoman MPH); Division of Human Nutrition, Wageningen University, Wageningen, Netherlands (J M Geleijnse PhD); Howard University, Washington, DC, USA (R F Gillum MD); University of Massachusetts Boston, Boston, MA, USA (Prof P N Gona PhD); Center of Check of Hospital Sirio Libanes, São Paulo, Brazil (A C Goulart PhD); Departments of Microbiology and Epidemiology & Biostatistics, Saint James School of Medicine, The Quarter, Anguilla (Prof H C Gugnani PhD); West Virginia Bureau for Public Health, Charleston, WV, USA (R Gupta MD); Eternal Heart Care Centre and Research Institute, Jaipur, India (R Gupta PhD); Western University, London, ON, Canada (Prof V Hachinski DSc); Arabian Gulf University, Manama, Bahrain (Prof R R Hamadeh DPhil); School of Medicine and Pharmacology, University of Western Australia, Perth, WA, Australia (Prof G J Hankey MD); Harry Perkins Institute of Medical Research, Nedlands, WA, Australia (Prof G J Hankey MD); Western Australian Neuroscience Research Institute, Nedlands, WA, Australia (Prof G J Hankey); Addis Ababa University, Addis Ababa, Ethiopia (H A Hareri MS, M A Woldu MS); Multiple Sclerosis Research Center, Neuroscience Institute (P Heydarpour MD, M A Sahraian MD), Hematology-Oncology and Stem Cell Transplantation Research Center (A Kasaeian PhD), Endocrinology and Metabolism Population Sciences Institute (A Kasaeian PhD), Digestive Diseases Research Institute (Prof R Malekzadeh MD, G Roshandel PhD, S G Sepanlou PhD), Tehran U

science Institute (P Heydarpour MD, M A Sahraian MD), Hematology-Oncology and Stem Cell Transplantation Research Center (A Kasaeian PhD), Endocrinology and Metabolism Population Sciences Institute (A Kasaeian PhD), Digestive Diseases Research Institute (Prof R Malekzadeh MD, G Roshandel PhD, S G Sepanlou PhD), Tehran U niversity of Medical Sciences, Tehran, Iran; College of Medicine, Baylor University, Houston, TX, USA (P J Hotez PhD); Faculty of Medical Sciences, University of Kragujevac, Kragujevac, Serbia (Prof M [B] Jakovljevic); University of Aberdeen, Aberdeen, UK (M Javanbakht PhD); Centre for Chronic Disease Control, New Delhi, India (P Jeemon PhD); Department of Ophthalmology, Medical Faculty Mannheim, Ruprecht-Karls-University Heidelberg, Mannheim, Germany (Prof J B Jonas MD); Society for Education, Action and Research in Community Health, Gadchiroli, India (Y Kalkonde MD); University at Buffalo, Buffalo, NY, USA (A Kandel MBBS); Epidemiological and Statistical Methods Research Group, Helmholtz Centre for Infection Research, Braunschweig, Germany (A Karch MD); Hannover-Braunschweig Site, German Center for Infection Research, Braunschweig, Germany (A Karch); International Institute for Population Sciences, Mumbai, India (A Kastor MPhil, M H U Rahman MPhil); Institute of Tropical and Infectious Diseases, Nairobi, Kenya (P N Keiyoro PhD); School of Continuing and Distance Education, Nairobi, Kenya (P N Keiyoro); Department of Community Medicine, Public Health and Family Medicine, Jordan University of Science and Technology, Irbid, Jordan (Prof Y S Khader ScD); Health Services Academy, Islamabad, Pakistan (E A Khan MD); Department of Health Policy and Management, Seoul National University College of Medicine, Seoul, South Korea (Prof Y Khang MD); Institute of Health Policy and Management, Seoul National University Medical Center, Seoul, South Korea (Prof Y Khang); Department of Public Health and Department of Family Medicine, College of Medicine, Mohammed Ibn Saudi University, Riyadh, Saudi Arabia (A T A Khoja MD); Johns Hopkins Bloomberg School of Public Health (A T A Khoja), Johns Hopkins University, Baltimore, MD, USA (B X Tran PhD); Department of Nutrition and Health Science, Ball State University, Muncie, IN, USA (J Khubchandani PhD); Department of Health Sciences, Northeastern University, Boston, MA, USA (Prof D Kim DrPH); School of Medicine, Xiamen University Malaysia Campus, Sepang, Malaysia (Y J Kim PhD); Department of Epidemiology and Public Health, University

ion and Health Science, Ball State University, Muncie, IN, USA (J Khubchandani PhD); Department of Health Sciences, Northeastern University, Boston, MA, USA (Prof D Kim DrPH); School of Medicine, Xiamen University Malaysia Campus, Sepang, Malaysia (Y J Kim PhD); Department of Epidemiology and Public Health, University College London, London, UK (Prof M Kivimaki PhD); Clinicum, Faculty of Medicine, University of Helsinki, Helsinki, Finland (Prof M Kivimaki); Department of Preventive Cardiology, National Cerebral and Cardiovascular Center, Suita, Japan (Y Kokubo PhD); Center for Community Empowerment, Health Policy and Humanities, National Institute of Health Research & Development, Jakarta, Indonesia (S Kosen MD); Research Center of Neurology, Moscow, Russia (M Kravchenko PhD, Prof M A Piradov DSc, Prof Y Y Varakin MD); Department of Social and Preventive Medicine, School of Public Health (Prof B Kuate Defo PhD); Rajrajeswari Medical College and Hospital, Bangalore, India (Prof C Kulkarni MD); Department of Demography and Public Health Research Institute, University of Montreal, Montreal, QC, Canada (Prof B Kuate Defo PhD); King George Medical University, Lucknow, India (Prof R Kumar MD); Department of Medical Sciences, Uppsala University, Uppsala, Sweden (Prof A Larsson PhD); Servicio de Neurologia, Clinica Alemana, Universidad del Desarrollo, Santiago, Chile (P M Lavados MD); San Francisco VA Medical Center, San Francisco, CA, USA (Y Li PhD); Chinese Center for Disease Control and Prevention, Beijing, China (Prof X Liang MD); Samara University, Samara, Ethiopia (M L Liben MPH); Departments of Pediatrics and Neurology, Ohio State University, Columbus, OH, USA (W D Lo MD); Nationwide Children's Hospital, Columbus, OH, USA (W D Lo MD); University of Bari, Bari, Italy (Prof G Logroscino PhD); The University of Sydney, Sydney, NSW, Australia (Prof C T Loy PhD); Royal Children's Hospital Melbourne, Melbourne, VIC, Australia (M T Mackay PhD); Department of Medicine (A Meretoja PhD), Institute of Health and Ageing (Prof C E I Szoeke PhD), University of Melbourne, Melbourne, VIC, Australia (M T Mackay PhD); Mansoura Faculty of Medicine, Mansoura, Egypt (H Magdy Abd El Razek MBBCh); Aswan University Hospital, Aswan Faculty of Medicine, Aswan, Egypt (H Magdy Abd El Razek); University of Milano Bicocca, Monza, Italy (Prof L G Mantovani DSc); Hospital Pedro Hispano/ULS Matosinhos, Matosinhos, Portugal (J Massano MD); Faculty of Medicine, University of Porto, Porto, Portugal (J Massano);

azek MBBCh); Aswan University Hospital, Aswan Faculty of Medicine, Aswan, Egypt (H Magdy Abd El Razek); University of Milano Bicocca, Monza, Italy (Prof L G Mantovani DSc); Hospital Pedro Hispano/ULS Matosinhos, Matosinhos, Portugal (J Massano MD); Faculty of Medicine, University of Porto, Porto, Portugal (J Massano); Key State Laboratory of Molecular Developmental Biology, Institute of Genetics and Developmental Biology, Chinese Academy of Sciences, Beijing, China (M Mazidi PhD); University Hospitals Bristol NHS Foundation Trust, Bristol, UK (C McAlinden PhD); Public Health Wales, Swansea, UK (C McAlinden PhD); Ipas Nepal, Kathmandu, Nepal (S Mehata PhD); Janakpuri Superspecialty Hospital, New Delhi, India (Prof M M Mehndiratta DM); Saudi Ministry of Health, Riyadh, Saudi Arabia (Prof Z A Memish MD); College of Medicine, Alfaisal University, Riyadh, Saudi Arabia (Prof Z A Memish); United Nations Population Fund, Lima, Peru (W Mendoza MD); Center for Translation Research and Implementation Science, National Heart, Lung, and Blood Institute, National Institutes of Health, Bethesda, MD, USA (G A Mensah MD); Department of Medicine (A Meretoja PhD), Institute of Health and Ageing (Prof C E I Szoeke PhD), University of Melbourne, Footscray, VIC, Australia (Prof T Wijeratne MD); Department of Neurology, Helsinki University Hospital, Helsinki, Finland (A Meretoja PhD); Pacific Institute for Research & Evaluation, Calverton, MD, USA (T R Miller PhD); Centre for Population Health, Curtin University, Perth, WA, Australia (T R Miller); Nepal Development Society, Chitwan, Nepal (S R Mishra MPH); Department of Medicine, Universiti Kebangsaan Malaysia Medical Center, Bandar Tun Razak, Malaysia (Prof N Mohamed Ibrahim MRCP); Neuroscience Research Center, Baqiyatallah University of Medical Sciences, Tehran, Iran (A Mohammadi PhD); Health Systems and Policy Research Unit, Ahmadu Bello University, Zaria, Nigeria (S Mohammed PhD); Department of Community Medicine, Preventive Medicine and Public Health Research Center, Gastrointestinal and Liver Disease Research Center (GILDRC), Iran University of Medical Sciences, Tehran, Iran (M Moradi-Lakeh MD); Gorgas Memorial Institute for Health Studies, Panama City, Panama (I Moreno Velasquez PhD); School of Medical Sciences, University of Science Malaysia, Kubang Kerian, Malaysia (K I Musa MD); University of Nairobi, Nairobi, Kenya (J W Ngunjiri PhD); Department of Public Health, Semarang State University, Semarang City, Indonesia (D N A Ningrum MPH); Gr

lth Studies, Panama City, Panama (I Moreno Velasquez PhD); School of Medical Sciences, University of Science Malaysia, Kubang Kerian, Malaysia (K I Musa MD); University of Nairobi, Nairobi, Kenya (J W Ngunjiri PhD); Department of Public Health, Semarang State University, Semarang City, Indonesia (D N A Ningrum MPH); Gr aduate Institute of Biomedical Informatics, College of Medical Science and Technology, Taipei Medical University, Taipei City, Taiwan (D N A Ningrum); Skane University Hospital, Department of Clinical Sciences Lund, Neurology, Lund University, Lund, Sweden (Prof B Norrving PhD); Department of Psychiatry (Prof D J Stein PhD), University of Cape Town, Cape Town, South Africa (J J N Noubiap MD); Medical Diagnostic Centre, Yaounde, Cameroon (J J N Noubiap); Centre for Health Research (F A Ogbo MPH), Western Sydney University, Penrith, NSW, Australia (Prof A M N Renzaho PhD); Department of Medicine, Ibadan, Nigeria (M O Owolabi Dr Med); Blossom Specialist Medical Center, Ibadan, Nigeria (M O Owolabi); Christian Medical College Ludhiana, Ludhiana, India (J D Pandian DM); Auckland University of Technology, Auckland, New Zealand (P G Parmar MSc); REQUIMTE/LAQV, LaboratÓrio de Farmacognosia, Departamento de Química, Faculdade de Farmácia, Universidade do Porto, Porto, Portugal (Prof D M Pereira PhD); Health Metrics Unit, University of Gothenburg, Gothenburg, Sweden (Prof M Petzold PhD); University of the Witwatersrand, Johannesburg, South Africa (Prof M Petzold); Shanghai Jiao Tong University School of Medicine, Shanghai, China (Prof M R Phillips MD); Emory University, Atlanta, GA, USA (Prof M R Phillips); University of Otago, Dunedin, New Zealand (Prof R G Poulton PhD); University of British Columbia, Vancouver, BC, Canada (F Pourmalek PhD); Non-Communicable Diseases Research Center, Alborz University of Medical Sciences, Karaj, Iran (M Qorbani PhD); Contech International Health Consultants, Lahore, Pakistan (A Rafay MS); Contech School of Public Health, Lahore, Pakistan (A Rafay); Research and Evaluation Division, BRAC, Dhaka, Bangladesh (M Rahman PhD); Society for Health and Demographic Surveillance, Suri, India (R K Rai MPH); ERAWEB Program, University for Health Sciences, Medical Informatics and Technology, Hall in Tirol, Austria (S Rajsic MD); Capital & Coast District Health Board, Wellington, New Zealand (A Ranta PhD); University of Otago, Wellington, New Zealand (A Ranta); Mazandaran University of Medical Sciences, Sari, Iran (M S Rezai MD); Golestan Research Cen

for Health Sciences, Medical Informatics and Technology, Hall in Tirol, Austria (S Rajsic MD); Capital & Coast District Health Board, Wellington, New Zealand (A Ranta PhD); University of Otago, Wellington, New Zealand (A Ranta); Mazandaran University of Medical Sciences, Sari, Iran (M S Rezai MD); Golestan Research Cen ter of Gastroenterology and Hepatology, Golestan University of Medical Sciences, Gorgan, Iran (G Roshandel PhD); Universidad Tecnica del Norte, Ibarra, Ecuador (E Rubagotti PhD); University of New South Wales, Randwick, NSW, Australia (Prof P Sachdev MD); Prince of Wales Hospital, Randwick, NSW, Australia (Prof P Sachdev); Managerial Epidemiology Research Center, Department of Public Health, School of Nursing and Midwifery, Maragheh University of Medical Sciences, Maragheh, Iran (S Safiri PhD); Universiti Kebangsaan Malaysia Medical Centre, Kuala Lumpur, Malaysia (R Sahathevan PhD); Ballarat Health Service, Ballarat, VIC, Australia (R Sahathevan); Faculty of Science, Ain Shams University, Cairo, Egypt (A M Samy PhD); Fondazione IRCCS Ospedale Maggiore Policlinico, Milan, Italy (P Santalucia MD); Public Health Medicine, School of Nursing and Public Health, University of KwaZulu-Natal, Durban, South Africa (Prof B Sartorius PhD); UKZN Gastrointestinal Cancer Research Centre, South African Medical Research Council, Durban, South Africa (Prof B Sartorius); Centre of Advanced Study in Psychology, Utkal University, Bhubaneswar, India (M Satpathy PhD); Department of Public Health, Marshall University, Huntington, WV, USA (M Sawhney PhD); Bayer Turkey, Istanbul, Turkey (M I Saylan PhD); Independent Consultant, Karachi, Pakistan (M A Shaikh MD); Department of Medical Surgical Nursing, School of Nursing and Midwifery, Hamadan University of Medical Sciences, Hamadan, Iran (M Shamsizadeh MPH); School of Medicine, Yale University, New Haven, CT, USA (K N Sheth MD); National Institute of Infectious Diseases, Tokyo, Japan (M Shigematsu PhD); Sandia National Laboratories, Albuquerque, NM, USA (M Shigematsu); Federal University of Santa Catarina, Florianopolis, Brazil (D A S Silva PhD); University of Yaoundé, Yaoundé, Cameroon (Prof E Sobngwi PhD); Yaoundé Central Hospital, Yaoundé, Cameroon (Prof E Sobngwi); Department of Clinical Neurological Sciences, Western University, London, ON, Canada (L A Sposato MD); South African Medical Research Council Unit on Anxiety & Stress Disorders, Cape Town, South Africa (Prof D J Stein PhD); Department of Neuroscience, Norwegian University o

Hospital, Yaoundé, Cameroon (Prof E Sobngwi); Department of Clinical Neurological Sciences, Western University, London, ON, Canada (L A Sposato MD); South African Medical Research Council Unit on Anxiety & Stress Disorders, Cape Town, South Africa (Prof D J Stein PhD); Department of Neuroscience, Norwegian University o f Science and Technology, Trondheim, Norway (Prof T J Steiner PhD, Prof L J Stovner PhD); Norwegian Advisory Unit on Headache, St.

Hospital, Yaoundé, Cameroon (Prof E Sobngwi); Department of Clinical Neurological Sciences, Western University, London, ON, Canada (L A Sposato MD); South African Medical Research Council Unit on Anxiety & Stress Disorders, Cape Town, South Africa (Prof D J Stein PhD); Department of Neuroscience, Norwegian University o f Science and Technology, Trondheim, Norway (Prof T J Steiner PhD, Prof L J Stovner PhD); Norwegian Advisory Unit on Headache, St. Olavs Hospital, Trondheim, Norway (Prof L J Stovner); Ministry of Health, Kingdom of Saudi Arabia, Riyadh, Saudi Arabia (R Suliankatchi Abdulkader MD); Chaim Sheba Medical Center, Tel Hashomer, Israel (Prof D Tanne MD); Tel Aviv University, Tel Aviv, Israel (Prof D Tanne); Department of Medicine, School of Clinical Sciences at Monash Health, Monash University, Melbourne, VIC, Australia (Prof A G Thrift PhD); Institute of Public Health, Faculty of Health Sciences, Jagiellonian University Medical College, Kraków, Poland (R Topor-Madry PhD); Faculty of Health Sciences, Wroclaw Medical University, Wroclaw, Poland (R Topor-Madry); Hanoi Medical University, Hanoi, Vietnam (B X Tran PhD); Department of Neurology, Rigshospitalet, University of Copenhagen, Copenhagen, Denmark (T Truelsen DMSc); Department of Internal Medicine, Federal Teaching Hospital, Abakaliki, Nigeria (K N Ukwaja MD); Warwick Medical School, University of Warwick, Coventry, UK (O A Uthman PhD); UKK Institute for Health Promotion Research, Tampere, Finland (Prof T Vasankari PhD); Raffles Neuroscience Centre, Raffles Hospital, Singapore, Singapore (N Venketasubramanian MBBS); National Research University Higher School of Economics, Moscow, Russia (Prof V V Vlassov MD); Wolaita Sodo University, Wolaita Sodo, Ethiopia (F Wadilo MS); VA Medical Center, Washington, DC, USA (M T Wallin MD); Neurology Department, Georgetown University, Washington, DC, USA (M T Wallin); Department of Research, Cancer Registry of Norway, Institute of Population-Based Cancer Research, Oslo, Norway (E Weiderpass PhD); Department of Community Medicine, Faculty of Health Sciences, University of Tromsø, The Arctic University of Norway, Tromsø, Norway (E Weiderpass); Genetic Epidemiology Group, Folkhälsan Research Center, Helsinki, Finland (E Weiderpass); Federal Institute for Population Research, Wiesbaden, Germany (R Westerman PhD); German National Cohort Consortium, Heidelberg, Germany (R Westerman); Western Health, Footscray, VIC, Australia (Prof T Wijeratne MD); South African Medical Research Council, Cochrane South Africa, Cape Town, South Africa (Prof C S Wiysonge PhD); Stellenbosch University, Cape Town, South Africa (Prof C

Westerman PhD); German National Cohort Consortium, Heidelberg, Germany (R Westerman); Western Health, Footscray, VIC, Australia (Prof T Wijeratne MD); South African Medical Research Council, Cochrane South Africa, Cape Town, South Africa (Prof C S Wiysonge PhD); Stellenbosch University, Cape Town, South Africa (Prof C S Wiysonge); Division of Health and Social Care Research, King's College London, London, UK (Prof C D Wolfe MD); National Institute for Health Research Comprehensive Biomedical Research Centre, Guy's & St Thomas' NHS Foundation Trust and King's College London, London, UK (Prof C D Wolfe); St. John's Medical College and Research Institute, Bangalore, India (Prof D Xavier MD); Department of Neurology, Jinling Hospital, Nanjing University School of Medicine, Nanjing, China (Prof G Xu PhD); Department of Preventive Medicine, Northwestern University, Chicago, IL, USA (Y Yano MD); Mizan Tepi University, Mizan Teferi, Ethiopia (H H Yimam MPH); Department of Biostatistics, School of Public Health, Kyoto University, Kyoto, Japan (N Yonemoto MPH); Department of Epidemiology and Biostatistics, School of Public Health (Prof C Yu PhD), Global Health Institute (Prof C Yu), Wuhan University, Wuhan, China; University Hospital of Setif, Setif, Algeria (Prof Z Zaidi DSc); and Faculty of Medicine, Mansoura University, Mansoura, Egypt (Prof M E Zaki PhD).

N Yonemoto MPH); Department of Epidemiology and Biostatistics, School of Public Health (Prof C Yu PhD), Global Health Institute (Prof C Yu), Wuhan University, Wuhan, China; University Hospital of Setif, Setif, Algeria (Prof Z Zaidi DSc); and Faculty of Medicine, Mansoura University, Mansoura, Egypt (Prof M E Zaki PhD). Contributors VLF prepared the first draft. TV analysed the data and edited the first draft and final versions of the manuscript. VLF and TV reviewed all drafts, finalised all drafts, and approved the final version of the manuscript. All other authors provided data, developed models, reviewed results, provided guidance on methodology, or reviewed the manuscript, and approved the final version of the manuscript. Declaration of interests JÄ reports personal fees from AstraZeneca. PJ reports a clinical and public health intermediate fellowship from the Wellcome Trust and Department of Biotechnology, India Alliance. MIS reports other funding from Bayer AG and is an employee of Bayer Pharmaceuticals. CEIS reports grants from the National Medical Health Research Council during the conduct of the study, grants from Lundbeck and Alzheimer's Association, and has a patent issued with the Melbourne Health and Murdoch Children's Research Institute (PCT/AU2008/001556). All other authors declare no competing interests.

Introduction The course of Alzheimer's disease is characterised by progressive accumulation of neuropathology over decades. The initial asymptomatic phase continues into a prodromal phase with mild, but noticeable, cognitive and functional impairment,1 and eventually progression to dementia. This gradual progression creates a window of opportunity for interventions in early disease stages.2 Specific criteria to define the prodromal phase of Alzheimer's disease have been proposed using biomarkers and clinical criteria,3, 4, 5 but no pharmacological treatment is currently available for individuals with prodromal Alzheimer's disease. Development of safe and effective interventions in early Alzheimer's disease stages remains imperative. Prevention trials from the past 2 years have shown promising results with multimodal, non-pharmacological approaches, including dietary interventions.6, 7

available for individuals with prodromal Alzheimer's disease. Development of safe and effective interventions in early Alzheimer's disease stages remains imperative. Prevention trials from the past 2 years have shown promising results with multimodal, non-pharmacological approaches, including dietary interventions.6, 7 Diet is an important modifiable risk factor for dementia,8 and a nutrient intervention in mild cognitive impairment showed effects on brain atrophy.9 LipiDiDiet is a research consortium, which studies the preclinical and clinical impact of nutrition in Alzheimer's disease. This research resulted in experimental dietary interventions, which contributed to the development of the medical food Souvenaid (Nutricia; Zoetermeer, the Netherlands). The active component of Souvenaid is the multinutrient combination (Fortasyn Connect), which contains docosahexaenoic acid (DHA); eicosapentaenoic acid (EPA); uridine monophosphate; choline; vitamins B12, B6, C, E, and folic acid; phospholipids; and selenium.10 These nutrients were selected based on their established biological and neuroprotective properties, and specifically combined to enhance efficacy compared with individual nutrients. The aim was to provide neuroprotection by targeting disease processes in early Alzheimer's disease—ie, by supplying rate-limiting compounds for brain phospholipid synthesis and addressing multiple Alzheimer's disease-related pathological processes in vivo.11, 12, 13, 14, 15, 16, 17 Results from animal studies showed that this multinutrient combination improved neuronal membrane composition; increased the formation of synapses, cholinergic neurotransmission, and cerebral blood flow and perfusion; preserved neuronal integrity; restored hippocampal neurogenesis; reduced β-amyloid pathology; and improved cognition.15, 16, 17, 18, 19, 20, 21 Concentrations of these nutrients in plasma and CSF or the brain were also found to be lower in patients with Alzheimer's disease.22 For clinical use, Fortasyn Connect was adapted to address nutritional requirements in the presence of Alzheimer's disease pathology.

ogy; and improved cognition.15, 16, 17, 18, 19, 20, 21 Concentrations of these nutrients in plasma and CSF or the brain were also found to be lower in patients with Alzheimer's disease.22 For clinical use, Fortasyn Connect was adapted to address nutritional requirements in the presence of Alzheimer's disease pathology. Two previous randomised clinical trials in patients with mild Alzheimer's disease dementia reported that daily intake of Fortasyn Connect for 3 or 6 months improved memory performance,23, 24 increased neurophysiological measures of synaptic activity, and enhanced functional connectivity in the brain.24, 25 A third randomised controlled trial26 in patients with mild-to-moderate Alzheimer's disease dementia did not report benefits, therefore, heterogeneity in the benefits of Fortasyn Connect exists in previous trials. All trials reported a positive safety profile23, 24, 27 and treatment was well tolerated in combination with Alzheimer's disease medications.26 An analysis of these trials indicated that Fortasyn Connect can achieve clinically detectable effects in patients with mild Alzheimer's disease dementia,28 but did not slow cognitive decline in mild-to-moderate Alzheimer's disease dementia.26 Given the hypothesis that earlier intervention might be more beneficial, the LipiDiDiet trial was designed to investigate the effects of Fortasyn Connect on cognition and related measures in prodromal Alzheimer's disease. Research in context Evidence before this study

Two previous randomised clinical trials in patients with mild Alzheimer's disease dementia reported that daily intake of Fortasyn Connect for 3 or 6 months improved memory performance,23, 24 increased neurophysiological measures of synaptic activity, and enhanced functional connectivity in the brain.24, 25 A third randomised controlled trial26 in patients with mild-to-moderate Alzheimer's disease dementia did not report benefits, therefore, heterogeneity in the benefits of Fortasyn Connect exists in previous trials. All trials reported a positive safety profile23, 24, 27 and treatment was well tolerated in combination with Alzheimer's disease medications.26 An analysis of these trials indicated that Fortasyn Connect can achieve clinically detectable effects in patients with mild Alzheimer's disease dementia,28 but did not slow cognitive decline in mild-to-moderate Alzheimer's disease dementia.26 Given the hypothesis that earlier intervention might be more beneficial, the LipiDiDiet trial was designed to investigate the effects of Fortasyn Connect on cognition and related measures in prodromal Alzheimer's disease. Research in context Evidence before this study We searched ClinicalTrials.gov, WHO's International Clinical Trial Registry Platform, and PubMed (Jan 1, 1950, to Dec 20, 2016) using the search terms “Alzheimer's disease” and “Fortasyn” or “Souvenaid”. There were no language restrictions. Only articles reporting clinical trials of Souvenaid or Fortasyn Connect in patients with Alzheimer's disease were included. We identified three completed 12-week to 24-week randomised controlled trials: Souvenir I (225 drug-naive patients with mild Alzheimer's disease dementia), Souvenir II (259 drug-naive patients with mild Alzheimer's disease dementia), and S-Connect (527 patients with mild-to-moderate Alzheimer's disease dementia treated with medications). Improved memory was reported in mild, but not mild-to-moderate, Alzheimer's disease. The three randomised controlled trials reported that the intervention was well tolerated and had a good safety profile, both alone and in combination with Alzheimer's disease medications. The LipiDiDiet trial differs from the previous trials of this multinutrient combination because it focuses on prodromal Alzheimer's disease and tests a longer duration of the intervention.

intervention was well tolerated and had a good safety profile, both alone and in combination with Alzheimer's disease medications. The LipiDiDiet trial differs from the previous trials of this multinutrient combination because it focuses on prodromal Alzheimer's disease and tests a longer duration of the intervention. Added value of this study LipiDiDiet is the first completed long-term randomised controlled trial focusing on prodromal Alzheimer's disease defined according to the International Working Group (IWG-1) criteria. Benefit was observed in relevant secondary cognitive-functional and brain atrophy outcome measures, but not in the primary neuropsychological test battery and other secondary measures including dementia diagnosis. Our findings support the hypothesis that intervening early in the disease continuum might achieve benefits more readily than late intervention. Implications of all the available evidence Our results emphasise the difficulty in finding adequately sensitive outcome measures for trials in prodromal Alzheimer's disease. The potential for impact on disease progression, combined with the feasibility aspects including the observed high long-term compliance, moderate costs of the intervention, the potentially relative ease of implementation in clinical practise, as well as the clear need for treatment, warrant further research on multinutrient intervention in early Alzheimer's disease.

combined with the feasibility aspects including the observed high long-term compliance, moderate costs of the intervention, the potentially relative ease of implementation in clinical practise, as well as the clear need for treatment, warrant further research on multinutrient intervention in early Alzheimer's disease. Methods Study design and participants The LipiDiDiet study was a 24-month randomised, controlled, double-blind, parallel-group, multicentre trial done in 11 study sites in Finland, Germany, the Netherlands, and Sweden (appendix) with one to four optional, 12-month, double-blind extension periods. Participants were primarily recruited from memory clinics and had routine assessments in the year before screening. The study was completed as planned. Here we report 24-month findings; extension studies are currently ongoing and will be reported later. We enrolled participants aged 55–85 years with a Mini-Mental State Examination (MMSE) score of 24 points or higher (≥20 if education level ≤6 years) who fulfilled criteria for prodromal Alzheimer's disease3 as defined by episodic memory disorder (performance below one standard deviation on two of eight cognitive tests [at least one on memory]) and evidence for underlying Alzheimer's disease pathology based on positive findings from at least one of the following diagnostic tests: CSF, MRI, and 18F fluorodeoxyglucose (18F-FDG) PET analysis (full list of inclusion criteria is in the appendix). We excluded participants with dementia according to Diagnostic and Statistical Manual of Mental Disorders, 4th Edition (DSM-IV); historical use of donepezil, rivastigmine, galantamine, or memantine, use of omega-3 preparations, alcohol or drug abuse, major depressive disorders (DSM-IV) or other concomitant serious conditions, intake of vitamins B6, B12, folic acid, vitamin C, or vitamin E of more than 200% of the recommended daily intake, those who participated in any other clinical trial in the last 30 days, and with an MRI or CT scan consistent with a diagnosis of stroke, intracranial bleeding, mass lesion, or normal pressure hydrocephalus (minimal white matter changes and up to two lacunar infarcts judged to be clinically insignificant were allowed). Participants who progressed to dementia during the trial could remain in the trial and start approved Alzheimer's disease medication, according to the clinician's judgment.

or normal pressure hydrocephalus (minimal white matter changes and up to two lacunar infarcts judged to be clinically insignificant were allowed). Participants who progressed to dementia during the trial could remain in the trial and start approved Alzheimer's disease medication, according to the clinician's judgment. The protocol was amended to allow participants who progressed to dementia to switch to the active product after it became generally available (appendix). The study protocol and consent forms were approved by the local ethical committees of all participating sites, and all participants provided written informed consent before study participation. The study was done in accordance with the Declaration of Helsinki and International Conference on Harmonization Good Clinical Practice guidelines.

nd consent forms were approved by the local ethical committees of all participating sites, and all participants provided written informed consent before study participation. The study was done in accordance with the Declaration of Helsinki and International Conference on Harmonization Good Clinical Practice guidelines. Randomisation and masking Eligible participants were randomly assigned (1:1) to receive either the active or control product once daily according to a randomisation list, which was computer generated by Nutricia Research, stratified by site, and in block sizes of four. Sealed opaque envelopes were available for each participant. After acceptance of a participant to the trial, the envelope with the lowest unused number was opened at the site, containing the code for that participant. The active and control products were isocaloric and similar in appearance and flavours (vanilla and strawberry). All study personnel and participants, including the investigators and study-site staff, were masked to treatment assignment. Only the trial-independent statistician and the independent data monitoring committee, who reviewed interim data for safety and efficacy purposes, were partially unmasked.

anilla and strawberry). All study personnel and participants, including the investigators and study-site staff, were masked to treatment assignment. Only the trial-independent statistician and the independent data monitoring committee, who reviewed interim data for safety and efficacy purposes, were partially unmasked. Procedures We enrolled eligible participants at a combined screening and baseline visit or during a separate baseline visit. Efficacy evaluations were done at baseline, 6, 12, and 24 months (appendix). Study sites received training on outcome assessments. Visits to the study nurse or physician were scheduled every 3 months during the first year and every 6 months thereafter. To maintain motivation, check compliance, and monitor safety, participants were contacted by phone throughout the trial (once per month during the first 6 months and every 2 months thereafter). Study products were dispensed to the participants every 3 months. Participants in the active group were given the medical food Souvenaid, a 125 mL once-a-day drink containing the specific nutrient combination Fortasyn Connect (appendix). Participants in the control group were given a 125 mL once-a-day control drink. The study product was produced by Nutricia (Zoetermeer, the Netherlands).

cipants in the active group were given the medical food Souvenaid, a 125 mL once-a-day drink containing the specific nutrient combination Fortasyn Connect (appendix). Participants in the control group were given a 125 mL once-a-day control drink. The study product was produced by Nutricia (Zoetermeer, the Netherlands). Outcomes The primary efficacy endpoint was the change over 24 months in a composite score of cognitive performance based on a neuropsychological test battery (NTB; appendix),29 assessed by study neuropsychologists at each site at baseline and months 6, 12, and 24. Based on advances in Alzheimer's disease research and results from a clinical trial with the active product,24 protocol amendments were made after the study started and before database lock to specify the composite scores of the NTB and to limit the number of secondary endpoints (appendix). The NTB primary endpoint was a composite Z score based on Consortium to Establish a Registry for Alzheimer's disease (CERAD) 10-word list learning immediate recall, CERAD 10-word delayed recall, CERAD 10-word recognition, category fluency, and letter digit substitution test (LDST). Secondary endpoints were NTB memory domain (composite Z score based on CERAD 10-word list learning immediate recall, delayed recall, and recognition), NTB executive function domain (composite Z score based on category fluency, Wechsler Memory Scale revised digit span total score, concept shifting test condition C [corrected for the zero trials], and LDST), and NTB total (composite Z score based on all 16 items of the NTB). Composite scores were calculated as Z scores standardised to the baseline mean and SD, with higher scores suggesting better performance.

Scale revised digit span total score, concept shifting test condition C [corrected for the zero trials], and LDST), and NTB total (composite Z score based on all 16 items of the NTB). Composite scores were calculated as Z scores standardised to the baseline mean and SD, with higher scores suggesting better performance. Other secondary endpoints, assessed at baseline, month 12, and month 24, unless stated otherwise, were change from baseline over 24 months in clinical dementia rating-sum of boxes (CDR-SB), brain volumes based on MRI (three-dimensional T1-weighted anatomical scans of total hippocampal, whole-brain, and ventricular volumes; details of MRI acquisition and central analysis are in the appendix), progression to dementia (according to criteria defined by DSM-IV, the National Institute of Neurological and Communicative Disorders and Stroke, and the Alzheimer's Disease and Related Disorders Association criteria for Alzheimer's disease), serum concentrations of HDL cholesterol and LDL cholesterol, plasma fatty acids (DHA and EPA, assessed at baseline and months 3, 6, 12, and 24; for laboratory analysis, see appendix), and DHA in CSF (CSF analysis not yet finalised). Safety assessments included adverse events, use of concomitant medications, consumption of nutritional supplements, study product compliance, vital signs (heart rate, systolic blood pressure, and diastolic blood pressure), and clinical safety laboratory tests. To monitor product compliance, we asked participants to record the amount of study product taken in a daily diary, which was collected at each visit. Study product compliance was defined as the percentage of study product used throughout the study period compared with the prescribed dosage. Compliance was calculated only for participants who completed the study product diary for at least 75% of their actual study time. An additional sensitivity calculation was done to include all available data up to the start of rescue medication (defined as use of active product or approved Alzheimer's disease medication after progression to dementia). In both calculations, missing diary intake entries were assumed to be 0. We coded adverse events with the Medical Dictionary for Regulatory Activities (version 18.0).

le data up to the start of rescue medication (defined as use of active product or approved Alzheimer's disease medication after progression to dementia). In both calculations, missing diary intake entries were assumed to be 0. We coded adverse events with the Medical Dictionary for Regulatory Activities (version 18.0). Statistical analyses Based on a t test and 5% significance level, we calculated that a sample size of 300 randomly assigned participants would be sufficient to provide 90% power to detect a 40% difference in NTB score change between groups at the end of the study. Based on results from a study in patients with mild Alzheimer's disease dementia,29 we expected the NTB Z score in the control group to decrease by −0·4 (SD 0·4) during 24 months. The sample size allowed for 20% dropout. We did a prespecified, blinded re-estimation of the SD to assess the adequacy of the calculated sample size. Representatives of the LipiDiDiet Consortium reviewed the SDs calculated from the interim dataset and concluded that they matched the estimated SDs in the protocol. Additionally, we amended the protocol to do an interim analysis for safety (occurrence of adverse events) and efficacy after approximately a third of participants completed the study. Between-group analyses on partially unmasked data were done by the trial-independent statistician and results were reviewed by the independent data monitoring committee, which recommended continuation of the study without modification.

vents) and efficacy after approximately a third of participants completed the study. Between-group analyses on partially unmasked data were done by the trial-independent statistician and results were reviewed by the independent data monitoring committee, which recommended continuation of the study without modification. We obtained NTB composite Z scores by averaging the individual NTB items' Z scores and weighting according to the number of NTB items available. The minimum number of NTB items required was set to four of five for NTB primary endpoint, three of three for NTB memory domain, three of four for NTB executive function domain, and 12 of 16 for NTB total. Analyses were done on the modified intention-to-treat (mITT) population of all participants randomly assigned, excluding data after the start of rescue medication. We did per-protocol analyses on all participants from the mITT population, excluding the respective visits of participants with major protocol deviations defined during a data review of masked data. The most common reason for exclusion from the per-protocol analysis was substantial irregular study product intake (appendix). All randomised participants who consumed at least one dose of study product were included in safety analyses. To allow for separate evaluation of safety data collected before and after a switch to active study product after progression to dementia, analyses were done in two safety phases: the double-blind treatment phase and active treatment phase.

o consumed at least one dose of study product were included in safety analyses. To allow for separate evaluation of safety data collected before and after a switch to active study product after progression to dementia, analyses were done in two safety phases: the double-blind treatment phase and active treatment phase. We analysed the primary endpoint and all secondary endpoints of a continuous type as prespecified in the statistical analysis plan, using a linear mixed model for longitudinal data with change from baseline as the response variable and linear time (days since baseline), baseline score, randomised treatment, and time × treatment as fixed effects. This is a multi-level model with three levels: measurements, participants, and sites. We used a random intercept with a variance components covariance structure within sites (small sites were pooled within country) and a random intercept and slope for time with an unstructured covariance structure within participants. Other covariance structures could be applied in case of converging issues. This model's estimated difference between the active and control groups in terms of mean change from baseline at month 24 was used as the primary indication of treatment effect during the 24-month intervention period. A baseline measurement or characteristic that showed an imbalance between the groups and was found to be a significant predictor for outcome parameters was considered a prognostic factor and included in the statistical models. We did a planned sensitivity analysis using a mixed model for repeated measures with change from baseline as the response variable and time as categorical (planned visit), baseline score, randomised treatment, and time × treatment as fixed effects. In this analysis, treatment effects were evaluated for each timepoint separately without assuming a treatment effect that increases linearly over time. We did an additional sensitivity analysis on participants who completed the 24-month study by using an analysis of covariance completer analysis with change from baseline as outcome, treatment as fixed factor, and baseline score as a covariate. We did an additional sensitivity analysis taking into account missing data due to dropout using a joint model. The joint model combined a mixed model comparable to our original model (mixed model) with a Cox proportional hazards model for time to dropout (appendix).

t as fixed factor, and baseline score as a covariate. We did an additional sensitivity analysis taking into account missing data due to dropout using a joint model. The joint model combined a mixed model comparable to our original model (mixed model) with a Cox proportional hazards model for time to dropout (appendix). We did a predefined subgroup analysis in participants with MMSE score 26 or higher at baseline using the same statistical models as for the primary and secondary endpoints. We aligned rules for visit windows between the statistical models to allow a visit window of 3 months (before and after scheduled visit date) for all visits. p values of less than 0·05 were deemed statistically significant in comparisons of efficacy and safety data. Statistical analyses were done with SAS software (version 9.4). The study is registered with the Dutch Trial Register, number NTR1705. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. HS, AS, PJV, SBH, KB, MK, and TH had full access to all the data in the study. The corresponding author had final responsibility for the decision to submit for publication.

funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. HS, AS, PJV, SBH, KB, MK, and TH had full access to all the data in the study. The corresponding author had final responsibility for the decision to submit for publication. Results Between April 20, 2009, and July 3, 2013, 311 participants of 382 screened were randomly assigned to either the active group (n=153) or the control group (n=158; figure 1). Dropout was 22% in the active group and 21% in the control group (p=0·891, Fisher's exact test), with no significant difference in time to dropout between groups. The main reasons for dropout were adverse events (n=15; appendix), withdrawal of informed consent (n=18), or other reasons (n=28). Five participants died during the study: four in the active group due to respiratory failure (n=2), bronchial carcinoma (n=1), and infection (n=1); and one in the control group due to sudden death with no apparent cause. All deaths were assessed as not related to the study product. Mean age was 71·0 years, and 154 (50%) of 311 participants were men. Further baseline characteristics of study participants are shown in table 1. Results and parameters used in assessment of eligibility for prodromal Alzheimer's disease at screening are summarised in the appendix, including classification according to the International Working Group (IWG)-1, IWG-2, and National Institute of Aging-Alzheimer Association (NIA-AA) criteria.3, 4, 5 The active and control groups were similar at baseline (table 1), except for MMSE score. Baseline MMSE was also found to be a significant predictor of outcome parameters, which made it a potential prognostic factor; therefore, it was included as a covariate in all statistical models except MMSE subgroup analyses.Figure 1 Trial profile

ontrol groups were similar at baseline (table 1), except for MMSE score. Baseline MMSE was also found to be a significant predictor of outcome parameters, which made it a potential prognostic factor; therefore, it was included as a covariate in all statistical models except MMSE subgroup analyses.Figure 1 Trial profile *Rescue medication was defined as the use of active product or approved Alzheimer's disease medication after progression to dementia. †All randomly assigned participants, excluding visit data after the start of rescue medication. ‡Respective visits of participants were additionally excluded in cases of major protocol deviations; number based on participants with at least one follow-up visit in the per-protocol dataset. Table 1 Baseline characteristics

*Rescue medication was defined as the use of active product or approved Alzheimer's disease medication after progression to dementia. †All randomly assigned participants, excluding visit data after the start of rescue medication. ‡Respective visits of participants were additionally excluded in cases of major protocol deviations; number based on participants with at least one follow-up visit in the per-protocol dataset. Table 1 Baseline characteristics Control (n=158) Active (n=153) Age (years) Mean (SD) 70·7 (6·2) 71·3 (7·0) Median (range) 71 (52–84) 72 (50–86) Sex Men 73 (46%) 81 (53%) Women 85 (54%) 72 (47%) Ethnic origin White 157 (99%) 152 (99%) Black 0 (0) 1 (1%) Other 1 (1%) 0 (0) Education (years) 10·7 (3·6) 10·6 (3·9) Mini-Mental State Examination 26·9 (1·9) 26·4 (2·1) APOE ɛ4 genotype* Carrier 90/143 (63%) 83/138 (60%) Non-carrier 53/143 (37%) 55/138 (40%) Cognitive measures (composite Z score) NTB primary endpoint 0·00 (0·68) −0·00 (0·70) NTB memory domain 0·03 (0·82) −0·02 (0·87) NTB executive function domain −0·01 (0·71) 0·01 (0·71) NTB total −0·02 (0·56) 0·02 (0·57) CDR-SB 1·75 (1·14) 1·87 (1·17) MRI brain volumes (cm3)† Total hippocampal volume 5·70 (1·25) 5·62 (1·10) Whole brain volume 1377·30 (84·08) 1370·56 (81·64) Ventricular volume 53·95 (25·31) 58·35 (26·66) CSF† Aβ42 (pg/mL) 401·1 (196·1) 426·9 (292·7) (Aβ42/Aβ40) × 10 0·62 (0·25) 0·65 (0·29) Total tau (pg/mL) 634·8 (287·7) 591·9 (260·9) Phosphorylated tau (pg/mL) 80·3 (30·6) 74·2 (25·8) Data are mean (SD), n (%), n/N (%), or median (range). NTB=neuropsychological test battery. CDR-SB=clinical dementia rating sum of boxes. Aβ=amyloid β.

L) 401·1 (196·1) 426·9 (292·7) (Aβ42/Aβ40) × 10 0·62 (0·25) 0·65 (0·29) Total tau (pg/mL) 634·8 (287·7) 591·9 (260·9) Phosphorylated tau (pg/mL) 80·3 (30·6) 74·2 (25·8) Data are mean (SD), n (%), n/N (%), or median (range). NTB=neuropsychological test battery. CDR-SB=clinical dementia rating sum of boxes. Aβ=amyloid β. * Data not available for all randomised participants. Percentages are calculated based on number of participants with available data. † Central analysis CSF data available for n=107 and MRI data for n=279 (appendix). Primary and main secondary endpoints are reported in table 2, in which higher scores indicate better performance for all endpoints except for CDR-SB and ventricular volume. Mean change from baseline to month 24 in the NTB primary endpoint was −0·108 (SD 0·528) in the control group, and −0·028 (SD 0·453) in the active group. The decline in the control group was lower than the prestudy estimate of −0·4 during 24 months. There was no statistically significant difference between groups for the primary endpoint (estimated mean treatment difference of 0·098, 95% CI −0·041 to 0·237; p=0·166). Similarly, there were no statistically significant differences when the analyses were done without adjustment for baseline MMSE (appendix) and in the sensitivity analysis (table 2).Table 2 Primary endpoint and main secondary endpoints

oint (estimated mean treatment difference of 0·098, 95% CI −0·041 to 0·237; p=0·166). Similarly, there were no statistically significant differences when the analyses were done without adjustment for baseline MMSE (appendix) and in the sensitivity analysis (table 2).Table 2 Primary endpoint and main secondary endpoints Control (n=158) Active (n=153) Difference Mixed model*, p value Sensitivity analysis†p value Effect size‡Cohen's d Mean (SD)§ n Mean (SD)§ n Estimate (95% CI)¶ Primary endpoint NTB primary endpoint (Z score) Modified intention-to-treat −0·108 (0·528) 141 −0·028 (0·453) 134 0·098 (−0·041 to 0·237) 0·166 0·214 0·17 Per-protocol −0·122 (0·570) 123 0·045 (0·414) 116 0·140 (−0·017 to 0·296) 0·080 0·043 0·24 Secondary endpoints NTB memory domain (Z score) Modified intention-to-treat −0·130 (0·619) 140 0·003 (0·569) 134 0·138 (−0·027 to 0·303) 0·101 0·112 0·17 Per-protocol −0·151 (0·663) 122 0·083 (0·532) 116 0·181 (−0·005 to 0·367) 0·057 0·026 0·25 NTB executive function domain (Z score) Modified intention-to-treat −0·039 (0·506) 141 −0·145 (0·445) 133 −0·043 (−0·180 to 0·095) 0·541 0·281 −0·08 Per-protocol −0·045 (0·546) 123 −0·090 (0·381) 115 0·009 (−0·137 to 0·155) 0·906 0·854 0·01 NTB total (Z score) Modified intention-to-treat −0·059 (0·400) 140 −0·047 (0·347) 134 0·027 (−0·078 to 0·132) 0·612 0·729 0·07 Per-protocol −0·061 (0·419) 122 −0·006 (0·317) 116 0·058 (−0·056 to 0·172) 0·316 0·352 0·15 CDR-SB|| Modified intention-to-treat 1·12 (1·72) 119 0·56 (1·32) 111 −0·60 (−1·01 to −0·19) 0·005 0·004 0·33 Per-protocol 1·07 (1·82) 98 0·40 (1·13) 94 −0·72 (−1·16 to −0·28) 0·002 0·002 0·43 MRI total hippocampal volume (cm3) Modified intention-to-treat −0·43 (0·33) 104 −0·30 (0·27) 96 0·12 (0·04 to 0·21) 0·005 0·005 0·22 Per-protocol −0·42 (0·32) 90 −0·28 (0·28) 86 0·12 (0·03 to 0·21) 0·010 0·008 0·20 MRI whole brain volume (cm3) Modified intention-to-treat −24·24 (20·93) 90 −20·27 (17·79) 83 3·66 (−2·81 to 10·14) 0·265 0·284 0·21 Per-protocol −23·88 (19·90) 77 −17·89 (16·88) 73 5·04 (−2·02 to 12·10) 0·160 0·137 0·29 MRI ventricular volume (cm3)|| Modified intention-to-treat 7·80 (5·53) 106 5·96 (4·66) 94 −1·36 (−2·70 to −0·03) 0·046 0·042 0·22 Per-protocol 7·40 (4·79) 92 5·39 (4·50) 83 −1·40 (−2·79 to −0·02) 0·046 0·042 0·20 n=number of participants with at least one post-baseline value in the mixed model. p values are for effect of intervention over 24 months. NTB=neuropsychological test battery. CDR-SB=clinical dementia rating sum of boxes.

·046 0·042 0·22 Per-protocol 7·40 (4·79) 92 5·39 (4·50) 83 −1·40 (−2·79 to −0·02) 0·046 0·042 0·20 n=number of participants with at least one post-baseline value in the mixed model. p values are for effect of intervention over 24 months. NTB=neuropsychological test battery. CDR-SB=clinical dementia rating sum of boxes. MMSE=Mini-Mental State Examination. * Mixed model: linear mixed model for longitudinal data with change from baseline as outcome, baseline score, and baseline MMSE as covariates, and real measurement time as a continuous variable. † Sensitivity analysis: mixed model for repeated measures with change from baseline as outcome, baseline score, and baseline MMSE as covariates, and planned visit time as a categorical variable. ‡ Cohen's d standardised effect size calculated based on the mean treatment difference over 24 months as estimated in the mixed model and the pooled SD; results are presented so that a positive effect size indicates improved performance in the active versus control group and vice versa. § Data for active and control groups are presented as observed mean change from baseline at month 24 (SD). ¶ Difference (active minus control) is calculated as based on least squares means for change from baseline over 24 months as estimated in the mixed model. || Higher scores indicate worse performance; for all other endpoints, higher scores indicate better performance.

§ Data for active and control groups are presented as observed mean change from baseline at month 24 (SD). ¶ Difference (active minus control) is calculated as based on least squares means for change from baseline over 24 months as estimated in the mixed model. || Higher scores indicate worse performance; for all other endpoints, higher scores indicate better performance. No statistically significant differences were observed for the secondary NTB composite scores. For CDR-SB, there was significantly less worsening in the active group than in the control group during 24 months (p=0·005; table 2, figure 2C). Similar results were obtained without adjustment for baseline MMSE (appendix) and in the sensitivity analysis (table 2). The worsening in CDR-SB was 45% less in the active group than in the control group, based on the estimated change from baseline over 24 months.Figure 2 Changes in main endpoints during the 24-month intervention (A) NTB primary endpoint. (B) NTB memory domain. (C) CDR-SB. (D) MRI total hippocampal volume. (E) MRI ventricular volume. (F) CDR-SB in subgroups defined by baseline MMSE. Data are observed mean change from baseline; error bars are SE. Sample size by baseline MMSE subgroup (control/active): ≥24: mITT 117/106 (PP 96/89), ≥25: 104/91 (86/75), ≥26: 95/79 (78/66), ≥27: 77/63 (66/53), ≥28: 55/43 (48/37), ≥29 29/21 (24/19). CDR-SB=clinical dementia rating-sum of boxes. mITT=modified intention-to-treat analysis. MMSE=Mini-Mental State Examination. NTB=neuropsychological test battery. PP=per-protocol analysis.

): ≥24: mITT 117/106 (PP 96/89), ≥25: 104/91 (86/75), ≥26: 95/79 (78/66), ≥27: 77/63 (66/53), ≥28: 55/43 (48/37), ≥29 29/21 (24/19). CDR-SB=clinical dementia rating-sum of boxes. mITT=modified intention-to-treat analysis. MMSE=Mini-Mental State Examination. NTB=neuropsychological test battery. PP=per-protocol analysis. We observed significantly less reduction in hippocampal volume (p=0·005) and less increase in ventricular volume (p=0·046) during 24 months in the active group than in the control group (table 2, figure 2D, 2E). The rates of deterioration were lower in the active group than in the control group, both for hippocampal (26%) and ventricular volumes (16%). Sensitivity analyses confirmed these observations for hippocampal volume and ventricular volume. No statistically significant differences between groups were found for changes in whole brain volume (table 2). During the 24-month trial period, 59 (37%) participants in the control group and 62 (41%) in the active group were diagnosed with dementia (p=0·642, Fisher's exact test; appendix). HDL cholesterol concentration significantly increased in the active group compared with the control group, but absolute changes were very small (<5%), and no differences between groups were found for changes in LDL cholesterol over 24 months (appendix).

diagnosed with dementia (p=0·642, Fisher's exact test; appendix). HDL cholesterol concentration significantly increased in the active group compared with the control group, but absolute changes were very small (<5%), and no differences between groups were found for changes in LDL cholesterol over 24 months (appendix). Differences in cognition-related scores between groups were more pronounced in per-protocol analyses than in the mITT analyses, particularly for the NTB primary endpoint and the NTB memory domain (table 2; appendix). There were no differences in baseline characteristics between active and control groups among participants included in the per-protocol analyses (appendix). Forest plots showing an overview of the mITT and per-protocol results from the different statistical models are provided in the appendix.

main (table 2; appendix). There were no differences in baseline characteristics between active and control groups among participants included in the per-protocol analyses (appendix). Forest plots showing an overview of the mITT and per-protocol results from the different statistical models are provided in the appendix. Predefined subgroup analyses (MMSE score ≥26) in mITT and per-protocol populations are shown in the appendix. We observed statistically significant differences between groups for CDR-SB and hippocampal volume in the mITT population, and for the NTB primary endpoint (mixed model p=0·131, sensitivity analysis p=0·044) and NTB memory domain (mixed model p=0·073, sensitivity analysis p=0·017) in the per-protocol population. Baseline MMSE was an effect modifier for CDR-SB in the per-protocol population (mixed model p=0·053 for interaction term treatment effect × baseline MMSE). Therefore, we did an exploratory analysis of CDR-SB performance across the spectrum of baseline MMSE (≥24 to ≥29), which suggested that the treatment effect on CDR-SB increased with higher baseline MMSE scores (figure 2F; appendix).

population (mixed model p=0·053 for interaction term treatment effect × baseline MMSE). Therefore, we did an exploratory analysis of CDR-SB performance across the spectrum of baseline MMSE (≥24 to ≥29), which suggested that the treatment effect on CDR-SB increased with higher baseline MMSE scores (figure 2F; appendix). Self-reported adherence to the intervention was high, both when calculated in all participants (mean 93·4% [SD 8·8] in both groups) and when calculated using all available data in the mITT (excluding data collected after starting rescue medication: mean 87·3% [SD 22·9] in active and 86·8% [23·4] in control). This adherence was confirmed by significant biochemical changes in plasma DHA and EPA during 24 months in the active group compared with no changes in the control group (p<0·0001; appendix). The incidences of adverse events and serious adverse events were similar between groups (p=0·864 and p=0·487; table 3), and among the 66 participants who dropped out (active vs control: 24 [73%] vs 22 [67%], p=0·789 and 8 [24%] vs three [9%], p=0·185). None of the serious adverse events were regarded as related to the study product and dropout due to adverse events was not significantly different between groups (nine [6%] in the active group vs six [4%] in the control group, p=0·437).Table 3 Summary of adverse events in all participants who were randomly assigned and on double-blind treatment

e events were regarded as related to the study product and dropout due to adverse events was not significantly different between groups (nine [6%] in the active group vs six [4%] in the control group, p=0·437).Table 3 Summary of adverse events in all participants who were randomly assigned and on double-blind treatment Control (n=157) Active (n=152) All events At least one adverse event 138 (88%) 132 (87%) At least one serious adverse event 30 (19%) 34 (22%) Most common serious adverse events* Myocardial infarction 2 (1%) 0 (0) Fall 1 (1%) 2 (1%) Intervertebral disc protrusion 2 (1%) 0 (0) Osteoarthritis 3 (2%) 0 (0) Syncope 0 (0) 3 (2%) (Major) depression 3 (2%) 1 (1%) Cardiac operation 2 (1%) 0 (0) Hospitalisation 0 (0) 2 (1%) Circulatory collapse 0 (0) 2 (1%) Most common adverse events† Vertigo 12 (8%) 6 (4%) Diarrhoea 14 (9%) 7 (5%) Cystitis 9 (6%) 4 (3%) Nasopharyngitis 16 (10%) 7 (5%) Respiratory tract infection 9 (6%) 7 (5%) Urinary tract infection 9 (6%) 7 (5%) Fall 8 (5%) 11 (7%) Arthralgia 9 (6%) 4 (3%) Back pain 5 (3%) 10 (7%) Headache 12 (8%) 9 (6%) Cough 10 (6%) 2 (1%) Data are n (%). Adverse events are presented by Medical Dictionary for Regulatory Activities preferred term. * Only those reported by at least two participants in either group are shown. † Only those reported by at least 5% of participants in either group are shown.

Control (n=157) Active (n=152) All events At least one adverse event 138 (88%) 132 (87%) At least one serious adverse event 30 (19%) 34 (22%) Most common serious adverse events* Myocardial infarction 2 (1%) 0 (0) Fall 1 (1%) 2 (1%) Intervertebral disc protrusion 2 (1%) 0 (0) Osteoarthritis 3 (2%) 0 (0) Syncope 0 (0) 3 (2%) (Major) depression 3 (2%) 1 (1%) Cardiac operation 2 (1%) 0 (0) Hospitalisation 0 (0) 2 (1%) Circulatory collapse 0 (0) 2 (1%) Most common adverse events† Vertigo 12 (8%) 6 (4%) Diarrhoea 14 (9%) 7 (5%) Cystitis 9 (6%) 4 (3%) Nasopharyngitis 16 (10%) 7 (5%) Respiratory tract infection 9 (6%) 7 (5%) Urinary tract infection 9 (6%) 7 (5%) Fall 8 (5%) 11 (7%) Arthralgia 9 (6%) 4 (3%) Back pain 5 (3%) 10 (7%) Headache 12 (8%) 9 (6%) Cough 10 (6%) 2 (1%) Data are n (%). Adverse events are presented by Medical Dictionary for Regulatory Activities preferred term. * Only those reported by at least two participants in either group are shown. † Only those reported by at least 5% of participants in either group are shown. Discussion Prodromal Alzheimer's disease is a new area of Alzheimer's disease research, with clinical research practices still under development. LipiDiDiet is the first randomised, controlled, double-blind, multicentre, international trial of a non-pharmacological intervention in prodromal Alzheimer's disease. No significant difference was found between groups for the NTB primary endpoint in the mITT analysis or on conversion to dementia. However, there was some evidence of a beneficial effect of the multinutrient intervention at the cognitive-functional level (detected by CDR-SB) and ameliorated structural brain changes (hippocampal and ventricular volume) shown on MRI scans.

r the NTB primary endpoint in the mITT analysis or on conversion to dementia. However, there was some evidence of a beneficial effect of the multinutrient intervention at the cognitive-functional level (detected by CDR-SB) and ameliorated structural brain changes (hippocampal and ventricular volume) shown on MRI scans. The LipiDiDiet study was initiated shortly after the first criteria for prodromal Alzheimer's disease were published.3 It has since become clear that changes in cognitive performance with currently used tests are not very pronounced in early Alzheimer's disease during time intervals close to 2 years.30, 31 The study design was based on a previous 12-month trial in Alzheimer's disease dementia;29 however, in our study, the control group had only a quarter of the projected 24-month decline on the NTB primary endpoint, possibly because of the earlier disease stage of the pre-dementia participants than those with Alzheimer's disease dementia. The lower than expected decline is in agreement with the previous observations of limited cognitive changes over 2 years. The small cognitive decline on the NTB primary endpoint in the control group was mainly due to stable performance during the first year of intervention, followed by a steeper decline in the second year. Conversely, CDR-SB scores had already significantly declined at month 12 in the control group. Therefore, in mITT analyses, the significant benefit on CDR-SB was combined with an absence of clear effects on NTB cognition endpoints, although benefits on the NTB primary endpoint and the NTB memory domain were suggested in the per-protocol analyses. Notably, the main reason for exclusion from the per-protocol analysis was no or irregular intake of study product, which emphasises the importance of sustained intake, as observed previously.27, 32

ints, although benefits on the NTB primary endpoint and the NTB memory domain were suggested in the per-protocol analyses. Notably, the main reason for exclusion from the per-protocol analysis was no or irregular intake of study product, which emphasises the importance of sustained intake, as observed previously.27, 32 The effect on CDR-SB in the LipiDiDiet trial differs from previous mild Alzheimer's disease dementia trials by adding a benefit at the cognitive-functional level.23, 24 The longer treatment duration and intervention at an earlier disease stage than in the previous dementia trials might be important reasons for this observation. The apparently more pronounced stabilisation of CDR-SB scores with increasing baseline MMSE observed in the active group indicates that early rather than late treatment within the prodromal stage might support better outcome with this cognitive-functional measure, in line with previous results from Fortasyn Connect trials,23, 24, 26 which showed that earlier intervention might increase the benefit. Within the disease continuum, early intervention might also be an important contributory factor to the similarity in progression to dementia in both treatment groups, because participants at baseline were only 2 years or less from advancing to dementia and therefore treatment efficacy might not have been sufficient to translate into less conversion to dementia. Moreover, dementia diagnosis was clustered at major study visits. Dementia diagnosis is a dichotomisation of the decline continuum, whereas CDR-SB is a sensitive measure of decline across a continuous scale. Therefore, CDR-SB might better reflect disease progression along our entire prodromal population.

n to dementia. Moreover, dementia diagnosis was clustered at major study visits. Dementia diagnosis is a dichotomisation of the decline continuum, whereas CDR-SB is a sensitive measure of decline across a continuous scale. Therefore, CDR-SB might better reflect disease progression along our entire prodromal population. No cognitive test is generally accepted as the gold standard for trials in prodromal Alzheimer's disease, although research has highlighted the potential usefulness of composite measures.33 Use of compound cognitive test batteries such as NTB combining performance on different validated tests have been suggested to help detection of more subtle changes that occur in pre-dementia disease stages.34 In the meantime, preliminary guidelines from regulatory agencies emphasise the importance of establishing the clinical value of treatment and suggest using a combined cognitive-functional measure such as CDR-SB, which showed reliability and validity in prodromal Alzheimer's disease and mild cognitive impairment due to Alzheimer's disease, and proposed CDR-SB as a single primary endpoint for efficacy.35, 36 However, this information was not available at the time of our trial design.

cognitive-functional measure such as CDR-SB, which showed reliability and validity in prodromal Alzheimer's disease and mild cognitive impairment due to Alzheimer's disease, and proposed CDR-SB as a single primary endpoint for efficacy.35, 36 However, this information was not available at the time of our trial design. Although it is generally difficult to translate performance on cognitive tests into clinical benefits, the CDR-SB is built on real-life items such as handling household emergencies, handling financial transactions, and forgetting a major event, which facilitates assessment of clinical benefit. For early Alzheimer's disease, a reduction by 0·5 or 1·0 in CDR-SB was proposed to capture both efficacy and clinical relevance.1 The current emphasis on using more sensitive cognitive or functional measures in ongoing trials in prodromal Alzheimer's disease is a major shift from the previous focus on progression to dementia used in unsuccessful trials in mild cognitive impairment.33

was proposed to capture both efficacy and clinical relevance.1 The current emphasis on using more sensitive cognitive or functional measures in ongoing trials in prodromal Alzheimer's disease is a major shift from the previous focus on progression to dementia used in unsuccessful trials in mild cognitive impairment.33 In addition to the CDR-SB cognition-function benefit, we observed benefits on progression of structural changes in the brain. The hippocampus is affected early in Alzheimer's disease, and the rate of hippocampal atrophy over time is considered a reliable measure of Alzheimer's disease progression.4 We noted 26% less reduction in hippocampal volume in the active group compared with the control group and the active group also had 16% less increase in ventricular volume, suggesting an interaction of the treatment with the disease process. Interaction could be hypothesised based on animal and mild cognitive impairment studies that showed effects on Alzheimer's disease-related brain pathologies.9, 10

the control group and the active group also had 16% less increase in ventricular volume, suggesting an interaction of the treatment with the disease process. Interaction could be hypothesised based on animal and mild cognitive impairment studies that showed effects on Alzheimer's disease-related brain pathologies.9, 10 Prodromal Alzheimer's disease was defined according to the IWG-1 criteria.3 Currently, three sets of research criteria are available for Alzheimer's disease diagnosis in people with mild cognitive impairment: IWG-1,3 IWG-2,4 and NIA-AA5 criteria. Comparative analysis showed that all three predict cognitive decline with reasonable accuracy.37 Baseline characteristics in our study were as expected for a prodromal Alzheimer's disease population, including the CSF biomarker profile; percentage of APOE ε4 carriers; and IWG-1, IWG-2, and NIA-AA criteria (appendix). The main differences between the IWG-1 and IWG-2 criteria are the definition of in-vivo evidence of Alzheimer's disease pathology (medial temporal lobe atrophy on MRI is included in IWG-1, but not in IWG-2) and the clinical Alzheimer's disease phenotypes (IWG-1 focuses on a typical amnestic phenotype, whereas IWG-2 also includes atypical, non-amnestic phenotypes). Thus, we cannot make inferences on intervention effects in prodromal Alzheimer's disease with atypical, non-amnestic phenotypes. This study indicates that populations within prodromal Alzheimer's disease exist who might benefit differently from early intervention. Baseline MMSE scores and decline in cognitive function contributed to different levels of benefit; further currently unknown factors might contribute as well. Identification of those individuals could aid in the ongoing process of refining prodromal Alzheimer's disease definition and prodromal Alzheimer's disease clinical trial design.

res and decline in cognitive function contributed to different levels of benefit; further currently unknown factors might contribute as well. Identification of those individuals could aid in the ongoing process of refining prodromal Alzheimer's disease definition and prodromal Alzheimer's disease clinical trial design. As expected from previous trials,23, 24, 26 study product compliance was high, and adverse events and serious adverse events were consistent with the studied population and the known safety profile of the active product.23, 24, 26 The proportion of participants with at least one serious adverse event (34 [22%] in the active group and 30 [19%] in the control group) and percentage of dropouts due to adverse events (nine [6%] in the active group and six [4%] in the control group) in our study are in the same range as those reported by Coric and colleagues38 for the control group (31 [23.7%] patients with at least one serious adverse event and 13 [9.9%] dropouts due to adverse events).

percentage of dropouts due to adverse events (nine [6%] in the active group and six [4%] in the control group) in our study are in the same range as those reported by Coric and colleagues38 for the control group (31 [23.7%] patients with at least one serious adverse event and 13 [9.9%] dropouts due to adverse events). Our study has some limitations. First, cognitive decline in this prodromal Alzheimer's disease population was much lower than expected, rendering the primary endpoint inadequately powered. Therefore, future trials aiming to implement this NTB endpoint might benefit from larger sample sizes and a longer duration of intervention than used in our study, or a cognitive composite designed for this pre-dementia population. Second, our 24-month trial was not designed with progression to dementia as a primary focus, thereby limiting the ability to draw conclusions on this outcome. Additionally, use of MRI as an alternative to CSF or PET amyloid assessments might have resulted in a somewhat more heterogeneous group of participants, because medial temporal atrophy on MRI can be both amyloid-related and non-amyloid-related.4 Finally, we included a demographically restricted population, largely comprising white participants from central European countries and Scandinavia. Participants completing the 24-month intervention were eligible to continue in the double-blind extension studies, which will provide additional data on long-term efficacy.

inally, we included a demographically restricted population, largely comprising white participants from central European countries and Scandinavia. Participants completing the 24-month intervention were eligible to continue in the double-blind extension studies, which will provide additional data on long-term efficacy. In conclusion, the multinutrient intervention had no significant effect on the NTB primary endpoint over 2 years in prodromal Alzheimer's disease, although potential benefits were seen on the cognitive-functional measure CDR-SB and brain atrophy measures. Further investigation of multinutrient approaches in early Alzheimer's disease stages is warranted. Supplementary Material Supplementary appendix

In conclusion, the multinutrient intervention had no significant effect on the NTB primary endpoint over 2 years in prodromal Alzheimer's disease, although potential benefits were seen on the cognitive-functional measure CDR-SB and brain atrophy measures. Further investigation of multinutrient approaches in early Alzheimer's disease stages is warranted. Supplementary Material Supplementary appendix Acknowledgments The research leading to these results was mainly funded by the European Commission under the 7th framework programme of the European Union (grant agreement number 211696). Additional funding was provided by the EU Joint Programme - Neurodegenerative Disease Research (MIND-AD grant); Kuopio University Hospital, Finland (EVO/VTR grant); and Academy of Finland (grant 287490). KB holds the Torsten Söderberg Professorship in Medicine at the Royal Swedish Academy of Sciences. MK holds the Stiftelsen Stockholms Sjukhems donation professor post and has received grants from Wallenberg Clinical Scholars, the Stockholm City Council and Center for Innovative Medicine at the Karolinska Institute, Sweden (CIMED and ALF grants). We thank all participants enrolled in the study and their families; Britta Theobald-Loeffler and Claudia Schacht (EURICE, Saarbrücken, Germany) for their support in management and coordination of the LipiDiDiet project; the members of the independent data monitoring committee (Roy W Jones, Dirk L Knol, and Craig W Ritchie) for their advice and time; Nico Rozendaal, Jose de Bont, and Anja Kerksiek; and all investigators and on-site study staff for their efforts in the conduct of the field work.

d coordination of the LipiDiDiet project; the members of the independent data monitoring committee (Roy W Jones, Dirk L Knol, and Craig W Ritchie) for their advice and time; Nico Rozendaal, Jose de Bont, and Anja Kerksiek; and all investigators and on-site study staff for their efforts in the conduct of the field work. Contributors HS, AS, PJV, KB, MK, and TH contributed to the design of the study. HS, AS, PJV, MK, and TH were members of the LipiDiDiet Trial Steering Committee; TH was the chairman. HS, AS, PJV, and MK contributed to data collection. PJV coordinated the central MRI analyses. KB coordinated the central CSF analyses. HS, AS, PJV, SBH, MK, and TH contributed to the development and implementation of the statistical analysis plan. SBH was responsible for data management and statistical analysis of the trial data. All authors were involved in data interpretation, drafting, review, and approval of the report, and the decision to submit for publication.

, MK, and TH contributed to the development and implementation of the statistical analysis plan. SBH was responsible for data management and statistical analysis of the trial data. All authors were involved in data interpretation, drafting, review, and approval of the report, and the decision to submit for publication. Declaration of interests HS, PJV, KB, MK, and TH were supported by a grant from the European Commission for this study (FP7-211696 LipiDiDiet). HS has received personal fees from ACImmune (adviser), outside the submitted work, and has been a principal investigator for drug trials (Lilly, Pfizer, Sanofi, Solvay, Wyeth, and Servier), without personal payment. PJV has received grants from Innovative Medicine Initiative (European Medical Information Framework, European prevention of Alzheimer's dementia consortium, and Real world outcomes across the Alzheimer's disease spectrum for better care: multi-modal data access platform projects), during the conduct of the study; non-financial support from GE Healthcare; grants from Biogen, and has served as a consultant for Eli Lilly and Janssen, outside the submitted work. SBH received payment for data management and statistical analysis from the LipiDiDiet Consortium during the conduct of the study, and is owner of Pentara Corporation, a company that consults with pharmaceutical, academic, and non-profit groups in the Alzheimer's space. KB has served as a consultant or at advisory boards for Fujirebio Europe, IBL International, and Roche Diagnostics, outside the submitted work. MK has received a grant from Innovative Medicine Initiative (European prevention of Alzheimer's dementia consortium project), during the conduct of the study. AS and TH declare no competing interests.

. Large-scale collaborations are required to refine and validate robust risk prediction scores. Meanwhile, our findings suggest that future risk scores to identify patients with stroke at risk of intracranial haemorrhage should include cerebral microbleeds as a neuroimaging biomarker in addition to clinical parameters. Our study has some important strengths. We prospectively studied a large inception cohort of patients at multiple hospital stroke units using predefined MRI sequences, rated for neuroimaging markers of small vessel disease using validated scales by a single trained observer. We followed up 97% of our cohort, and experienced observers adjudicated all primary events blinded to baseline cerebral microbleed presence. We undertook survival analysis to account for baseline confounding factors and varying follow-up, and followed a prespecified statistical analysis plan.

Introduction Atrial fibrillation increases the risk of ischaemic stroke by five times.1 In most individuals, oral anticoagulation with either vitamin K antagonists (VKAs) or direct oral anticoagulants (DOACs) is indicated because they reduce the risk of ischaemic stroke by about two thirds, with only a minimal increase in extracranial haemorrhage.2, 3 However, a devastating and unpredictable complication of oral anticoagulation is symptomatic intracranial haemorrhage, which has 42% in-hospital mortality and causes substantial disability in survivors.4 There is an unmet clinical need to reliably predict the risk of intracranial haemorrhage and to differentiate this from the risk of ischaemic stroke, to allow clinicians to assess the likely net clinical benefit of oral anticoagulation. Risk scores including clinical factors (eg, hypertension and age) have been developed to identify patients at high risk of bleeding on anticoagulation—including the HAS-BLED,5 HAEMORR2HAGES,6 and ATRIA7 scores—but these are of limited value in clinical decision making because they do not differentiate between prediction of ischaemic stroke and of intracranial haemorrhage. Research in context Evidence before this study

Introduction Atrial fibrillation increases the risk of ischaemic stroke by five times.1 In most individuals, oral anticoagulation with either vitamin K antagonists (VKAs) or direct oral anticoagulants (DOACs) is indicated because they reduce the risk of ischaemic stroke by about two thirds, with only a minimal increase in extracranial haemorrhage.2, 3 However, a devastating and unpredictable complication of oral anticoagulation is symptomatic intracranial haemorrhage, which has 42% in-hospital mortality and causes substantial disability in survivors.4 There is an unmet clinical need to reliably predict the risk of intracranial haemorrhage and to differentiate this from the risk of ischaemic stroke, to allow clinicians to assess the likely net clinical benefit of oral anticoagulation. Risk scores including clinical factors (eg, hypertension and age) have been developed to identify patients at high risk of bleeding on anticoagulation—including the HAS-BLED,5 HAEMORR2HAGES,6 and ATRIA7 scores—but these are of limited value in clinical decision making because they do not differentiate between prediction of ischaemic stroke and of intracranial haemorrhage. Research in context Evidence before this study We searched MEDLINE without language restrictions for publications regarding cerebral microbleeds, atrial fibrillation, and ischaemic stroke published from inception up to Nov 3, 2017. We used the search terms (Cerebral microbleed*.mp OR microbleed.mp) AND (atrial fibrillation/ OR anticoagula*.mp OR anticoagula* OR warfarin.mp OR rivaroxaban.mp OR apixaban.mp OR edoxaban.mp OR dabigatran.mp) AND (cerebral infarction/ OR brain ischemia/ or stroke/ or isch?emi*.mp/ or transient isch?emic attack). We found four published prospective studies that reported rates of intracerebral haemorrhage in relation to baseline cerebral microbleeds in patients with ischaemic stroke or transient ischaemic attack treated with anticoagulation for atrial fibrillation. The largest study, involving 550 patients from Korea, showed a significant association between cerebral microbleed presence and intracerebral haemorrhage after adjusting for age, sex, and previous haemorrhagic stroke (hazard ratio [HR] 3·8; 95% CI 1·1–13·1), but none of the other studies were sufficiently powered to confirm this association. A post-hoc aggregate data meta-analysis of data mainly from small retrospective and prospective cohorts, with variable completeness and follow-up duration, suggested that cerebral microbleeds are associated with increased intracerebral haemorrhage risk, but could not adjust for confounding factors or develop risk models for intracranial haemorrhage.

analysis of data mainly from small retrospective and prospective cohorts, with variable completeness and follow-up duration, suggested that cerebral microbleeds are associated with increased intracerebral haemorrhage risk, but could not adjust for confounding factors or develop risk models for intracranial haemorrhage. Added value of this study Our observational, predominantly UK-based, multicentre, prospective inception cohort study including 3366 patient-years of follow-up was designed and powered to determine whether cerebral microbleeds are independently associated with a higher risk of intracranial haemorrhage in patients with recent acute ischaemic stroke or transient ischaemic attack associated with atrial fibrillation and started for the first time on oral anticoagulation. We provide new evidence that in patients with ischaemic stroke or transient ischaemic attack and atrial fibrillation, cerebral microbleed presence is an independent risk factor for intracranial haemorrhage. We also show that the risk of intracranial haemorrhage increases as cerebral microbleed burden increases, but that the absolute event rate for ischaemic stroke remains higher than that of intracranial haemorrhage, even in patients with multiple cerebral microbleeds. We developed and internally validated a simple risk prediction score for intracranial haemorrhage, showing for the first time that the inclusion of cerebral microbleed presence as a neuroimaging biomarker improves the predictive value of a commonly used bleeding risk score based on clinical data alone (the HAS-BLED score).

eveloped and internally validated a simple risk prediction score for intracranial haemorrhage, showing for the first time that the inclusion of cerebral microbleed presence as a neuroimaging biomarker improves the predictive value of a commonly used bleeding risk score based on clinical data alone (the HAS-BLED score). Implications of all the new evidence Our study provides proof of concept that including a neuroimaging biomarker (cerebral microbleeds) improves the predictive ability of clinical risk scores for intracranial haemorrhage—a potentially deadly complication of oral anticoagulation—which could help clinicians and patients to make better informed anticoagulation decisions. Our findings support further pooled meta-analyses of individual participant data from large prospective cohorts to increase the precision of risk estimates for intracranial haemorrhage, to determine whether high cerebral microbleed counts can identify patients who will experience net harm from oral anticoagulation, and to refine and validate intracranial haemorrhage risk scores incorporating clinical and neuroimaging factors including cerebral microbleeds.

risk estimates for intracranial haemorrhage, to determine whether high cerebral microbleed counts can identify patients who will experience net harm from oral anticoagulation, and to refine and validate intracranial haemorrhage risk scores incorporating clinical and neuroimaging factors including cerebral microbleeds. Cerebral microbleeds are small, hypointense, round or ovoid areas identified on blood-sensitive MRI sequences (T2*-weighted gradient-recalled echo [GRE] or susceptibility-weighted imaging).8, 9 In most cases, cerebral microbleeds correspond pathologically to small clusters of haemosiderin-laden macrophages resulting from small self-limiting haemorrhages.10, 11 Thus, cerebral microbleeds are a promising radiological biomarker of the cerebral small vessel diseases that are prone to bleeding and cause most spontaneous intracerebral haemorrhages,9 so might be a specific and clinically useful predictor of anticoagulant-related intracranial haemorrhage. With the increasing use of blood-sensitive MRI, cerebral microbleeds can be detected in about 30% of patients with ischaemic stroke and atrial fibrillation,12 generating uncertainty about the risk–benefit balance of anticoagulation in patients with cerebral microbleeds.

f anticoagulant-related intracranial haemorrhage. With the increasing use of blood-sensitive MRI, cerebral microbleeds can be detected in about 30% of patients with ischaemic stroke and atrial fibrillation,12 generating uncertainty about the risk–benefit balance of anticoagulation in patients with cerebral microbleeds. We did an observational, prospective, multicentre, inception cohort study to determine whether cerebral microbleeds are independently associated with an increased risk of symptomatic intracranial haemorrhage in patients with recent acute ischaemic stroke or transient ischaemic attack with atrial fibrillation treated with anticoagulation. We developed and internally validated risk prediction scores for symptomatic intracranial haemorrhage including cerebral microbleed presence as a neuroimaging biomarker in addition to clinical risk factors.

ecent acute ischaemic stroke or transient ischaemic attack with atrial fibrillation treated with anticoagulation. We developed and internally validated risk prediction scores for symptomatic intracranial haemorrhage including cerebral microbleed presence as a neuroimaging biomarker in addition to clinical risk factors. Methods Study design and participants CROMIS-2 is an observational, multicentre, prospective, inception cohort study that recruited adults (ie, ≥18 years of age) with electrocardiogram-confirmed non-valvular atrial fibrillation who presented to one of 80 participating hospitals (79 in the UK and one in the Netherlands) with ischaemic stroke or transient ischaemic attack and were identified by their treating physician for anticoagulation treatment. We did not strictly control the timing of oral anticoagulation, which depended on best clinical judgment according to standard practice. We excluded patients if they could not undergo MRI, had a definite contraindication to anticoagulation, or had previously received therapeutic anticoagulation. CROMIS-2 was approved by the UK National Health Service Research Ethics Committee. Patients with capacity gave informed written consent. When patients could not consent, we obtained written informed consent from a proxy as defined by relevant local legislation.

Methods Study design and participants CROMIS-2 is an observational, multicentre, prospective, inception cohort study that recruited adults (ie, ≥18 years of age) with electrocardiogram-confirmed non-valvular atrial fibrillation who presented to one of 80 participating hospitals (79 in the UK and one in the Netherlands) with ischaemic stroke or transient ischaemic attack and were identified by their treating physician for anticoagulation treatment. We did not strictly control the timing of oral anticoagulation, which depended on best clinical judgment according to standard practice. We excluded patients if they could not undergo MRI, had a definite contraindication to anticoagulation, or had previously received therapeutic anticoagulation. CROMIS-2 was approved by the UK National Health Service Research Ethics Committee. Patients with capacity gave informed written consent. When patients could not consent, we obtained written informed consent from a proxy as defined by relevant local legislation. Procedures All patients underwent baseline brain MRI according to a predefined protocol parameter range, designed to detect relevant markers of cerebrovascular disease13 (see appendix), which required T2*-weighted GRE (echo time 10–45 ms), axial T1-weighted, axial T2-weighted, coronal fluid-attenuated inversion recovery, and diffusion-weighted imaging with apparent diffusion coefficient maps. MRIs were analysed for markers of cerebral small vessel disease defined according to consensus definitions9 using validated scales where available. We used the Microbleed Anatomical Rating Scale14 to identify and classify cerebral microbleeds as lobar or non-lobar (ie, deep, including the basal ganglia, thalamus, deep white matter, brainstem, and cerebellum). We rated white matter hyperintensities using the Fazekas and age-related white matter changes (ARWMC) scales,15, 16 and defined cortical superficial siderosis using consensus criteria.17, 18, 19 All neuroimaging ratings were done by a clinical research fellow (DW) trained by a professor of neuroradiology with cerebrovascular expertise (HRJ). A second trained clinical research fellow (GB) rated a random 10% sample for cerebral microbleed presence; we quantified intra-rater and inter-rater reliability for cerebral microbleed presence using Cohen's κ coefficient.

search fellow (DW) trained by a professor of neuroradiology with cerebrovascular expertise (HRJ). A second trained clinical research fellow (GB) rated a random 10% sample for cerebral microbleed presence; we quantified intra-rater and inter-rater reliability for cerebral microbleed presence using Cohen's κ coefficient. We obtained screening logs to assess selection bias. We obtained detailed clinical and demographic baseline data. From these data we calculated CHA₂DS₂VASc and HAS-BLED scores, designed to predict the risks of ischaemic stroke and major bleeding, respectively, in patients with non-valvular atrial fibrillation. We obtained follow-up information from patients and general practitioners at 6 months, 12 months, and 24 months via standardised structured postal questionnaires or telephone interviews. We obtained National Health Service digital data regarding hospital admissions or death during follow-up. For reported outcome events, we obtained additional clinical and radiological details from treating clinical teams and medical records to allow central adjudication, blinded to baseline neuroimaging findings.

iews. We obtained National Health Service digital data regarding hospital admissions or death during follow-up. For reported outcome events, we obtained additional clinical and radiological details from treating clinical teams and medical records to allow central adjudication, blinded to baseline neuroimaging findings. Outcomes The primary outcome was symptomatic intracranial haemorrhage, defined as brain-imaging evidence of non-traumatic spontaneous intracranial haemorrhage with appropriate clinical symptoms, at any time before the final follow-up at 24 months. The secondary outcomes were recurrent ischaemic stroke and death of any cause. Further secondary outcomes not reported in this paper were transient ischaemic attack, cardiac ischaemic events (defined by dynamic electrocardiogram changes or troponin rise), subdivisions of intracranial haemorrhage (intracerebral [reported], subarachnoid, subdural, and extradural haemorrhage), major bleeding (defined as intracranial bleeding or extracranial bleeding in either a critical area or requiring hospitalisation and two units of blood transfusion20), quality of life, and long-term physical disability. A composite outcome of death, ischaemic stroke, and symptomatic inracranial haemorrhage was prespecified by the Steering Committee prior to the end of recruitment and data analysis, but was not prespecified in the statistical analysis plan.

ts of blood transfusion20), quality of life, and long-term physical disability. A composite outcome of death, ischaemic stroke, and symptomatic inracranial haemorrhage was prespecified by the Steering Committee prior to the end of recruitment and data analysis, but was not prespecified in the statistical analysis plan. Two professors of vascular neurology (DJW and MMB) and a clinical research fellow (DW) adjudicated all primary outcome events. A trained clinical research fellow (DW) adjudicated all ischaemic stroke outcomes; a random 10% of these were adjudicated by a professor of vascular neurology (DJW) and a professor of neuroradiology (HRJ). All adjudication was blinded to baseline cerebral microbleed ratings. In cases of disagreement, we reached consensus after discussion. Statistical analysis We followed a prespecified published statistical analysis plan, which is provided in full in the appendix. We calculated a planned sample size of 1425 participants to detect a relative risk of 4·0 for intracranial haemorrhage associated with cerebral microbleeds, assuming an annual incidence of intracranial haemorrhage of 1·25% in those without cerebral microbleeds and that 20% of our population would have cerebral microbleeds; these estimates were derived from previous smaller studies.13 We included patients in the final analysis if they had undergone MRI with T2*-weighted GRE sequences of adequate technical quality to rate cerebral microbleeds.

hose without cerebral microbleeds and that 20% of our population would have cerebral microbleeds; these estimates were derived from previous smaller studies.13 We included patients in the final analysis if they had undergone MRI with T2*-weighted GRE sequences of adequate technical quality to rate cerebral microbleeds. We compared baseline demographics and risk factor profiles between those with and without cerebral microbleeds, and between those with and without our primary outcome event (symptomatic intracranial haemorrhage). We used appropriate statistical measures for categorical and continuous measures. We visually inspected the distributions of continuous variables using histograms, summarised as means with SDs or medians with IQRs. Groups were compared using the Mann-Whitney U test if not normally distributed or the t test if normally distributed; categorical variables were compared between groups with the χ2 test or, where appropriate, Fisher's exact test. Univariate Kaplan-Meier survival probabilities were estimated for those with and without cerebral microbleeds; we used the log-rank test to compare groups. We did univariable and multivariable Cox regression (adjusted for age and history of hypertension, as documented in our statistical analysis plan). We did three further multivariable Cox regression sensitivity analyses: first, including variables strongly associated with intracranial haemorrhage in univariate analysis; second, including cerebral microbleed presence and the HAS-BLED clinical bleeding risk score;21 and third, including cerebral microbleed presence and other neuroimaging markers of small vessel disease. We assessed the proportional hazards assumption through visual inspection of log-log plots of the log cumulative hazard against log time. We calculated absolute event rates per 1000 patient-years for the primary and the main secondary outcomes. For recurrent ischaemic stroke multivariable analysis, we adjusted for variables that differed between those with and without recurrent ischaemic stroke at the 20% level.

he log cumulative hazard against log time. We calculated absolute event rates per 1000 patient-years for the primary and the main secondary outcomes. For recurrent ischaemic stroke multivariable analysis, we adjusted for variables that differed between those with and without recurrent ischaemic stroke at the 20% level. We developed two prediction models using Cox regression: first, including all predictors associated with intracranial haemorrhage at the 20% level in univariable analysis; and second, including cerebral microbleed presence and HAS-BLED score. We assessed calibration using the Cox calibration slopes, and quantified discrimination using Harrell's C-index. For bootstrapping validation, the models were re-fitted in 1000 bootstrap samples and applied to the original dataset. For each model, we then calculated the calibration slope and optimism-adjusted C-index.22 We also fitted these models using the lasso method23 to investigate possible overfitting. We did all statistical analysis using Stata version 12.0. This study is registered with ClinicalTrials.gov, number NCT02513316. Role of the funding source Neither the funders nor the sponsor had input into study design; data collection, data analysis, data interpretation; writing of the report; or the decision to submit the paper for publication. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Research in context Evidence before this study We did a systematic review of studies of the risk of intracerebral haemorrhage growth, and associations with it, published in OVID MEDLINE (from Jan 1, 1970, to Dec 31, 2015) using a comprehensive search strategy, limited to humans, combining terms for intracerebral haemorrhage (“exp basal ganglia hemorrhage/”, “intracranial hemorrhages/”, “cerebral hemorrhage/”, “intracranial hemorrhage, hypertensive/”, and other text words) with text words suggesting growth (“expansion”, “growth”, or “enlargement”), with no language restrictions. When we updated the search to March 1, 2018, we identified reports of five new cohorts, representing a maximum of a 10% increase in the number of eligible patients compared with those from the 36 cohorts that provided individual patient data in this meta-analysis. We did not include these five new cohorts in our analyses. Intracerebral haemorrhage growth risk is known to be highest soon after intracerebral haemorrhage symptom onset, but its absolute risks over time and by baseline volume are unclear. Studies have identified several risk factors associated with intracerebral haemorrhage growth, but many associations are not consistent across studies, and the predictive values of these risk factors remain to be determined. Added value of this study

ers nor the sponsor had input into study design; data collection, data analysis, data interpretation; writing of the report; or the decision to submit the paper for publication. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between Aug 3, 2011, and July 31, 2015, 1686 potentially eligible patients consented from 79 centres across the UK and one centre in the Netherlands. After neuroimaging quality assurance, our final analysis included 1490 participants (1294 [87%] with 1·5 Tesla and 196 [13%] with 3 Tesla MRI scans); patient flow through the study is shown in figure 1. We found no significant differences in demographics, stroke risk factors, or stroke severity between patients included in the final analysis compared with those who gave consent and were screened but were ineligible or excluded. We collected screening logs from 26 sites (1120 patients) to assess selection bias; compared with the 506 patients who were included in the final analysis, the 614 patients who were eligible but did not consent were older (mean age 80 years [SD 11] vs 75 years [10]; p<0·0001), more likely to be female (252 [55%] of 460 patients vs 213 [42%] of 506 patients; p<0·0001), and had more severe strokes (median baseline National Institutes of Health Stroke Score 8 [IQR 3–16] vs 5 [2–10], p<0·0001).Figure 1 Participant flow

ot consent were older (mean age 80 years [SD 11] vs 75 years [10]; p<0·0001), more likely to be female (252 [55%] of 460 patients vs 213 [42%] of 506 patients; p<0·0001), and had more severe strokes (median baseline National Institutes of Health Stroke Score 8 [IQR 3–16] vs 5 [2–10], p<0·0001).Figure 1 Participant flow Of 1490 patients included, 1447 (97%) had follow-up information available from any time during the 24 months (figure 1). The 43 patients without follow-up did not differ from those with follow-up in age (76 years [SD 10] vs 73 years [11]; p=0·166), hypertension (23 [58%] of 40 patients vs 907 [64%] of 1427 patients; p=0·43), or cerebral microbleed prevalence (seven [16%] of 43 patients vs 304 [21%] of 1447 patients; p=0·452). Cerebral microbleeds were present in 311 (21%) of 1490 participants included in the final analyses, with a median cerebral microbleed count of one (IQR 1–3; table 1). Intra-rater and inter-rater reliability for the presence of cerebral microbleeds were excellent (intra-rater κ 0·93, 95% CI 0·86–1·00 and inter-rater 0·85, 0·74–0·96). Cerebral microbleeds were strictly lobar in 116 patients, strictly non-lobar (deep) in 120 patients, and mixed in 75 patients. 46 (3%) patients fulfilled the modified Boston criteria for cerebral amyloid angiopathy.18 Cortical superficial siderosis was identified in five patients, of which one was considered to be disseminated (table 1). 432 (29%) patients had severe white matter hyperintensities (ARWMC score15 ≥2 in either basal ganglia or white matter regions).Table 1 Baseline characteristics

ria for cerebral amyloid angiopathy.18 Cortical superficial siderosis was identified in five patients, of which one was considered to be disseminated (table 1). 432 (29%) patients had severe white matter hyperintensities (ARWMC score15 ≥2 in either basal ganglia or white matter regions).Table 1 Baseline characteristics All patients (n=1490) Patients with cerebral microbleeds (n=311) Patients without cerebral microbleeds (n=1179) Age, years 76 (10) 78 (10) 75 (10) Sex Female 631 (42%) 129 (41%) 502 (43%) Male 859 (58%) 182 (59%) 677 (57%) Hypertension 930/1467 (63%) 212/303 (70%) 718/1164 (62%) Hyperlipidaemia 661/1469 (45%) 145/307 (47%) 516/1162 (44%) Diabetes 251 (17%) 55 (18%) 196 (17%) Ischaemic heart disease 243 (16%) 66 (21%) 177 (15%) Previous ischaemic stroke 142 (10%) 41 (13%) 101 (9%) Previous intracerebral haemorrhage 8 (1%) 3 (1%) 5 (<1%) Alcohol use Units per week 2 (0–9) 2 (0–7) 2 (0–10) >14 units per week 213/1384 (15%) 43 (15%) 170/1092 (16%) Congestive heart failure 60 (4%) 20 (6%) 40 (3%) Abnormal renal function 174 (12%) 46 (15%) 128 (11%) Ethnicity White 1414 (95%) 290 (95%) 1124 (95%) Asian* 33 (2%) 10 (3%) 23 (2%) Black 20 (1%) 5 (2%) 15 (1%) C-reactive protein, mg/L 4·6 (2·0–12·0) 4·4 (2·0–12·0) 4·9 (2·0–11·2) Platelet count 221 (185–265) 221 (185–265) 222 (183–265) HAS-BLED score 3 (2–3) 3 (2–4) 3 (3–4) CHA2DS2VASc score 5 (4–6) 5 (4–6) 5 (4–6) Anticoagulation started 1436 (96%) 300 (96%) 1136 (96%) Anticoagulant used DOAC 542/1436 (37%) 121/300 (40%) 421/1136 (37%) VKA 894/1436 (62%) 179/300 (60%) 715/1136 (63%) Concurrent antiplatelet use 57/894 (6%) 9 (3%) 48 (4%) Poor time in therapeutic range† 133/894 (15%) 24/179 (13%) 109/715 (15%) Anticoagulation stopped during follow-up 55/1436 (4%) 13/300 (4%) 42/1136 (4%) Total white matter hyperintensity (ARWMC) score 1 (0–3) 2 (1–4) 1 (0–3) Cerebral microbleeds .. 1 (1–3);range 1–107 NA cSS presence 5 (<1%) 1 (<1%) 4 (<1%) Data are n (%), n/N (%), mean (SD), or median (IQR). DOAC=direct oral anticoagulant; cSS=cortical superficial siderosis. ARWMC=age-related white matter changes. NA=not applicable. VKA=vitamin K antagonist.

score 1 (0–3) 2 (1–4) 1 (0–3) Cerebral microbleeds .. 1 (1–3);range 1–107 NA cSS presence 5 (<1%) 1 (<1%) 4 (<1%) Data are n (%), n/N (%), mean (SD), or median (IQR). DOAC=direct oral anticoagulant; cSS=cortical superficial siderosis. ARWMC=age-related white matter changes. NA=not applicable. VKA=vitamin K antagonist. * Asian denotes Indian, Pakistani, Bangladeshi, or “any other Asian background”. † Poor time in therapeutic range for VKA use was defined as <60%. The 1447 patients with follow-up data available provided 3366 patient-years of follow-up data (mean follow-up 850 days, SD 373). In this population, there were 14 symptomatic intracranial haemorrhages: 11 intracerebral haemorrhages, two subdural haemorrhages, and one subarachnoid haemorrhage. Compared with those who remained free of intracranial haemorrhage, patients who had a symptomatic intracranial haemorrhage during follow-up had a higher prevalence of diabetes, were more likely to have been treated with a VKA than a DOAC, and more likely to have cerebral microbleeds and cortical superficial siderosis (table 2). In the seven patients with a documented international normalised ratio at the time of the intracranial haemorrhage, the median international normalised ratio was 1·9 (IQR 1·4–4·0, range 1·1–4·8).Table 2 Characteristics of patients with and without symptomatic intracranial haemorrhage at follow-up

l siderosis (table 2). In the seven patients with a documented international normalised ratio at the time of the intracranial haemorrhage, the median international normalised ratio was 1·9 (IQR 1·4–4·0, range 1·1–4·8).Table 2 Characteristics of patients with and without symptomatic intracranial haemorrhage at follow-up Patients with symptomatic intracranial haemorrhage (n=14) Patients without symptomatic intracranial haemorrhage (n=1433) p value Age, years 79 (10) 76 (10) 0·322 Sex .. .. 0·620 Female 5 (36%) 606 (42%) .. Male 9 (64%) 827 (58%) .. Hypertension 8 (57%) 898/1411 (64%) 0·615 Hyperlipidaemia 8 (57%) 629/1413 (45%) 0·344 Diabetes 6 (43%) 236 (16%) 0·0086 Ischaemic heart disease 1 (7%) 238 (17%) 0·343 Previous ischaemic stroke 2 (14%) 138 (10%) 0·500 Previous intracerebral haemorrhage 0 8 (1%) 1·0 Alcohol use Units per week 1·5 (0·0–5·0) 2 (0–9) 0·515 >14 units per week 1/12 (8%) 205/1339 (15%) 0·496 Congestive heart failure 0 59 (4%) 0·440 Abnormal renal function 2 (14%) 169 (12%) 0·774 Ethnicity White 14 (100%) 1356 (95%) .. Non-white 0 46 (3%) 0·492 Asian* 0 29 (2%) .. Black 0 17 (1%) .. C-reactive protein, mg/L 5·5 (4·6–16·2) 4·4 (2·0–12·0) 0·113 Platelet count 212 (167–225) 220 (185–264) 0·252 CHA2DS2VASc score 6 (4–6) 5 (4–6) 0·224 HAS-BLED score 2 (2–3) 3 (2–3) 0·144 Anticoagulation started 14 (100%) 1385 (97%) 0·485 Anticoagulant used .. .. 0·071 DOAC 2 (14%) 523/1385 (38%) .. VKA 12 (86%) 862/1385 (62%) .. Concurrent antiplatelets 1 (7%) 56 (4%) 0·536 Poor time in therapeutic range† 0 133/862 (15%) 0·145 Total white matter hyperintensity (ARWMC) score 1·5 (0·0–5·0) 1 (0·0–3·0) 0·968 Cerebral microbleed presence 7 (50%) 297 (21%) 0·0075 Median 0·5 (0·0–3·0) 0 (0·0–0·0) 0·0034 Range 0–12 0–107 NA cSS presence 1 (7%) 4 (<1%) <0·0001 Data are n (%), n/N (%), mean (SD), or median (IQR). Follow-up was at any time during the 24 months after enrolment, with a minimum of 6 months. DOAC=direct oral anticoagulant. cSS=cortical superficial siderosis. ARWMC=age-related white matter changes. NA=not applicable. VKA=vitamin K antagonist.

1 (7%) 4 (<1%) <0·0001 Data are n (%), n/N (%), mean (SD), or median (IQR). Follow-up was at any time during the 24 months after enrolment, with a minimum of 6 months. DOAC=direct oral anticoagulant. cSS=cortical superficial siderosis. ARWMC=age-related white matter changes. NA=not applicable. VKA=vitamin K antagonist. * Asian denotes Indian, Pakistani, Bangladeshi, or “any other Asian background”. † Poor time in the therapeutic range for VKA use was defined as <60%. The symptomatic intracranial haemorrhage event rate in patients with cerebral microbleeds was 9·8 per 1000 patient-years (95% CI 4·0–20·3) compared with 2·6 per 1000 patient-years (95% CI 1·1–5·4) in those without cerebral microbleeds; the absolute rate increase associated with cerebral microbleeds was 7·2 per 1000 patient-years (95% CI 2·9–14·9; table 3).Table 3 Absolute event rates, absolute risks, and univariable and multivariable hazard ratios for symptomatic intracranial haemorrhage and recurrent ischaemic stroke during follow-up, according to baseline presence and burden of cerebral microbleeds

microbleeds was 7·2 per 1000 patient-years (95% CI 2·9–14·9; table 3).Table 3 Absolute event rates, absolute risks, and univariable and multivariable hazard ratios for symptomatic intracranial haemorrhage and recurrent ischaemic stroke during follow-up, according to baseline presence and burden of cerebral microbleeds Absolute event rate* Rate per 1000 patient-years (95% CI) Absolute rate increase per 1000 patient-years (95% CI) Univariable hazard ratio (95% CI) Adjusted hazard ratio (95% CI)† Symptomatic intracranial haemorrhage No cerebral microbleeds 7/2654 2·6 (1·1 to 5·4) 1 (ref) 1 (ref) 1 (ref) Cerebral microbleeds present 7/712 9·8 (4·0 to 20·3) 7·2 (2·9 to 14·9) 3·73 (1·31 to 10·64) 3·67 (1·27 to 10·60) 1 cerebral microbleed 2/367 5·4 (0·7 to 19·7) 2·8 (−0·4 to 14·3) 2·04 (0·42 to 9·84) 2·03 (0·42 to 9·83) ≥2 cerebral microbleeds 5/345 14·4 (4·7 to 33·8) 11·8 (3·6 to 28·4) 5·58 (1·77 to 17·58) 5·46 (1·70 to 17·51) Recurrent ischaemic stroke No cerebral microbleeds 39/2608 15·0 (10·6 to 20·4) 1 (ref) 1 (ref) 1 (ref) Cerebral microbleeds present 17/704 24·1 (14·1 to 38·7) 9·1 (3·5 to 18·3) 1·62 (0·92 to 2·87) 1·53 (0·85 to 2·76) 1 cerebral microbleed 9/362 24·9 (11·4 to 47·2) 9·9 (0·8 to 32·2) 1·68 (0·82 to 3·47) 1·75 (0·84 to 3·65) ≥2 cerebral microbleeds 8/341 23·4 (10·1 to 46·2) 8·4 (−0·5 to 25·8) 1·56 (0·73 to 3·35) 1·32 (0·60 to 2·93) Data are calculated on the 1447 participants with follow-up data available. * Calculated as number of events/patient-years.

Absolute event rate* Rate per 1000 patient-years (95% CI) Absolute rate increase per 1000 patient-years (95% CI) Univariable hazard ratio (95% CI) Adjusted hazard ratio (95% CI)† Symptomatic intracranial haemorrhage No cerebral microbleeds 7/2654 2·6 (1·1 to 5·4) 1 (ref) 1 (ref) 1 (ref) Cerebral microbleeds present 7/712 9·8 (4·0 to 20·3) 7·2 (2·9 to 14·9) 3·73 (1·31 to 10·64) 3·67 (1·27 to 10·60) 1 cerebral microbleed 2/367 5·4 (0·7 to 19·7) 2·8 (−0·4 to 14·3) 2·04 (0·42 to 9·84) 2·03 (0·42 to 9·83) ≥2 cerebral microbleeds 5/345 14·4 (4·7 to 33·8) 11·8 (3·6 to 28·4) 5·58 (1·77 to 17·58) 5·46 (1·70 to 17·51) Recurrent ischaemic stroke No cerebral microbleeds 39/2608 15·0 (10·6 to 20·4) 1 (ref) 1 (ref) 1 (ref) Cerebral microbleeds present 17/704 24·1 (14·1 to 38·7) 9·1 (3·5 to 18·3) 1·62 (0·92 to 2·87) 1·53 (0·85 to 2·76) 1 cerebral microbleed 9/362 24·9 (11·4 to 47·2) 9·9 (0·8 to 32·2) 1·68 (0·82 to 3·47) 1·75 (0·84 to 3·65) ≥2 cerebral microbleeds 8/341 23·4 (10·1 to 46·2) 8·4 (−0·5 to 25·8) 1·56 (0·73 to 3·35) 1·32 (0·60 to 2·93) Data are calculated on the 1447 participants with follow-up data available. * Calculated as number of events/patient-years. † Adjusted for age and hypertension for symptomatic intracranial haemorrhage, and adjusted for age, sex, hypertension, diabetes, previous ischaemic stroke, and age-related white matter hyperintensities score for recurrent ischaemic stroke.

Absolute event rate* Rate per 1000 patient-years (95% CI) Absolute rate increase per 1000 patient-years (95% CI) Univariable hazard ratio (95% CI) Adjusted hazard ratio (95% CI)† Symptomatic intracranial haemorrhage No cerebral microbleeds 7/2654 2·6 (1·1 to 5·4) 1 (ref) 1 (ref) 1 (ref) Cerebral microbleeds present 7/712 9·8 (4·0 to 20·3) 7·2 (2·9 to 14·9) 3·73 (1·31 to 10·64) 3·67 (1·27 to 10·60) 1 cerebral microbleed 2/367 5·4 (0·7 to 19·7) 2·8 (−0·4 to 14·3) 2·04 (0·42 to 9·84) 2·03 (0·42 to 9·83) ≥2 cerebral microbleeds 5/345 14·4 (4·7 to 33·8) 11·8 (3·6 to 28·4) 5·58 (1·77 to 17·58) 5·46 (1·70 to 17·51) Recurrent ischaemic stroke No cerebral microbleeds 39/2608 15·0 (10·6 to 20·4) 1 (ref) 1 (ref) 1 (ref) Cerebral microbleeds present 17/704 24·1 (14·1 to 38·7) 9·1 (3·5 to 18·3) 1·62 (0·92 to 2·87) 1·53 (0·85 to 2·76) 1 cerebral microbleed 9/362 24·9 (11·4 to 47·2) 9·9 (0·8 to 32·2) 1·68 (0·82 to 3·47) 1·75 (0·84 to 3·65) ≥2 cerebral microbleeds 8/341 23·4 (10·1 to 46·2) 8·4 (−0·5 to 25·8) 1·56 (0·73 to 3·35) 1·32 (0·60 to 2·93) Data are calculated on the 1447 participants with follow-up data available. * Calculated as number of events/patient-years. † Adjusted for age and hypertension for symptomatic intracranial haemorrhage, and adjusted for age, sex, hypertension, diabetes, previous ischaemic stroke, and age-related white matter hyperintensities score for recurrent ischaemic stroke. Using the log-rank test for equality of survivor functions, we found that symptomatic intracranial haemorrhages were more frequent in patients with cerebral microbleeds compared with those without (p=0·0081). In univariable Cox regression, the hazard of symptomatic intracranial haemorrhage for patients with cerebral microbleeds was more than three times higher than that for patients without cerebral microbleeds; this risk was maintained in multivariable Cox regression analysis adjusted for hypertension and age (figure 2; table 3). The risk of symptomatic intracranial haemorrhage increased with increasing cerebral microbleed burden (overall p=0·017 from adjusted Cox regression for categories 0, 1, and ≥2 cerebral microbleeds and overall p=0·032 from unadjusted Cox regression for categories 0, 1, 2–4, and ≥5 cerebral microbleeds; table 3; appendix). We explored cerebral microbleed distribution and rates of symptomatic intracranial haemorrhage, but there were too few events within each category to draw reliable conclusions (appendix).Figure 2 Probability of symptomatic intracranial haemorrhage according to the presence or absence of cerebral microbleeds

appendix). We explored cerebral microbleed distribution and rates of symptomatic intracranial haemorrhage, but there were too few events within each category to draw reliable conclusions (appendix).Figure 2 Probability of symptomatic intracranial haemorrhage according to the presence or absence of cerebral microbleeds The hazard ratio (HR) and 95% CI are derived from the model adjusted for hypertension and age. Of the 1490 patients recruited and identified to start anticoagulation, 1436 (96%) did so (table 1); 54 patients did not start because 12 had died, 13 refused or did not attend their anticoagulation clinic appointments, 17 had medical contraindications, and for 12 patients the reason was not specified. The median time from stroke symptoms until starting anticoagulation was 11 days (IQR 4–17); 894 (60%) of 1490 patients started a VKA and 542 (36%) of 1490 patients started a DOAC. Repeat analyses including only the 1436 anticoagulated participants did not affect our main result (univariable hazard ratio [HR] for cerebral microbleed presence 3·73; 95% CI 1·31–10·63). The type of anticoagulant (DOAC or VKA) did not significantly affect the hazard of symptomatic intracranial haemorrhage associated with cerebral microbleed presence (HR interaction term 0·88; 95% CI 0·04–17·13 p=0·92).

main result (univariable hazard ratio [HR] for cerebral microbleed presence 3·73; 95% CI 1·31–10·63). The type of anticoagulant (DOAC or VKA) did not significantly affect the hazard of symptomatic intracranial haemorrhage associated with cerebral microbleed presence (HR interaction term 0·88; 95% CI 0·04–17·13 p=0·92). There were 56 recurrent ischaemic strokes during 3312 patient-years of follow-up. We observed an increased ischaemic stroke rate of 9·1 per 1000 patient-years (95% CI 3·5–18·3) associated with cerebral microbleeds (table 3). However, cerebral microbleed presence was not associated with recurrent ischaemic stroke in univariable or multivariable analyses (table 3; appendix). Mortality following symptomatic intracranial haemorrhage during follow-up was higher than that of recurrent ischaemic stroke (seven [50%, 95% CI 23–77] of 14 patients vs 12 [21%, 12–34] of 56 patients; p=0·041). In multivariable analysis, when adjusting for age and hypertension, cerebral microbleed presence was associated with symptomatic intracerebral haemorrhage (HR 4·24; 95% CI 1·27–14·08) but not death or the composite outcome (appendix).

seven [50%, 95% CI 23–77] of 14 patients vs 12 [21%, 12–34] of 56 patients; p=0·041). In multivariable analysis, when adjusting for age and hypertension, cerebral microbleed presence was associated with symptomatic intracerebral haemorrhage (HR 4·24; 95% CI 1·27–14·08) but not death or the composite outcome (appendix). In the first prediction model, we included variables that were significant at the 20% level in univariable analyses: cerebral microbleed presence, diabetes, DOAC use, and HAS-BLED score. We excluded cortical superficial siderosis owing to its rarity, and time in therapeutic range for VKA because it is captured within HAS-BLED. Missing alcohol scores for HAS-BLED were imputed using multiple imputation with chained equations24 (ten imputations). Fitting a model with all four predictors (cerebral microbleed presence, diabetes, DOAC use, and HAS-BLED score) produced an optimism-adjusted C-index of 0·74 (95% CI 0·60–0·88). In the second model, we included cerebral microbleed presence and HAS-BLED score (imputed as above), which produced an optimism-adjusted C-index of 0·66 (95% CI 0·53–0·80; see appendix for Cox calibration slopes). Compared with the HAS-BLED score alone (C-index 0·41; 95% CI 0·29–0·53), the first model (C-index difference 0·33, 0·14–0·51; p=0·00059) and the second model (C-index difference 0·25, 0·07–0·43; p=0·0065) were both better in predicting symptomatic intracranial haemorrhage.

·80; see appendix for Cox calibration slopes). Compared with the HAS-BLED score alone (C-index 0·41; 95% CI 0·29–0·53), the first model (C-index difference 0·33, 0·14–0·51; p=0·00059) and the second model (C-index difference 0·25, 0·07–0·43; p=0·0065) were both better in predicting symptomatic intracranial haemorrhage. We undertook three sensitivity analyses to confirm a robust independent association of cerebral microbleed presence with symptomatic intracranial haemorrhage. Because we observed only 14 symptomatic intracranial haemorrhages, we included a maximum of two predictor variables in each analysis. Cerebral microbleed presence remained an independent predictor of intracranial haemorrhage as follows: first, when adjusted for the two strongest univariable predictors (diabetes and anticoagulant type, but not cortical superficial siderosis because of its rarity): HR 3·63; 95% CI 1·27–10·38; second, when adjusted for HAS-BLED score: 5·64, 1·79–17·80; and third, when adjusted for other neuroimaging markers of small vessel disease: HR adjusted for total age-related white matter hyperintensities score 3·69, 1·26–10·74; HR adjusted for any cortical superficial siderosis 4·12, 1·42–11·97 (appendix). For each model, visual inspection of the log-log plots suggested that the proportional hazards assumption was satisfactory.

ers of small vessel disease: HR adjusted for total age-related white matter hyperintensities score 3·69, 1·26–10·74; HR adjusted for any cortical superficial siderosis 4·12, 1·42–11·97 (appendix). For each model, visual inspection of the log-log plots suggested that the proportional hazards assumption was satisfactory. Discussion Our prospective, observational, multicentre cohort of patients anticoagulated after recent ischaemic stroke or transient ischaemic attack associated with atrial fibrillation shows that baseline cerebral microbleed presence is independently associated with an increased risk of symptomatic intracranial haemorrhage, but not of recurrent ischaemic stroke. However, the absolute rate of recurrent ischaemic stroke was much higher than the absolute rate of intracranial haemorrhage, even in those with cerebral microbleeds. We also show that the addition of a neuroimaging biomarker (cerebral microbleed presence) improves the predictive ability of a clinical bleeding risk score (HAS-BLED), which could help clinicians better identify patients at high risk of intracranial haemorrhage.

ial haemorrhage, even in those with cerebral microbleeds. We also show that the addition of a neuroimaging biomarker (cerebral microbleed presence) improves the predictive ability of a clinical bleeding risk score (HAS-BLED), which could help clinicians better identify patients at high risk of intracranial haemorrhage. Our results are consistent with a smaller hospital-based cohort study in Korea of 550 patients with ischaemic stroke and atrial fibrillation25 that reported an increased risk of intracerebral haemorrhage associated with cerebral microbleeds (HR 3·8, 95% CI 1·1–13·1), as well as with a recent aggregate data meta-analysis.12 Our finding that diabetes is independently associated with symptomatic intracranial haemorrhage has not, to the best of our knowledge, been previously reported in ischaemic stroke cohorts. However, a large community-based study reported that diabetes was associated with intracerebral haemorrhage risk (1·59, 1·26–2·02),26 whereas another study27 of older patients (≥75 years of age; median age 82 years) with atrial fibrillation attending an anticoagulation clinic found an association between diabetes and major bleeding (mostly intracranial haemorrhage; odds ratio 4·4, 95% CI 1·3–14·7).

intracerebral haemorrhage risk (1·59, 1·26–2·02),26 whereas another study27 of older patients (≥75 years of age; median age 82 years) with atrial fibrillation attending an anticoagulation clinic found an association between diabetes and major bleeding (mostly intracranial haemorrhage; odds ratio 4·4, 95% CI 1·3–14·7). Our finding that cerebral microbleed presence was not associated with recurrent ischaemic stroke differs from our recent meta-analysis28 of patients with recent ischaemic stroke or transient ischaemic attack, probably because the meta-analysis included mostly patients without atrial fibrillation and treated with antiplatelet therapy. The association of cerebral microbleed presence with future symptomatic intracranial haemorrhage but not ischaemic stroke risk in our cohort supports the hypothesis that cerebral microbleeds are a neuroimaging biomarker of a bleeding-prone arteriopathy specifically relevant for intracranial haemorrhage associated with anticoagulation. However, the relationship between cerebral microbleed presence and recurrent ischaemic stroke risk, while not statistically significant, also favoured a positive association. Thus, cerebral microbleeds, as a marker of overall vascular fragility, might not reliably discriminate between intracranial bleeding and ischaemic stroke risks, but this important question requires further study. Indeed, the absolute event rate of ischaemic stroke in patients with cerebral microbleeds (24·1 per 1000 patient-years) was much higher than the absolute event rate of symptomatic intracranial haemorrhage (9·8 per 1000 patient-years). By contrast with cerebral microbleeds, white matter hyperintensities were not associated with symptomatic intracranial haemorrhage in our study, in keeping with data from two previous smaller similar cohort studies.25, 29

than the absolute event rate of symptomatic intracranial haemorrhage (9·8 per 1000 patient-years). By contrast with cerebral microbleeds, white matter hyperintensities were not associated with symptomatic intracranial haemorrhage in our study, in keeping with data from two previous smaller similar cohort studies.25, 29 Although recent meta-analyses of ischaemic stroke and transient ischaemic attack cohorts have explored the risk of intracerebral haemorrhage in patients with five or more cerebral microbleeds,12, 28 we did not present hazard ratios for this subgroup because of the very low number of participants with high cerebral microbleed counts and of symptomatic intracranial haemorrhage events, which could lead to statistically unreliable results and over-interpretation. Thus, although we found that the rate of symptomatic intracranial haemorrhage increased as cerebral microbleed burden increased (and the rate of recurrent ischaemic stroke remained stable), we could not establish whether a cerebral microbleed burden threshold exists at which the absolute event rate of intracranial haemorrhage exceeds that of ischaemic stroke (ie, where anticoagulation might be associated with net harm as judged by absolute event rates). We found that having a single cerebral microbleed was not associated with a higher hazard of symptomatic intracranial haemorrhage, possibly because one cerebral microbleed reflects only minor small vessel disease, or because of limited inter-rater and intra-rater reliability for one cerebral microbleed.14, 30

t rates). We found that having a single cerebral microbleed was not associated with a higher hazard of symptomatic intracranial haemorrhage, possibly because one cerebral microbleed reflects only minor small vessel disease, or because of limited inter-rater and intra-rater reliability for one cerebral microbleed.14, 30 Most currently available bleeding risk scores (which include clinical risk factors but not neuroimaging biomarkers) show only modest predictive value for intracranial haemorrhage with C-indexes of about 0·5,31 although a post-hoc analysis32 of the ROCKET-AF study suggested that including more detailed quantitative factors (eg, platelet count, albumin, diastolic blood pressure) might also improve the predictive performance. Our findings suggest that adding cerebral microbleed presence as a neuroimaging biomarker to a widely used clinical risk score (HAS-BLED) might improve specificity and sensitivity in identifying ischaemic stroke and transient ischaemic attack patients at high risk of intracranial haemorrhage; this knowledge should allow better informed counselling, closer follow-up of high-risk individuals, rational anticoagulant choice, consideration of non-anticoagulant treatment options (eg, left atrial appendage occlusion), and more aggressive management of modifiable risk factors for intracranial haemorrhage (eg, hypertension, anticoagulant monitoring, and compliance). Large-scale collaborations are required to refine and validate robust risk prediction scores. Meanwhile, our findings suggest that future risk scores to identify patients with stroke at risk of intracranial haemorrhage should include cerebral microbleeds as a neuroimaging biomarker in addition to clinical parameters.

a single trained observer. We followed up 97% of our cohort, and experienced observers adjudicated all primary events blinded to baseline cerebral microbleed presence. We undertook survival analysis to account for baseline confounding factors and varying follow-up, and followed a prespecified statistical analysis plan. We also acknowledge our study's limitations. Our cohort is likely to be affected by selection bias because patients with more severe strokes were less likely to be enrolled. Nevertheless, our cohort is likely to be representative of patients considered for anticoagulation soon after ischaemic stroke. All local investigators agreed to a policy of making anticoagulation decisions without considering cerebral microbleeds, but it was not possible to mandate and monitor blinding at the 79 participating centres. Although bias remains possible owing to the absence of formal blinding, 96% of all recruited patients with satisfactory MRI sequences were started on anticoagulants regardless of cerebral microbleed status. The proportion of DOAC use was similar in patients with cerebral microbleeds (40%) and without cerebral microbleeds (36%), suggesting that cerebral microbleeds did not influence the choice of VKA or DOAC. Because most participants in CROMIS-2 took VKA, our findings might not be generalisable to health-care settings where DOACs are the most widely used anticoagulant. Our study had a low rate of symptomatic intracranial haemorrhage, limiting our ability to adjust for multiple confounders, the robustness of our risk prediction models and, importantly, our ability to determine how increasing cerebral microbleed burden might relate to intracranial haemorrhage risk. Although we standardised parameters for MRI, different scanners with different magnetic field strengths were used, which can influence cerebral microbleed detection.33 Furthermore, T2*-weighted GRE MRI sequences are less sensitive to cerebral microbleeds than is susceptibility-weighted imaging,34 so our interpretation of cerebral microbleed-related risk might not generalise to susceptibility-weighted imaging data. Treatment decisions might be influenced by clinical nihilism about intracranial haemorrhage compared with ischaemic stroke; judgment of different apparent severities of incident intracranial haemorrhage compared with ischaemic stroke might in part be artefacts of clinical behaviour.

usceptibility-weighted imaging data. Treatment decisions might be influenced by clinical nihilism about intracranial haemorrhage compared with ischaemic stroke; judgment of different apparent severities of incident intracranial haemorrhage compared with ischaemic stroke might in part be artefacts of clinical behaviour. The low incidence of symptomatic intracranial haemorrhage in patients with ischaemic stroke or transient ischaemic attack anticoagulated for atrial fibrillation makes randomised controlled trials in this field challenging. However, large-scale international pooled collaborative observational cohort analyses should help to refine risk prediction and determine whether high cerebral microbleed counts might be associated with an increased risk of intracranial haemorrhage sufficient to clearly identify patients at risk of net harm from oral anticoagulation. Supplementary Material Supplementary appendix Acknowledgments CROMIS-2 was jointly funded by the Stroke Association and the British Heart Foundation and supported by researchers at the National Institute for Health Research (NIHR) University College London Hospitals Biomedical Research Centre. University College London acted as the sponsor for CROMIS-2, with responsibility for study conduct and management. We thank the principal investigators, research practitioners and patients involved in the CROMIS-2 study, which was supported by the NIHR Clinical Research Network.

on Hospitals Biomedical Research Centre. University College London acted as the sponsor for CROMIS-2, with responsibility for study conduct and management. We thank the principal investigators, research practitioners and patients involved in the CROMIS-2 study, which was supported by the NIHR Clinical Research Network. Contributors MMB, HRJ, and DJW conceived the study. GA, CS, MMB, RA-SS, GYHL, HC, HH, TAY, HRJ, and DJW contributed to study design. GA oversaw the statistical analysis plan. KWM chaired the steering committee. CS, AC, and DJW contributed to study set-up. DW, CS, AC, MJW, KH, EF, NS, LJS, EW, and DJW contributed to data acquisition. MMB, DJW, DW, and HRJ contributed to event adjudication. MJW contributed to data management and storage. DW contributed to data quality assurance and data quality analysis. GA, AC, GB, HRJ, and DJW contributed to data analysis. DW, GA, MMB, AC, RA-SS, GYHL, HC, GB, HH, TAY, KWM, HRJ, and DJW contributed to data interpretation. DW drafted the manuscript and all remaining authors critically revised the manuscript. All authors gave final approval for publication.

e and data quality analysis. GA, AC, GB, HRJ, and DJW contributed to data analysis. DW, GA, MMB, AC, RA-SS, GYHL, HC, GB, HH, TAY, KWM, HRJ, and DJW contributed to data interpretation. DW drafted the manuscript and all remaining authors critically revised the manuscript. All authors gave final approval for publication. Declaration of interests GB reports grants from the Rosetrees Trust. MMB reports grants from The Stroke Association. HC reports grants and honoraria (diverted to local charity) from Bayer. GYHL reports consultancy and speaker fees to his institution from Bayer, Bayer/Janssen, Bristol-Myers Squibb/Pfizer, Biotronik, Medtronic, Boehringer Ingelheim, Microlife, Roche, and Daiichi-Sankyo, outside of the submitted work. DJW reports personal fees from Bayer and Amgen. TAY reports grants from the UK Medical Research Council, MS Society of Great Britain and Northern Ireland, Stroke Association, British Heart Foundation, and Wellcome Trust; personal fees for consultancy or travel from Biogen Idec, IXICO Technologies, European School of Radiology, and Hikma; and has participated in clinical trials funded by Biogen Idec, GlaxoSmithKline, Novartis, Merck, and MS Society Australia. All other authors declare no competing interests.

We did a systematic review of studies of the risk of intracerebral haemorrhage growth, and associations with it, published in OVID MEDLINE (from Jan 1, 1970, to Dec 31, 2015) using a comprehensive search strategy, limited to humans, combining terms for intracerebral haemorrhage (“exp basal ganglia hemorrhage/”, “intracranial hemorrhages/”, “cerebral hemorrhage/”, “intracranial hemorrhage, hypertensive/”, and other text words) with text words suggesting growth (“expansion”, “growth”, or “enlargement”), with no language restrictions. When we updated the search to March 1, 2018, we identified reports of five new cohorts, representing a maximum of a 10% increase in the number of eligible patients compared with those from the 36 cohorts that provided individual patient data in this meta-analysis. We did not include these five new cohorts in our analyses. Intracerebral haemorrhage growth risk is known to be highest soon after intracerebral haemorrhage symptom onset, but its absolute risks over time and by baseline volume are unclear. Studies have identified several risk factors associated with intracerebral haemorrhage growth, but many associations are not consistent across studies, and the predictive values of these risk factors remain to be determined. Added value of this study Our systematic review led to the pooling of 5435 eligible patients from 36 cohorts, which is, to the best of our knowledge, the largest patient-level meta-analysis to explore the absolute risk and predictors of intracerebral haemorrhage growth. We found that the risks of growth over time and by baseline intracerebral haemorrhage volume were not linear. The sample size enabled us to model these associations with good precision and construct and validate multivariable models adjusted for 13 categorical or continuous covariates. Four predictors (time from symptom onset to baseline imaging, intracerebral haemorrhage volume on baseline imaging, antiplatelet use, and anticoagulant use) were independent predictors of intracerebral haemorrhage growth (C-index 0·78, 95% CI 0·75–0·82). Addition of information about the presence of spot sign on CT angiography to the model increased the C-index by just 0·05 (95% CI 0·03–0·07).

haemorrhage volume on baseline imaging, antiplatelet use, and anticoagulant use) were independent predictors of intracerebral haemorrhage growth (C-index 0·78, 95% CI 0·75–0·82). Addition of information about the presence of spot sign on CT angiography to the model increased the C-index by just 0·05 (95% CI 0·03–0·07). Implications of all the available evidence Models using four or five predictors that are simple to collect had acceptable to good discrimination for predicting intracerebral haemorrhage growth, which was slightly improved by the addition of information on spot sign from CT angiography. These models could guide the monitoring of patients at risk of clinical deterioration as well as the interpretation and investigation of treatment effects in randomised trials. Introduction Haemorrhagic stroke is responsible for around 11% of strokes in high-income countries but 22% of strokes in low-income and middle-income countries,1 where 75% of deaths due to haemorrhagic stroke occur.2 Spontaneous (non-traumatic) intracerebral haemorrhage is the most frequent type of haemorrhagic stroke and has the worst outcome: almost half of patients die within the first month and 80% of survivors are dependent on a caregiver.3

me and middle-income countries,1 where 75% of deaths due to haemorrhagic stroke occur.2 Spontaneous (non-traumatic) intracerebral haemorrhage is the most frequent type of haemorrhagic stroke and has the worst outcome: almost half of patients die within the first month and 80% of survivors are dependent on a caregiver.3 Intracerebral haemorrhage volume increases after vessel rupture and growth can continue after intracerebral haemorrhage is first diagnosed on brain imaging. Intracerebral haemorrhage growth is associated with poor clinical outcome.4 Therefore, immediately after confirmation of intracerebral haemorrhage diagnosis on brain imaging, accurate prediction of the risk of later intracerebral haemorrhage growth could help to target patients' monitoring, treatment and transfer to specialist care, and the design and interpretation of randomised trials of treatments to limit intracerebral haemorrhage growth.5

Introduction The prevalence of patent foramen ovale (PFO) is increased in cryptogenic transient ischaemic attack and stroke,1, 2, 3 and three recent trials4, 5, 6 showed better outcomes after percutaneous closure than after medical treatment alone. These trials included patients aged 60 years or younger, with mostly non-disabling events and PFO or an associated interatrial septal aneurysm (appendix).4, 5, 6 Clinical trials in older patients would be justified if an association between PFO and cryptogenic cerebrovascular events was shown at older ages, but existing evidence is conflicting,2, 7, 8 and the need for more data from older patients has been highlighted,9 along with a need for a better evidence base on the most appropriate screening strategy. Research in context Evidence before this study

Introduction The prevalence of patent foramen ovale (PFO) is increased in cryptogenic transient ischaemic attack and stroke,1, 2, 3 and three recent trials4, 5, 6 showed better outcomes after percutaneous closure than after medical treatment alone. These trials included patients aged 60 years or younger, with mostly non-disabling events and PFO or an associated interatrial septal aneurysm (appendix).4, 5, 6 Clinical trials in older patients would be justified if an association between PFO and cryptogenic cerebrovascular events was shown at older ages, but existing evidence is conflicting,2, 7, 8 and the need for more data from older patients has been highlighted,9 along with a need for a better evidence base on the most appropriate screening strategy. Research in context Evidence before this study We searched MEDLINE for articles published before Sept 1, 2017, with no language restrictions, reporting on age-specific prevalence of patent foramen ovale (PFO) in cryptogenic stroke compared with strokes of known cause and including older patients. We used the terms “stroke”, “CVA”, “cryptogenic stroke”, “undetermined stroke”, “stroke of undetermined aetiology”, “embolic stroke of undetermined source”, “patent foramen ovale”, “PFO”, “atrial septal abnormality”, “interatrial septal abnormality”, and “right-to-left shunt”. We also hand-searched international registries, reference lists of systematic reviews, and appropriate journals. We found that published evidence on the older population (age range 40 to ≥55 years) was exclusively hospital-based, focusing mainly on major stroke, and was mostly based on transoesophageal echocardiography, which is not the ideal screening method as up to a third of older stroke patients cannot undergo this procedure. Overall, evidence on the association between PFO and cryptogenic events in older patients was heterogeneous, especially when pooling smaller studies. No population-based data were found, and only two small studies of 210 and 203 patients used contrast-enhanced transcranial Doppler (bubble-TCD) as a screening method.

e. Overall, evidence on the association between PFO and cryptogenic events in older patients was heterogeneous, especially when pooling smaller studies. No population-based data were found, and only two small studies of 210 and 203 patients used contrast-enhanced transcranial Doppler (bubble-TCD) as a screening method. Added value of this study This is the first population-based study investigating the association between PFO and cryptogenic events in a large number of patients, irrespective of age, but with clinical characteristics otherwise more similar to patients enrolled in trials on PFO closure (ie, mainly non-disabling events) than in previous stroke unit-based studies. Our study showed that bubble-TCD is possible in most older patients with transient ischaemic attack or non-disabling stroke, providing a feasible PFO screening strategy to reduce the need for transoesophageal echocardiography. We showed that PFO is significantly associated with cryptogenic events in older patients, consistent with previous evidence from transoesophageal echocardiography-based studies. Extrapolation from our results suggests that the burden of large PFO-related events at the population level is high, with as many as 8477 patients in the UK having large PFOs and cryptogenic transient ischaemic attack or non-disabling stroke per year, most of whom (61%, 5951 patients) are older than 60 years. Implications of all the available evidence

This is the first population-based study investigating the association between PFO and cryptogenic events in a large number of patients, irrespective of age, but with clinical characteristics otherwise more similar to patients enrolled in trials on PFO closure (ie, mainly non-disabling events) than in previous stroke unit-based studies. Our study showed that bubble-TCD is possible in most older patients with transient ischaemic attack or non-disabling stroke, providing a feasible PFO screening strategy to reduce the need for transoesophageal echocardiography. We showed that PFO is significantly associated with cryptogenic events in older patients, consistent with previous evidence from transoesophageal echocardiography-based studies. Extrapolation from our results suggests that the burden of large PFO-related events at the population level is high, with as many as 8477 patients in the UK having large PFOs and cryptogenic transient ischaemic attack or non-disabling stroke per year, most of whom (61%, 5951 patients) are older than 60 years. Implications of all the available evidence In view of the results of published trials of PFO closure in younger patients showing an advantage of closure over medical therapy alone in preventing recurrent strokes and our results showing that right-to-left shunt was associated with cryptogenic events in older patients, age restrictions on access to diagnostic or therapeutic procedures in older patients with cryptogenic transient ischaemic attack or stroke should not prevent the necessary further research and randomised trials of PFO closure at older ages.

t right-to-left shunt was associated with cryptogenic events in older patients, age restrictions on access to diagnostic or therapeutic procedures in older patients with cryptogenic transient ischaemic attack or stroke should not prevent the necessary further research and randomised trials of PFO closure at older ages. Transoesophageal echocardiography has been considered the gold standard for diagnosis of PFO,10, 11 and has been used in several recent trials.4, 5, 6 However, up to a third of older patients (>60 years) with stroke cannot undergo this procedure because of severity of stroke, dysphagia, excessive gag reflex, or refusal of consent.7, 12 Alternatively, contrast-enhanced transcranial Doppler (bubble-TCD) is a non-invasive, bedside, repeatable technique with high sensitivity and specificity for PFO detection in younger patients (≤60 years) and controls,13 through identification of right-to-left shunt (RLS). A small proportion of RLSs are due to non-PFO sources, usually pulmonary shunts, but the majority (around 95%) of TCD-detected RLSs are shown to be due to PFO on transoesophageal echocardiography.13, 14 An unsuitable temporal bone window limits the use of bubble-TCD in 10% of cases, but transoccipital insonation of the posterior circulation is a sensitive alternative to minimise screening failure and detect RLS.15

(around 95%) of TCD-detected RLSs are shown to be due to PFO on transoesophageal echocardiography.13, 14 An unsuitable temporal bone window limits the use of bubble-TCD in 10% of cases, but transoccipital insonation of the posterior circulation is a sensitive alternative to minimise screening failure and detect RLS.15 We previously reported an apparently low prevalence of markers of possible cardioembolic cause in patients with cryptogenic transient ischaemic attack and stroke at all ages,16 but we did not systematically screen for RLS. Therefore, we did a bubble-TCD-based screening study for RLS in patients with transient ischaemic attack or non-disabling stroke in a large population-based cohort, and extrapolated data to the UK population. We also did a systematic review of published studies reporting the prevalence of PFO in cryptogenic transient ischaemic attack or stroke compared with other causes stratified by age, and pooled our data in a meta-analysis.

-disabling stroke in a large population-based cohort, and extrapolated data to the UK population. We also did a systematic review of published studies reporting the prevalence of PFO in cryptogenic transient ischaemic attack or stroke compared with other causes stratified by age, and pooled our data in a meta-analysis. Methods Study design and participants Our study was nested in the Oxford Vascular Study (OXVASC),17 an ongoing population-based study of the incidence and outcome of all acute vascular events in a population of 92 728 individuals, irrespective of age, registered with 100 primary care physicians in nine practices in Oxfordshire, UK. Multiple methods of ascertainment are used to enrol patients with transient ischaemic attack or stroke, as detailed elsewhere (appendix).17 These methods include a daily, rapid-access transient ischaemic attack and stroke clinic, to which participating physicians and the local emergency department refer individuals with suspected transient ischaemic attack or non-disabling stroke. As part of the OXVASC phenotyped cohort, consecutive eligible patients attending this clinic with an acute event, or at 1-month follow-up after an inpatient admission, were enrolled in PFO screening. Patients were eligible for PFO screening if they attended the transient ischaemic attack or stroke clinic or the 1-month follow-up clinic, had a diagnosis of transient ischaemic attack or stroke, and were able to undergo bubble-TCD.

, or at 1-month follow-up after an inpatient admission, were enrolled in PFO screening. Patients were eligible for PFO screening if they attended the transient ischaemic attack or stroke clinic or the 1-month follow-up clinic, had a diagnosis of transient ischaemic attack or stroke, and were able to undergo bubble-TCD. The OXVASC study and TCD assessment were approved by the local ethics committee and written informed consent was obtained from all participants, or assent was obtained from relatives in the case of cognitive impairment or speech difficulty.

, or at 1-month follow-up after an inpatient admission, were enrolled in PFO screening. Patients were eligible for PFO screening if they attended the transient ischaemic attack or stroke clinic or the 1-month follow-up clinic, had a diagnosis of transient ischaemic attack or stroke, and were able to undergo bubble-TCD. The OXVASC study and TCD assessment were approved by the local ethics committee and written informed consent was obtained from all participants, or assent was obtained from relatives in the case of cognitive impairment or speech difficulty. Procedures Patients were assessed by a neurologist or stroke physician and all presentations and investigations were reviewed by the senior study neurologist (PMR). Demographic data, atherosclerotic risk factors (ie, male sex, history of hypertension, diabetes, smoking, or hypercholesterolaemia), and history of coronary or peripheral vascular disease were recorded during face-to-face interviews and cross-referenced with primary care records.16 Patients routinely had 12-lead electrocardiography (ECG) and routine blood testing (ie, full blood count, clotting, C-reactive protein, erythrocyte sedimentation rate, liver function, renal function, thyroid function, electrolytes, and lipid profile) after the event. All patients had MRI brain and vascular imaging when not contraindicated (3T MRI with time-of-flight magnetic resonance angiography [MRA] of the intracranial vessels and a contrast-enhanced MRA of the large neck arteries), or brain CT with contrast-enhanced CT angiography, or Duplex ultrasound if MRI was contraindicated. Patients with cryptogenic transient ischaemic attack or stroke, or those younger than 55 years, also had thrombophilia screening, vasculitis screening, and genetic tests when clinically indicated. Clinical investigation was completed with 5-day ambulatory ECG recording (R test) and transthoracic echocardiography.16

ed. Patients with cryptogenic transient ischaemic attack or stroke, or those younger than 55 years, also had thrombophilia screening, vasculitis screening, and genetic tests when clinically indicated. Clinical investigation was completed with 5-day ambulatory ECG recording (R test) and transthoracic echocardiography.16 Cause of ischaemic events was classified according to the Trial of Org 10172 in Acute Stroke Treatment (TOAST) criteria.18 We classified events as cryptogenic if the diagnostic investigation included at least brain imaging, ECG, transthoracic echocardiography, and complete vascular imaging, and no clear cause was found. Events of known cause included cardioembolic events, large artery disease, small vessel disease, events of other cause, or events of multiple causes. We did not consider PFO alone as a criterion for cardioembolic stroke.16

ECG, transthoracic echocardiography, and complete vascular imaging, and no clear cause was found. Events of known cause included cardioembolic events, large artery disease, small vessel disease, events of other cause, or events of multiple causes. We did not consider PFO alone as a criterion for cardioembolic stroke.16 Contrast-enhanced TCD (bubble-TCD) sonography (Doppler Box; Compumedics DWL, Singen, Germany) was done according to the Consensus Conference of Venice19 by one of two experienced operators (SM and LL), who were masked to the patient's clinical presentation. Agitated saline with addition of 0·5 mL of the patient's blood and 0·5 mL of air was used as a contrast agent in all cases according to accepted guidelines (appendix).19, 20 A large RLS was defined as a shunt with 20 or more microbubbles recorded. Since Nov 15, 2015, if a temporal bone window was not suitable for monitoring, the basilar artery was monitored through a transoccipital approach.15 Designation of RLS status was made at the time of assessment, and recordings were archived.

9, 20 A large RLS was defined as a shunt with 20 or more microbubbles recorded. Since Nov 15, 2015, if a temporal bone window was not suitable for monitoring, the basilar artery was monitored through a transoccipital approach.15 Designation of RLS status was made at the time of assessment, and recordings were archived. Search strategy and selection criteria of the systematic review, and data extraction We did a systematic review and meta-analysis according to the Meta-analysis of Observational Studies in Epidemiology (MOOSE) criteria.21 We aimed to include case-control studies, cohort studies, and population-based studies that included older patients and reported on age-specific prevalence of PFO in cryptogenic stroke and strokes of known cause, irrespective of imaging modality (transthoracic echocardiography, transoesophageal echocardiography, or bubble-TCD). We excluded case reports, but did not limit the search to English language studies.

ncluded older patients and reported on age-specific prevalence of PFO in cryptogenic stroke and strokes of known cause, irrespective of imaging modality (transthoracic echocardiography, transoesophageal echocardiography, or bubble-TCD). We excluded case reports, but did not limit the search to English language studies. We searched MEDLINE using the terms “stroke”, “CVA”, “cryptogenic stroke”, “undetermined stroke”, “stroke of undetermined aetiology”, “embolic stroke of undetermined source”, “patent foramen ovale”, “PFO”, “atrial septal abnormality”, “interatrial septal abnormality”, and “right-to-left shunt” for articles published before Sept 1, 2017. We also hand-searched reference lists of all articles identified in the electronic search, the publications related to the component databases of the Risk of Paradoxical Embolism (RoPE) study,22 and any previous systematic reviews. We also contacted experts in the field, and all screened abstracts and selected papers were in English. Two researchers (SM and LL) independently did the search, and each identified eligible studies. Any discrepancy in relation to inclusion was resolved by majority decision by a third reviewer (PMR). Two researchers (SM and LL) independently extracted data from eligible papers.

tracts and selected papers were in English. Two researchers (SM and LL) independently did the search, and each identified eligible studies. Any discrepancy in relation to inclusion was resolved by majority decision by a third reviewer (PMR). Two researchers (SM and LL) independently extracted data from eligible papers. For studies published more than once (ie, duplicates), we included only the report with the most informative and complete data. In the extracted data, PFO was defined according to the protocol of each individual study. In general, for studies using transthoracic echocardiography or transoesophageal echocardiography, the definition of PFO was based on demonstration of RLS by appearance of contrast microbubbles in the left atrium within three to five cardiac cycles after right atrium opacification.2, 7, 12 For studies using bubble-TCD, a definition based on the current consensus was used.19, 20 Extracted information included PFO screening modality (transthoracic echocardiography, transoesophageal echocardiography, or bubble-TCD), type of event (stroke or transient ischaemic attack), setting, case enrolment (consecutive vs non-consecutive), cryptogenic event out of all events ratio, stroke subtype classification (TOAST vs other), mean age of the population, age stratification, and excluded cases with reasons for exclusion, when provided.

e-TCD), type of event (stroke or transient ischaemic attack), setting, case enrolment (consecutive vs non-consecutive), cryptogenic event out of all events ratio, stroke subtype classification (TOAST vs other), mean age of the population, age stratification, and excluded cases with reasons for exclusion, when provided. Statistical analysis In the OXVASC cohort, baseline characteristics and prevalence of PFO of any size and of large size were compared between cryptogenic events and events of known cause for all patients and for those older than 60 years, using the χ2 test for categorical variables and t test for continuous variables. The numbers of large PFO in patients with cryptogenic events were reported by age and extrapolated to the UK population (data from mid-2016) based on age-specific rates.23 For each study in the systematic review, we established the odds ratio (OR) for PFO of any size in cryptogenic events compared with events of known cause, stratified by screening modality (transthoracic echocardiography, transoesophageal echocardiography, or bubble-TCD). We also compared the detection rate (%) of PFO using different screening modalities stratified by stroke cause (cryptogenic vs known cause) and by age (<70 vs ≥70 years). We derived pooled ORs (95% CI) by meta-analysis, including the OXVASC data, using the Mantel-Haenszel-Peto method (random-effects), with χ2 tests to assess heterogeneity between studies. Statistical analyses were done using Review Manager 5·3, Stata 15, and SPSS 22. The study protocol is registered with PROSPERO, number CRD42018087074.

We derived pooled ORs (95% CI) by meta-analysis, including the OXVASC data, using the Mantel-Haenszel-Peto method (random-effects), with χ2 tests to assess heterogeneity between studies. Statistical analyses were done using Review Manager 5·3, Stata 15, and SPSS 22. The study protocol is registered with PROSPERO, number CRD42018087074. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data in the study and PMR had final responsibility for the decision to submit for publication. Results Among 572 consecutive patients with transient ischaemic attack or non-disabling ischaemic stroke between Sept 1, 2014, and Oct 9, 2017 (439 [77%] patients aged >60 years, mean age 70·0 years [SD 13·7]), bubble-TCD was feasible in 523 (91%) patients. Of the 49 (9%) patients who did not undergo bubble-TCD (appendix), two (4%) were deemed unsuitable for contrast (agitated saline) injection because of pregnancy, 24 (49%) could not tolerate testing (too frail, could not tolerate supine position, or too anxious); seven (14%; between Sept 1, 2014, and Nov 15, 2015, before the start of the transoccipital approach) did not have a suitable temporal bone window, and 16 (33%) had issues related to cannulation (refused or failed cannulation). One patient with a large RLS reported an episode of visual aura without headache, similar to his usual migraine aura, shortly after the injection of microbubbles,24 but no other complications were reported.

a suitable temporal bone window, and 16 (33%) had issues related to cannulation (refused or failed cannulation). One patient with a large RLS reported an episode of visual aura without headache, similar to his usual migraine aura, shortly after the injection of microbubbles,24 but no other complications were reported. Of the 523 patients who underwent bubble-TCD, 397 (76%) were aged older than 60 years and 264 (50%) had cryptogenic events. Patients with cryptogenic events had fewer vascular risk factors, lower prevalence of comorbid atherosclerotic disease, and were more likely to have presented with a transient ischaemic attack than were patients with events of known cause (table 1).Table 1 Baseline characteristics in patients with cryptogenic events versus events of known cause

events had fewer vascular risk factors, lower prevalence of comorbid atherosclerotic disease, and were more likely to have presented with a transient ischaemic attack than were patients with events of known cause (table 1).Table 1 Baseline characteristics in patients with cryptogenic events versus events of known cause Overall Age >60 years Cryptogenic (n=264) Known cause (n=259) p value Cryptogenic (n=190) Known cause (n=207) p value Age 67·3 (7·3) 71·9 (8·1) <0·0001 74·0 (6·9) 77·3 (8·3) <0·0001 Male sex 132 (50%) 151 (58%) 0·06 84 (44%) 118 (57%) 0·01 Index event 0·0001 0·002 Transient ischaemic attack 199 (75%) 154 (59%) .. 150 (79%) 134 (65%) .. Ischaemic stroke 65 (25%) 105 (41%) .. 40 (21%) 73 (35%) .. Previous vascular event Myocardial infarction 10 (4%) 29 (11%) 0·002 9 (5%) 27 (13%) 0·004 Peripheral vascular disease 3 (1%) 21 (8%) 0·001 2 (1%) 20 (10%) 0·0002 Transient ischaemic attack 16 (6%) 39 (15%) 0·001 14 (7%) 39 (19%) 0·001 Stroke 21 (8%) 42 (16%) 0·005 18 (9%) 37 (18%) 0·02 Known vascular risk factors Hypertension 139 (53%) 164 (63%) 0·01 120 (63%) 144 (70%) 0·18 Diabetes 33 (13%) 44 (17%) 0·15 28 (15%) 32 (15%) 0·84 Hyperlipidaemia 95 (36%) 102 (39%) 0·42 81 (43%) 89 (43%) 0·94 Valvular heart disease 9 (3%) 25 (10%) 0·006 8 (4%) 23 (11%) 0·01 Cardiac failure 3 (1%) 17 (7%) 0·005 3 (2%) 16 (8%) 0·004 Venous thrombosis 8 (3%) 14 (5%) 0·18 6 (3%) 14 (7%) 0·10 Atrial fibrillation* 1 (<1%) 111 (43%) <0·0001 1 (1%)* 102 (49%) <0·0001 History of smoking 135 (51%) 148 (57%) 0·17 94 (49%) 113 (55%) 0·31 Current smoker† 35 (13%) 41 (16%) 0·4 17 (9%) 16 (8%)† 0·67 Data are mean (SD) or n (%). Data are stratified by age.

·004 Venous thrombosis 8 (3%) 14 (5%) 0·18 6 (3%) 14 (7%) 0·10 Atrial fibrillation* 1 (<1%) 111 (43%) <0·0001 1 (1%)* 102 (49%) <0·0001 History of smoking 135 (51%) 148 (57%) 0·17 94 (49%) 113 (55%) 0·31 Current smoker† 35 (13%) 41 (16%) 0·4 17 (9%) 16 (8%)† 0·67 Data are mean (SD) or n (%). Data are stratified by age. * Including both history of atrial fibrillation and new atrial fibrillation detected after the index event. One patient with previous history of atrial fibrillation had successful ablation and was in sinus rhythm in repeated 5-day ambulatory cardiac monitoring. † Data missing for one patient.

·004 Venous thrombosis 8 (3%) 14 (5%) 0·18 6 (3%) 14 (7%) 0·10 Atrial fibrillation* 1 (<1%) 111 (43%) <0·0001 1 (1%)* 102 (49%) <0·0001 History of smoking 135 (51%) 148 (57%) 0·17 94 (49%) 113 (55%) 0·31 Current smoker† 35 (13%) 41 (16%) 0·4 17 (9%) 16 (8%)† 0·67 Data are mean (SD) or n (%). Data are stratified by age. * Including both history of atrial fibrillation and new atrial fibrillation detected after the index event. One patient with previous history of atrial fibrillation had successful ablation and was in sinus rhythm in repeated 5-day ambulatory cardiac monitoring. † Data missing for one patient. Overall, we found RLS in 157 (30%) of 523 patients, and large RLS in 68 (13%) patients. Compared with patients with transient ischaemic attack or stroke of known cause, cryptogenic events had a higher prevalence of RLS overall (OR 1·93, 95% CI 1·32–2·82; p=0·001; table 2). Results were consistent when stratified by type of presenting event (transient ischaemic attack: 1·90, 1·19–3·04; p=0·01; ischaemic stroke: 1·98, 1·00–3·90; p=0·05; ischaemic stroke or transient ischaemic attack with corresponding acute lesion on brain imaging: 2·18, 1·18–4·02; p=0·01). We found the same association when analysis was restricted to patients aged older than 60 years (2·06, 1·32–3·23; p=0·001; table 2); this association was independent of RLS size (large size: 2·10, 95% CI 1·27–3·46; small size: 1·77, 1·07–2·92; pdifference=0·76; patients aged >60 years with large RLS: 2·67, 1·40–5·09; small RLS: 1·72, 0·99–2·97; pdifference=0·78) and remained so when cryptogenic events were compared separately with events of cardioembolic, large vessel, or small vessel cause (appendix).Table 2 Prevalence of RLS in patients with cryptogenic events compared with patients with events of known cause

LS: 2·67, 1·40–5·09; small RLS: 1·72, 0·99–2·97; pdifference=0·78) and remained so when cryptogenic events were compared separately with events of cardioembolic, large vessel, or small vessel cause (appendix).Table 2 Prevalence of RLS in patients with cryptogenic events compared with patients with events of known cause Cryptogenic Known cause Odds ratio (95% CI) p value RLS of any size Age ≤60 years 29/74 (39%) 16/52 (31%) 1·45 (0·68–3·07) 0·33 Age >60 years 68/190 (36%) 44/207 (21%) 2·06 (1·32–3·23) 0·001 Total 97/264 (37%) 60/259 (23%) 1·93 (1·32–2·82) 0·001 Large RLS only Age ≤60 years 16/74 (22%) 12/52 (23%) 0·92 (0·39–2·15) 0·85 Age >60 years 25/190 (13%) 15/207 (7%) 1·94 (0·99–3·80) 0·05 Total 41/264 (16%) 27/259 (10%) 1·58 (0·94–2·66) 0·08 Data are n/N (%) unless otherwise indicated. Data are stratified by age and size of the shunt. RLS=right-to-left shunt. If extrapolated to the UK population (data from mid-2016), 41 patients with transient ischaemic attack and non-disabling cryptogenic stroke and a large RLS (25 [61%] of whom were older than 60 years) projected 8477 cases annually in the UK, 5951 (70·2%) of whom would be older than 60 years (appendix). Of the 976 records identified in the systematic review (appendix), we selected 30 potentially relevant papers, of which eight met our criteria for inclusion in our meta-analysis (table 3),7, 8, 12, 25, 26, 27, 28, 29 reporting age-specific prevalence of PFO in cryptogenic stroke and strokes of known cause in a population including older patients.Table 3 Studies included in the meta-analysis

ected 30 potentially relevant papers, of which eight met our criteria for inclusion in our meta-analysis (table 3),7, 8, 12, 25, 26, 27, 28, 29 reporting age-specific prevalence of PFO in cryptogenic stroke and strokes of known cause in a population including older patients.Table 3 Studies included in the meta-analysis Screening modality Event (% stroke) Setting Consecutive cases Cryptogenic events/total events TOAST criteria Mean age (SD), years Age groups, years Excluded cases Hospital-based Di Tullio et al (1992)25 Transthoracic echocardiography Stroke Neurology department No 45/146 (31%) No 61·8 (15·3) <55 and ≥55 34% not referred for transthoracic echocardiography; 6·8% with inadequate transthoracic echocardiography Hausmann et al (1992)26 Transoesophageal echocardiography Transient ischaemic attack or stroke (59·2%) Not reported Not reported 65/103 (63%) No 52 (10) <40 and ≥40 Not reported Jones et al (1994)27 Transoesophageal echocardiography Transient ischaemic attack or stroke (90·5%) Hospital admission Yes 71/220 (33%) No 66 (13) <50 and 50–69 27·6% Handke et al (2007)7 Transoesophageal echocardiography Stroke Stroke unit or intensive care unit Yes 227/503 (45%) Yes 62·2 (13·1) <55 and ≥55 15·6% De Castro et al (2010)12 Transoesophageal echocardiography Transient ischaemic attack or stroke (62·1% major stroke) Stroke unit Yes 403/660 (61%) Yes 64·4 (13·5) <55 and ≥55 38·9% Force et al (2008)8 Transoesophageal echocardiography Transient ischaemic attack or stroke Stroke unit Not reported 62/132 (47%) Not reported 70·7 (8·6) ≥55 Not reported Yeung et al (1996)28 Bubble-TCD Transient ischaemic attack or stroke (70·5%) Hospital admission Yes 116/210 (55%) No Men: 65 (range 12–86); Women: 63 (23–86) ≤50, >50, and ≥70 51% Serena et al (1998)29 Bubble-TCD Transient ischaemic attack or stroke (71·2%) Neurology department Yes 53/203 (26%) No 64·8 (12·3) <50, 51 to <70, and ≥70 20·9% (no temporal bone window) plus 1·9% (no Valsalva done) Population-based OXVASC (2017) Bubble-TCD Transient ischaemic attack or non-disabling stroke (32·5%) Stroke or transient ischaemic attack clinic service Yes 264/523 (50%) Yes 69·6 (13·4) ≤60, >60, <70, and ≥70 8·6% TOAST=Trial of Org 10172 in Acute Stroke Treatment. Bubble-TCD=contrast-enhanced transcranial Doppler. OXVASC=Oxford Vascular Study.

17) Bubble-TCD Transient ischaemic attack or non-disabling stroke (32·5%) Stroke or transient ischaemic attack clinic service Yes 264/523 (50%) Yes 69·6 (13·4) ≤60, >60, <70, and ≥70 8·6% TOAST=Trial of Org 10172 in Acute Stroke Treatment. Bubble-TCD=contrast-enhanced transcranial Doppler. OXVASC=Oxford Vascular Study. Existing studies were exclusively hospital-based and predominantly based on transoesophageal echocardiography; overall, when pooling with results from the OXVASC study, an association between PFO and cryptogenic events was consistently shown for all screening modalities (figure 1). The association between RLS and cryptogenic events in patients older than 60 years found in the OXVASC study (based on 112 detected PFOs) was consistent with two previous smaller bubble-TCD studies of older age groups (based on 4428 and 3129 PFOs; figure 1), yielding a highly significant pooled estimate (OR 2·35, 95% CI 1·42–3·90; p=0·0009; pheterogeneity=0·15). This estimate was also consistent with that derived from studies of transoesophageal echocardiography (2·20, 1·15–4·22; p=0·02; figure 1), although the transoesophageal echocardiography estimate was heterogeneous (pheterogeneity=0·02).Figure 1 Prevalence of PFO in patients with cryptogenic events compared with patients with events of known cause

istent with that derived from studies of transoesophageal echocardiography (2·20, 1·15–4·22; p=0·02; figure 1), although the transoesophageal echocardiography estimate was heterogeneous (pheterogeneity=0·02).Figure 1 Prevalence of PFO in patients with cryptogenic events compared with patients with events of known cause Meta-analyses of the prevalence of PFO in patients with cryptogenic events compared with patients with events of known cause, stratified by imaging modalities, (A) overall and (B) in older patients, according to study author's definition. PFO=patent foramen ovale. OR=odds ratio. Bubble-TCD=contrast-enhanced transcranial Doppler. OXVASC=Oxford Vascular Study. *Age cutoff points for the older group in different studies ranged between 40 and 60 years. The prevalence of RLS suggested by bubble-TCD was higher than in studies based on transoesophageal echocardiography (0·27, 95% CI 0·20–0·33 vs 0·17, 0·13–0·23; p<0·0001; figure 2), both in patients with cryptogenic events (pooled prevalence 0·38, 0·33–0·43, for bubble-TCD studies vs 0·24, 0·16–0·32, for transoesophageal echocardiography studies; p<0·0001) and in strokes of known cause (0·17, 0·10–0·25, for bubble-TCD studies vs 0·11, 0·07–0·16, for transoesophageal echocardiography; p<0·0001; appendix). This difference in detected prevalence was also present in patients older than 70 years in OXVASC and other studies8, 27, 29 that reported data (appendix).Figure 2 Meta-analyses of the prevalence of PFO stratified by screening modality

dies vs 0·11, 0·07–0·16, for transoesophageal echocardiography; p<0·0001; appendix). This difference in detected prevalence was also present in patients older than 70 years in OXVASC and other studies8, 27, 29 that reported data (appendix).Figure 2 Meta-analyses of the prevalence of PFO stratified by screening modality PFO=patent foramen ovale. Bubble-TCD=contrast-enhanced transcranial Doppler. OXVASC=Oxford Vascular Study. Discussion We showed that bubble-TCD was feasible in most patients in a consecutive series of relatively unselected older individuals with transient ischaemic attack or non-disabling stroke, and that there was a significant association between RLS and cryptogenic events in the older population. We found few published studies concerning the association between PFO and cryptogenic events at older ages, mostly based on transoesophageal echocardiography. Although the absolute prevalence varied from study to study, the association of PFO with cryptogenic transient ischaemic attack or stroke was reasonably consistent, particularly in bubble-TCD studies.28, 29 No previous study reported data for large PFOs in patients with TOAST-defined cryptogenic events compared with other events, but we confirmed the association for large RLS in patients aged older than 60 years, with more than one in ten patients with a cryptogenic event having a large RLS.

y in bubble-TCD studies.28, 29 No previous study reported data for large PFOs in patients with TOAST-defined cryptogenic events compared with other events, but we confirmed the association for large RLS in patients aged older than 60 years, with more than one in ten patients with a cryptogenic event having a large RLS. To our knowledge, no evidence has been reported from randomised trials that closure of PFO is effective in secondary prevention of stroke in patients older than 60 years. However, there is evidence that presence of a PFO is associated with increased risk of recurrent stroke in this age group.9 Our findings on the potential burden of transient ischaemic attack or stroke associated with a large RLS (ie, about 6000 patients aged >60 years with cryptogenic transient ischaemic attack and non-disabling stroke every year in the UK) show that large cohort studies with older patients are feasible, and that recruitment into subsequent randomised trials of PFO closure would probably be achievable. However, in this research it will be important that older patients with transient ischaemic attack or stroke are included, and that we avoid the age-related under-investigation and treatment that has been shown previously for other interventions in routine practice.30, 31, 32 Although trials might show that PFO closure is less effective and carries a higher risk of complications in older patients than in younger patients, concerns about the balance of risk and benefit at older ages have often proved unfounded in relation to other interventions.33, 34

rventions in routine practice.30, 31, 32 Although trials might show that PFO closure is less effective and carries a higher risk of complications in older patients than in younger patients, concerns about the balance of risk and benefit at older ages have often proved unfounded in relation to other interventions.33, 34 Large cohort studies and trials in older patients with PFO will only be possible if patients undergo screening in routine clinical practice, although current clinical guidelines usually advocate PFO screening in younger patients with cryptogenic transient ischaemic attack or stroke,35 and routine screening in other populations is discouraged.36, 37 However, although bubble-TCD is not a substitute for transoesophageal echocardiography, which is still necessary to confirm the site of RLS and associated interatrial septal anatomical features, the latter procedure is invasive, often requires sedation, is costly, and has rare but serious complications.38 We have shown that bubble-TCD is a feasible screening method to identify the smaller subset of patients in whom transoesophageal echocardiography might be indicated. Not all patients were able to undergo screening, although this happened randomly and was not a proper source of selection. Bubble-TCD is much better tolerated than transoesophageal echocardiography and very few of our patients were unsuitable because of lack of a temporal bone window, particularly after we introduced the transoccipital approach in these cases.15 Moreover, the few patients in our study who did not undergo bubble-TCD would also not be suitable for transoesophageal echocardiography because of respiratory problems, frailty, or issues with cannulation. Overall, by using bubble-TCD as a screening tool, we showed that only 41 (16%) of 264 patients with cryptogenic transient ischaemic attack or stroke had a large RLS. Given the difference in cost of the procedures (National Health Service tariff of £506·30 for transoesophageal echocardiography with sedation vs £45·97 for bubble-TCD done by a technician), the cost of bubble-TCD prescreening followed by transoesophageal echocardiography in selected cases would be £32 784, compared with £133 663 for transoesophageal echocardiography screening in all 264 patients with cryptogenic events and RLS in our study.

phy with sedation vs £45·97 for bubble-TCD done by a technician), the cost of bubble-TCD prescreening followed by transoesophageal echocardiography in selected cases would be £32 784, compared with £133 663 for transoesophageal echocardiography screening in all 264 patients with cryptogenic events and RLS in our study. Transoesophageal echocardiography might sometimes identify other cardiac abnormalities in patients with crypotogenic transient ischaemic attack or stroke, but few health-care systems offer routine transoesophageal echocardiography in patients older than 60 years.

phy with sedation vs £45·97 for bubble-TCD done by a technician), the cost of bubble-TCD prescreening followed by transoesophageal echocardiography in selected cases would be £32 784, compared with £133 663 for transoesophageal echocardiography screening in all 264 patients with cryptogenic events and RLS in our study. Transoesophageal echocardiography might sometimes identify other cardiac abnormalities in patients with crypotogenic transient ischaemic attack or stroke, but few health-care systems offer routine transoesophageal echocardiography in patients older than 60 years. In our systematic review of all studies of PFO screening at older ages, we showed that studies using bubble-TCD reported RLS rates that were about 40–50% higher than studies using transoesophageal echocardiography, which contrasts with previous direct comparisons of the two techniques.13, 14 Although no studies in our review systematically screened patients using both techniques, and some differences in rates might be in part explained by patient selection, the higher RLS rate in the bubble-TCD studies might have other causes. First, transoesophageal echocardiography is less sensitive to latent shunts than is bubble-TCD because of the difficulty some older patients have doing a Valsalva manoeuvre under sedation. Second, the prevalence of non-cardiac RLS probably increases with age and so bubble-TCD could give a so-called false-positive for PFO in such cases; however, non-cardiac RLS might still be relevant to the cause of the stroke. We found a lower than expected prevalence of RLS in younger patients compared with some previous studies, which might reflect a lack of power due to the relatively small number of young patients enrolled, and our focus on transient ischaemic attack and non-disabling stroke. In the studies included in our analysis, there was heterogeneity between the age cutoff points for the younger and older group, ranging between 40 and 60 years of age (table 3). However, we found consistent results in analyses with cutoff points at more than 60 years and more than 70 years. We chose to classify the OXVASC study patients as older when aged over 60 years and younger when aged 60 years or less to be consistent with the recent trials, which included patients either younger than 60 years5 or aged 60 years and younger.4, 6

nalyses with cutoff points at more than 60 years and more than 70 years. We chose to classify the OXVASC study patients as older when aged over 60 years and younger when aged 60 years or less to be consistent with the recent trials, which included patients either younger than 60 years5 or aged 60 years and younger.4, 6 Our study has several strengths. First, we studied a consecutive series of unselected older patients. Second, we focused mainly on patients with non-disabling events, in line with trials in which PFO closure was shown to be beneficial.4, 5, 6 The other studies in our systematic review were based on admission to stroke units or intensive care units, and included many patients with major and disabling strokes.7, 12 However, our study does have some limitations. First, patients with RLS on bubble-TCD did not systematically undergo transoesophageal echocardiography, because of the absence of evidence on the benenfit of PFO closure at the time of our study. Second, for the same reason, we did not identify the presence of interatrial septal aneurysm. However, a policy of post-screening transoesophageal echocardiography in selected patients would identify interatrial septal aneurysm, although eligibility for closure of a large PFO was not dependent on the presence of interatrial septal aneurysm in any of the trials. Third, although we found large RLS to be associated with cryptogenic stroke at older ages, in the absence of further trials, it does not follow that routine closure would be justified. Moreover, not all trials of closure at younger ages showed benefit.39, 40 However, given that the diameter of PFO and the prevalence of venous thrombosis increase with age,41, 42 older patients might be more susceptible to paradoxical embolism associated with RLS, and some evidence suggests that the presence of PFO significantly increases the risk of recurrent ischaemic stroke or death in older patients with cryptogenic events than in younger patients with cryptogenic events.9

e with age,41, 42 older patients might be more susceptible to paradoxical embolism associated with RLS, and some evidence suggests that the presence of PFO significantly increases the risk of recurrent ischaemic stroke or death in older patients with cryptogenic events than in younger patients with cryptogenic events.9 In conclusion, we found that bubble-TCD is feasible in most older patients with transient ischaemic attack or non-disabling stroke, with a higher rate of RLS than is usually reported in studies of transoesophageal echocardiography. We showed that large RLS is commonly associated with cryptogenic transient ischaemic attack or stroke and might be causal in some cases, such that cohort studies and trials of PFO closure in older patients are justified. Routine bubble-TCD is a feasible first-line screening modality for the detection of possible PFO and could facilitate further research. Supplementary Material Supplementary appendix Acknowledgments The Oxford Vascular Study is funded by the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre, Wellcome Trust, and Wolfson Foundation. The views expressed are those of the author(s) and not necessarily those of the National Health Service, the NIHR, or the Department of Health. Requests for data will be considered by PMR.

ded by the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre, Wellcome Trust, and Wolfson Foundation. The views expressed are those of the author(s) and not necessarily those of the National Health Service, the NIHR, or the Department of Health. Requests for data will be considered by PMR. Contributors SM provided neurosonological expertise. SM and PMR drafted the manuscript. SM, LL, and LB acquired the data. SM, LL, LB, and PMR revised the manuscript and analysed and interpreted the data. PMR came up the with the concept and design for the study and was responsible for study supervision and funding. Declaration of interests PMR reports personal fees from Bayer for serving on the Executive Committee of the ARRIVE trial and personal fees from Bristol-Myers Squibb for attending an advisory board meeting. All other authors declare no competing interests.

erebral haemorrhage diagnosis on brain imaging, accurate prediction of the risk of later intracerebral haemorrhage growth could help to target patients' monitoring, treatment and transfer to specialist care, and the design and interpretation of randomised trials of treatments to limit intracerebral haemorrhage growth.5 The timing of the first brain imaging done after intracerebral haemorrhage onset and the intracerebral haemorrhage volume found on imaging are two consistently identified risk factors for intracerebral haemorrhage growth, although the association of other potential risk factors has been inconsistent in many small observational studies. Interest has grown in whether a so-called spot sign due to contrast extravasation on additional angiography at the time of diagnostic imaging is a predictor of intracerebral haemorrhage growth.6 There are several multivariable prediction models for intracerebral haemorrhage growth,7, 8, 9, 10, 11 but the identified predictors have varied across models, and several have relied on CT angiography,12 which is not readily available in low-income and middle-income countries. Identifying more accurate predictors of intracerebral haemorrhage growth is recognised to be a research priority.13 Therefore, we aimed to identify the risk and predictors of acute intracerebral haemorrhage growth, develop and validate prediction models that could be used worldwide, and evaluate the added value of CT angiography.

The timing of the first brain imaging done after intracerebral haemorrhage onset and the intracerebral haemorrhage volume found on imaging are two consistently identified risk factors for intracerebral haemorrhage growth, although the association of other potential risk factors has been inconsistent in many small observational studies. Interest has grown in whether a so-called spot sign due to contrast extravasation on additional angiography at the time of diagnostic imaging is a predictor of intracerebral haemorrhage growth.6 There are several multivariable prediction models for intracerebral haemorrhage growth,7, 8, 9, 10, 11 but the identified predictors have varied across models, and several have relied on CT angiography,12 which is not readily available in low-income and middle-income countries. Identifying more accurate predictors of intracerebral haemorrhage growth is recognised to be a research priority.13 Therefore, we aimed to identify the risk and predictors of acute intracerebral haemorrhage growth, develop and validate prediction models that could be used worldwide, and evaluate the added value of CT angiography. Methods Search strategy and selection criteria We conducted a systematic review to identify studies of intracerebral haemorrhage growth that would share individual patient data for a patient-level meta-analysis of the absolute risks and predictors of intracerebral haemorrhage growth.14 A prespecified protocol (finalised on June 20, 2013, and not registered; appendix) guided our data collection and analyses.

y studies of intracerebral haemorrhage growth that would share individual patient data for a patient-level meta-analysis of the absolute risks and predictors of intracerebral haemorrhage growth.14 A prespecified protocol (finalised on June 20, 2013, and not registered; appendix) guided our data collection and analyses. One author (JF) identified potentially eligible cohorts by searching OVID MEDLINE from Jan 1, 1970, to Dec 31, 2015, using a comprehensive search strategy (appendix); hand-searching relevant studies' bibliographies; contacting authors of collaborating studies; and accessing patient-level data from eligible cohorts in the Virtual International Stroke Trials Archive. We included the largest single report of any observational or randomised cohort—regardless of language of publication—if it included at least ten eligible patients with acute intracerebral haemorrhage who had brain imaging (by CT with or without angiography or by MRI) to diagnose intracerebral haemorrhage and used a predefined protocol for repeat imaging (done regardless of clinical need), which would minimise the risks of selection and information biases about intracerebral haemorrhage growth.

acerebral haemorrhage who had brain imaging (by CT with or without angiography or by MRI) to diagnose intracerebral haemorrhage and used a predefined protocol for repeat imaging (done regardless of clinical need), which would minimise the risks of selection and information biases about intracerebral haemorrhage growth. We included patients from these cohorts if they were aged 18 years or older; had non-traumatic intracerebral haemorrhage that was probably due to cerebral small vessel disease and not secondary to an underlying structural cause identified by brain imaging; had data available from brain imaging initially done 0·5–24 h and repeated fewer than 6 days after symptom onset; had baseline intracerebral haemorrhage volume of less than 150 mL; and did not undergo acute treatment that might reduce intracerebral haemorrhage volume (ie, surgical evacuation,15 haemostatic therapy,5 or blood pressure lowering16). We excluded patients if the time from symptom onset to baseline imaging was not known in hours or if they had not been included in the published report of their cohort.

rgo acute treatment that might reduce intracerebral haemorrhage volume (ie, surgical evacuation,15 haemostatic therapy,5 or blood pressure lowering16). We excluded patients if the time from symptom onset to baseline imaging was not known in hours or if they had not been included in the published report of their cohort. We emailed our protocol and an invitation to collaborate to the corresponding authors of cohorts that were eligible for inclusion, followed by one reminder. We included cohorts if corresponding authors of studies reporting them confirmed their eligibility and provided patient-level data on eligibility criteria and other variables at baseline, information on type and timing of baseline and repeat brain imaging, intracerebral haemorrhage characteristics (location, volume on baseline and repeat imaging, presence of intraventricular haemorrhage), and the presence of the spot sign on CT angiography if done (appendix). Research ethics committees or other entities overseeing the use of patients' data had approved the collaborating cohorts. Cohorts shared only anonymised data, so neither individual consent nor specific approval for this individual patient data meta-analysis were required.

We emailed our protocol and an invitation to collaborate to the corresponding authors of cohorts that were eligible for inclusion, followed by one reminder. We included cohorts if corresponding authors of studies reporting them confirmed their eligibility and provided patient-level data on eligibility criteria and other variables at baseline, information on type and timing of baseline and repeat brain imaging, intracerebral haemorrhage characteristics (location, volume on baseline and repeat imaging, presence of intraventricular haemorrhage), and the presence of the spot sign on CT angiography if done (appendix). Research ethics committees or other entities overseeing the use of patients' data had approved the collaborating cohorts. Cohorts shared only anonymised data, so neither individual consent nor specific approval for this individual patient data meta-analysis were required. Data analysis We used reports of the included cohorts to categorise their method of intracerebral haemorrhage volume measurement as a cohort-level characteristic into either the manual ABC/2 method17 or an automated or semi-automated planimetric method.18 We assessed risk of bias across cohorts by identifying the studies that did not meet our eligibility criteria, did not share data, or did not provide data on a sufficient number of the variables of interest (appendix). We checked data completeness and consistency within each cohort and resolved any queries directly with the relevant collaborators. We standardised the format, coding, and units of measurement of variables to maximise the number available for analysis in all cohorts. We did not use or request aggregate data from cohorts that did not share patient-level data.

tency within each cohort and resolved any queries directly with the relevant collaborators. We standardised the format, coding, and units of measurement of variables to maximise the number available for analysis in all cohorts. We did not use or request aggregate data from cohorts that did not share patient-level data. We prespecified that the primary outcome measure of intracerebral haemorrhage growth would be an increase in intracerebral haemorrhage volume between baseline and repeat imaging of more than 6 mL; we chose an absolute measure of intracerebral haemorrhage growth in volume because such measures seem to have higher positive predictive values for more severe clinical outcomes than does the combination of absolute or relative increases in intracerebral haemorrhage volume (eg, >33%).19

epeat imaging of more than 6 mL; we chose an absolute measure of intracerebral haemorrhage growth in volume because such measures seem to have higher positive predictive values for more severe clinical outcomes than does the combination of absolute or relative increases in intracerebral haemorrhage volume (eg, >33%).19 We prespecified the variables that might be predictors of intracerebral haemorrhage growth in our protocol (appendix) on the basis of their clinical relevance, likelihood of being associated with outcome, and reliability and accuracy of measurement (appendix). To these variables, we added history of liver disease and history of stroke; we also added CT angiography spot sign in view of the increasing interest in its role as a predictor since the protocol had originally been written (appendix).6 Of these prespecified variables, we selected potential predictors on the basis of their completeness and availability at the time of diagnosis in the available cohorts and the extent to which their selection maximised the total sample size available for multivariable analyses. Many cohorts excluded patients taking anticoagulant therapy at onset and only a few cohorts conducted CT angiography, so we took a hierarchical approach to investigating univariable and multivariable associations and predictors of intracerebral haemorrhage growth.

total sample size available for multivariable analyses. Many cohorts excluded patients taking anticoagulant therapy at onset and only a few cohorts conducted CT angiography, so we took a hierarchical approach to investigating univariable and multivariable associations and predictors of intracerebral haemorrhage growth. First, we analysed patients not taking anticoagulant therapy at intracerebral haemorrhage symptom onset because they constituted the vast majority of the included cohorts. In this dataset, we examined the associations between intracerebral haemorrhage growth and a subset of the variables, which were chosen on the basis of their completeness and availability at the time of intracerebral haemorrhage diagnosis in the participating cohorts. We visually inspected plots of cohort-specific estimates of association for each variable to exclude major heterogeneity. We then used a one-stage approach to meta-analysis to obtain unadjusted and adjusted estimates pooled across the cohorts using logistic regression models with random intercepts and random coefficients. For all continuous predictors, we used either a linear term or, where there was strong evidence (p<0·01) of non-linearity on the log-odds scale, a fractional polynomial. We described the univariable associations between intracerebral haemorrhage growth and two of the continuous variables (time to baseline imaging and intracerebral haemorrhage volume at baseline) by plotting the predicted probability of intracerebral haemorrhage growth derived from the model against the predictor. For the remaining continuous variables, we quantified the unadjusted and adjusted associations using the odds ratio for the upper quartile compared with the lower quartile based on the fitted linear or fractional polynomial terms in the logistic regression model. We had a sufficient sample size to split those patients who were not taking anticoagulant therapy by contributing cohort into two datasets: one to develop a prediction model and another to validate its performance. We did this temporal validation with patients from earlier cohorts (1994–2007) allocated to the development dataset and patients from more recent cohorts (2008–15) allocated to the validation dataset.

py by contributing cohort into two datasets: one to develop a prediction model and another to validate its performance. We did this temporal validation with patients from earlier cohorts (1994–2007) allocated to the development dataset and patients from more recent cohorts (2008–15) allocated to the validation dataset. We chose a subset of potential predictors for entry into a multivariable model on the basis of their combined availability in the development dataset and the number of patients with intracerebral haemorrhage growth (to avoid overfitting), without considering the results of the unadjusted and adjusted associations between each predictor and intracerebral haemorrhage growth. We did not examine interactions between other covariates and these associations. We derived a prediction index for intracerebral haemorrhage growth with the predictors that remained in a multivariable logistic regression model after backwards elimination. We assessed the performance of the prediction model using calibration plots of predicted versus observed probabilities, receiver operating characteristic curves, and the C-index to assess discrimination in both the development and validation datasets and in patients from cohorts that included patients taking anticoagulant therapy at intracerebral haemorrhage onset. Second, we assessed the performance of the prediction model in patients taking anticoagulant therapy at intracerebral haemorrhage onset.

We chose a subset of potential predictors for entry into a multivariable model on the basis of their combined availability in the development dataset and the number of patients with intracerebral haemorrhage growth (to avoid overfitting), without considering the results of the unadjusted and adjusted associations between each predictor and intracerebral haemorrhage growth. We did not examine interactions between other covariates and these associations. We derived a prediction index for intracerebral haemorrhage growth with the predictors that remained in a multivariable logistic regression model after backwards elimination. We assessed the performance of the prediction model using calibration plots of predicted versus observed probabilities, receiver operating characteristic curves, and the C-index to assess discrimination in both the development and validation datasets and in patients from cohorts that included patients taking anticoagulant therapy at intracerebral haemorrhage onset. Second, we assessed the performance of the prediction model in patients taking anticoagulant therapy at intracerebral haemorrhage onset. Third, we split by contributing cohort those patients from cohorts that included at least some patients taking anticoagulant therapy at intracerebral haemorrhage onset into one dataset to develop a prediction model and another to validate its performance (using temporal validation, as described above). We considered the same subset of potential predictors as for the first prediction model, with the addition of anticoagulant therapy use at intracerebral haemorrhage onset. We derived a prediction index for intracerebral haemorrhage growth and assessed its performance using the same approaches as for the first prediction model.

dered the same subset of potential predictors as for the first prediction model, with the addition of anticoagulant therapy use at intracerebral haemorrhage onset. We derived a prediction index for intracerebral haemorrhage growth and assessed its performance using the same approaches as for the first prediction model. Fourth, in cohorts that included at least some patients with data available on the spot sign identified by CT angiography and that also included and distinguished patients taking anticoagulant therapy at onset, we assessed whether spot sign presence was independently associated with intracerebral haemorrhage growth and the predictive performance when it was added to the predictors in the second prediction model. We did a prespecified sensitivity analysis to compare our findings using a definition of intracerebral haemorrhage growth as an absolute increase of more than 6 mL versus an absolute increase of more than 6 mL or a relative increase of more than 33% in intracerebral haemorrhage volume. We did post-hoc sensitivity analyses to compare associations between time from intracerebral haemorrhage symptom onset to baseline brain imaging and intracerebral haemorrhage volume on baseline imaging with intracerebral haemorrhage growth in cohorts using ABC/2 versus planimetric methods of measuring intracerebral haemorrhage volume and in cohorts from earlier versus later time periods. Analyses were done using SAS software version 9.4 (SAS Institute) and Stata version 12.1 (StataCorp).

We did a prespecified sensitivity analysis to compare our findings using a definition of intracerebral haemorrhage growth as an absolute increase of more than 6 mL versus an absolute increase of more than 6 mL or a relative increase of more than 33% in intracerebral haemorrhage volume. We did post-hoc sensitivity analyses to compare associations between time from intracerebral haemorrhage symptom onset to baseline brain imaging and intracerebral haemorrhage volume on baseline imaging with intracerebral haemorrhage growth in cohorts using ABC/2 versus planimetric methods of measuring intracerebral haemorrhage volume and in cohorts from earlier versus later time periods. Analyses were done using SAS software version 9.4 (SAS Institute) and Stata version 12.1 (StataCorp). Role of the funding source The study sponsors had no role in study design; in the collection, analysis, and interpretation of data; in the writing of the report; and in the decision to submit the paper for publication. The data were available to all authors on request. The corresponding author had final responsibility for the decision to submit for publication.

rs had no role in study design; in the collection, analysis, and interpretation of data; in the writing of the report; and in the decision to submit the paper for publication. The data were available to all authors on request. The corresponding author had final responsibility for the decision to submit for publication. Results We screened 4191 studies identified by our searches, assessed 167 for eligibility, invited 77 eligible cohorts to share data, and obtained patient-level data from 36 (47%) cohorts18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51 involving 6428 patients with repeat brain imaging after intracerebral haemorrhage between 1985 and 2015 (no data up to 1984 were obtained; figure 1; appendix).20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55 Countries classified as high income by the World Bank contributed to 26 (72%) of 36 collaborating cohorts versus 30 (73%) of 41 eligible cohorts that did not collaborate. Planimetric methods of measuring intracerebral haemorrhage volume were used by 19 (53%) of 36 collaborating cohorts versus six (15%) of 41 eligible cohorts that did not collaborate.Figure 1 Study selection *Excluded studies and cohorts are listed in the appendix.

Results We screened 4191 studies identified by our searches, assessed 167 for eligibility, invited 77 eligible cohorts to share data, and obtained patient-level data from 36 (47%) cohorts18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51 involving 6428 patients with repeat brain imaging after intracerebral haemorrhage between 1985 and 2015 (no data up to 1984 were obtained; figure 1; appendix).20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55 Countries classified as high income by the World Bank contributed to 26 (72%) of 36 collaborating cohorts versus 30 (73%) of 41 eligible cohorts that did not collaborate. Planimetric methods of measuring intracerebral haemorrhage volume were used by 19 (53%) of 36 collaborating cohorts versus six (15%) of 41 eligible cohorts that did not collaborate.Figure 1 Study selection *Excluded studies and cohorts are listed in the appendix. After confirming the integrity of the data from eligible cohorts and excluding patients who were ineligible, we created a dataset of 5435 patients (appendix), from which we identified four groups of patients for further analysis: 5076 patients not taking anticoagulant therapy at intracerebral haemorrhage onset, 351 patients taking anticoagulant therapy at intracerebral haemorrhage onset, 3550 patients from cohorts that included at least some patients taking anticoagulant therapy at intracerebral haemorrhage onset, and 868 patients for whom both information about anticoagulant therapy at intracerebral haemorrhage onset and spot sign on acute CT angiography were known (table 1; appendix). The availability of potential predictors varied between the collaborating cohorts such that their overall completeness was 86% in the patients not taking anticoagulant therapy at intracerebral haemorrhage onset, 88% in the patients taking anticoagulant therapy at intracerebral haemorrhage onset, 89% in patients from cohorts that included at least some patients taking anticoagulant therapy at intracerebral haemorrhage onset, and 91% in the patients with information about anticoagulant therapy at intracerebral haemorrhage onset and spot sign on acute CT angiography. More than 80% of patients in all groups had repeat imaging done within 48 h of intracerebral haemorrhage onset and less than 2% of patients had repeat imaging done more than 4 days after intracerebral haemorrhage onset (appendix).Table 1 Characteristics of patients included in the four datasets for meta-analysis

raphy. More than 80% of patients in all groups had repeat imaging done within 48 h of intracerebral haemorrhage onset and less than 2% of patients had repeat imaging done more than 4 days after intracerebral haemorrhage onset (appendix).Table 1 Characteristics of patients included in the four datasets for meta-analysis Not taking anticoagulant therapy (n=5076) Taking anticoagulant therapy (n=351) From cohorts with some patients taking anticoagulant therapy (n=3550) CT angiography (n=868) Sex Female 1971/4884 (40%) 135 (38%) 1449 (41%) 379 (44%) Male 2913/4884 (60%) 216 (62%) 2101 (59%) 489 (56%) Age, years 67 (56–76) 76 (69–82) 69 (58–78) 70 (57–79) Previous stroke 607/4560 (13%) 77/317 (24%) 481/3308 (15%) 97/829 (12%) Previous intracerebral haemorrhage* 179/2753 (7%) 9/246 (4%) 113/2051 (6%) 29/805 (4%) Previous ischaemic stroke* 246/2755 (9%) 53/246 (22%) 213/2051 (10%) 70/805 (9%) History of hypertension 3787/5050 (75%) 291 (83%) 2739/3547 (77%) 616/866 (71%) History of diabetes mellitus 727/4197 (17%) 82/343 (24%) 626/3475 (18%) 137/807 (17%) History of liver disease 256/3360 (8%) 14/220 (6%) 96/1946 (5%) 15/360 (4%) History of excessive alcohol consumption† 568/3091 (18%) 20/177 (11%) 221/1455 (15%) 73/554 (13%) Antiplatelet therapy at symptom onset 913/5030 (18%) 102 (29%) 855/3543 (24%) 225/837 (27%) Anticoagulant therapy at symptom onset 0 351 (100%) 349/3547 (10%) 87/841 (10%) Systolic blood pressure at presentation, mm Hg 177 (158–198); n=4882 170 (147–190); n=320 177 (157–197); n=3333 175 (150–200); n=860 Blood glucose at presentation, mmol/L 7·0 (5·9–8·7); n=4265 7·4 (6·0–9·3); n=340 7·0 (5·9–8·7); n=3417 7·3 (6·1–8·9); n=864 Platelet count (×109/L) at presentation 221 (181–266); n=3857 209 (174–260); n=289 222 (185–267); n=2284 227 (181–273); n=862 Glasgow Coma Scale score at presentation 3–6 285/4564 (6%) 42/342 (12%) 248/3502 (7%) 73/831 (9%) 7–12 1033/4564 (23%) 81/342 (24%) 824/3502 (24%) 193/831 (23%) 13–14 1157/4564 (25%) 70/342 (20%) 830/3502 (24%) 151/831 (18%) 15 2089/4564 (46%) 149/342 (44%) 1600/3502 (46%) 414/831 (50%) NIHSS score at presentation 12 (7–18); n=2661 13 (7–17); n=126 12 (7–17); n=2014 14 (6–18); n=325 Time from symptom onset to baseline imaging, h 2·4 (1·3–4·7) 3·3 (1·7–6·4) 2·2 (1·3–4·2) 2·9 (1·5–5·1) Intracerebral haemorrhage volume on baseline imaging, mL 13·2 (6·3–30·0) 16·0 (6·4–39·0) 13·4 (6·6–30·3) 15·0 (6·6–34·1) Lobar location of intracerebral haemorrhage on baseline imaging 1080/4920 (22%) 129/344 (38%) 907/3439 (26%) 267/866 (

t to baseline imaging, h 2·4 (1·3–4·7) 3·3 (1·7–6·4) 2·2 (1·3–4·2) 2·9 (1·5–5·1) Intracerebral haemorrhage volume on baseline imaging, mL 13·2 (6·3–30·0) 16·0 (6·4–39·0) 13·4 (6·6–30·3) 15·0 (6·6–34·1) Lobar location of intracerebral haemorrhage on baseline imaging 1080/4920 (22%) 129/344 (38%) 907/3439 (26%) 267/866 ( 31%) Intraventricular haemorrhage present on baseline imaging 1834/4980 (37%) 157/348 (45%) 1265/3452 (37%) 344 (40%) CT angiogram spot sign present .. .. .. 204 (24%) >6 mL intracerebral haemorrhage growth 1009 (20%) 110 (31%) 771 (22%) 177 (20%) >6 mL or >33% intracerebral haemorrhage growth 1301 (26%) 139 (40%) 986 (28%) 219 (25%) Data are n (%), n/N (%), or median (IQR). NIHSS=National Institutes of Health Stroke Scale. * Available in a subgroup of cohorts that quantified the subtype of previous stroke. Not all cohorts that quantified the subtype included both intracerebral haemorrhage and ischaemic stroke. † Definition of excessive consumption varied across cohorts.

.. .. 204 (24%) >6 mL intracerebral haemorrhage growth 1009 (20%) 110 (31%) 771 (22%) 177 (20%) >6 mL or >33% intracerebral haemorrhage growth 1301 (26%) 139 (40%) 986 (28%) 219 (25%) Data are n (%), n/N (%), or median (IQR). NIHSS=National Institutes of Health Stroke Scale. * Available in a subgroup of cohorts that quantified the subtype of previous stroke. Not all cohorts that quantified the subtype included both intracerebral haemorrhage and ischaemic stroke. † Definition of excessive consumption varied across cohorts. When assessing the two variables with non-linear associations, we found that in patients not taking anticoagulant therapy at intracerebral haemorrhage onset, the predicted probability of intracerebral haemorrhage growth declined with increasing time from intracerebral haemorrhage symptom onset to baseline imaging: the rate of decline was steepest 0·5–3 h after intracerebral haemorrhage symptom onset (figure 2A). The predicted probability of intracerebral haemorrhage growth increased with increasing intracerebral haemorrhage volume on baseline brain imaging and peaked at about 75 mL, above which it declined (figure 2B). We aimed to quantify the associations between 17 additional variables and the occurrence of intracerebral haemorrhage growth (appendix). There were too few patients with data for six variables (previous intracerebral haemorrhage, previous ischaemic stroke, history of liver disease, history of excessive alcohol consumption, platelet count at presentation, and National Institutes of Health Stroke Scale [NIHSS] score at presentation). Therefore, we selected 13 of the 19 variables as potential predictors for a multivariable model in patients not taking anticoagulant therapy, on the basis of maximising the number of predictors being considered while also maximising the number of patients with complete data for all the predictors chosen for the subset: time from symptom onset to baseline imaging, intracerebral haemorrhage volume on baseline imaging, sex, age, previous stroke, history of hypertension, history of diabetes, antiplatelet therapy at symptom onset, systolic blood pressure at presentation, blood glucose at presentation, Glasgow Coma Scale score at presentation, intracerebral haemorrhage location on baseline scan, and intraventricular haemorrhage on baseline scan. We restricted all further analyses to datasets of patients with complete data on these 13 potential predictors.Figure 2 Predicted probability of intracerebral haemorrhage growth >6 mL

oma Scale score at presentation, intracerebral haemorrhage location on baseline scan, and intraventricular haemorrhage on baseline scan. We restricted all further analyses to datasets of patients with complete data on these 13 potential predictors.Figure 2 Predicted probability of intracerebral haemorrhage growth >6 mL Data calculated on 5076 patients who were not taking anticoagulant therapy at symptom onset. (A) Predicted probability by time from intracerebral haemorrhage symptom onset to baseline imaging, and (B) according to intracerebral haemorrhage volume on baseline imaging. The solid line indicates predicted probability and the shaded region indicates the 95% CIs. 3479 patients who were not taking anticoagulant therapy at intracerebral haemorrhage onset had data available for the 13 predictors. We developed a prediction model for intracerebral haemorrhage growth using a dataset of 2534 (73%) of these patients from 18 earlier cohorts (ie, 1994–2007; appendix). From the 13 potential predictors considered, three significant predictors constituted the final model (table 2): Predicted probability of intracerebral haemorrhage growth=1(1+e−PI) where the predictive index (PI) is given byTable 2 Multivariable models of predictors of intracerebral haemorrhage growth >6 mL

appendix). From the 13 potential predictors considered, three significant predictors constituted the final model (table 2): Predicted probability of intracerebral haemorrhage growth=1(1+e−PI) where the predictive index (PI) is given byTable 2 Multivariable models of predictors of intracerebral haemorrhage growth >6 mL Comparison Odds ratio (95% CI) p value Patients not taking anticoagulant therapy at symptom onset* Time from symptom onset to baseline imaging, h† 3·4 vs 1·2 0·65 (0·51–0·82) 0·0003 Intracerebral haemorrhage volume on baseline imaging, mL† 28 vs 7 4·73 (3·81–5·87) <0·0001 Antiplatelet therapy at symptom onset Yes vs no 1·38 (1·06–1·79) 0·016 Patients from cohorts including at least some patients taking anticoagulant therapy at symptom onset‡ Time from symptom onset to baseline imaging, h† 3·5 vs 1·2 0·59 (0·42–0·82) 0·0021 Intracerebral haemorrhage volume on baseline imaging, mL† 29 vs 7 4·81 (3·82–6·05) <0·0001 Antiplatelet therapy at symptom onset Yes vs no 1·36 (1·04–1·78) 0·023 Anticoagulant therapy at symptom onset Yes vs no 2·91 (1·97–4·26) <0·0001 * Data were calculated on 2534 patients from 18 cohorts (appendix). † The odds ratios for time from symptom onset to baseline imaging and intracerebral haemorrhage volume on baseline imaging are for upper quartile compared with lower quartile. ‡ Data were calculated on 2381 patients from ten cohorts (appendix).

Comparison Odds ratio (95% CI) p value Patients not taking anticoagulant therapy at symptom onset* Time from symptom onset to baseline imaging, h† 3·4 vs 1·2 0·65 (0·51–0·82) 0·0003 Intracerebral haemorrhage volume on baseline imaging, mL† 28 vs 7 4·73 (3·81–5·87) <0·0001 Antiplatelet therapy at symptom onset Yes vs no 1·38 (1·06–1·79) 0·016 Patients from cohorts including at least some patients taking anticoagulant therapy at symptom onset‡ Time from symptom onset to baseline imaging, h† 3·5 vs 1·2 0·59 (0·42–0·82) 0·0021 Intracerebral haemorrhage volume on baseline imaging, mL† 29 vs 7 4·81 (3·82–6·05) <0·0001 Antiplatelet therapy at symptom onset Yes vs no 1·36 (1·04–1·78) 0·023 Anticoagulant therapy at symptom onset Yes vs no 2·91 (1·97–4·26) <0·0001 * Data were calculated on 2534 patients from 18 cohorts (appendix). † The odds ratios for time from symptom onset to baseline imaging and intracerebral haemorrhage volume on baseline imaging are for upper quartile compared with lower quartile. ‡ Data were calculated on 2381 patients from ten cohorts (appendix). −4·254−0·196time−0·0754volume+1·186√volume+0·320antiplatelet with time measured in hours, volume measured in mL, and antiplatelet an indicator variable for antiplatelet therapy at intracerebral haemorrhage onset taking values 1 for yes and 0 for no.

† The odds ratios for time from symptom onset to baseline imaging and intracerebral haemorrhage volume on baseline imaging are for upper quartile compared with lower quartile. ‡ Data were calculated on 2381 patients from ten cohorts (appendix). −4·254−0·196time−0·0754volume+1·186√volume+0·320antiplatelet with time measured in hours, volume measured in mL, and antiplatelet an indicator variable for antiplatelet therapy at intracerebral haemorrhage onset taking values 1 for yes and 0 for no. This first prediction model had good calibration (appendix) and its discrimination was good in both the development dataset (C-index 0·75, 95% CI 0·72–0·77) and the temporal validation dataset of 945 (27%) patients from six later cohorts (0·76, 0·73–0·79). This prediction model, derived in patients who were not taking anticoagulant therapy at symptom onset, underestimated the probability of intracerebral haemorrhage growth in the 351 patients in 21 cohorts who were taking anticoagulant therapy at symptom onset (appendix), but its discrimination remained good (0·73, 0·68–0·79). We also developed a prediction model for intracerebral haemorrhage growth using a dataset of 2381 patients from ten cohorts that included at least some patients taking anticoagulant therapy at intracerebral haemorrhage onset (appendix). From the 13 potential predictors plus anticoagulant therapy at intracerebral haemorrhage symptom onset, four predictors constituted the final model (table 2), where PI is given by

1 patients from ten cohorts that included at least some patients taking anticoagulant therapy at intracerebral haemorrhage onset (appendix). From the 13 potential predictors plus anticoagulant therapy at intracerebral haemorrhage symptom onset, four predictors constituted the final model (table 2), where PI is given by −4·426−0·230time   −0·0776volume+1·196√volume   +0·310antiplatelet+1·065anticoagulant where anticoagulant is an indicator variable for anticoagulant therapy at intracerebral haemorrhage onset taking values 1 for yes and 0 for no. This second prediction model was well calibrated (appendix) and its discrimination was good in both the development dataset (C-index 0·75, 95% CI 0·73–0·78) and the validation dataset of 895 patients from five cohorts (0·74, 0·71–0·78). Finally, to assess the additional predictive value of spot sign on CT angiography, we assessed the performance of a third prediction model in the 837 patients from six cohorts with available data on all covariates (appendix), where PI is given by −4·954−0·138time−0·0769volume   +1·139√volume+0·370antiplatelet   +1·028anticoagulant+1·496spot where spot is an indicator variable for presence of CT angiography spot sign taking values 1 for present and 0 for absent.

Finally, to assess the additional predictive value of spot sign on CT angiography, we assessed the performance of a third prediction model in the 837 patients from six cohorts with available data on all covariates (appendix), where PI is given by −4·954−0·138time−0·0769volume   +1·139√volume+0·370antiplatelet   +1·028anticoagulant+1·496spot where spot is an indicator variable for presence of CT angiography spot sign taking values 1 for present and 0 for absent. The presence of the spot sign was strongly and independently associated with the occurrence of intracerebral haemorrhage growth (table 3) and improved the C-index of the prediction model by 0·05 (95% CI 0·03–0·07) from 0·78 (0·75–0·82) to 0·83 (0·80–0·86; figure 3).Table 3 Multivariable models of predictors of intracerebral haemorrhage growth >6 mL in patients with assessment of CT angiography spot sign, data on antiplatelet therapy, and data on anticoagulant therapy use at symptom onset

ediction model by 0·05 (95% CI 0·03–0·07) from 0·78 (0·75–0·82) to 0·83 (0·80–0·86; figure 3).Table 3 Multivariable models of predictors of intracerebral haemorrhage growth >6 mL in patients with assessment of CT angiography spot sign, data on antiplatelet therapy, and data on anticoagulant therapy use at symptom onset Comparison Four predictors Four predictors with the addition of CT angiography spot sign Odds ratio (95% CI) p value Odds ratio (95% CI) p value Time from symptom onset to baseline imaging, h* 5·1 vs 1·5 0·50 (0·36–0·70) <0·0001 0·61 (0·44–0·84) 0·0030 Intracranial haemorrhage volume on baseline imaging, mL* 33 vs 6 7·18 (4·46–11·56) <0·0001 5·35 (3·25–8·81) <0·0001 Antiplatelet therapy at symptom onset Yes vs no 1·68 (1·06–2·66) 0·026 1·45 (0·89–2·35) 0·13 Anticoagulant therapy at symptom onset Yes vs no 3·48 (1·96–6·16) <0·0001 2·80 (1·53–5·10) 0·0008 CT angiography spot sign Present vs absent .. .. 4·46 (2·95–6·75) <0·0001 Data were calculated on 837 patients from six cohorts (appendix). * Odds ratios for time from symptom onset to baseline imaging and intracranial haemorrhage volume on baseline imaging are for upper quartile vs lower quartile. Figure 3 Receiver operating characteristic curves for the predicted probability of intracerebral haemorrhage growth >6 mL

Comparison Four predictors Four predictors with the addition of CT angiography spot sign Odds ratio (95% CI) p value Odds ratio (95% CI) p value Time from symptom onset to baseline imaging, h* 5·1 vs 1·5 0·50 (0·36–0·70) <0·0001 0·61 (0·44–0·84) 0·0030 Intracranial haemorrhage volume on baseline imaging, mL* 33 vs 6 7·18 (4·46–11·56) <0·0001 5·35 (3·25–8·81) <0·0001 Antiplatelet therapy at symptom onset Yes vs no 1·68 (1·06–2·66) 0·026 1·45 (0·89–2·35) 0·13 Anticoagulant therapy at symptom onset Yes vs no 3·48 (1·96–6·16) <0·0001 2·80 (1·53–5·10) 0·0008 CT angiography spot sign Present vs absent .. .. 4·46 (2·95–6·75) <0·0001 Data were calculated on 837 patients from six cohorts (appendix). * Odds ratios for time from symptom onset to baseline imaging and intracranial haemorrhage volume on baseline imaging are for upper quartile vs lower quartile. Figure 3 Receiver operating characteristic curves for the predicted probability of intracerebral haemorrhage growth >6 mL Data calculated on 837 patients with assessment of CT angiography spot sign, data on antiplatelet therapy, and data on anticoagulant therapy use at symptom onset. Receiver operating characteristic curves used four predictors (time from symptom onset to baseline imaging [h], intracerebral haemorrhage volume on baseline imaging [mL], antiplatelet therapy at symptom onset, and anticoagulant therapy at symptom onset) and four predictors plus CT angiography spot sign.

symptom onset. Receiver operating characteristic curves used four predictors (time from symptom onset to baseline imaging [h], intracerebral haemorrhage volume on baseline imaging [mL], antiplatelet therapy at symptom onset, and anticoagulant therapy at symptom onset) and four predictors plus CT angiography spot sign. We assessed the performance of the second and third prediction models at different thresholds of predicted probability of intracerebral haemorrhage growth and found very few significant differences in sensitivity, specificity, positive predictive value, and negative predictive value (appendix). In a prespecified sensitivity analysis, when we defined intracerebral haemorrhage growth as an absolute increase of more than 6 mL or a relative increase of more than 33% in intracerebral haemorrhage volume between baseline and follow-up imaging, the direction, strength, and significance of the adjusted associations between almost all predictors and intracerebral haemorrhage growth remained the same (appendix), and the C-index of our second prediction model improved from 0·71 (95% CI 0·67–0·75) to 0·76 (0·72–0·80) with the addition of information from CT angiography (appendix). In a post-hoc sensitivity analysis, we found no evidence that the risk of intracerebral haemorrhage growth according to time from symptom onset to baseline imaging or according to intracerebral haemorrhage volume on baseline imaging differed by cohort epoch or volumetric method used (appendix).

angiography (appendix). In a post-hoc sensitivity analysis, we found no evidence that the risk of intracerebral haemorrhage growth according to time from symptom onset to baseline imaging or according to intracerebral haemorrhage volume on baseline imaging differed by cohort epoch or volumetric method used (appendix). Discussion This collaborative meta-analysis evaluated 19 covariates in one or more analyses of predictors of intracerebral haemorrhage growth from 5435 eligible patients in 36 cohorts. We identified novel non-linear associations between the probability of intracerebral haemorrhage growth and both the time from symptom onset to baseline imaging and baseline intracerebral haemorrhage volume. We showed that only four predictors that are simple to collect (time from symptom onset to baseline imaging, intracerebral haemorrhage volume on baseline imaging, antiplatelet use, and anticoagulant use) were independently associated with intracerebral haemorrhage growth in multivariable models, and a prediction model that we developed using these predictors not only had good calibration and discrimination but also done well in an external validation dataset. The addition of information about the presence of spot sign on CT angiography to this prediction model gave a small increase in discrimination.

e models, and a prediction model that we developed using these predictors not only had good calibration and discrimination but also done well in an external validation dataset. The addition of information about the presence of spot sign on CT angiography to this prediction model gave a small increase in discrimination. Although many studies have investigated unadjusted and adjusted associations between a wide variety of clinical, blood, genetic, imaging, and pharmacological factors and the occurrence of intracerebral haemorrhage growth, only a few prediction models have been developed and the predictors used have varied considerably.7, 8, 9, 10, 11, 51 Since 2011, there has been growing interest in use of the spot sign on CT angiography for predicting intracerebral haemorrhage growth,10, 30 but the clinical utility of the small increase in discrimination that resource-intensive advanced vascular imaging adds to simple clinical and imaging predictors that are available worldwide is unclear. The strengths of this study include its large sample size and availability of many predictors from geographically diverse cohorts to develop and externally validate prediction models involving simple predictors that could be used in any health-care setting, as well as the added value of CT angiography in high-income countries. We minimised the risk of selection and information biases by restricting eligibility to cohorts that had defined when they would repeat brain imaging soon after intracerebral haemorrhage onset in all survivors and not according to clinical need alone.

well as the added value of CT angiography in high-income countries. We minimised the risk of selection and information biases by restricting eligibility to cohorts that had defined when they would repeat brain imaging soon after intracerebral haemorrhage onset in all survivors and not according to clinical need alone. Although our study was large, only half of the investigators of the available cohorts shared patient-level data. Most cohorts were assembled in high-income countries. A shortage of data on the following variables precluded their inclusion in our prediction models: previous intracerebral haemorrhage, previous ischaemic stroke, history of liver disease, history of excessive alcohol consumption, platelet count at presentation, and NIHSS score at presentation. Since the end of the literature search that defined inclusion in our analyses, our update of the search to March 1, 2018, identified reports of five new cohorts involving 669 patients, representing a maximum of a 10% increase over the 6428 patients from 36 cohorts that provided individual patient data. Nonetheless, the sample size we achieved allowed us to develop and validate prediction models using a large number of widely available predictors, without omitting any predictors that had been identified by previous prediction models. Included cohorts with data collected in the 1990s might not have used multiple-row detector array technology and digitisation, which might have affected their accuracy of intracerebral haemorrhage volume measurement, although there was no evidence that our findings differed by cohort epoch in sensitivity analyses. 19 (53%) of 36 cohorts used planimetric methods to estimate intracerebral haemorrhage volume but 17 (47%) of 36 cohorts used the ABC/2 method (which can marginally overestimate intracerebral haemorrhage volume18), although we found no evidence that our findings differed by volumetric method in sensitivity analyses. Since these cohorts were studied, a variety of new imaging signs (eg, density, irregularity, fluid level, hypodensity, island, satellite, swirl,56 blend,37 and black hole57) have been described, but we were unable to evaluate them because they were not collected by the collaborating cohorts and we could not re-evaluate patients' imaging. However, our simple prediction models provide the basis upon which the added value of these new signs can be assessed, as we have done for the CT angiography spot sign.

ribed, but we were unable to evaluate them because they were not collected by the collaborating cohorts and we could not re-evaluate patients' imaging. However, our simple prediction models provide the basis upon which the added value of these new signs can be assessed, as we have done for the CT angiography spot sign. We found that the rate of decline in the probability of intracerebral haemorrhage growth was steepest during the 0·5–3 h after intracerebral haemorrhage symptom onset and that the predicted probability of intracerebral haemorrhage growth peaked at an intracerebral haemorrhage volume of about 75 mL. These findings could in part explain the neutral results of recent randomised trials of acute interventions designed to limit intracerebral haemorrhage growth, which enrolled many patients towards or beyond the time of greatest risk of intracerebral haemorrhage growth and most patients had small intracerebral haemorrhages at low probability of growth. For example, the average time to randomisation after intracerebral haemorrhage symptom onset and average intracerebral haemorrhage volume were 3·7 h and 13 mL in TICH2,58 3·7 h and 11 mL in INTERACT2,22 3·1 h and 10 mL in ATACH2,59 and 2·7 h and 22–24 mL in FAST.40 In particular, our findings about the association between time after intracerebral haemorrhage symptom onset and the probability of intracerebral haemorrhage growth emphasise the importance of extremely rapid assessment, investigation, and randomisation in future trials of therapies to improve outcome by limiting intracerebral haemorrhage growth.

about the association between time after intracerebral haemorrhage symptom onset and the probability of intracerebral haemorrhage growth emphasise the importance of extremely rapid assessment, investigation, and randomisation in future trials of therapies to improve outcome by limiting intracerebral haemorrhage growth. The prediction models that we have developed could be useful in clinical practice for predicting the risk of intracerebral haemorrhage growth, which is recommended in the emergency assessment of acute intracerebral haemorrhage. The clinically useful threshold for the predicted probability of intracerebral haemorrhage growth will vary according to its desired accuracy (appendix), the clinical setting, and future therapeutic advances, such that our models might help in determining patients' place of care and frequency of observation.60 Supplementary Material Supplementary appendix Acknowledgments This study was funded by the UK Medical Research Council (senior clinical fellowship G1002605 and the Edinburgh Hub for Trials Methodology Research G0800803) and the British Heart Foundation (travel fellowship FS/13/72/30531).

The prediction models that we have developed could be useful in clinical practice for predicting the risk of intracerebral haemorrhage growth, which is recommended in the emergency assessment of acute intracerebral haemorrhage. The clinically useful threshold for the predicted probability of intracerebral haemorrhage growth will vary according to its desired accuracy (appendix), the clinical setting, and future therapeutic advances, such that our models might help in determining patients' place of care and frequency of observation.60 Supplementary Material Supplementary appendix Acknowledgments This study was funded by the UK Medical Research Council (senior clinical fellowship G1002605 and the Edinburgh Hub for Trials Methodology Research G0800803) and the British Heart Foundation (travel fellowship FS/13/72/30531). Contributors RA-SS conceived and designed the project. JF designed the literature search strategies and searched the literature. RA-SS, JF, and RJL co-wrote the protocol and arbitrated cohort eligibility. JF described included studies and communicated with coauthors. RJL processed data and did the data analyses, with oversight from RA-SS. RA-SS, JF, and RJL wrote the first draft of the manuscript. PDL, TWKB, AMA, JNG, SAM, TS, XW, HA, HH, MO, DAG, LM, DD, DR-L, CAM, D-KJ, AD, JCa, XY, JCl, BV, SK, YO, SF, KT, QL, JK, PD, JÁS, MH-G, LP-S, CC, MPK, RM, CV, MND, YI, HW, WCZ, CDd'E, RIA, PR, YM, ARZ, KSB, SMS, JCG, JM-F, JM, JB, HY, DS, ESC, MS, RL, BHM, AMD, MDN, YF, CSA, and JR acquired data, revised the work critically for important intellectual content where required, approved the final version to be published, and agreed to be accountable for all aspects of the work.

MND, YI, HW, WCZ, CDd'E, RIA, PR, YM, ARZ, KSB, SMS, JCG, JM-F, JM, JB, HY, DS, ESC, MS, RL, BHM, AMD, MDN, YF, CSA, and JR acquired data, revised the work critically for important intellectual content where required, approved the final version to be published, and agreed to be accountable for all aspects of the work. Declaration of interests RJL reports grants from the UK Medical Research Council, during the conduct of the study. JNG reports personal fees from CSL Behring and Octapharma; and grants from Pfizer, Boehringer Ingelheim, and Portola, outside of the submitted work. HA reports personal fees from Asuka, Bayer, Daiichi-Sankyo, and Takeda, outside of the submitted work. JCl reports grants from the DANA Foundation and personal fees from SAGE Therapeutics, outside of the submitted work. BV reports personal fees from Pfizer/Bristol-Myers Squibb and Bayer, outside of the submitted work. JK reports grants from the National Institutes of Health/National Institute of Neurological Disorders and Stroke, during the conduct of the study. CV reports grants from the Neurocritical Care Society, during the conduct of the study. MS reports grants from National Institutes of Health/National Institute of Neurological Disorders and Stroke and the American Heart Association, outside of the submitted work. CSA reports grants from the National Health and Medical Research Council of Australia, during the conduct of the study; and personal fees from Takeda and Amgen, outside of the submitted work. JR reports grants from the National Institutes of Health and personal fees from Boehringer Ingelheim and Pfizer, outside of the submitted work. All remaining authors declare no competing interests.

Introduction Neurological disorders are now the leading source of disability globally.1 Among neurological disorders examined in the Global Burden of Disease, Injuries, and Risk Factors Study (GBD) 2015, Parkinson's disease was the fastest growing in prevalence, disability, and deaths. In that study,1 the overall number of people affected by the disease was estimated to have more than doubled globally from 1990 to 2015. Previous studies have examined the epidemiology of Parkinson's disease for different parts of the world,2 including systematic reviews on the prevalence of Parkinson's disease.3, 4 However, none has examined change in prevalence, disability, and deaths in detail over the past generation for the entire world and across all countries. In GBD 2015, we identified larger variation in Parkinson's disease death rate estimates over time and between countries than we observed in prevalence estimates.1 This pattern suggested that coding practices rather than real changes over time and location were responsible, similar to what was observed for dementia.1

s. In GBD 2015, we identified larger variation in Parkinson's disease death rate estimates over time and between countries than we observed in prevalence estimates.1 This pattern suggested that coding practices rather than real changes over time and location were responsible, similar to what was observed for dementia.1 The prevalence of a disease reflects both the incidence and the duration of disease. The incidence of Parkinson's disease is linked to risk and protective factors.2, 5, 6 The most important risk factor is age, but the risk of Parkinson's disease also appears to be associated with industrial chemicals and pollutants, such as pesticides,7 solvents,7 and metals.8, 9 Conversely, smoking is associated with a decreased risk of Parkinson's disease,10 but whether this association is causal is debatable.11 The factors that affect disease duration are less well known, but increasing longevity also translates into longer disease duration.3, 12 Therefore, as ageing and industrialisation increase globally and smoking decreases in some regions, the prevalence of Parkinson's disease seems poised to increase.13, 14 Detailed estimates of the disease burden can help to evaluate the effect of these risk factors and inform efforts to prevent the disease and to care for and treat those affected by the condition. Research in context Evidence before this study

The prevalence of a disease reflects both the incidence and the duration of disease. The incidence of Parkinson's disease is linked to risk and protective factors.2, 5, 6 The most important risk factor is age, but the risk of Parkinson's disease also appears to be associated with industrial chemicals and pollutants, such as pesticides,7 solvents,7 and metals.8, 9 Conversely, smoking is associated with a decreased risk of Parkinson's disease,10 but whether this association is causal is debatable.11 The factors that affect disease duration are less well known, but increasing longevity also translates into longer disease duration.3, 12 Therefore, as ageing and industrialisation increase globally and smoking decreases in some regions, the prevalence of Parkinson's disease seems poised to increase.13, 14 Detailed estimates of the disease burden can help to evaluate the effect of these risk factors and inform efforts to prevent the disease and to care for and treat those affected by the condition. Research in context Evidence before this study The Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2015 examined the epidemiology of Parkinson's disease for different parts of the world and showed that the number of people affected by the condition had more than doubled globally from 1990 to 2015. The increase in deaths from Parkinson's disease was greater than the increase in prevalence, and the large variation in death rates between countries was suggestive of a change in coding practices rather than greater death rates among Parkinson's disease cases. For pragmatic reasons, systematic reviews for Parkinson's disease are scheduled every other iteration of GBD. For GBD 2016, we updated our GBD 2013 PubMed search without language restrictions using the terms (((“Parkinson disease” AND “epidemiology”) AND (“2011/01/01”[PDat]: “2015/12/31”[PDat])) AND (“Parkinson disease” AND “epidemiology”)) to identify articles published between Jan, 1, 2011, and Dec, 31, 2015. Papers were selected if representative of the general population and identification of cases was based on our reference case definition (the presence of at least two of four primary symptoms: rest tremor, bradykinesia, stiffness of limbs and torso, and postural instability) or alternative case definitions (UK Parkinson's Disease Society Brain Bank criteria, and doctor's diagnosis based on International Classification of Diseases codes and prescription of medications specifically for Parkinson's disease).

s: rest tremor, bradykinesia, stiffness of limbs and torso, and postural instability) or alternative case definitions (UK Parkinson's Disease Society Brain Bank criteria, and doctor's diagnosis based on International Classification of Diseases codes and prescription of medications specifically for Parkinson's disease). Added value of this study We used the results of this search to obtain the data needed to estimate global, regional, and country-specific prevalence and years lived with disability for Parkinson's disease from 1990 to 2016. To address the possible measurement error in Parkinson's disease death rates as reported by vital registration systems, we used a method that was previously applied to dementia in GBD 2015. In a natural history modelling approach, we assume a constant risk of death in Parkinson's disease cases over time and between locations and let the death rates be determined by variations in prevalence. Although the assumption of similar mortality risk in all time periods and countries is problematic, it produces less error than the large variation in death rates estimated previously. We also explored variation in the burden by age, sex, country, region, and Socio-demographic Index. This study showed that counts of prevalence, mortality, and disability-adjusted life-years more than doubled from 1990 to 2016, and that this increase was not solely due to increasing numbers of older people because age-standardised rates also increased in most regions. In addition, the burden of Parkinson's disease increased with increasing Socio-demographic Index.

tality, and disability-adjusted life-years more than doubled from 1990 to 2016, and that this increase was not solely due to increasing numbers of older people because age-standardised rates also increased in most regions. In addition, the burden of Parkinson's disease increased with increasing Socio-demographic Index. Implications of all the available evidence Neurological disorders are now the leading source of disability in the world, and Parkinson's disease is the fastest growing of these disorders. As the population ages and life expectancy increases, the number of individuals with Parkinson's disease will continue to increase as well as the duration of the disease, leading to more patients with advanced Parkinson's disease. To address this burden, primary prevention strategies based on the underlying causes of Parkinson's disease and more effective treatments than are currently available are required. Additional incidence and prevalence studies are needed, especially in areas in which little data are available, to examine time trends and the factors that drive them. As part of GBD 2016, we aimed to examine the changes from 1990 to 2016 in counts and age-standardised rates of Parkinson's disease for prevalence, disability, and deaths by location and by the Socio-demographic Index (SDI), a composite measure of income per capita, education, and fertility.15

Neurological disorders are now the leading source of disability in the world, and Parkinson's disease is the fastest growing of these disorders. As the population ages and life expectancy increases, the number of individuals with Parkinson's disease will continue to increase as well as the duration of the disease, leading to more patients with advanced Parkinson's disease. To address this burden, primary prevention strategies based on the underlying causes of Parkinson's disease and more effective treatments than are currently available are required. Additional incidence and prevalence studies are needed, especially in areas in which little data are available, to examine time trends and the factors that drive them. As part of GBD 2016, we aimed to examine the changes from 1990 to 2016 in counts and age-standardised rates of Parkinson's disease for prevalence, disability, and deaths by location and by the Socio-demographic Index (SDI), a composite measure of income per capita, education, and fertility.15 Methods Overview The general methods for the studies on the global, regional, and national burden of neurological disorders have been published previously,1 and key aspects are summarised in the appendix. Additional information on derivation of non-fatal and fatal estimates are provided in the appendix as well as on Global Health Data Exchange.

al methods for the studies on the global, regional, and national burden of neurological disorders have been published previously,1 and key aspects are summarised in the appendix. Additional information on derivation of non-fatal and fatal estimates are provided in the appendix as well as on Global Health Data Exchange. Data sources The International Classification of Diseases ninth revision (ICD-9) codes used in cause of death analyses for Parkinson's disease are 332 (Parkinson's disease), 332.0 (paralysis agitans), and 332.1 (secondary parkinsonism), and the corresponding ICD-10 codes are G20 (Parkinson's disease), G21 (secondary parkinsonism), and G22 (parkinsonism in diseases classified elsewhere). The reference case definition for Parkinson's disease used in many epidemiological studies is the presence of at least two of the four primary symptoms: rest tremor, bradykinesia, stiffness of the limbs and torso, and postural instability. We also accepted alternative definitions, including the UK Parkinson's Disease Society Brain Bank criteria,16 in addition to ICD-9 and ICD-10 codes, a doctor's diagnosis of Parkinson's disease, and the prescription of Parkinson's disease-specific medications. We also added 3 years of medical claims data (years 2000, 2010, and 2012) from the USA; for a disease such as Parkinson's disease, the data from claims sources would be expected to match true prevalence under the expectation that most patients would be receiving medical attention each year. If datapoints from epidemiological studies spanned ages of more than 20 years, we split the datapoints using the age pattern from the USA, for which the most detailed age data were available.

ld be expected to match true prevalence under the expectation that most patients would be receiving medical attention each year. If datapoints from epidemiological studies spanned ages of more than 20 years, we split the datapoints using the age pattern from the USA, for which the most detailed age data were available. Disease model For Parkinson's disease, we have seen large inconsistencies between the cause of death data and the non-fatal data. For example, US vital registration data show a greater than three times increase in the age-standardised rates of death from Parkinson's disease since 1980 without a corresponding increase in the prevalence data over the same time period (appendix). Likewise, we found a greater than 25 times difference across different countries in age-standardised mortality rates for the most recent year of vital registration data available at the time of GBD 2016 (see Causes of Death Visualization), and we did not see such heterogeneity between countries in our non-fatal data. Therefore, these differences are probably the result of changes and inconsistencies in coding practices for certifying deaths from Parkinson's disease. To correct for this bias, we jointly modelled the prevalence and mortality from Parkinson's disease. First, we ran an initial cause of death model using CODEm, the cause of death ensemble model used throughout GBD, and an initial non-fatal model using DisMod-MR 2.1, the Bayesian meta-regression tool developed for GBD. The initial CODEm model used 14 990 site-years of data (ie, a unique combination of calendar year and country) as well as predictive covariates of SDI,15 cumulative cigarette consumption, health-care access and quality,17 education, and mean cholesterol level (a full list of predictive covariates is in the appendix). The initial DisMod-MR 2.1 model included settings of no remission (ie, no cure) and no incidence before the age of 20 years because the disease is exceptional before that age. We excluded all incidence data from the model, since we saw inconsistencies between the available prevalence and incidence data, and we considered measurement error less likely to occur with prevalence data than with incidence data. We let DisMod-MR 2.1 adjust medical claims data to correct for any systematic under-reporting and datapoints with case definitions that differed from the reference. Smoking prevalence and SDI were used as predictive covariates in the model.

ement error less likely to occur with prevalence data than with incidence data. We let DisMod-MR 2.1 adjust medical claims data to correct for any systematic under-reporting and datapoints with case definitions that differed from the reference. Smoking prevalence and SDI were used as predictive covariates in the model. We used these initial model results to identify countries with high-quality vital registration systems, age-standardised prevalence of more than five per 10 000, and a population of more than 1 million that also had the highest ratios of cause-specific mortality to prevalence, or highest likelihood to code to Parkinson's disease as a cause of death per prevalent case in the most recent year of estimates. For GBD 2016, these countries were Austria, Finland, and the USA. We then used the log-transformed ratios of cause-specific mortality rate to prevalence in 2016 to run a fixed-effects regression with dummy variables on age and sex. Because the ratio between cause-specific mortality rate and prevalence is equivalent to an excess mortality rate or excess rate of dying among people with Parkinson's disease compared with the general population, we used the results of this regression as input data for a second DisMod-MR 2.1 model. The excess mortality data obtained from the regression model were used as data for the entire 1990–2016 period and for every country except for the three used in the regression model, which retained their own data for 2016, and data for these countries were assumed to be constant over the entire time series. Apart from this addition of excess mortality data, the second DisMod-MR 2.1 model was identical to the initial model and used the same settings and covariates. We used the cause-specific mortality and prevalence results from this model as final outputs because they ensured consistency between the available non-fatal input data and the excess mortality rate in 2016 from the three countries most likely to code to Parkinson's disease as a cause of death.

settings and covariates. We used the cause-specific mortality and prevalence results from this model as final outputs because they ensured consistency between the available non-fatal input data and the excess mortality rate in 2016 from the three countries most likely to code to Parkinson's disease as a cause of death. Severity and years lived with disability To calculate years lived with disability (YLDs) for Parkinson's disease, we split the overall prevalence from the second DisMod-MR 2.1 model into three severity categories using data reporting on the Hoehn and Yahr stages.18, 19 We used 30 unique sources, covering nine of 21 world regions, and equated a score of 2·0 or less on the Hoehn and Yahr scale to mild Parkinson's disease, a score of 2·5–3·0 to moderate Parkinson's disease, and a score of 4·0–5·0 to severe Parkinson's disease (appendix). These data informed meta-analyses of the proportion of Parkinson's disease that is mild, moderate, and severe. We then used these proportions to split the overall prevalence of Parkinson's disease into the severity categories. Finally, we multiplied the prevalence of each severity category by severity-specific disability weights20 (see appendix for a detailed description) to calculate YLDs. YLDs were then corrected for comorbidity with a simulation that assigned all non-fatal outcomes to hypothetical individuals and adjusted disability in patients who had multiple conditions.

ce of each severity category by severity-specific disability weights20 (see appendix for a detailed description) to calculate YLDs. YLDs were then corrected for comorbidity with a simulation that assigned all non-fatal outcomes to hypothetical individuals and adjusted disability in patients who had multiple conditions. Risk estimation Of the 84 risks quantified in GBD 2016,21 only smoking was judged to have sufficient evidence for a relationship with Parkinson's disease, with smoking associated with decreased risk.10 The main sources of exposure data were population-based surveys. We used the estimates of exposure, relative risk, and a theoretical minimum level of exposure of zero lifetime cigarettes smoked to calculate population attributable fractions. Further information on risk factor calculations can be found in the GBD 2016 risk factors paper.21 Compilation of results We calculated years of life lost (YLLs) by multiplying the number of deaths in an age group by the remaining life expectancy in that age group, taken from the GBD standard life table.22 Disability-adjusted life-years (DALYs) were then calculated as the sum of YLLs and YLDs. Through each computational step, uncertainty was propagated by sampling 1000 draws, which allowed us to combine the uncertainty from input data, corrections to the data, and residual non-sampling error. Uncertainty intervals (UIs) were defined as the 25th and 975th values of the ordered draws.

Compilation of results We calculated years of life lost (YLLs) by multiplying the number of deaths in an age group by the remaining life expectancy in that age group, taken from the GBD standard life table.22 Disability-adjusted life-years (DALYs) were then calculated as the sum of YLLs and YLDs. Through each computational step, uncertainty was propagated by sampling 1000 draws, which allowed us to combine the uncertainty from input data, corrections to the data, and residual non-sampling error. Uncertainty intervals (UIs) were defined as the 25th and 975th values of the ordered draws. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to the study data and had final responsibility for the decision to submit for publication.

Through each computational step, uncertainty was propagated by sampling 1000 draws, which allowed us to combine the uncertainty from input data, corrections to the data, and residual non-sampling error. Uncertainty intervals (UIs) were defined as the 25th and 975th values of the ordered draws. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to the study data and had final responsibility for the decision to submit for publication. Results The results of our analyses can be downloaded from the Global Health Data Exchange and Institute for Health Metrics and Evaluation (Seattle, WA, USA) results tools. Through the systematic analysis we identified 127 data sources on Parkinson's disease, including 91 sources on prevalence covering 16 of the 21 GBD world regions, 34 sources on incidence covering nine world regions, and 11 sources on mortality risk covering two world regions (appendix). The 11 sources on mortality risk are used in the non-fatal modelling process and therefore are marked as belonging to “non-fatal” in the Global Health Data Exchange input data tool. The 11 sources with data on mortality risk are easily identified by the words “mortality” or “survival” in the title. For prevalence, 40 (44·0%) studies were done in western Europe, nine (9·9%) in east Asia, seven (7·7%) each in high-income Asia Pacific, high-income North America, and North Africa and Middle East. 21 (23·0%) studies were from other regions, except for central Asia, central Latin America, tropical Latin America, central sub-Saharan Africa, and southern sub-Saharan Africa, for which no data were available.

, seven (7·7%) each in high-income Asia Pacific, high-income North America, and North Africa and Middle East. 21 (23·0%) studies were from other regions, except for central Asia, central Latin America, tropical Latin America, central sub-Saharan Africa, and southern sub-Saharan Africa, for which no data were available. In 2016, 6·1 million (95% UI 5·0–7·3) individuals worldwide had Parkinson's disease, of whom 2·9 million (47·5%) were women and 3·2 million (52·5%) were men. 2·1 million (34·4%) of these individuals were from high SDI countries, 3·1 million (50·8%) from high-middle or middle SDI countries, and 0·9 million (14·8%) from low-middle or low SDI countries (table). The number of individuals with Parkinson's disease in 2016 was 2·4 times higher than in 1990 (2·5 million, 95% UI 2·0–3·0). In 1990, 1·1 million (44·0%) cases were in high SDI countries, 1·1 million (44·0%) in high-middle or middle SDI countries, and 0·3 million (12·0%) in low-middle or low SDI countries. The increase in the number of patients with Parkinson's disease worldwide between 1990 and 2016 was not explained exclusively by an increasing number of older people, because global age-standardised prevalence rates increased by 21·7% (95% UI 18·1–25·3) from 1990 to 2016 compared with an increase of 74·3% (69·2–79·6) for crude prevalence rates. The increase in the number of patients with Parkinson's disease between 1990 and 2016 was less pronounced in high SDI countries (9·2%, 95% UI 5·5–13·2) than in other countries, and the largest increase was seen in middle SDI countries (59·8%, 53·2–66·1). The increase in age-standardised prevalence rates between 1990 and 2016 was similar in men (21·4%, 95% UI 17·6–24·9) and women (19·3%, 15·7–22·7). Age-standardised prevalence rates of Parkinson's disease by country varied greater than five times, with the highest rates generally in high-income North America and lowest rates in sub-Saharan Africa (figure 1).Table Deaths, prevalence, and DALYs for Parkinson's disease in 2016 and percentage change between 1990 and 2016 in age-standardised rates by location

inson's disease by country varied greater than five times, with the highest rates generally in high-income North America and lowest rates in sub-Saharan Africa (figure 1).Table Deaths, prevalence, and DALYs for Parkinson's disease in 2016 and percentage change between 1990 and 2016 in age-standardised rates by location Deaths Prevalence DALYs 2016 counts Percentage change in age-standardised rates, 1990–2016 2016 counts Percentage change in age-standardised rates, 1990–2016 2016 counts Percentage change in age-standardised rates, 1990–2016 Global 211 296 (167 771 to 265 160) 19·5% (15·6 to 23·3) 6 062 893 (4 971 461 to 7 324 997) 21·7% (18·1 to 25·3) 3 234 514 (2 563 609 to 4 012 766) 22·1% (18·2 to 25·8) High SDI 84 911 (69 795 to 103 772) 11·3% (7·0 to 16·2) 2 052 069 (1 739 363 to 2 406 677) 9·2% (5·5 to 13·2) 1 128 768 (923 886 to 1 359 135) 9·9% (6·0 to 14·4) High-middle SDI 44 111 (33 506 to 56 880) 13·8% (6·3 to 21·6) 1 328 576 (1 056 629 to 1 639 499) 20·3% (16·4 to 24·2) 682 750 (523 447 to 864 241) 17·1% (10·3 to 24·3) Middle SDI 54 709 (42 446 to 69 607) 49·3% (42·7 to 55·8) 1 778 180 (1 434 399 to 2 166 529) 59·8% (53·2 to 66·1) 942 921 (738 016 to 1 174 509) 57·2% (50·4 to 63·8) Low-middle SDI 23 409 (17 892 to 30 263) 45·4% (36·0 to 56·0) 786 869 (624 622 to 970 524) 31·6% (28·3 to 34·8) 409 620 (316 011 to 515 880) 42·6% (36·0 to 50·5) Low SDI 4061 (3088 to 5273) 34·9% (27·0 to 43·4) 112 859 (88 680 to 141 337) 20·8% (18·3 to 23·5) 68 638 (52 894 to 87 401) 32·4% (26·4 to 39·5) High-income North America 30 461 (27 651 to 33 532) 25·7% (14·0 to 38·9) 811 354 (749 201 to 873 720) 12·8% (2·8 to 23·8) 414 699 (366 012 to 460 392) 18·7% (8·1 to 30·8) Canada 4343 (3131 to 5477) 45·4% (15·1 to 71·6) 103 903 (78 532 to 126 685) 43·0% (16·5 to 67·0) 58 911 (43 090 to 73 670) 44·5% (15·1 to 69·0) Greenland 1 (1 to 2) 12·8% (−4·7 to 33·6) 45 (35 to 57) 12·7% (6·8 to 19·2) 24 (17 to 32) 13·0% (−3·5 to 32·3) USA 26 117 (24 057 to 28 472) 22·4% (8·9 to 36·8) 707 158 (664 026 to 753 627) 9·5% (−0·8 to 21·8) 355 735 (320 852 to 392 813) 15·0% (3·4 to 28·7) Australasia 1979 (1510 to 2588) 16·1% (6·9 to 26·4) 47 265 (37 446 to 58 360) 9·0% (3·5 to 14·2) 27 109 (20 483 to 34 716) 10·9% (3·0 to 19·8) Australia 1721 (1314 to 2260) 15·5% (4·9 to 27·3) 41 016 (32 614 to 50 769) 8·2% (1·8 to 14·4) 23 497 (17 742 to 30 279) 10·1% (1·4 to 19·9) New Zealand 258 (196 to 338) 18·5% (5·5 to 33·0) 6249 (4876 to 7728) 13·6% (7·5 to 19·8) 3612 (2736 to 4638) 14·9% (3·9 to 27·1) High-income Asia

83 to 34 716) 10·9% (3·0 to 19·8) Australia 1721 (1314 to 2260) 15·5% (4·9 to 27·3) 41 016 (32 614 to 50 769) 8·2% (1·8 to 14·4) 23 497 (17 742 to 30 279) 10·1% (1·4 to 19·9) New Zealand 258 (196 to 338) 18·5% (5·5 to 33·0) 6249 (4876 to 7728) 13·6% (7·5 to 19·8) 3612 (2736 to 4638) 14·9% (3·9 to 27·1) High-income Asia Pacific 13 181 (10 069 to 17 172) 11·9% (6·2 to 18·3) 316 347 (248 589 to 395 456) 21·2% (18·6 to 24·0) 174 232 (132 665 to 223 694) 16·5% (11·2 to 22·2) Brunei 5 (4 to 7) 17·9% (3·1 to 31·7) 180 (144 to 221) 12·5% (6·0 to 18·1) 95 (73 to 121) 17·0% (4·0 to 29·7) Japan 10 936 (8270 to 14 260) 10·2% (5·7 to 14·8) 256 455 (201 529 to 321 563) 21·3% (18·6 to 24·2) 141 226 (107 717 to 181 551) 15·7% (11·7 to 19·5) Singapore 165 (119 to 220) 11·3% (−8·3 to 35·9) 4166 (3324 to 5180) 16·2% (10·6 to 22·3) 2270 (1709 to 2923) 12·7% (−4·1 to 34·0) South Korea 2075 (1459 to 2871) 24·6% (−3·9 to 59·9) 55 545 (43 464 to 68 533) 21·0% (14·9 to 28·1) 30 642 (22 126 to 41 397) 21·4% (−3·1 to 50·9) Western Europe 38 233 (29 203 to 50 366) 9·0% (0·5 to 17·3) 828 703 (652 541 to 1 036 222) 8·4% (0·5 to 15·1) 487 578 (369 862 to 629 443) 8·2% (0·0 to 15·7) Andorra 8 (6 to 10) 15·9% (−7·2 to 43·8) 154 (122 to 195) 12·1% (6·5 to 17·7) 94 (70 to 125) 14·1% (−7·2 to 38·0) Austria 744 (543 to 1011) 15·1% (5·4 to 26·9) 15 891 (12 441 to 19 948) 14·2% (8·2 to 20·0) 9574 (6988 to 12 832) 14·6% (5·8 to 24·5) Belgium 975 (732 to 1276) 15·5% (3·1 to 28·7) 20 862 (16 363 to 26 229) 12·4% (6·7 to 18·4) 12 338 (9301 to 15 866) 13·6% (3·2 to 25·3) Cyprus 58 (43 to 75) 4·9% (−4·4 to 15·6) 1245 (982 to 1558) 16·3% (10·4 to 22·3) 757 (580 to 978) 6·0% (−3·0 to 15·4) Denmark 411 (312 to 541) 56·7% (30·8 to 88·6) 9068 (7118 to 11 235) 45·9% (29·4 to 66·9) 5463 (4161 to 7061) 51·3% (28·8 to 79·4) Finland 525 (384 to 714) 15·7% (2·6 to 28·8) 10 258 (8074 to 12 834) 5·8% (−3·6 to 13·7) 6935 (5102 to 9185) 10·3% (−0·8 to 22·0) France 5798 (4370 to 7590) −5·1% (−17·3 to 7·5) 120 455 (94 861 to 151 896) −2·2% (−13·0 to 8·1) 70 410 (53 916 to 90 225) −4·9% (−16·4 to 6·3) Germany 7306 (5402 to 9675) 14·5% (1·9 to 28·9) 162 246 (126 379 to 203 964) 11·5% (2·8 to 19·7) 96 664 (73 054 to 127 109) 12·6% (0·7 to 25·0) Greece 1066 (801 to 1387) 11·6% (1·2 to 22·6) 22 837 (17 855 to 28 778) 13·2% (7·8 to 18·9) 13 376 (10 072 to 17 216) 11·5% (3·0 to 21·3) Iceland 22 (17 to 29) 20·4% (8·2 to 33·4) 474 (375 to 591) 13·4% (8·2 to 18·5) 287 (217 to 373) 17·4% (6·7 to 28·3) Ireland 251 (187 to 335) 18·0% (3·8 to 34·1) 6

12·6% (0·7 to 25·0) Greece 1066 (801 to 1387) 11·6% (1·2 to 22·6) 22 837 (17 855 to 28 778) 13·2% (7·8 to 18·9) 13 376 (10 072 to 17 216) 11·5% (3·0 to 21·3) Iceland 22 (17 to 29) 20·4% (8·2 to 33·4) 474 (375 to 591) 13·4% (8·2 to 18·5) 287 (217 to 373) 17·4% (6·7 to 28·3) Ireland 251 (187 to 335) 18·0% (3·8 to 34·1) 6 001 (4712 to 7491) 17·2% (9·9 to 24·0) 3451 (2607 to 4524) 16·4% (3·8 to 30·5) Israel 411 (304 to 544) −4·8% (−22·5 to 13·6) 9395 (7477 to 11 858) −4·1% (−10·7 to 2·4) 5338 (4016 to 6964) −5·6% (−21·1 to 10·5) Italy 6520 (4878 to 8605) −5·5% (−20·8 to 12·0) 144 606 (113 316 to 180 277) −3·4% (−16·1 to 9·5) 82 834 (62 455 to 108 059) −4·6% (−19·5 to 12·2) Luxembourg 40 (30 to 53) 18·1% (5·9 to 31·4) 873 (681 to 1091) 13·4% (7·9 to 19·9) 520 (395 to 673) 16·0% (5·5 to 27·4) Malta 29 (21 to 39) 15·3% (−2·2 to 36·3) 720 (561 to 906) 15·4% (10·4 to 22·8) 418 (311 to 550) 15·4% (−0·2 to 33·6) Netherlands 1467 (1099 to 1920) −6·0% (−22·7 to 11·7) 33 297 (25 931 to 41 654) −7·5% (−22·1 to 6·1) 19 621 (14 640 to 25 313) −6·7% (−23·9 to 9·2) Norway 342 (258 to 447) 93·0% (54·2 to 141·0) 7517 (5900 to 9463) 87·1% (54·9 to 122·6) 4412 (3338 to 5691) 93·9% (54·9 to 137·2) Portugal 842 (634 to 1119) 34·3% (13·1 to 59·4) 18 496 (14 530 to 23 206) 31·9% (11·5 to 54·4) 10 902 (8249 to 14 165) 32·9% (11·2 to 56·9) Spain 4363 (3322 to 5775) 0·6% (−12·6 to 15·7) 92 971 (73 044 to 116 691) 8·0% (−4·3 to 20·5) 54 175 (41 280 to 70 222) 4·9% (−8·2 to 19·3) Sweden 921 (679 to 1202) 18·3% (6·1 to 33·8) 19 776 (15 538 to 24 631) 13·6% (9·0 to 18·9) 11 805 (8933 to 15 111) 15·1% (4·4 to 27·7) Switzerland 695 (499 to 942) 13·9% (−8·9 to 42·4) 14 979 (11 761 to 18 750) 10·3% (4·4 to 16·0) 8857 (6553 to 11 759) 10·8% (−7·9 to 34·4) UK 5438 (4194 to 7099) 32·5% (28·5 to 36·9) 115 846 (91 722 to 144 139) 22·3% (20·0 to 24·7) 69 262 (53 335 to 88 203) 26·3% (23·0 to 30·1) Southern Latin America 4149 (3109 to 5407) 3·1% (−9·1 to 14·7) 100 190 (77 965 to 125 392) 5·4% (−4·1 to 14·2) 57 932 (43 522 to 73 989) 4·3% (−6·6 to 14·9) Argentina 2798 (2099 to 3623) 0·1% (−13·4 to 13·8) 68 048 (52 574 to 85 157) 2·5% (−8·5 to 13·5) 39 297 (29 708 to 49 971) 1·8% (−10·2 to 14·6) Chile 1064 (756 to 1437) 16·5% (−7·1 to 43·0) 25 845 (20 232 to 32 298) 19·9% (12·8 to 27·5) 14 860 (10 643 to 19 579) 17·1% (−2·7 to 40·4) Uruguay 287 (216 to 375) 7·1% (−2·9 to 18·9) 6289 (4860 to 7952) 10·9% (2·8 to 18·7) 3775 (2870 to 4861) 8·2% (−1·2 to 19·2) Eastern Europe 12 866 (9222 to 17 122) 8·5% (−9·1 to 29·9) 365 078 (

56 to 1437) 16·5% (−7·1 to 43·0) 25 845 (20 232 to 32 298) 19·9% (12·8 to 27·5) 14 860 (10 643 to 19 579) 17·1% (−2·7 to 40·4) Uruguay 287 (216 to 375) 7·1% (−2·9 to 18·9) 6289 (4860 to 7952) 10·9% (2·8 to 18·7) 3775 (2870 to 4861) 8·2% (−1·2 to 19·2) Eastern Europe 12 866 (9222 to 17 122) 8·5% (−9·1 to 29·9) 365 078 ( 282 400 to 459 433) 6·9% (2·4 to 11·4) 197 660 (141 621 to 259 341) 8·6% (−6·4 to 26·8) Belarus 599 (431 to 820) 10·8% (−6·2 to 29·9) 16 588 (12 933 to 20 675) 9·0% (3·6 to 14·1) 9005 (6559 to 11 905) 10·8% (−3·2 to 26·8) Estonia 121 (88 to 158) 5·6% (−10·3 to 22·0) 3078 (2372 to 3877) 1·5% (−12·1 to 12·6) 1722 (1259 to 2228) 3·2% (−12·0 to 18·6) Latvia 178 (127 to 237) 7·7% (−4·6 to 21·1) 4613 (3553 to 5819) 7·6% (2·3 to 14·2) 2586 (1905 to 3373) 8·0% (−3·1 to 20·0) Lithuania 266 (196 to 349) 9·2% (−0·4 to 19·6) 6775 (5234 to 8545) 8·8% (3·1 to 14·9) 3795 (2819 to 4927) 9·6% (0·9 to 18·7) Moldova 161 (120 to 212) 5·9% (−6·0 to 20·5) 4825 (3785 to 6006) 6·5% (1·6 to 12·3) 2575 (1932 to 3332) 6·9% (−3·4 to 19·3) Russia 8516 (5836 to 12 064) 9·0% (−16·5 to 41·7) 244 559 (189 528 to 307 585) 7·1% (0·9 to 13·6) 131 691 (91 670 to 180 532) 9·0% (−12·2 to 36·2) Ukraine 3025 (2094 to 4191) 6·9% (−12·3 to 32·3) 84 640 (65 434 to 106 094) 5·7% (0·6 to 10·7) 46 286 (32 447 to 62 092) 7·1% (−9·3 to 27·4) Central Europe 9061 (6794 to 11 883) 8·6% (3·9 to 13·4) 231 329 (179 712 to 291 637) 10·2% (7·6 to 12·8) 131 027 (99 413 to 168 732) 9·4% (5·1 to 13·7) Albania 150 (111 to 202) 10·9% (−4·3 to 27·1) 4067 (3131 to 5154) 14·0% (7·6 to 20·5) 2295 (1709 to 3026) 12·4% (−0·7 to 26·3) Bosnia and Herzegovina 253 (186 to 334) 16·0% (−0·9 to 36·2) 6631 (5127 to 8317) 19·5% (13·7 to 25·5) 3739 (2788 to 4814) 18·0% (2·9 to 35·5) Bulgaria 665 (485 to 898) −5·3% (−21·7 to 11·5) 16 915 (13 099 to 21 383) 0·8% (−10·7 to 11·1) 9672 (7100 to 12 864) −2·3% (−17·9 to 12·8) Croatia 402 (300 to 533) 7·9% (−6·3 to 21·6) 9662 (7485 to 12 183) 7·9% (2·5 to 13·7) 5653 (4220 to 7339) 8·6% (−3·8 to 20·9) Czech Republic 866 (647 to 1133) 7·6% (−1·5 to 17·6) 22 651 (17 359 to 28 690) 9·3% (4·0 to 15·4) 12 719 (9558 to 16 483) 7·5% (−0·9 to 16·5) Hungary 822 (608 to 1074) 8·3% (−3·9 to 21·5) 20 908 (16 195 to 26 302) 9·5% (4·3 to 16·1) 11 898 (8915 to 15 343) 9·0% (−1·5 to 21·0) Macedonia 108 (80 to 140) 5·9% (−3·1 to 15·7) 3021 (2340 to 3787) 7·7% (2·3 to 13·4) 1705 (1280 to 2213) 6·8% (−2·0 to 15·8) Montenegro 39 (29 to 51) 10·9% (−2·5 to 25·1) 1035 (799 to 1307) 8·4% (2·7 to 14·3) 582 (440 to 752) 10

5) 20 908 (16 195 to 26 302) 9·5% (4·3 to 16·1) 11 898 (8915 to 15 343) 9·0% (−1·5 to 21·0) Macedonia 108 (80 to 140) 5·9% (−3·1 to 15·7) 3021 (2340 to 3787) 7·7% (2·3 to 13·4) 1705 (1280 to 2213) 6·8% (−2·0 to 15·8) Montenegro 39 (29 to 51) 10·9% (−2·5 to 25·1) 1035 (799 to 1307) 8·4% (2·7 to 14·3) 582 (440 to 752) 10 ·3% (−1·2 to 22·4) Poland 2943 (2195 to 3902) 13·7% (2·5 to 25·5) 74 905 (58 130 to 93 685) 14·2% (8·6 to 20·4) 41 955 (31 278 to 54 100) 14·0% (4·0 to 23·8) Romania 1605 (1201 to 2110) 8·1% (−2·4 to 20·3) 40 517 (31 427 to 50 995) 10·2% (4·5 to 15·5) 23 144 (17 467 to 30 057) 9·2% (0·2 to 20·0) Serbia 650 (482 to 832) 8·7% (−1·5 to 20·8) 16 702 (12 943 to 20 877) 7·6% (2·5 to 12·6) 9540 (7187 to 12 219) 8·9% (−0·1 to 19·5) Slovakia 358 (264 to 466) 8·0% (−4·4 to 21·2) 9523 (7411 to 11 952) 9·7% (4·7 to 14·9) 5368 (3965 to 6965) 8·6% (−2·2 to 20·0) Slovenia 200 (150 to 264) 6·7% (−7·5 to 20·9) 4792 (3697 to 6028) 9·7% (4·0 to 15·6) 2755 (2061 to 3601) 7·4% (−5·0 to 20·5) Central Asia 1833 (1353 to 2401) 10·5% (4·6 to 16·9) 56 062 (44 137 to 69 550) 10·4% (7·9 to 12·9) 29 509 (22 429 to 37 368) 10·7% (5·4 to 16·4) Armenia 142 (106 to 188) 13·5% (1·8 to 27·3) 3727 (2918 to 4629) 10·7% (4·1 to 17·1) 2058 (1565 to 2663) 13·1% (2·8 to 24·7) Azerbaijan 235 (165 to 318) 15·1% (−1·6 to 35·6) 7307 (5748 to 9051) 11·9% (7·0 to 17·2) 3836 (2790 to 4989) 13·8% (−1·0 to 31·1) Georgia 236 (171 to 315) 8·6% (−7·2 to 24·9) 5900 (4565 to 7397) 5·2% (0·5 to 11·3) 3370 (2496 to 4368) 8·3% (−5·7 to 23·2) Kazakhstan 397 (288 to 530) 5·5% (−9·4 to 24·9) 13 372 (10 530 to 16 557) 9·0% (3·4 to 14·5) 6764 (4973 to 8711) 6·3% (−6·9 to 22·6) Kyrgyzstan 90 (67 to 117) 6·7% (−2·8 to 17·7) 2802 (2209 to 3474) 5·2% (0·3 to 10·6) 1465 (1107 to 1854) 6·9% (−1·5 to 17·1) Mongolia 40 (29 to 53) 5·2% (−9·0 to 21·2) 1353 (1067 to 1694) 11·2% (5·4 to 17·7) 704 (523 to 916) 5·9% (−6·7 to 19·6) Tajikistan 92 (69 to 121) 12·0% (−1·9 to 30·7) 2884 (2269 to 3573) 8·9% (3·5 to 14·9) 1533 (1164 to 1961) 12·0% (−0·8 to 28·4) Turkmenistan 79 (60 to 103) 14·2% (4·5 to 24·9) 2711 (2144 to 3363) 18·3% (12·3 to 23·9) 1370 (1044 to 1751) 16·2% (7·3 to 25·4) Uzbekistan 522 (387 to 681) 14·0% (2·4 to 25·7) 16 006 (12 659 to 19 798) 15·3% (9·6 to 21·8) 8409 (6379 to 10 644) 15·5% (4·7 to 26·4) Central Latin America 4246 (3249 to 5442) 13·2% (9·2 to 17·2) 129 124 (102 593 to 159 008) 16·6% (14·6 to 18·5) 67 023 (51 781 to 84 193) 14·8% (11·1 to 18·5) Colombia 796 (607 to 1022) 12·2% (1·4 to 25·2) 25 930 (20 527 to 32

25·7) 16 006 (12 659 to 19 798) 15·3% (9·6 to 21·8) 8409 (6379 to 10 644) 15·5% (4·7 to 26·4) Central Latin America 4246 (3249 to 5442) 13·2% (9·2 to 17·2) 129 124 (102 593 to 159 008) 16·6% (14·6 to 18·5) 67 023 (51 781 to 84 193) 14·8% (11·1 to 18·5) Colombia 796 (607 to 1022) 12·2% (1·4 to 25·2) 25 930 (20 527 to 32  111) 15·5% (9·8 to 20·8) 13 140 (10 059 to 16 860) 13·4% (3·5 to 25·1) Costa Rica 110 (83 to 143) 12·8% (2·4 to 24·4) 3230 (2532 to 4003) 15·3% (10·0 to 21·2) 1700 (1300 to 2192) 14·0% (4·6 to 24·3) El Salvador 128 (97 to 168) 13·8% (0·9 to 28·3) 3436 (2680 to 4268) 18·6% (12·2 to 25·1) 1902 (1440 to 2462) 13·7% (2·9 to 26·5) Guatemala 182 (130 to 244) 15·7% (−8·2 to 41·6) 5194 (4068 to 6433) 18·0% (12·3 to 23·8) 2825 (2078 to 3721) 15·7% (−5·8 to 37·6) Honduras 107 (76 to 146) 19·8% (−3·2 to 47·4) 2741 (2166 to 3383) 18·4% (12·8 to 24·4) 1626 (1168 to 2140) 18·7% (−2·0 to 43·7) Mexico 2299 (1752 to 2952) 14·4% (10·3 to 18·6) 68 715 (54 711 to 83 874) 17·7% (15·9 to 19·6) 35 633 (27 612 to 44 856) 16·4% (12·8 to 20·0) Nicaragua 77 (58 to 102) 13·8% (−1·7 to 32·1) 2273 (1799 to 2788) 15·9% (10·6 to 22·0) 1185 (908 to 1519) 14·8% (1·6 to 30·4) Panama 83 (62 to 108) 11·4% (−3·3 to 29·4) 2369 (1878 to 2906) 14·3% (8·5 to 19·4) 1255 (949 to 1629) 12·2% (−0·6 to 26·9) Venezuela 464 (335 to 613) 9·3% (−6·2 to 28·6) 15 235 (12 126 to 18 584) 14·0% (6·9 to 19·9) 7758 (5734 to 10 079) 11·0% (−2·9 to 29·0) Andean Latin America 1093 (832 to 1450) 15·3% (3·7 to 28·5) 30 717 (24 372 to 37 972) 13·0% (9·1 to 16·4) 16 698 (12 634 to 21 390) 14·8% (4·7 to 26·3) Bolivia 202 (147 to 269) 21·1% (3·1 to 43·7) 5114 (4011 to 6330) 14·1% (8·8 to 19·8) 3003 (2237 to 3871) 19·2% (3·0 to 38·5) Ecuador 295 (224 to 383) 10·8% (0·7 to 22·5) 8688 (6872 to 10 735) 13·6% (8·5 to 19·3) 4589 (3519 to 5833) 11·2% (2·3 to 21·8) Peru 596 (436 to 806) 15·9% (−4·6 to 37·7) 16 915 (13 398 to 20 887) 12·3% (6·8 to 18·0) 9106 (6795 to 11 952) 15·2% (−2·4 to 34·8) Caribbean 1169 (883 to 1515) 10·6% (3·6 to 17·5) 31 751 (25 123 to 39 315) 11·4% (7·6 to 14·9) 17 253 (13 197 to 22 090) 11·3% (5·5 to 17·4) Antigua and Barbuda 2 (1 to 3) 14·5% (0·9 to 30·7) 57 (45 to 70) 12·3% (6·8 to 18·0) 31 (23 to 39) 13·3% (1·7 to 27·1) The Bahamas 9 (7 to 12) 11·5% (−1·8 to 25·6) 262 (206 to 325) 9·4% (3·4 to 14·1) 145 (109 to 184) 10·5% (−1·1 to 22·2) Barbados 12 (9 to 17) 18·6% (5·6 to 32·2) 314 (247 to 392) 11·3% (5·5 to 16·7) 180 (136 to 235) 16·2% (5·3 to 27·7) Belize 4 (3 to 5) 30·4% (12·1 to 48·3) 105 (82 to 130)

23 to 39) 13·3% (1·7 to 27·1) The Bahamas 9 (7 to 12) 11·5% (−1·8 to 25·6) 262 (206 to 325) 9·4% (3·4 to 14·1) 145 (109 to 184) 10·5% (−1·1 to 22·2) Barbados 12 (9 to 17) 18·6% (5·6 to 32·2) 314 (247 to 392) 11·3% (5·5 to 16·7) 180 (136 to 235) 16·2% (5·3 to 27·7) Belize 4 (3 to 5) 30·4% (12·1 to 48·3) 105 (82 to 130) 20·1% (13·8 to 25·9) 60 (45 to 76) 28·9% (13·0 to 46·0) Bermuda 2 (1 to 3) 12·8% (−3·1 to 30·5) 52 (41 to 64) 10·7% (6·0 to 15·9) 30 (22 to 38) 11·9% (−2·2 to 26·1) Cuba 507 (381 to 662) 7·9% (−4·4 to 20·2) 12 678 (9985 to 15 908) 9·0% (2·0 to 16·5) 7203 (5459 to 9294) 9·0% (−2·0 to 19·9) Dominica 2 (1 to 3) 17·2% (3·5 to 33·8) 51 (40 to 63) 15·5% (10·1 to 21·0) 29 (22 to 37) 16·5% (4·1 to 31·3) Dominican Republic 202 (149 to 264) 13·8% (−3·2 to 30·9) 5456 (4317 to 6690) 17·5% (11·5 to 23·6) 2999 (2282 to 3831) 14·6% (−0·3 to 30·6) Grenada 2 (2 to 3) 29·3% (12·9 to 45·8) 58 (46 to 72) 22·7% (17·1 to 28·6) 34 (26 to 43) 28·3% (14·1 to 43·2) Guyana 8 (6 to 10) 16·8% (3·8 to 30·2) 297 (235 to 367) 15·1% (10·0 to 20·3) 152 (116 to 193) 16·0% (4·7 to 28·1) Haiti 103 (74 to 139) 21·7% (4·3 to 41·3) 3025 (2368 to 3805) 14·5% (8·4 to 20·2) 1738 (1258 to 2333) 20·7% (4·3 to 39·2) Jamaica 81 (60 to 107) 20·8% (3·1 to 42·6) 1949 (1544 to 2425) 13·7% (7·8 to 19·2) 1123 (827 to 1467) 19·6% (3·4 to 38·1) Puerto Rico 181 (136 to 235) 14·1% (2·4 to 26·9) 4300 (3392 to 5366) 12·2% (7·1 to 18·6) 2478 (1872 to 3177) 13·5% (3·0 to 24·7) Saint Lucia 5 (4 to 7) 21·3% (9·6 to 33·1) 128 (101 to 160) 18·0% (12·8 to 25·4) 73 (56 to 93) 20·8% (10·7 to 31·8) Saint Vincent and the Grenadines 2 (2 to 3) 17·1% (4·5 to 31·4) 62 (49 to 77) 18·6% (12·6 to 24·6) 34 (26 to 44) 17·9% (6·7 to 30·3) Suriname 10 (8 to 13) 20·1% (8·4 to 33·9) 285 (225 to 351) 14·0% (8·5 to 20·1) 161 (123 to 206) 17·8% (6·9 to 29·5) Trinidad and Tobago 32 (24 to 41) 10·8% (1·2 to 20·9) 987 (779 to 1221) 13·2% (7·2 to 18·8) 522 (398 to 663) 11·3% (2·3 to 20·6) Virgin Islands 5 (4 to 7) 10·3% (−1·4 to 24·1) 143 (112 to 181) 12·8% (7·5 to 18·7) 83 (62 to 108) 10·3% (−0·5 to 22·9) Tropical Latin America 4132 (3149 to 5331) 15·2% (11·3 to 19·9) 131 748 (104 807 to 162 882) 16·5% (14·3 to 18·8) 67 778 (52 122 to 85 976) 15·2% (11·8 to 19·2) Brazil 4033 (3074 to 5199) 15·0% (11·1 to 19·6) 128 836 (102 469 to 159 395) 16·4% (14·2 to 18·7) 66 204 (50 914 to 84 027) 15·0% (11·6 to 19·1) Paraguay 98 (74 to 129) 23·4% (8·0 to 39·4) 2912 (2300 to 3602) 19·4% (13·7 to 25·2) 1573 (1189 to 2015) 23·1% (10·7 to 37·0) East Asia 41 584 (3

85 976) 15·2% (11·8 to 19·2) Brazil 4033 (3074 to 5199) 15·0% (11·1 to 19·6) 128 836 (102 469 to 159 395) 16·4% (14·2 to 18·7) 66 204 (50 914 to 84 027) 15·0% (11·6 to 19·1) Paraguay 98 (74 to 129) 23·4% (8·0 to 39·4) 2912 (2300 to 3602) 19·4% (13·7 to 25·2) 1573 (1189 to 2015) 23·1% (10·7 to 37·0) East Asia 41 584 (3 2 292 to 53 090) 77·4% (64·4 to 89·0) 1 451 650 (1 162 770 to 1 793 630) 109·4% (94·2 to 123·8) 734 156 (567 971 to 920 610) 93·8% (80·1 to 107·0) China 40 012 (31 132 to 51 074) 82·7% (68·3 to 95·3) 1 407 701 (1 126 570 to 1 738 833) 115·7% (99·5 to 131·0) 710 041 (549 539 to 890 276) 100·4% (85·3 to 114·9) North Korea 418 (310 to 545) 15·1% (0·1 to 35·4) 14 098 (11 067 to 17 482) 6·0% (1·2 to 10·4) 7527 (5649 to 9693) 14·4% (1·5 to 30·5) Taiwan (province of China) 1154 (868 to 1540) −20·1% (−32·7 to −6·6) 29 851 (23 788 to 36 740) 2·0% (−7·8 to 9·2) 16 588 (12 664 to 21 336) −17·2% (−29·5 to −5·5) Southeast Asia 11 900 (9275 to 14 966) 34·2% (26·6 to 43·6) 409 655 (338 902 to 493 631) 24·7% (21·1 to 29·1) 213 332 (169 151 to 262 446) 31·4% (25·2 to 38·5) Cambodia 172 (130 to 217) 57·6% (38·1 to 88·1) 5768 (4623 to 7126) 26·6% (20·8 to 32·6) 3302 (2563 to 4119) 46·9% (30·9 to 68·6) Indonesia 3490 (2682 to 4487) 55·1% (39·6 to 76·0) 146 236 (117 531 to 178 755) 21·7% (19·1 to 24·2) 70 145 (54 835 to 88 069) 43·4% (32·7 to 56·7) Laos 62 (47 to 81) 48·0% (32·0 to 71·5) 2305 (1851 to 2826) 26·0% (19·9 to 32·3) 1190 (911 to 1512) 42·8% (29·5 to 60·6) Malaysia 514 (386 to 672) 19·9% (8·5 to 32·7) 19 586 (15 697 to 23 908) 26·4% (20·9 to 32·7) 9694 (7546 to 12 275) 19·1% (9·4 to 29·0) Maldives 6 (4 to 8) 25·8% (2·7 to 53·4) 187 (152 to 227) 25·1% (19·7 to 31·1) 96 (71 to 123) 25·8% (6·3 to 51·0) Mauritius 34 (26 to 45) 19·8% (5·9 to 35·2) 1224 (971 to 1492) 21·5% (16·0 to 27·0) 606 (465 to 765) 18·1% (6·2 to 31·8) Myanmar 1129 (859 to 1471) 45·7% (29·6 to 64·5) 28 152 (22 550 to 34 703) 29·4% (23·4 to 35·9) 19 732 (15 161 to 25 096) 41·2% (26·7 to 57·4) Philippines 1132 (851 to 1462) 17·0% (3·1 to 31·4) 45 978 (37 005 to 56 836) 16·9% (11·3 to 22·8) 22 431 (17 069 to 28 411) 17·6% (5·5 to 30·5) Sri Lanka 558 (402 to 758) 8·2% (−11·3 to 30·0) 17 814 (14 263 to 21 949) 20·5% (15·0 to 26·8) 9475 (6993 to 12 384) 11·9% (−5·4 to 31·9) Seychelles 2 (2 to 3) 16·0% (3·2 to 29·1) 76 (62 to 92) 19·0% (13·5 to 25·1) 39 (30 to 49) 15·1% (4·0 to 26·6) Thailand 2400 (1917 to 2980) 25·2% (8·7 to 46·6) 76 568 (66 494 to 86 991) 31·7% (16·3 to 52·6) 40 432 (32 376 to 48 959) 27·6% (12

63 to 21 949) 20·5% (15·0 to 26·8) 9475 (6993 to 12 384) 11·9% (−5·4 to 31·9) Seychelles 2 (2 to 3) 16·0% (3·2 to 29·1) 76 (62 to 92) 19·0% (13·5 to 25·1) 39 (30 to 49) 15·1% (4·0 to 26·6) Thailand 2400 (1917 to 2980) 25·2% (8·7 to 46·6) 76 568 (66 494 to 86 991) 31·7% (16·3 to 52·6) 40 432 (32 376 to 48 959) 27·6% (12 ·3 to 46·0) Timor-Leste 13 (9 to 18) 50·6% (21·4 to 93·9) 454 (357 to 565) 26·6% (21·1 to 33·4) 250 (180 to 334) 44·8% (19·1 to 77·9) Vietnam 2389 (1819 to 3083) 29·9% (11·8 to 51·0) 64 452 (52 241 to 78 194) 25·2% (19·6 to 30·9) 35 840 (28 180 to 44 837) 27·0% (10·9 to 44·2) Oceania 62 (48 to 80) 6·5% (−1·8 to 16·2) 2687 (2124 to 3321) 14·8% (11·7 to 18·2) 1243 (963 to 1560) 9·1% (1·1 to 18·8) American Samoa 1 (0 to 1) −0·8% (−13·9 to 15·6) 21 (17 to 26) 11·5% (6·4 to 16·8) 10 (8 to 13) 1·3% (−11·4 to 15·4) Federated States of Micronesia 1 (1 to 2) 13·4% (−3·6 to 36·1) 32 (25 to 39) 19·1% (12·6 to 25·1) 20 (15 to 26) 14·5% (−2·5 to 37·3) Fiji 10 (7 to 14) 5·4% (−14·9 to 28·7) 386 (306 to 478) 14·7% (9·6 to 20·1) 195 (145 to 256) 7·5% (−11·3 to 30·6) Guam 4 (3 to 5) 5·3% (−7·1 to 18·3) 116 (91 to 144) 9·8% (5·2 to 15·8) 64 (48 to 82) 5·8% (−5·6 to 17·4) Kiribati 1 (1 to 1) 28·3% (11·2 to 47·1) 26 (20 to 32) 15·9% (10·5 to 22·2) 18 (13 to 23) 25·6% (9·7 to 41·2) Marshall Islands 0 (0 to 1) −4·7% (−16·4 to 8·8) 17 (14 to 22) 13·3% (6·7 to 20·2) 8 (6 to 10) −1·9% (−12·6 to 10·5) Northern Mariana Islands 0 (0 to 1) 8·1% (−9·6 to 30·3) 21 (16 to 26) 14·0% (8·9 to 20·2) 9 (7 to 12) 9·5% (−6·6 to 29·2) Papua New Guinea 34 (25 to 44) 18·9% (4·1 to 37·3) 1414 (1114 to 1760) 15·7% (10·8 to 20·9) 692 (526 to 885) 16·9% (3·3 to 34·0) Samoa 3 (2 to 4) 8·5% (−6·4 to 26·0) 67 (53 to 84) 13·2% (7·8 to 19·4) 44 (33 to 56) 9·1% (−4·9 to 25·1) Solomon Islands 4 (3 to 5) 12·7% (−1·1 to 29·3) 120 (94 to 148) 12·9% (6·9 to 19·4) 74 (55 to 95) 13·2% (−0·2 to 28·8) Tonga 2 (1 to 2) 9·8% (−6·4 to 28·9) 41 (32 to 51) 13·9% (7·8 to 20·5) 25 (19 to 32) 11·0% (−3·8 to 29·3) Vanuatu 2 (2 to 3) 14·6% (−0·4 to 31·0) 69 (54 to 85) 14·4% (8·8 to 20·9) 42 (31 to 54) 14·8% (0·6 to 30·5) North Africa and Middle East 9460 (7348 to 12 233) 41·9% (33·2 to 51·7) 297 861 (239 654 to 367 829) 41·4% (37·3 to 46·0) 153 897 (120 636 to 193 371) 42·8% (35·2 to 51·0) Afghanistan 183 (139 to 235) 37·5% (22·1 to 57·9) 5813 (4574 to 7275) 27·0% (20·2 to 35·4) 3436 (2641 to 4334) 36·2% (22·9 to 52·7) Algeria 943 (714 to 1243) 42·4% (27·1 to 59·6) 24 250 (19 187 to 29 987) 38·0% (31·1 to 46·1) 14 128 (10 823 to 18

9) 41·4% (37·3 to 46·0) 153 897 (120 636 to 193 371) 42·8% (35·2 to 51·0) Afghanistan 183 (139 to 235) 37·5% (22·1 to 57·9) 5813 (4574 to 7275) 27·0% (20·2 to 35·4) 3436 (2641 to 4334) 36·2% (22·9 to 52·7) Algeria 943 (714 to 1243) 42·4% (27·1 to 59·6) 24 250 (19 187 to 29 987) 38·0% (31·1 to 46·1) 14 128 (10 823 to 18  044) 42·4% (28·8 to 57·3) Bahrain 11 (8 to 15) 26·4% (3·5 to 54·3) 480 (383 to 603) 29·0% (23·0 to 35·2) 226 (170 to 296) 25·9% (5·4 to 49·7) Egypt 1436 (1075 to 1860) 40·2% (22·7 to 61·2) 48 694 (39 464 to 59 862) 40·7% (33·0 to 49·8) 24 460 (18 678 to 31 133) 41·0% (25·5 to 59·9) Iran 1811 (1343 to 2381) 61·2% (35·4 to 96·4) 59 590 (48 749 to 73 996) 58·4% (49·1 to 71·0) 30 138 (23 315 to 38 332) 62·5% (39·2 to 92·4) Iraq 299 (224 to 391) 23·3% (1·4 to 43·6) 9777 (7735 to 12 063) 22·7% (16·2 to 29·7) 5204 (3940 to 6679) 23·5% (3·1 to 42·8) Jordan 83 (59 to 112) 27·1% (1·4 to 58·2) 2589 (2084 to 3172) 28·7% (18·8 to 37·9) 1353 (999 to 1763) 26·0% (2·9 to 53·6) Kuwait 27 (19 to 37) 42·8% (9·0 to 86·7) 1280 (1019 to 1622) 35·7% (28·4 to 44·1) 572 (417 to 770) 40·5% (10·8 to 78·1) Lebanon 168 (125 to 224) 19·7% (−1·0 to 46·3) 5114 (4051 to 6360) 33·9% (26·5 to 41·8) 2524 (1920 to 3280) 20·8% (2·1 to 44·6) Libya 105 (80 to 136) 38·0% (19·4 to 58·7) 3109 (2499 to 3866) 42·2% (31·4 to 54·7) 1718 (1322 to 2203) 40·6% (23·3 to 59·6) Morocco 695 (528 to 908) 57·3% (37·5 to 102·4) 20 893 (16 569 to 26 080) 39·7% (32·4 to 46·9) 10 968 (8401 to 13 728) 54·2% (38·0 to 84·4) Oman 39 (30 to 50) 63·4% (43·1 to 87·5) 1496 (1200 to 1859) 74·3% (65·0 to 84·5) 724 (567 to 910) 67·9% (49·3 to 89·6) Palestine 34 (25 to 43) 18·2% (2·9 to 35·0) 1143 (907 to 1419) 19·8% (13·7 to 26·0) 595 (461 to 745) 19·7% (5·9 to 34·6) Qatar 11 (7 to 16) 32·5% (−1·3 to 76·5) 562 (441 to 714) 34·0% (26·8 to 41·6) 255 (179 to 354) 33·7% (3·2 to 75·8) Saudi Arabia 337 (263 to 432) 65·7% (43·7 to 94·7) 12 853 (10 251 to 15 936) 65·1% (58·8 to 71·4) 6126 (4814 to 7779) 67·5% (48·4 to 92·1) Sudan 349 (261 to 455) 53·2% (38·6 to 72·5) 11 758 (9351 to 14 608) 38·2% (31·8 to 46·6) 6097 (4668 to 7754) 50·4% (37·8 to 65·3) Syria 233 (178 to 303) 41·8% (27·4 to 56·7) 7409 (5926 to 9203) 40·9% (33·7 to 47·8) 3803 (2945 to 4844) 42·6% (29·4 to 55·3) Tunisia 321 (237 to 430) 40·0% (16·2 to 65·5) 8450 (6790 to 10 486) 44·9% (36·3 to 54·3) 4697 (3568 to 6052) 43·0% (21·4 to 65·8) Turkey 2160 (1603 to 2865) 25·6% (6·8 to 48·0) 63 708 (50 912 to 79 341) 33·8% (27·0 to 40·9) 32 482 (24 631 to 42 095) 27·8% (11·0 to 46·

7 to 47·8) 3803 (2945 to 4844) 42·6% (29·4 to 55·3) Tunisia 321 (237 to 430) 40·0% (16·2 to 65·5) 8450 (6790 to 10 486) 44·9% (36·3 to 54·3) 4697 (3568 to 6052) 43·0% (21·4 to 65·8) Turkey 2160 (1603 to 2865) 25·6% (6·8 to 48·0) 63 708 (50 912 to 79 341) 33·8% (27·0 to 40·9) 32 482 (24 631 to 42 095) 27·8% (11·0 to 46· 7) United Arab Emirates 39 (29 to 51) 48·6% (18·7 to 86·5) 2498 (1953 to 3181) 41·4% (34·5 to 49·4) 1101 (809 to 1473) 50·5% (21·1 to 84·5) Yemen 178 (134 to 230) 60·9% (41·2 to 86·1) 6160 (4866 to 7739) 45·6% (38·0 to 54·0) 3262 (2521 to 4155) 58·1% (40·0 to 79·7) South Asia 21 007 (15 942 to 27 123) 48·3% (36·8 to 62·4) 696 108 (552 987 to 860 047) 29·2% (25·6 to 32·8) 364 282 (281 631 to 460 552) 43·8% (35·9 to 53·4) Bangladesh 1501 (1138 to 1953) −10·2% (−22·6 to 6·2) 54 198 (42 488 to 67 532) 25·0% (18·8 to 31·7) 25 363 (19 435 to 32 110) −3·8% (−15·8 to 9·8) Bhutan 11 (8 to 15) 43·5% (21·0 to 73·5) 288 (227 to 357) 35·8% (28·9 to 43·7) 171 (129 to 222) 41·1% (21·1 to 66·1) India 17 539 (13 317 to 22 637) 55·8% (42·3 to 71·8) 575 946 (458 316 to 712 213) 29·7% (25·9 to 33·5) 305 274 (235 390 to 385 725) 49·6% (40·9 to 60·1) Nepal 319 (236 to 419) 66·7% (38·1 to 107·3) 9445 (7390 to 11 806) 29·5% (23·2 to 36·4) 5449 (4071 to 7009) 54·5% (33·8 to 82·1) Pakistan 1637 (1213 to 2148) 43·8% (23·6 to 70·0) 56 231 (43 998 to 70 068) 27·9% (21·7 to 35·7) 28 025 (21 124 to 36 119) 39·5% (22·1 to 59·7) Southern sub-Saharan Africa 756 (577 to 988) 25·6% (16·4 to 36·8) 20 980 (16 480 to 26 027) 14·3% (12·0 to 16·6) 11 750 (9090 to 14 869) 23·5% (16·2 to 33·0) Botswana 15 (8 to 23) 26·3% (−23·0 to 69·2) 460 (361 to 569) 22·0% (15·3 to 28·8) 264 (159 to 384) 25·3% (−18·4 to 63·7) Lesotho 14 (10 to 19) 17·2% (−5·3 to 44·2) 397 (312 to 501) 18·0% (11·6 to 24·5) 223 (161 to 295) 16·8% (−4·2 to 41·0) Namibia 16 (10 to 23) 25·7% (−8·4 to 55·2) 442 (349 to 551) 20·2% (13·4 to 28·0) 272 (178 to 374) 23·6% (−6·4 to 50·5) South Africa 594 (454 to 787) 21·3% (12·5 to 31·6) 17 305 (13 650 to 21 444) 14·2% (11·7 to 16·7) 9245 (7128 to 11 715) 19·9% (12·7 to 28·2) Swaziland 7 (4 to 10) 11·2% (−15·0 to 40·7) 215 (168 to 270) 15·4% (8·7 to 22·4) 118 (77 to 164) 11·5% (−11·3 to 38·8) Zimbabwe 110 (78 to 146) 55·9% (25·6 to 143·8) 2162 (1701 to 2685) 8·5% (2·7 to 14·3) 1628 (1189 to 2120) 47·0% (21·1 to 117·4) Western sub-Saharan Africa 1578 (1170 to 2059) 22·6% (13·5 to 32·1) 44 230 (34 637 to 55 905) 15·9% (12·6 to 19·3) 27 359 (20 483 to 35 211) 21·7% (13·2 to 30·5) Benin 47 (35 to 62)

3 to 38·8) Zimbabwe 110 (78 to 146) 55·9% (25·6 to 143·8) 2162 (1701 to 2685) 8·5% (2·7 to 14·3) 1628 (1189 to 2120) 47·0% (21·1 to 117·4) Western sub-Saharan Africa 1578 (1170 to 2059) 22·6% (13·5 to 32·1) 44 230 (34 637 to 55 905) 15·9% (12·6 to 19·3) 27 359 (20 483 to 35 211) 21·7% (13·2 to 30·5) Benin 47 (35 to 62) 15·6% (2·3 to 31·1) 1196 (932 to 1541) 14·1% (7·5 to 20·2) 811 (610 to 1058) 16·2% (4·0 to 29·4) Burkina Faso 59 (44 to 79) 11·6% (−0·9 to 29·1) 1537 (1199 to 1939) 14·4% (8·1 to 21·8) 1023 (777 to 1308) 11·3% (0·1 to 26·4) Cameroon 147 (105 to 195) 15·5% (−1·4 to 34·6) 3077 (2420 to 3870) 11·6% (6·1 to 16·9) 2268 (1646 to 2958) 15·3% (−0·7 to 33·6) Cape Verde 6 (4 to 8) 31·6% (16·4 to 47·4) 127 (100 to 157) 24·5% (18·2 to 31·0) 78 (59 to 101) 30·3% (16·0 to 44·6) Chad 51 (38 to 67) 11·3% (−0·6 to 24·3) 1284 (1005 to 1642) 12·0% (6·4 to 18·3) 839 (635 to 1087) 11·3% (−0·4 to 24·6) Côte d'Ivoire 110 (81 to 144) 20·8% (8·7 to 35·4) 2747 (2142 to 3475) 13·7% (8·5 to 19·8) 1923 (1440 to 2501) 20·4% (9·0 to 34·6) The Gambia 8 (6 to 10) 18·2% (3·2 to 35·1) 186 (146 to 237) 9·6% (3·8 to 15·8) 126 (96 to 162) 17·0% (3·2 to 32·0) Ghana 171 (126 to 225) 28·4% (12·5 to 50·3) 4048 (3183 to 5118) 13·5% (7·8 to 19·4) 2768 (2085 to 3568) 26·5% (11·5 to 45·2) Guinea 58 (43 to 78) 17·9% (3·1 to 33·5) 1441 (1121 to 1836) 11·0% (5·8 to 16·8) 993 (733 to 1315) 17·5% (2·8 to 32·3) Guinea-Bissau 8 (6 to 11) 16·2% (2·1 to 33·4) 217 (169 to 274) 13·1% (8·0 to 18·8) 148 (111 to 191) 15·5% (2·7 to 30·5) Liberia 19 (14 to 25) 17·6% (4·5 to 32·5) 508 (400 to 648) 10·6% (5·2 to 16·6) 332 (249 to 427) 17·1% (5·6 to 30·8) Mali 65 (47 to 86) 23·1% (6·3 to 44·5) 1567 (1224 to 2024) 13·2% (7·4 to 18·9) 1049 (767 to 1395) 20·0% (4·4 to 38·7) Mauritania 23 (17 to 31) 20·5% (0·3 to 42·0) 565 (441 to 708) 15·7% (9·5 to 21·7) 381 (278 to 501) 18·9% (0·4 to 38·8) Niger 62 (45 to 82) 14·0% (−1·4 to 34·1) 1732 (1351 to 2224) 8·0% (2·4 to 14·5) 1105 (809 to 1450) 13·1% (−1·6 to 31·5) Nigeria 613 (429 to 835) 27·8% (6·4 to 48·8) 20 851 (16 306 to 26 258) 19·8% (13·4 to 26·6) 11 302 (8080 to 15 109) 26·5% (7·3 to 45·9) São Tomé and Príncipe 1 (1 to 2) 22·2% (4·1 to 44·0) 27 (22 to 34) 13·8% (8·2 to 20·1) 17 (13 to 23) 19·5% (2·9 to 38·4) Senegal 80 (60 to 105) 24·8% (13·2 to 39·5) 1688 (1322 to 2154) 11·2% (5·5 to 17·1) 1274 (965 to 1646) 22·2% (11·4 to 35·5) Sierra Leone 20 (15 to 26) 15·4% (2·2 to 31·0) 637 (498 to 814) 12·7% (6·4 to 18·7) 384 (286 to 496) 15·6% (2·7 to 29·6) Togo 31 (23 to 40) 17·5% (4

(8·2 to 20·1) 17 (13 to 23) 19·5% (2·9 to 38·4) Senegal 80 (60 to 105) 24·8% (13·2 to 39·5) 1688 (1322 to 2154) 11·2% (5·5 to 17·1) 1274 (965 to 1646) 22·2% (11·4 to 35·5) Sierra Leone 20 (15 to 26) 15·4% (2·2 to 31·0) 637 (498 to 814) 12·7% (6·4 to 18·7) 384 (286 to 496) 15·6% (2·7 to 29·6) Togo 31 (23 to 40) 17·5% (4 ·3 to 33·1) 794 (619 to 1007) 15·1% (8·8 to 21·6) 538 (403 to 695) 17·9% (5·0 to 32·3) Eastern sub-Saharan Africa 1975 (1493 to 2591) 34·6% (25·8 to 44·6) 46 489 (36 657 to 57 613) 21·7% (18·6 to 25·0) 30 752 (23 631 to 39 371) 31·5% (24·0 to 40·0) Burundi 42 (30 to 57) 25·4% (8·5 to 43·7) 1115 (880 to 1407) 15·9% (9·4 to 23·1) 683 (511 to 896) 23·0% (7·5 to 41·0) Comoros 4 (3 to 5) 31·0% (14·4 to 48·5) 101 (80 to 126) 23·0% (16·6 to 29·8) 65 (50 to 84) 29·1% (14·5 to 45·8) Djibouti 6 (5 to 9) 38·2% (14·5 to 63·0) 153 (119 to 191) 27·3% (20·3 to 35·4) 101 (74 to 135) 35·6% (13·5 to 57·3) Eritrea 19 (14 to 26) 43·7% (26·1 to 63·1) 531 (419 to 668) 26·9% (19·7 to 34·1) 344 (256 to 442) 39·7% (24·6 to 57·2) Ethiopia 556 (405 to 750) 40·7% (21·9 to 63·6) 12 384 (9654 to 15 735) 24·9% (17·8 to 32·4) 8580 (6437 to 11 338) 36·9% (19·5 to 56·8) Kenya 232 (169 to 309) 40·9% (23·6 to 67·9) 6557 (5228 to 8065) 22·0% (19·7 to 24·7) 3714 (2782 to 4800) 36·4% (22·8 to 54·2) Madagascar 99 (73 to 134) 17·2% (0·3 to 34·9) 2940 (2333 to 3665) 14·1% (7·6 to 20·2) 1657 (1221 to 2127) 15·9% (0·2 to 32·0) Malawi 97 (69 to 130) 28·3% (3·4 to 60·4) 2112 (1650 to 2683) 19·0% (11·6 to 25·6) 1497 (1091 to 1973) 26·9% (3·1 to 56·7) Mozambique 157 (114 to 212) 21·0% (4·1 to 42·4) 3376 (2626 to 4246) 20·7% (13·1 to 27·9) 2420 (1783 to 3224) 20·7% (4·6 to 40·7) Rwanda 61 (44 to 81) 52·5% (32·3 to 74·8) 1376 (1089 to 1714) 24·6% (17·9 to 31·6) 903 (660 to 1176) 45·8% (27·7 to 68·3) Somalia 42 (31 to 57) 19·5% (4·9 to 37·2) 1107 (865 to 1405) 13·8% (7·4 to 19·7) 700 (517 to 919) 17·8% (4·3 to 34·4) South Sudan 52 (37 to 71) 34·2% (13·5 to 58·2) 1421 (1115 to 1789) 15·9% (9·2 to 23·1) 852 (616 to 1125) 30·5% (11·5 to 51·7) Tanzania 353 (266 to 460) 35·2% (17·6 to 56·2) 7443 (6048 to 9048) 24·0% (13·4 to 34·4) 5302 (4081 to 6708) 31·5% (15·5 to 51·4) Uganda 167 (123 to 226) 29·6% (13·2 to 49·0) 3873 (3046 to 4804) 23·9% (17·1 to 30·6) 2553 (1929 to 3310) 28·6% (13·3 to 46·5) Zambia 88 (63 to 119) 33·8% (6·1 to 70·2) 1963 (1538 to 2453) 20·4% (14·2 to 27·3) 1376 (1009 to 1844) 33·1% (6·5 to 66·0) Central sub-Saharan Africa 474 (346 to 621) 22·0% (12·7 to 33·8) 13 564 (10 608 to 17 192) 10·1% (5·9

3·2 to 49·0) 3873 (3046 to 4804) 23·9% (17·1 to 30·6) 2553 (1929 to 3310) 28·6% (13·3 to 46·5) Zambia 88 (63 to 119) 33·8% (6·1 to 70·2) 1963 (1538 to 2453) 20·4% (14·2 to 27·3) 1376 (1009 to 1844) 33·1% (6·5 to 66·0) Central sub-Saharan Africa 474 (346 to 621) 22·0% (12·7 to 33·8) 13 564 (10 608 to 17 192) 10·1% (5·9 to 14·5) 8083 (6104 to 10 365) 19·4% (11·3 to 29·8) Angola 88 (62 to 123) 50·2% (21·6 to 89·2) 2615 (2034 to 3271) 19·7% (13·2 to 26·4) 1555 (1125 to 2107) 41·8% (16·8 to 74·1) Central African Republic 24 (17 to 32) 10·9% (−4·9 to 26·3) 717 (559 to 915) 8·8% (2·7 to 14·9) 401 (293 to 525) 10·5% (−3·8 to 25·4) Congo (Brazzaville) 30 (21 to 41) 23·6% (3·5 to 45·1) 789 (623 to 986) 16·4% (9·7 to 23·5) 490 (361 to 648) 21·5% (2·7 to 42·5) Democratic Republic of the Congo 306 (224 to 404) 16·0% (4·8 to 30·1) 8845 (6920 to 11 285) 6·7% (1·0 to 12·2) 5244 (3904 to 6686) 14·4% (4·8 to 26·2) Equatorial Guinea 5 (3 to 8) 58·3% (11·8 to 108·9) 159 (126 to 197) 42·3% (33·8 to 50·8) 91 (59 to 131) 52·4% (12·3 to 97·9) Gabon 21 (15 to 28) 28·3% (9·4 to 51·0) 439 (346 to 550) 18·1% (12·4 to 24·8) 302 (223 to 395) 25·8% (7·9 to 46·6) 95% uncertainty intervals are in parentheses. DALYs=disability-adjusted life-years. SDI=Socio-demographic Index. For more details about the rationale for this classification of countries see reference 23. Figure 1 Age-standardised prevalence of Parkinson's disease per 100 000 population by location for both sexes, 2016

to 14·5) 8083 (6104 to 10 365) 19·4% (11·3 to 29·8) Angola 88 (62 to 123) 50·2% (21·6 to 89·2) 2615 (2034 to 3271) 19·7% (13·2 to 26·4) 1555 (1125 to 2107) 41·8% (16·8 to 74·1) Central African Republic 24 (17 to 32) 10·9% (−4·9 to 26·3) 717 (559 to 915) 8·8% (2·7 to 14·9) 401 (293 to 525) 10·5% (−3·8 to 25·4) Congo (Brazzaville) 30 (21 to 41) 23·6% (3·5 to 45·1) 789 (623 to 986) 16·4% (9·7 to 23·5) 490 (361 to 648) 21·5% (2·7 to 42·5) Democratic Republic of the Congo 306 (224 to 404) 16·0% (4·8 to 30·1) 8845 (6920 to 11 285) 6·7% (1·0 to 12·2) 5244 (3904 to 6686) 14·4% (4·8 to 26·2) Equatorial Guinea 5 (3 to 8) 58·3% (11·8 to 108·9) 159 (126 to 197) 42·3% (33·8 to 50·8) 91 (59 to 131) 52·4% (12·3 to 97·9) Gabon 21 (15 to 28) 28·3% (9·4 to 51·0) 439 (346 to 550) 18·1% (12·4 to 24·8) 302 (223 to 395) 25·8% (7·9 to 46·6) 95% uncertainty intervals are in parentheses. DALYs=disability-adjusted life-years. SDI=Socio-demographic Index. For more details about the rationale for this classification of countries see reference 23. Figure 1 Age-standardised prevalence of Parkinson's disease per 100 000 population by location for both sexes, 2016 ATG=Antigua and Barbuda. FSM=Federated States of Micronesia. Isl=Islands. LCA=Saint Lucia. TLS=Timor-Leste. TTO=Trinidad and Tobago. VCT=Saint Vincent and the Grenadines.

DALYs=disability-adjusted life-years. SDI=Socio-demographic Index. For more details about the rationale for this classification of countries see reference 23. Figure 1 Age-standardised prevalence of Parkinson's disease per 100 000 population by location for both sexes, 2016 ATG=Antigua and Barbuda. FSM=Federated States of Micronesia. Isl=Islands. LCA=Saint Lucia. TLS=Timor-Leste. TTO=Trinidad and Tobago. VCT=Saint Vincent and the Grenadines. Globally, Parkinson's disease caused 211 296 deaths (95% UI 167 771–265 160; 93 512, 95% UI 73 702–118 421, in women and 117 784, 93 729–147 607, in men) and 3·2 million DALYs (95% UI 2·6–4·0; 1·4 million, 95% UI 1·1–1·7, in women; 1·8 million, 1·5–2·3, in men) in 2016. Of these, high SDI countries accounted for 84 911 (40·2%) deaths and 1·1 million (34·4%) DALYs, high-middle or middle SDI countries for 98 820 (46·8%) deaths and 1·6 million (50·0%) DALYs, and low-middle or low SDI countries for 27 470 (13·0%) deaths and 0·5 million (15·6%) DALYs. The number of deaths was 2·6 times higher and the number of DALYs was 2·5 times higher in 2016 than in 1990. These increases were not explained exclusively by an increasing number of older people, because age-standardised rates increased from 1990 to 2016 for both deaths and DALYs by about 20% (table). Similar to prevalence, the increases in deaths and DALYs were lowest in high SDI countries and highest in middle SDI countries, and were seen in both men and women.

usively by an increasing number of older people, because age-standardised rates increased from 1990 to 2016 for both deaths and DALYs by about 20% (table). Similar to prevalence, the increases in deaths and DALYs were lowest in high SDI countries and highest in middle SDI countries, and were seen in both men and women. Parkinson's disease was uncommon before 50 years of age. Prevalence in 2016 increased with age thereafter and peaked between 85 years and 89 years (1·7% for men; 1·2% for women) and decreased after that age (figure 2). Age-standardised prevalence of Parkinson's disease in 2016 was 1·40 times (95% UI 1·36–1·43) higher in men than in women; the male-to-female ratio was similar in 1990 (1·37, 95% UI 1·34–1·40). A similar pattern was seen for the rates of YLLs and YLDs according to age, although the age-related increase was considerably steeper for YLLs than for YLDs, suggesting that Parkinson's disease-related case fatality rises with age (figure 3).Figure 2 Global prevalence of Parkinson's disease by age and sex, 2016 Prevalence is expressed as the percentage of the population that is affected by the disease. Shading indicates 95% uncertainty intervals. Figure 3 Global YLD and YLL rates per 100 000 population due to Parkinson's disease by age, 2016 Shading indicates 95% uncertainty intervals. YLDs=years lived with disability. YLLs=years of life lost.

Prevalence is expressed as the percentage of the population that is affected by the disease. Shading indicates 95% uncertainty intervals. Figure 3 Global YLD and YLL rates per 100 000 population due to Parkinson's disease by age, 2016 Shading indicates 95% uncertainty intervals. YLDs=years lived with disability. YLLs=years of life lost. The age-standardised rate of DALYs of Parkinson's disease by the 21 GBD world regions generally increased with SDI (figure 4). Sub-Saharan Africa, Latin America and the Caribbean (with the exception of southern Latin America), south Asia, and high-income Asia Pacific had lower age-standardised DALY rates than other regions with similar SDI. Southern Latin America and high-income North America were the regions with highest estimates relative to expected values based on SDI.Figure 4 Age-standardised DALY rates for Parkinson's disease by 21 Global Burden of Disease regions by Socio-demographic Index, 1990–2016 Expected values based on Socio-demographic Index and disease rates in all locations are shown as the black line. The black line represents expected values of age-standardised DALY rates for each value of Socio-demographic Index and is based on a Gaussian process regression of results for all Global Burden of Disease locations over the entire 1990–2016 estimation period. DALY=disability-adjusted life-year. Smoking was found to have a small, protective effect on Parkinson's disease, and would have been expected to prevent 461 194 DALYs (95% UI 324 745–599 845) globally in 2016 if the association was truly causal.

Expected values based on Socio-demographic Index and disease rates in all locations are shown as the black line. The black line represents expected values of age-standardised DALY rates for each value of Socio-demographic Index and is based on a Gaussian process regression of results for all Global Burden of Disease locations over the entire 1990–2016 estimation period. DALY=disability-adjusted life-year. Smoking was found to have a small, protective effect on Parkinson's disease, and would have been expected to prevent 461 194 DALYs (95% UI 324 745–599 845) globally in 2016 if the association was truly causal. Discussion Over the past generation, the number of individuals with Parkinson's disease globally has more than doubled to over 6 million. Of all the neurological disorders included in GBD 2015,1 Parkinson's disease was the fastest growing. Ageing populations contributed to much of that growth as crude prevalence rates increased by about 74% from 1990 to 2016 and age-standardised prevalence rates increased by about 22%. However, because age-standardised prevalence, DALYs, and death rates all increased from 1990 to 2016, additional factors are probably important.

populations contributed to much of that growth as crude prevalence rates increased by about 74% from 1990 to 2016 and age-standardised prevalence rates increased by about 22%. However, because age-standardised prevalence, DALYs, and death rates all increased from 1990 to 2016, additional factors are probably important. First, changes in study methods, availability of higher-quality studies, and greater awareness of diagnosis24 might have led to better estimates of prevalence, DALYs, and deaths since 1990.3 For example, door-to-door studies are less likely to miss individuals who have never been diagnosed and would be missed in health records.25 Our DisMod-MR 2.1 model did not show evidence in favour of a systematic bias between door-to-door surveys and studies based on administrative records; however, establishing such evidence in a model with relatively sparse data is difficult. Although many regions have seen improvements in study methods, these changes alone are probably insufficient to explain the rising burden of Parkinson's disease. For example, the rates of Parkinson's disease have also increased in high-income countries without substantial changes in study methodology.

ata is difficult. Although many regions have seen improvements in study methods, these changes alone are probably insufficient to explain the rising burden of Parkinson's disease. For example, the rates of Parkinson's disease have also increased in high-income countries without substantial changes in study methodology. Second, increasing life expectancy is probably contributing to longer disease duration in individuals with Parkinson's disease and thus to higher prevalence, even if incidence remains constant and individuals with Parkinson's disease show the same time trends in mortality as the general population.12 Indeed, in a meta-analysis26 of ten studies, recent cohorts showed longer disease duration, with an increase of 2·5 years per decade. That study26 showed no clear evidence that the introduction of levodopa and improvements in Parkinson's disease care have led to improvement in survival of individuals with Parkinson's disease compared with similar individuals without Parkinson's disease. A recent study12 estimated the burden of Parkinson's disease in France from 2010 to 2030 under a constant incidence scenario and assuming that the relative risk of death of individuals with Parkinson's disease relative to controls had not changed over time. It showed that the life expectancy of individuals with Parkinson's disease would be expected to increase by approximately 3 years and the age-standardised and sex-standardised prevalence rate by 12% over 20 years. As patients live with Parkinson's disease for more years and the number of individuals with advanced Parkinson's disease increases, studies will be needed to inform the distribution of the severity of the disease in representative samples with simple instruments such as the Hoehn and Yahr scale.

e by 12% over 20 years. As patients live with Parkinson's disease for more years and the number of individuals with advanced Parkinson's disease increases, studies will be needed to inform the distribution of the severity of the disease in representative samples with simple instruments such as the Hoehn and Yahr scale. Third, the increase in Parkinson's disease burden might be linked to environmental factors tied to the growing industrialisation of the world. In general, better health is positively associated with socioeconomic level.27, 28, 29 However, with Parkinson's disease, the opposite is true; age-standardised DALY rates due to Parkinson's disease increased with SDI. The reason for this association is not clear. Some environmental exposures tied to industrialisation, including pesticides,7 solvents,7 or metals,8, 9 which are also more common in high SDI countries, might contribute to the increased incidence of Parkinson's disease. For example, in China (a middle SDI country), which has undergone rapid industrial growth since 1990, the age-adjusted prevalence rates of Parkinson's disease more than doubled between 1990 and 2016, the largest increase worldwide. If environmental factors related to industrialisation played a part, an increase in incidence over time would be expected. A few studies have examined time trends in incidence of Parkinson's disease, with inconsistent findings. A study in the USA30 suggested that incidence increased by 24% (95% CI 8–43) per decade between 1976 and 2005 in men but not in women. By contrast, in the Netherlands, the Rotterdam Study31 reported a substantial decrease in Parkinson's disease incidence between 1990 and 2011, without any obvious explanation. One study in Canada32 and another in the USA33 study showed no significant time trends. In Finland, using the Finnish National Prescription Register, the incidence of Parkinson's disease increased between 1997 and 2014 both in rural and urban regions.34 High-quality prospective cohort studies and detailed registries are needed to survey time trends in the worldwide incidence of Parkinson's disease more accurately and understand the factors that might be driving time trends. Alternative explanations for the positive association between the burden of Parkinson's disease and SDI include better ascertainment of Parkinson's disease in higher SDI countries through better study methods or health-care access and disease recognition.

ly and understand the factors that might be driving time trends. Alternative explanations for the positive association between the burden of Parkinson's disease and SDI include better ascertainment of Parkinson's disease in higher SDI countries through better study methods or health-care access and disease recognition. However, in lower SDI countries, we included door-to-door studies, when available, that are considered to be less prone to underestimation.25

ly and understand the factors that might be driving time trends. Alternative explanations for the positive association between the burden of Parkinson's disease and SDI include better ascertainment of Parkinson's disease in higher SDI countries through better study methods or health-care access and disease recognition. However, in lower SDI countries, we included door-to-door studies, when available, that are considered to be less prone to underestimation.25 Fourth, declining smoking rates in some countries,35 although a global health boon, might contribute to higher incidence of Parkinson's disease.14 The risk of Parkinson's disease is decreased by approximately 40% among smokers.10 Whether this association is truly causal or explained by reverse causation or other biases is still debated.11 If the association between smoking and Parkinson's disease were causal, decreasing smoking rates would lead to an increase in the incidence of Parkinson's disease in the future. Assuming a causal inverse association and a 10-year lag to account for the temporal effect of smoking on the incidence of Parkinson's disease, one study in the USA14 estimated that declining smoking rates in the country might increase the projected burden of Parkinson's disease in 2040 by 10%. However, because the lag time between exposure and the actual effect on disease risk is unknown and might actually be longer than 10 years,36 the timing of the potential effect of declining smoking rates on Parkinson's disease incidence is uncertain, and additional studies are needed to examine the effects of changing smoking habits in different parts of the world with different smoking rates and time trends. Regardless of the results of such studies, the adverse health consequences of smoking far outweigh any potential benefit on Parkinson's disease. Finally, changes in the prevalence of other known (eg, head trauma37) or unknown risk or protective factors that were not included in GBD 2016 might contribute to changing incidence rates of Parkinson's disease.

udies, the adverse health consequences of smoking far outweigh any potential benefit on Parkinson's disease. Finally, changes in the prevalence of other known (eg, head trauma37) or unknown risk or protective factors that were not included in GBD 2016 might contribute to changing incidence rates of Parkinson's disease. This study confirms that Parkinson's disease is about 1·4 times more frequent in men than women, and this ratio did not change substantially over the study period. Environmental (eg, occupational) exposures to which men are more frequently exposed might contribute to this pattern. The prevalence of Parkinson's disease increased with age. Underascertainment at older ages owing to underdiagnosis, comorbidities, or institutional care might explain the decrease seen in the oldest age groups after the peak between 85 years and 89 years.

en are more frequently exposed might contribute to this pattern. The prevalence of Parkinson's disease increased with age. Underascertainment at older ages owing to underdiagnosis, comorbidities, or institutional care might explain the decrease seen in the oldest age groups after the peak between 85 years and 89 years. The current estimates for the global burden of Parkinson's disease are generated from imperfect data and models that are refined in each iteration of the GBD study. Estimates from GBD studies can vary from year to year as revised estimates are generated on the basis of refined methods and inclusion of more and higher-quality studies that are less likely to underestimate the true burden of the disease. Nonetheless, high-quality epidemiological studies (especially on incidence and disease severity) are still rare for large portions of the world, especially in low-income regions, where such studies are needed to understand trends and guide efforts to reduce the disease burden. Methodological differences for determining prevalence and study shortcomings might result in estimates that vary considerably and underestimate the true burden of Parkinson's disease.38 This under-reporting is well known for Parkinson's disease in studies based on death certificates,39, 40, 41, 42, 43 whereas population-based door-to-door studies are considered a better approach because they are able to capture undiagnosed cases.25 However, disease frequency estimates from population-based studies might be affected by selection bias resulting from non-response, particularly if individuals affected by the disease under investigation are less likely to participate. Non-response is an important issue as participation rates in epidemiological studies have considerably decreased over the past 30 years.44 Another limitation is that because we lack strong predictors for the occurrence of Parkinson's disease, some of the variation between countries is probably due to measurement error that we have been unable to correct. Because we rely on prevalence data to derive our cause of death estimates, any residual measurement error in the prevalence estimates is transposed onto the death estimates for Parkinson's disease. Although the the bias in mortality is an unwanted property, it is less than would have been the case if we had based our estimates on the observed rates of death with Parkinson's disease as the underlying cause from vital registration data.

lence estimates is transposed onto the death estimates for Parkinson's disease. Although the the bias in mortality is an unwanted property, it is less than would have been the case if we had based our estimates on the observed rates of death with Parkinson's disease as the underlying cause from vital registration data. The large variation in death rates over time within the same countries and the even larger variation between countries are implausible and probably explained by changing death coding practices. Neurological disorders are now the leading source of disability in the world, and Parkinson's disease is the fastest growing of these disorders.1 As the population ages and life expectancy increases, the doubling of the number of individuals with Parkinson's disease between 1990 and 2016 is projected to occur again in the coming generation.12, 13, 14, 45 To address this great health challenge will require action aimed at preventing the disease where feasible and improving the lives of those affected by the condition.46 Among the potential responses available are preventing the disease (eg, by increasing physical activity earlier in adulthood47 and reducing exposure to pesticides7), improving worldwide access to care and effective treatments (eg, levodopa), increasing funding for research (eg, to understand the underlying causes), and development of new therapies. Supplementary Material Supplementary appendix

Neurological disorders are now the leading source of disability in the world, and Parkinson's disease is the fastest growing of these disorders.1 As the population ages and life expectancy increases, the doubling of the number of individuals with Parkinson's disease between 1990 and 2016 is projected to occur again in the coming generation.12, 13, 14, 45 To address this great health challenge will require action aimed at preventing the disease where feasible and improving the lives of those affected by the condition.46 Among the potential responses available are preventing the disease (eg, by increasing physical activity earlier in adulthood47 and reducing exposure to pesticides7), improving worldwide access to care and effective treatments (eg, levodopa), increasing funding for research (eg, to understand the underlying causes), and development of new therapies. Supplementary Material Supplementary appendix GBD 2016 Parkinson's Disease Collaborators E Ray Dorsey*, Alexis Elbaz*, Emma Nichols, Foad Abd-Allah, Ahmed Abdelalim, Jose C Adsuar, Mustafa Geleto Ansha, Carol Brayne, Jee-Young J Choi, Daniel Collado-Mateo, Nabila Dahodwala, Huyen Phuc Do, Dumessa Edessa, Matthias Endres, Seyed-Mohammad Fereshtehnejad, Kyle J Foreman, Fortune Gbetoho Gankpe, Rahul Gupta, Graeme J Hankey, Simon I Hay, Mohamed I Hegazy, Desalegn T Hibstu, Amir Kasaeian, Yousef Khader, Ibrahim Khalil, Young-Ho Khang, Yun Jin Kim, Yoshihiro Kokubo, Giancarlo Logroscino, João Massano, Norlinah Mohamed Ibrahim, Mohammed A Mohammed, Alireza Mohammadi, Maziar Moradi-Lakeh, Mohsen Naghavi, Binh Thanh Nguyen, Yirga Legesse Nirayo, Felix Akpojene Ogbo, Mayowa Ojo Owolabi, David M Pereira, Maarten J Postma, Mostafa Qorbani, Muhammad Aziz Rahman, Kedir T Roba, Hosein Safari, Saeid Safiri, Maheswar Satpathy, Monika Sawhney, Azadeh Shafieesabet, Mekonnen Sisay Shiferaw, Mari Smith, Cassandra E I Szoeke, Rafael Tabarés-Seisdedos, Nu Thi Truong, Kingsley Nnanna Ukwaja, Narayanaswamy Venketasubramanian, Santos Villafaina, Kidu Gidey Weldegwergs, Ronny Westerman, Tissa Wijeratne, Andrea S Winkler, Bach Tran Xuan, Naohiro Yonemoto, Valery L Feigin, Theo Vos, Christopher J L Murray.

iferaw, Mari Smith, Cassandra E I Szoeke, Rafael Tabarés-Seisdedos, Nu Thi Truong, Kingsley Nnanna Ukwaja, Narayanaswamy Venketasubramanian, Santos Villafaina, Kidu Gidey Weldegwergs, Ronny Westerman, Tissa Wijeratne, Andrea S Winkler, Bach Tran Xuan, Naohiro Yonemoto, Valery L Feigin, Theo Vos, Christopher J L Murray. *Contributed equally.

iferaw, Mari Smith, Cassandra E I Szoeke, Rafael Tabarés-Seisdedos, Nu Thi Truong, Kingsley Nnanna Ukwaja, Narayanaswamy Venketasubramanian, Santos Villafaina, Kidu Gidey Weldegwergs, Ronny Westerman, Tissa Wijeratne, Andrea S Winkler, Bach Tran Xuan, Naohiro Yonemoto, Valery L Feigin, Theo Vos, Christopher J L Murray. *Contributed equally. Affiliations Department of Neurology, University of Rochester, USA (E R Dorsey MD); INSERM U1018-CESP, Hôpital Paul Brousse, Villejuif cedex, France (A Elbaz MD); Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA, USA (E Nichols BA, K J Foreman PhD, Prof S I Hay DSc, I Khalil MD, Prof M Naghavi PhD, M Smith MPA, Prof T Vos PhD, Prof C J L Murray DPhil); Department of Neurology, Cairo University, Cairo, Egypt (Prof F Abd-Allah MD, Prof A Abdelalim MD, Prof M I Hegazy PhD); Faculty of Sport Science (J C Adsuar PhD, S Villafaina MSc), and Departamento de Didáctica de la Expresión Musical, Plástica y Corporal (D Collado-Mateo MSc), University of Extremadura, Spain; Public Health, Debre Berhan University, Debre Berhan, Ethiopia (M G Ansha MPH); Public Health and Primary Care, University of Cambridge, UK (Prof C Brayne MD); Biochemistry, Biomedical Science, Seoul National University Hospital, Seoul, South Korea (J-Y J Choi PhD); Facultad de Educación Universidad Autónoma de Chile, Talca, Chile (D Collado-Mateo); Department of Neurology, University of Pennsylvania, Philadelphia, PA, USA (Prof N Dahodwala MD); Institute for Global Health Innovations, Duy Tan University, Hanoi, Vietnam (H P Do PhD, B T Nguyen MPH, N T Truong BHlthSci); School of Pharmacy (D Edessa MPharm, M S Shiferaw MSc), and School of Nursing and Midwifery (K T Roba PhD), Haramaya University, Harar, Ethiopia; Department of Neurology, Charité – Universitätsmedizin Berlin, Berlin, Germany (Prof M Endres MD); Department of Neurology and Neurosurgery, Montreal Neurological Institute, McGill University, Montreal, QC, Canada (S-M Fereshtehnejad PhD); Department of Neurobiology, Care Sciences and Society, Karolinska Institutet, Huddinge, Sweden (S-M Fereshtehnejad); Department of Epidemiology and Biostatistics, Imperial College London, UK (K J Foreman); Neurosurgery Department, Faculty of Medicine and Pharmacy of Fez, Fez, Morocco (F G Gankpe MD); Non-Communicable Disease Department, Laboratoire D'etudes et de Recherche-action en Santé (leras Afrique), Porto Novo, Benin (F G Gankpe); West Virginia Bureau for Public Health, Charleston, WV, USA (Prof R Gupta MD); Health Policy,

gery Department, Faculty of Medicine and Pharmacy of Fez, Fez, Morocco (F G Gankpe MD); Non-Communicable Disease Department, Laboratoire D'etudes et de Recherche-action en Santé (leras Afrique), Porto Novo, Benin (F G Gankpe); West Virginia Bureau for Public Health, Charleston, WV, USA (Prof R Gupta MD); Health Policy, Management and Leadership, West Virginia University School of Public Health, Morgantown, WV, USA (Prof R Gupta); Medical School, University of Western Australia, Perth, WA, Australia (Prof G J Hankey MD); Department of Reproductive Health, Hawassa University, Hawassa, Ethiopia (D T Hibstu MPH); Hematology-oncology and Stem Cell Transplantation Research Center, Hematologic Malignancies Research Center, Tehran University of Medical Sciences, Tehran, Iran (A Kasaeian PhD); Public Health and Community Medicine, Jordan University of Science and Technology, Alramtha, Jordan (Prof Y Khader PhD); Department of Health Policy and Management, College of Medicine, Institute of Health Policy and Management, SNU Medical Research Center, Seoul National University, Seoul, South Korea (Prof Y-H Khang MD); School of Medicine, Xiamen University Malaysia, Sepang, Malaysia (Prof Y J Kim PhD); Department of Preventive Cardiology, National Cerebral and Cardiovascular Center, Suita, Japan (Prof Y Kokubo PhD); University of Bari Aldo Moro, Bari, Italy (Prof G Logroscino PhD); Department of Clinical Neurosciences and Mental Health, Faculty of Medicine (J Massano MD), and REQUIMTE/LAQV, Laboratório de Farmacognosia, Departamento de Química, Faculdade de Farmácia (Prof D Pereira), University of Porto, Porto, Portugal (J Massano MD); Department of Medicine, Universiti Kebangsaan Malaysia Medical Centre, Bandar Tun Razak, Malaysia (Prof N M Ibrahim MD); Department of Public Health, Jigjiga Unviersity, Jigjiga, Ethiopia (M A Mohammed PhD); University of Sydney, Sydney, NSW, Australia (M A Mohammed); Neuroscience Research Center, Baqiyatallah University of Medical Science, Tehran, Iran (Prof A Mohammadi PhD); Community and Family Medicine, Preventive Medicine and Public Health Research Center, Iran University of Medical Sciences, Tehran, Iran (M Moradi-Lakeh MD); Department of Clinical Pharmacy, Mekelle University, Mekelle, Ethiopia (Y L Nirayo MS, K G Weldegwergs MS); Western Sydney University, Penrith, NSW, Australia (F A Ogbo PhD); Department of Medicine, University of Ibadan, Ibadan, Nigeria (Prof M O Owolabi DrM); Department of Pharmacy, University Medical Center Groningen, University of

linical Pharmacy, Mekelle University, Mekelle, Ethiopia (Y L Nirayo MS, K G Weldegwergs MS); Western Sydney University, Penrith, NSW, Australia (F A Ogbo PhD); Department of Medicine, University of Ibadan, Ibadan, Nigeria (Prof M O Owolabi DrM); Department of Pharmacy, University Medical Center Groningen, University of Groningen, Groningen, Netherlands (Prof M J Postma PhD); Non-communicable Diseases Research Center, Alborz University of Medical Sciences, Karaj, Iran (M Qorbani PhD); Austin Clinical School of Nursing, La Trobe University, Heidelberg, VIC, Australia (M A Rahman PhD); School of Medicine, Deakin University, Waurn Ponds, VIC, Australia (M A Rahman); Department of Neurosurgery, Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran (Prof H Safari MD); Department of Public Health, Managerial Epidemiology Research Center, School of Nursing and Midwifery, Maragheh University of Medical Sciences, Maragheh, Iran (S Safiri PhD); Centre of Advanced Study in Psychology, Utkal University, Bhubaneswar, India (Prof M Satpathy PhD); Department of Public Health Sciences, University of North Carolina at Charlotte, Charlotte, NC, USA (M Sawhney PhD); Langone Medical Center, New York University, New York, NY, USA (A Shafieesabet MD); Department of Medicine, Dentistry and Health Science (Prof C E I Szoeke PhD) and Department of Medicine (Prof T Wijeratne MD), University of Melbourne, Melbourne, VIC, Australia (Prof C E I Szoeke PhD); Department of Medicine, University of Valencia, Valencia, Spain (Prof R Tabarés-Seisdedos PhD); Department of Internal Medicine, Federal Teaching Hospital, Abakaliki, Nigeria (K N Ukwaja MD); Raffles Neuroscience Centre, Raffles Hospital, Singapore, Singapore (Prof N Venketasubramanian FRCP); Yong Loo Lin School of Medicine, National University of Singapore, Singapore, Singapore (Prof N Venketasubramanian); Competence Center of Mortality-follow-up, German National Cohort, Federal Institute for Population Research, Wiesbaden, Germany (R Westerman DSc); Department of Psychology, La Trobe University, Melbourne, Australia (Prof T Wijeratne MD); Department of Neurology, Technical University of Munich, Munich, Germany (Prof A S Winkler PhD); Institute for Health and Society, University of Oslo, Oslo, Norway (Prof A S Winkler PhD); Department of Health Economics, Hanoi Medical University, Hanoi, Vietnam (Prof B T Xuan PhD); Department of Biostatistics, School of Public Health, Kyoto University, Kyoto, Japan (Prof N Yonemoto MPH); National Institute fo

PhD); Institute for Health and Society, University of Oslo, Oslo, Norway (Prof A S Winkler PhD); Department of Health Economics, Hanoi Medical University, Hanoi, Vietnam (Prof B T Xuan PhD); Department of Biostatistics, School of Public Health, Kyoto University, Kyoto, Japan (Prof N Yonemoto MPH); National Institute fo r Stroke and Applied Neurosciences, Auckland University of Technology, Auckland, New Zealand (V L Feigin PhD). Contributors ERD and AE prepared the first draft. EN, VLF, and TV analysed the data and edited the first draft and final versions of the manuscript. ERD and AE finalised all drafts, and approved the final version of the manuscript. All other authors provided data, developed models, reviewed results, provided guidance on methods, or reviewed the manuscript, and approved the final version of the manuscript. Declaration of interests MJP has grants or personal fees from Sigma Tau, Merck Sharp and Dohme, GlaxoSmithKline, Pfizer, Mundipharma, Boehringer Ingelheim, Novavax, Ingress Health, Quintiles, Bayer, Bristol-Myers Squibb, AbbVie, AstraZeneca, Sanofi, Astellas, Mapi, Optumlnsight, Advice, Research & Training in Health Economics Groningen, AscA, Novartis, Swedish Orphan, Innoval, Jansen, Intercept, and Pharmerit. MJP holds stocks in Ingress Health and Pharmacoeconomics Advice Groningen (PAG Ltd). CEIS reports grants from Denmark National Medical Health Research Council, Lundbeck, and Alzheimer's Association, and has a patent PCT/AU2008/001556 issued. All other authors declare no competing interests.

Introduction Migraine and other headache disorders are among the most prevalent disorders worldwide,1 but recognition of their importance for public health has come only since 2000. This delay has occurred in part because headache is not fatal and does not result in permanent or objective disability, and in part because headaches are experienced by most people from time to time, which has hindered the realisation that headache disorders are debilitating for a relatively large minority of the people who are affected. The Global Burden of Diseases, Injuries, and Risk Factors (GBD) studies have as one of their main aims the evaluation not only of mortality but also of non-fatal health outcomes. GBD offers a method of quantifying health loss in time units, enabling comparisons over time and across conditions, cultures, and countries, and, from 2016, at subnational levels in some countries. GBD has become an important tool for priority setting and planning of health services by international health organisations and governments.

d of quantifying health loss in time units, enabling comparisons over time and across conditions, cultures, and countries, and, from 2016, at subnational levels in some countries. GBD has become an important tool for priority setting and planning of health services by international health organisations and governments. Migraine was not included in GBD 1990, but was added in GBD 2000. In GBD 2010, tension-type headache was added and in GBD 2013, medication overuse headache was included. In GBD 2000, migraine was ranked as the 19th cause of disability globally.2 For the GBD 2000 study, data were absent for more than half of the world's population. When new data came from big countries like Russia, China, India, and some parts of Africa, and with tension-type headache and medication overuse headache also taken into account, headache disorders were collectively the third cause of disability in people under 50 years of age in GBD 2015.3 Since GBD 2010, prevalence and burden of disability have been re-estimated for the full time period from 1990 until the most recent year for which data are available, each time incorporating new data sources and any updates to methods. Research in context Evidence before this study

Migraine was not included in GBD 1990, but was added in GBD 2000. In GBD 2010, tension-type headache was added and in GBD 2013, medication overuse headache was included. In GBD 2000, migraine was ranked as the 19th cause of disability globally.2 For the GBD 2000 study, data were absent for more than half of the world's population. When new data came from big countries like Russia, China, India, and some parts of Africa, and with tension-type headache and medication overuse headache also taken into account, headache disorders were collectively the third cause of disability in people under 50 years of age in GBD 2015.3 Since GBD 2010, prevalence and burden of disability have been re-estimated for the full time period from 1990 until the most recent year for which data are available, each time incorporating new data sources and any updates to methods. Research in context Evidence before this study Since 2000, the Global Burden of Diseases, Injuries, and Risk Factors (GBD) studies have produced estimates of prevalence and burden of migraine. Since GBD 2010, tension-type headache and medication overuse headache have been added and estimates have been made by country spanning the period from 1990 to the most recent year for which data are available. Headache disorders, and in particular, migraine, have been found to be highly prevalent and a cause of large burden. To date, no research article has focused on the detailed methods and results of headache estimates from GBD. With the present study, we updated a previous systematic review covering 1980–2001 by doing a review that searched PubMed for articles using the terms “migraine”, “tension”, “headache”, “medication”, and “epidemiology” from Jan 1, 2001, until Dec 31, 2015. There were no language restrictions.

adache estimates from GBD. With the present study, we updated a previous systematic review covering 1980–2001 by doing a review that searched PubMed for articles using the terms “migraine”, “tension”, “headache”, “medication”, and “epidemiology” from Jan 1, 2001, until Dec 31, 2015. There were no language restrictions. Added value of this study In 2016, of all GBD causes of disease, tension-type headache was the third most prevalent, and migraine the sixth. In terms of years of life lived with disability, migraine ranked second globally, and was among the ten most disabling disorders in each of the 21 GBD regions. It was particularly burdensome among young and middle-aged women. Unlike many other diseases and injuries quantified in GBD studies, headache showed no clear relation to sociodemographic development, as measured by the Socio-demographic Index. No risk factors have yet been established in the GBD studies for headache disorders, and headache epidemiological studies are absent in many countries and regions. Implications of all the available evidence

In 2016, of all GBD causes of disease, tension-type headache was the third most prevalent, and migraine the sixth. In terms of years of life lived with disability, migraine ranked second globally, and was among the ten most disabling disorders in each of the 21 GBD regions. It was particularly burdensome among young and middle-aged women. Unlike many other diseases and injuries quantified in GBD studies, headache showed no clear relation to sociodemographic development, as measured by the Socio-demographic Index. No risk factors have yet been established in the GBD studies for headache disorders, and headache epidemiological studies are absent in many countries and regions. Implications of all the available evidence Through the GBD studies, headache disorders, and in particular, migraine, have been shown to be among the most disabling disorders worldwide. Many fatal and disabling disorders decrease with socioeconomic development, but this does not seem to be true for migraine and tension-type headache. Hence, their relative importance is likely to increase in the future. More high-quality headache epidemiological studies and studies aiming to identify modifiable risk factors should be done. Effective strategies to modify the course of headaches and alleviate pain exist, but many people affected by headache are not benefiting from this knowledge.

nce is likely to increase in the future. More high-quality headache epidemiological studies and studies aiming to identify modifiable risk factors should be done. Effective strategies to modify the course of headaches and alleviate pain exist, but many people affected by headache are not benefiting from this knowledge. Given the importance of headache disorders for global public health, which has become evident through GBD, we wanted to inform an audience of headache specialists about these studies. The aims of the present Article are to provide an overview of the GBD methods as applied to headache, to present detailed results of the update for 1990–2016 on headache burden in different world regions and with time trends, and to discuss the implications of these results both for future iterations of GBD and for health policies around the world. Methods Overview The main elements of the GBD methods, both general and pertaining to migraine and tension-type headache, are described in the appendix. In the main text of this Article, we concentrate on methods pertaining to estimation of the burden of migraine and tension-type headache. A flowchart of the different steps in these methods is shown in the appendix.

thods, both general and pertaining to migraine and tension-type headache, are described in the appendix. In the main text of this Article, we concentrate on methods pertaining to estimation of the burden of migraine and tension-type headache. A flowchart of the different steps in these methods is shown in the appendix. In the GBD cause hierarchy, migraine and tension-type headache are individual disorders on Level 3, under neurological disorders (Level 2) and non-communicable diseases (Level 1). No further subdivision exists for headaches, so each reappears at Level 4. In GBD 2013 and GBD 2015, medication overuse headache was treated as a separate disorder, but in GBD 2016 it was considered a sequela of either migraine or tension-type headache. The burden of medication overuse headache was therefore added to the burdens estimated for these headache types according to a meta-analysis of three studies reporting the proportions of medication overuse headache resulting from migraine (73·4%, 95% uncertainty interval [UI] 63·9–82·0) or tension-type headache (26·6%, 18·0–36·1).4, 5, 6

ation overuse headache was therefore added to the burdens estimated for these headache types according to a meta-analysis of three studies reporting the proportions of medication overuse headache resulting from migraine (73·4%, 95% uncertainty interval [UI] 63·9–82·0) or tension-type headache (26·6%, 18·0–36·1).4, 5, 6 In GBD, disease burden is estimated in disability-adjusted life-years (DALYs), which are the sum of years of life lost (YLLs) to premature mortality and years of life lived with disability (YLDs). Because GBD does not estimate any deaths from headache disorders as the underlying cause, DALYs for headaches are equivalent to YLDs. YLDs for each headache disorder are calculated from its prevalence and the mean time patients spend with that type of headache multiplied by the associated disability weight. The determination of headache disability weights through population and internet surveys was on the basis of lay descriptions (appendix).7, 8 The disability weight for migraine was 0·434, meaning that during an attack the affected person experiences health loss of 43·4% compared with a person in full health. The disability weight for medication overuse headache was 0·223 and for tension-type headache was 0·037. After all diseases were estimated separately, an adjustment was made to YLDs to account for comorbidity by use of simulation methods assuming a multiplicative, rather than additive, model. This adjustment led to a downward correction for YLDs for migraine in women and children by factors ranging from 2·1% (at ages 5–9 years) to 20·6% (at ages ≥95 years), reflecting a strong correlation between comorbidity and age. The corresponding figures in males were 2·1% and 20·7%, respectively.

than additive, model. This adjustment led to a downward correction for YLDs for migraine in women and children by factors ranging from 2·1% (at ages 5–9 years) to 20·6% (at ages ≥95 years), reflecting a strong correlation between comorbidity and age. The corresponding figures in males were 2·1% and 20·7%, respectively. Data sources For headaches, the data sources were mostly published population-based studies of prevalence; however, survey data for which we had the individual record data were also included. The PubMed search terms are shown in the appendix. Reference lists in published articles were reviewed and data were solicited from our GBD network of more than 2500 collaborators.

stly published population-based studies of prevalence; however, survey data for which we had the individual record data were also included. The PubMed search terms are shown in the appendix. Reference lists in published articles were reviewed and data were solicited from our GBD network of more than 2500 collaborators. Only studies diagnosing the headaches according to the International Classification of Headache Disorders (ICHD) were considered. This classification was first published in 1988, updated in 2004 (ICHD-2),9 and updated again in 2018 (ICHD-3).10 In the three ICHD versions, no major differences exist with regard to diagnostic criteria of migraine, tension-type headache, and medication overuse headache. For evaluation of headache burden, these diagnoses were selected because they are, by far, the most common and of the greatest public health importance.11 Cluster headache causes severe pain and disability, but with a lifetime prevalence of approximately 0·2% the effect on public health is much less.1 Diagnosis of the many secondary headaches (eg, due to infections or brain tumours) is difficult in epidemiological studies, and the burden of these headaches should be attributed to the underlying disorder.11 An exception is secondary headache caused by medication overuse, which occurs almost exclusively in patients with either migraine or tension-type headache.

g, due to infections or brain tumours) is difficult in epidemiological studies, and the burden of these headaches should be attributed to the underlying disorder.11 An exception is secondary headache caused by medication overuse, which occurs almost exclusively in patients with either migraine or tension-type headache. For GBD 2016, data on migraine were extracted from 135 studies, covering 16 of the 21 GBD world regions. For tension-type headache, data were extracted from 76 studies covering 16 GBD regions. All data sources are available online. For medication overuse headache, data were extracted from 37 studies from ten regions (for details about the numbers of studies from each region see the appendix). In addition, hospital claims data from the USA on migraine and tension-type headache for 3 years (2000, 2010, and 2012) were used.12, 13, 14 Sets of claims data were included because, owing to their size, they provide more detailed information on age patterns and time trends than published epidemiological studies, and also provide estimates on subnational locations. Because claims data might not be nationally representative, do not refer to the ICHD criteria, and capture only those individuals who are able to seek health care, they are evaluated for systematic biases compared with data sources of high quality (appendix).

tudies, and also provide estimates on subnational locations. Because claims data might not be nationally representative, do not refer to the ICHD criteria, and capture only those individuals who are able to seek health care, they are evaluated for systematic biases compared with data sources of high quality (appendix). Approximately half of the 135 studies from which data were extracted were from three of the 21 GBD regions (western Europe, high-income North America, and high-income Asia Pacific), and no data on any of the headaches were available from the Caribbean, central sub-Saharan Africa, southern sub-Saharan Africa, or Oceania (see appendix) regions. In all regions and countries, prevalence was estimated with a Bayesian meta-regression model (DisMod-MR 2.1), and estimates were obtained in this way also for countries and regions where no relevant headache studies had been done. Calculation of proportion of time in the symptomatic state Headache disorders are modelled as chronic episodic conditions. The prevalence reflects the individuals in the population who have had at least one episode in the past 12 months fulfilling ICHD criteria. To calculate the average proportion of time with headache (ie, in the symptomatic state) necessary for YLD calculation, 13 population-based studies were identified that had data on frequency and duration of migraine attacks (appendix). From these studies, we estimated the average number of hours migraineurs spend in attacks, and expressed this as a proportion of a year, which was 8·5% (95% UI 5·8–11·2).

e) necessary for YLD calculation, 13 population-based studies were identified that had data on frequency and duration of migraine attacks (appendix). From these studies, we estimated the average number of hours migraineurs spend in attacks, and expressed this as a proportion of a year, which was 8·5% (95% UI 5·8–11·2). For tension-type headache, seven studies on duration and frequency of attacks showed that affected people spend, on average, 4·7% (95% UI 1·3–8·0) of their time with headache. For both migraine and tension-type headache, frequency and duration were reported most commonly in categories, and the midpoint was assumed to represent each category. For medication overuse headache, only one study included in GBD 2016 (from Russia) gave data on time in symptomatic state, reporting a mean headache frequency of 23·1 (SD 6·7) days per month.15 According to the ICHD-3 definition, medication overuse headache is present on more than 15 days per month for more than 3 months.10

edication overuse headache, only one study included in GBD 2016 (from Russia) gave data on time in symptomatic state, reporting a mean headache frequency of 23·1 (SD 6·7) days per month.15 According to the ICHD-3 definition, medication overuse headache is present on more than 15 days per month for more than 3 months.10 Modelling of prevalence In the mathematical modelling, the mortality due to headache was set at zero,16 as was occurrence below age 5 years.17 In the sources used for this study, prevalence rates vary, but the degree to which this reflects real variation across borders and time, or methodological differences, is mostly not known. Method most probably plays a large role because results can be substantially influenced by relatively minor differences, such as variations in the screening question.18 To adjust for differences in methodological quality, all prevalence studies included in GBD are scored according to a modified version (dichotomised variables) of published methodological quality criteria for headache epidemiological studies,11 taking into account the representativeness of the population of interest (representative of country or community vs selected population), quality of sampling (random sample of the population of interest vs not random sample), recall period (1-year prevalence vs other recall period), participation rate (≥70% vs <70%), survey method (face to face with headache expert or trained interviewer vs other), validation of diagnostic instrument (sensitivity or specificity ≥70% vs <70% or no validation), and application of ICHD criteria (strict criteria or reasonable modification of criteria vs other modification of criteria). In DisMod-MR 2.1, these methodological variables were evaluated for a systematic difference and corrected accordingly (appendix).

ument (sensitivity or specificity ≥70% vs <70% or no validation), and application of ICHD criteria (strict criteria or reasonable modification of criteria vs other modification of criteria). In DisMod-MR 2.1, these methodological variables were evaluated for a systematic difference and corrected accordingly (appendix). Socio-demographic Index (SDI) We examined the relationships between migraine and tension-type headache DALYs and SDI, a composite measure of income per capita, education, and fertility.19 We also present results by groupings of countries into quintiles (high SDI, middle-high SDI, middle SDI, middle-low SDI, and low SDI) based on their 2016 SDI value. Additional details on the SDI methods can be found online. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation or the writing of the report. All authors had full access to the data in the study and had final responsibility for the decision to submit for publication.