Browse the corpus

Introduction Diabetes prevalence and diabetes-related deaths are rising in most parts of the world, at least partly fuelled by the worldwide increase in excess weight and adiposity.1, 2, 3, 4, 5 This trend has created concerns about the health and functional consequences for patients, and costs for health systems.6, 7, 8 Tracking the epidemic and the progress of programmes aimed at reducing diabetes and its complications requires consistent and comparable measurement of the prevalence of diabetes and the coverage of drug and lifestyle interventions that slow diabetes progression and decrease the risk of complications.

h systems.6, 7, 8 Tracking the epidemic and the progress of programmes aimed at reducing diabetes and its complications requires consistent and comparable measurement of the prevalence of diabetes and the coverage of drug and lifestyle interventions that slow diabetes progression and decrease the risk of complications. Different biomarkers have been used to define diabetes, including fasting plasma glucose (FPG), 2-h plasma glucose in an oral glucose tolerance test (2hOGTT), and, more recently, HbA1c.9, 10, 11, 12, 13, 14, 15 Population-based health surveys in different countries and at different times have also used different biomarkers for glycaemia and diabetes, and thus define diabetes differently. The variety of biomarkers and definitions creates a challenge in consistently analysing diabetes prevalence across countries and over time, and in measuring what proportion of people with diabetes are diagnosed and receive effective treatments for diabetes and its complications.1, 16, 17 Therefore, there is a need to understand how the use of different biomarkers and definitions affects the identification of diabetes cases and the resulting estimates of population prevalence. This need is particularly pressing because two of the nine global targets for non-communicable diseases set after the 2011 United Nations high-level meeting on non-communicable diseases require estimates of diabetes prevalence: to halt the rise in the prevalence of diabetes, and to achieve a 50% coverage of drug treatment and counselling, including glycaemic control, to prevent coronary heart disease and stroke in people at high risk of cardiovascular disease.4, 18 Diabetes is also one of the four main non-communicable diseases for which there is a global target of 25% reduction in premature mortality by 2025 compared with 2010.4, 18

nt and counselling, including glycaemic control, to prevent coronary heart disease and stroke in people at high risk of cardiovascular disease.4, 18 Diabetes is also one of the four main non-communicable diseases for which there is a global target of 25% reduction in premature mortality by 2025 compared with 2010.4, 18 Research in context Evidence before this study

nt and counselling, including glycaemic control, to prevent coronary heart disease and stroke in people at high risk of cardiovascular disease.4, 18 Diabetes is also one of the four main non-communicable diseases for which there is a global target of 25% reduction in premature mortality by 2025 compared with 2010.4, 18 Research in context Evidence before this study We reviewed studies included in the NCD Risk Factor Collaboration databases for comparisons of various diabetes definitions. We also searched PubMed with the term ((A1c[Title/Abstract]) AND Sensitivity[Title/Abstract]) AND Specificity[Title/Abstract]) on April 13, 2015. We also searched the references of recent reviews and guidelines. We found some studies on the classification of individuals as having diabetes or on comparison of prevalence estimates based on different definitions in specific cohorts, especially for HbA1c compared with either fasting plasma glucose (FPG) or 2-h oral glucose tolerance test (2hOGTT). Most of these analyses were based on a single cohort and very few covered different world regions. Two pooled analyses of Asian and European cohorts, and a study in the Pacific and Indian Ocean islands, assessed how the prevalence of diabetes and the classification of individuals as having diabetes versus not having diabetes changed depending whether diabetes was based on FPG or on 2hOGTT. There is no pooling study for HbA1c and we identified only one review of data from six countries. Other studies compared different diabetes definitions among people with specific pre-existing diseases—eg, heart disease and tuberculosis. We also found some prospective studies that assessed how HbA1c predicts future incidence of diabetes or cardiovascular diseases with mixed results.

eview of data from six countries. Other studies compared different diabetes definitions among people with specific pre-existing diseases—eg, heart disease and tuberculosis. We also found some prospective studies that assessed how HbA1c predicts future incidence of diabetes or cardiovascular diseases with mixed results. Added value of this study This study is the first pooling of a large number of population-based data from different world regions that addresses how different definitions of diabetes affect both the total prevalence, and the identification of previously undiagnosed individuals. By pooling a large number of data sources, the overall meta-analytical finding overcomes between-study variation, which can be probed in meta-regressions. Furthermore, by having a large number of studies, and age–sex groups within each study, we were able to develop regressions to convert across different diabetes definitions, which is essential for enhancing comparability over time and across countries in surveillance. Implications of all the available evidence

This study is the first pooling of a large number of population-based data from different world regions that addresses how different definitions of diabetes affect both the total prevalence, and the identification of previously undiagnosed individuals. By pooling a large number of data sources, the overall meta-analytical finding overcomes between-study variation, which can be probed in meta-regressions. Furthermore, by having a large number of studies, and age–sex groups within each study, we were able to develop regressions to convert across different diabetes definitions, which is essential for enhancing comparability over time and across countries in surveillance. Implications of all the available evidence The use of HbA1c in surveillance requires further consideration in terms of how it predicts, and helps prevent, diabetes complications and sequelae. As such studies are done, and to maximise comparability of results across surveys, the best approach in population-based health surveys is to measure FPG and define diabetes as FPG 7·0 mmol/L or more or history of diagnosis with diabetes or using insulin or oral hypoglycaemic drugs, as used in the global monitoring framework for prevention and control of non-communicable diseases. When HbA1c is used, it would be valuable to also measure FPG in a subsample of participants to provide information about how the two tests relate. The conversion regressions developed here can be used to convert prevalence based on FPG to that based on FPG-or-2hOGTT.

prevention and control of non-communicable diseases. When HbA1c is used, it would be valuable to also measure FPG in a subsample of participants to provide information about how the two tests relate. The conversion regressions developed here can be used to convert prevalence based on FPG to that based on FPG-or-2hOGTT. Some studies have analysed the classification of individuals as having diabetes or compared prevalence estimates based on different definitions in specific cohorts, especially for HbA1c compared with either FPG or 2hOGTT.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, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61 Most of these analyses were based on a single cohort and very few covered different geographical regions. Two pooled analyses of Asian and European cohorts, and a study in the Pacific and Indian Ocean islands, assessed how the prevalence of diabetes and the classification of individuals as having diabetes versus not having diabetes changed depending on whether diabetes was defined by FPG or 2hOGTT.62, 63, 64, 65, 66 There is no pooling study for HbA1c, which can be measured easily in population-based surveys without the need for overnight fasting and has been approved by the American Diabetes Association and WHO as a diagnostic test for diabetes.11, 14 However, a review of data from six countries reported that the sensitivity of diabetes diagnosis based on HbA1c compared with FPG ranged from 17% to 78%,67 raising concerns about ethnic variation of HbA1c-based definition.17

approved by the American Diabetes Association and WHO as a diagnostic test for diabetes.11, 14 However, a review of data from six countries reported that the sensitivity of diabetes diagnosis based on HbA1c compared with FPG ranged from 17% to 78%,67 raising concerns about ethnic variation of HbA1c-based definition.17 We assessed the effect of diagnostic definitions both on the identification of diabetes in previously undiagnosed individuals and on the population prevalence estimates for diabetes in a pooled analysis of data from population-based health examination surveys in different world regions. Methods Study design We aimed to answer two questions. First, how does the estimated prevalence of diabetes in a population change when the new definition of diabetes based on HbA1c is used compared with earlier definitions based on blood glucose? Second, how does the new definition of diabetes based on HbA1c compare with earlier definitions in identifying previously undiagnosed people with diabetes, as measured by the sensitivity and specificity of the new definition with respect to the previous ones? We further assessed whether sensitivity varied by the characteristics of the study population, because this possible variation is a source of concern about the generalisability of HbA1c as a diagnostic and surveillance measure.17, 67, 68, 69, 70

y and specificity of the new definition with respect to the previous ones? We further assessed whether sensitivity varied by the characteristics of the study population, because this possible variation is a source of concern about the generalisability of HbA1c as a diagnostic and surveillance measure.17, 67, 68, 69, 70 For the HbA1c-based definition of diabetes, we used HbA1c of 6·5% or more, or history of diagnosis with diabetes or using insulin or oral hypoglycaemic drugs.11 For definitions based on blood glucose, we used either the American Diabetes Association definition of FPG of 7·0 mmol/L or more, or history of diagnosis with diabetes or using insulin or oral hypoglycaemic drugs (which is also used in the global monitoring framework for prevention and control of non-communicable diseases),12, 18 or the WHO definition of FPG of 7·0 mmol/L or more, or 2hOGTT of 11·1 mmol/L or more, or history of diabetes or using insulin or oral hypoglycaemic drugs.9, 10

using insulin or oral hypoglycaemic drugs (which is also used in the global monitoring framework for prevention and control of non-communicable diseases),12, 18 or the WHO definition of FPG of 7·0 mmol/L or more, or 2hOGTT of 11·1 mmol/L or more, or history of diabetes or using insulin or oral hypoglycaemic drugs.9, 10 Data sources We used population-based data collated by the NCD Risk Factor Collaboration (NCD-RisC), a worldwide network of health researchers and practitioners who, together with WHO, have collated a large database of population-based health examination surveys and epidemiological studies of cardiometabolic risk factors. All data sources were checked by at least two independent reviewers as being representative of a national, subnational, or community population, and for study quality indicators such as fasting duration and the protocol for OGTT. We excluded surveys that had not used a standard glucose load for OGTT. Within each survey, we included participants aged 18 years and older who were not pregnant and had fasted at least for 6 h before measurement as a part of the survey instructions. We excluded HbA1c data from before the year 2000 to minimise the use of non-standard assays.71 We also excluded surveys that had measured a biomarker only among participants with a high value of another—eg, studies in which FPG was only measured in participants with HbA1c above a prespecified value, because the relation between the two measurements might be different in this prescreened group compared with the whole sample. The appendix shows details of individual surveys.

ong participants with a high value of another—eg, studies in which FPG was only measured in participants with HbA1c above a prespecified value, because the relation between the two measurements might be different in this prescreened group compared with the whole sample. The appendix shows details of individual surveys. We restricted the analysis of sensitivity and specificity to people without a history of diabetes diagnosis, because previous diagnosis and the use of drug treatments probably affect the concentrations of biomarkers used to diagnose diabetes. History of diabetes diagnosis was established with survey-specific questions, such as “have you ever been told by a doctor or other health professional that you have diabetes?” or the combination of “do you now have, or have you ever had diabetes?” and “were you told by a doctor that you had diabetes?”. We also excluded follow-up surveys of closed cohorts from the analysis of sensitivity and specificity because active surveillance within a cohort shifts participants from undiagnosed to diagnosed status at each follow-up, thus affecting the composition of undiagnosed cases.

you told by a doctor that you had diabetes?”. We also excluded follow-up surveys of closed cohorts from the analysis of sensitivity and specificity because active surveillance within a cohort shifts participants from undiagnosed to diagnosed status at each follow-up, thus affecting the composition of undiagnosed cases. Statistical analysis We calculated diabetes prevalence by sex and age group, taking into account complex survey design and survey sample weights when relevant. We excluded age–sex groups with fewer than 25 participants when calculating prevalence because the sampling error of estimated prevalence can bias the associations between prevalences based on different definitions. Some surveys had measured HbA1c or FPG in all participants, but had not measured 2hOGTT among people with diagnosed diabetes. These previously diagnosed participants were included in calculation of diabetes prevalence because their exclusion would underestimate diabetes prevalence. Furthermore, some surveys measured 2hOGTT in only a subset of people without history of diabetes diagnosis, generally for logistical or cost reasons. Simply combining these participants with previously diagnosed participants might overestimate diabetes prevalence based on 2hOGTT. To account for these missing measurements, and to avoid overestimation of diabetes prevalence, we recalculated the survey sample weights for these participants as the original sample weights divided by weighted proportion of non-diabetic participants with data. This approach is similar to that used in the US National Health and Nutrition Examination Survey for their 2hOGTT sample weights.72 A similar approach was taken in a few surveys that had measured HbA1c in all participants, but had not measured FPG among people diagnosed with diabetes.

n-diabetic participants with data. This approach is similar to that used in the US National Health and Nutrition Examination Survey for their 2hOGTT sample weights.72 A similar approach was taken in a few surveys that had measured HbA1c in all participants, but had not measured FPG among people diagnosed with diabetes. We compared graphically the prevalences of diabetes using different definitions. We also did regression analyses of the relation between diabetes defined (1) on the basis of FPG-or-2hOGTT versus on the basis of FPG only and (2) on the basis of HbA1c versus on the basis of FPG. We did not do a regression for diabetes prevalence based on HbA1c versus prevalence based on FPG-or-2hOGTT because very few surveys had data for both 2hOGTT and HbA1c, leading to unstable regression coefficients. We probit-transformed diabetes prevalence because it provided better fit to the data and it avoids predicting prevalences that are less than 0 or greater than 1. We considered regression models with alternative covariates and specifications, and chose the best model using the Bayesian information criterion, which measures the relative goodness of fit of a model; it rewards how well the model fits the data but discourages overfitting.73 The regressions included age (mean age of each age–sex group); the years over which each survey collected data (as the midyear of the period of data collection; appendix); national income (natural logarithm of per person gross domestic product) in the survey country and year; whether the study was representative of a national, subnational, or community population; and mean BMI for each age–sex group. Sex was excluded from the regressions on the basis of the Bayesian information criterion. The regression of diabetes prevalence based on HbA1c against diabetes prevalence based on FPG, for which there were more data, also included terms for geographical region as random effects on the basis of Bayesian information criterion; these random effects account for differences in the relationship by region. Two regions consisted of high-income countries, as in previous global analyses5, 74—high-income Asia Pacific (consisting of Japan, Singapore, and South Korea) and high-income western countries (consisting of countries in Australasia, North America, and western Europe).

unt for differences in the relationship by region. Two regions consisted of high-income countries, as in previous global analyses5, 74—high-income Asia Pacific (consisting of Japan, Singapore, and South Korea) and high-income western countries (consisting of countries in Australasia, North America, and western Europe). The other countries were divided based on their geography into central and eastern Europe; central Asia, Middle East and north Africa; east and southeast Asia; south Asia; Latin America and the Caribbean; and sub-Saharan Africa. We plotted the residuals of the regression models against the main independent variable (probit-transformed FPG-based prevalence), and found no evidence of heteroscedasticity in the residuals. We also report the univariate and semipartial R2 for each of the variables in the regression model. Univariate R2 measures how much of the variance is explained by each independent variable. Semipartial R2 measures the contribution of each variable to the total explained variance, conditional on the presence of the other model variables.75

ivariate and semipartial R2 for each of the variables in the regression model. Univariate R2 measures how much of the variance is explained by each independent variable. Semipartial R2 measures the contribution of each variable to the total explained variance, conditional on the presence of the other model variables.75 We calculated sensitivity and specificity of diagnosis separately in each survey, and then pooled the sensitivities and specificities across surveys with a random-effects model.76 We examined the sources of heterogeneity in sensitivity and specificity with metaregressions and a-priori selected study characteristics: mean age, proportion of male participants, midyear of study data collection period; sample size; prevalence of undiagnosed diabetes in the survey; whether the survey was representative of a national, subnational, or community population; geographical region; national income in the survey country and year; and mean haemoglobin concentration in the survey country and year. We did the analyses with Stata (version 12.2) and R (version 3.0.3). Role of the funding source The funders had no role in study design, data collection, analysis, or interpretation, or writing of the report. SF, YL, and BZ had full access to all the data. ME was responsible for submitting the Article for publication.

We calculated sensitivity and specificity of diagnosis separately in each survey, and then pooled the sensitivities and specificities across surveys with a random-effects model.76 We examined the sources of heterogeneity in sensitivity and specificity with metaregressions and a-priori selected study characteristics: mean age, proportion of male participants, midyear of study data collection period; sample size; prevalence of undiagnosed diabetes in the survey; whether the survey was representative of a national, subnational, or community population; geographical region; national income in the survey country and year; and mean haemoglobin concentration in the survey country and year. We did the analyses with Stata (version 12.2) and R (version 3.0.3). Role of the funding source The funders had no role in study design, data collection, analysis, or interpretation, or writing of the report. SF, YL, and BZ had full access to all the data. ME was responsible for submitting the Article for publication. Results After exclusions, we included 96 population-based health examination surveys of 331 288 participants (figure 1). 46 surveys were from Australia, USA, and western Europe; 18 from east and southeast Asia; ten from Latin America and the Caribbean; seven from Oceania; six from sub-Saharan Africa; five from south Asia; three from the Middle East and north Africa; and one from central and eastern Europe. All 96 studies measured FPG; 47 also measured 2hOGTT and 63 measured HbA1c (appendix). 14 of these studies measured all three biomarkers. All but three studies of the 47 studies used for comparing prevalence based on FPG alone versus based on FPG-or-2hOGTT measured FPG in a laboratory; two of the remaining studies used a portable unit, and we did not have information for the remaining study. All studies measured 2hOGTT in a laboratory. All but one of the 63 studies used for comparing glucose-based and HbA1c-based prevalences measured glucose in a laboratory; the remaining study measured FPG with a portable unit. An enzymatic method was used to measure FPG in 65 of the 92 studies that had measured FPG in a laboratory, but we had no information for the remaining 27 studies. In all 63 studies, HbA1c was measured in a laboratory; in 40 of these studies, the measurements were done by chromatography or immunoassay. No information was available for the remaining 23. Such a dominance of laboratory-based measurements prevented us from assessing the role of measurement method as a source of variation because laboratory-based methods are equally acceptable, especially for glucose.77

urements were done by chromatography or immunoassay. No information was available for the remaining 23. Such a dominance of laboratory-based measurements prevented us from assessing the role of measurement method as a source of variation because laboratory-based methods are equally acceptable, especially for glucose.77 Diabetes prevalence ranged from 0% in people younger than 40 years of age in some surveys to about 70% in middle-aged and older adults in Nauru (figure 2). Prevalence of diabetes based on FPG alone was lower than that based on FPG-or-2hOGTT, by 2–6 percentage points at different prevalence levels, although prevalences estimated using these two glucose-based measures were highly correlated (r=0·98; figure 2). Table 1, Table 2 show results of the regression analyses. After accounting for prevalence based on FPG, prevalence based on FPG-or-2hOGTT increased with age—ie, prevalence based on FPG-or-2hOGTT rose more sharply with age than did prevalence based on FPG only.65, 80, 81

asures were highly correlated (r=0·98; figure 2). Table 1, Table 2 show results of the regression analyses. After accounting for prevalence based on FPG, prevalence based on FPG-or-2hOGTT increased with age—ie, prevalence based on FPG-or-2hOGTT rose more sharply with age than did prevalence based on FPG only.65, 80, 81 HbA1c-based prevalences were lower than those based on FPG for 42·8% of age–sex–survey groups and higher in another 41·6%; in the other 15·6%, the two definitions gave similar prevalences (figure 3). In the regression analysis, prevalence based on HbA1c was on average slightly lower than prevalence based on FPG (table 2). The most important determinant of variation between these two prevalences was age, with some effect from national income, mean BMI, year of survey, and whether the survey was representative of a national, subnational, or community population. After accounting for prevalence based on FPG, prevalence based on HbA1c increased with age, national income, mean BMI, and the year of survey. After accounting for prevalence based on FPG, HbA1c-based prevalence was higher in south Asia than in other regions, and was lower in high-income regions than in other regions (appendix).

unting for prevalence based on FPG, prevalence based on HbA1c increased with age, national income, mean BMI, and the year of survey. After accounting for prevalence based on FPG, HbA1c-based prevalence was higher in south Asia than in other regions, and was lower in high-income regions than in other regions (appendix). Diabetes defined as HbA1c of 6·5% or more had a pooled sensitivity of 52·8% (95% CI 51·3–54·3) compared with a definition of FPG of 7·0 mmol/L or more for diagnosing participants without a previous diagnosis of diabetes. This finding suggests that 47·2% of participants without a previous diagnosis of diabetes who would be considered to have diabetes based on their FPG concentration would not be considered to have diabetes with an HbA1c test (table 3). The sensitivity of HbA1c varied substantially across studies (I2 of 97·6%), ranging from 13·0% to 93·2% (appendix pp 11–12). HbA1c had even lower sensitivity when compared with defining diabetes based on FPG-or-2hOGTT (30·5%, 95% CI 28·7–32·3). None of the preselected study-level characteristics explained the heterogeneity in the sensitivity of HbA1c versus FPG (all p values >0·06; table 4). Pooled specificity of HbA1c was 99·74% (95% CI 99·71–99·78) relative to FPG and 99·69% (99·63–99·76) relative to FPG-or-2hOGTT, suggesting few false positives compared with glucose-based definitions.

-level characteristics explained the heterogeneity in the sensitivity of HbA1c versus FPG (all p values >0·06; table 4). Pooled specificity of HbA1c was 99·74% (95% CI 99·71–99·78) relative to FPG and 99·69% (99·63–99·76) relative to FPG-or-2hOGTT, suggesting few false positives compared with glucose-based definitions. Lowering the threshold for diabetes by HbA1c from 6·5% to 6·3% (a cutoff suggested by some studies49, 50) increased sensitivity compared with the FPG-based definition from 52·8% to 64·3% while maintaining a high specificity at 99·53%. Lowering it further to 6·1% increased sensitivity to 72·8% but the specificity would drop to 99·08%, resulting in more false positives. Follow-up studies are needed to establish how these cutoffs predict complications and sequelae in newly diagnosed patients.83, 84

64·3% while maintaining a high specificity at 99·53%. Lowering it further to 6·1% increased sensitivity to 72·8% but the specificity would drop to 99·08%, resulting in more false positives. Follow-up studies are needed to establish how these cutoffs predict complications and sequelae in newly diagnosed patients.83, 84 Discussion In this large international pooled analysis of population-based health examination surveys, we found that the use of different biomarkers and definitions for diabetes can lead to different estimates of population prevalence of diabetes, with the highest prevalence when diabetes is defined on the basis of FPG-or-2hOGTT and the lowest when based on HbA1c alone. For example, at an FPG-based prevalence of 10%, similar to the age-standardised global prevalence of diabetes in adults aged 25 years and older in 2008,1 prevalence based on FPG-or-2hOGTT would be about 13% according to the relation in figure 2. The variation across studies in the relation between glucose-based and HbA1c-based prevalences was partly related to age, followed by national income, mean BMI, the year of survey, and whether the survey population was national, subnational, or from specific communities. The reasons for additional regional effects—higher HbA1c-based prevalence in south Asia and lower prevalence in high-income regions than in other regions after accounting for prevalence based on FPG—are unknown, but they might be a result of true physiological differences; for example, related to red blood cell turnover (itself related to anaemia and iron status), which affects HbA1c, or related to glucose dysregulation during fasting and non-fasting which are captured by HbA1c.85 Establishing these reasons requires multicentre studies with consistent methods and protocols and data for phenotypical factors that might affect the relation between glucose and HbA1c. For now, they are unexplained empirical results that should be taken into account when using surveys from different regions.

HbA1c.85 Establishing these reasons requires multicentre studies with consistent methods and protocols and data for phenotypical factors that might affect the relation between glucose and HbA1c. For now, they are unexplained empirical results that should be taken into account when using surveys from different regions. Similarly, different definitions identified different people without a previous diagnosis as having diabetes. Specifically, use of an HbA1c-based definition would not identify almost half of the undiagnosed cases that could be detected with a FPG test, and more than three-quarters of undiagnosed cases that would be detected by FPG and 2hOGTT combined, but it would lead to few false positives compared with glucose-based definitions. Inversely, using a glucose-based test alone would not identify some people who would be considered as having diabetes with HbA1c.

G test, and more than three-quarters of undiagnosed cases that would be detected by FPG and 2hOGTT combined, but it would lead to few false positives compared with glucose-based definitions. Inversely, using a glucose-based test alone would not identify some people who would be considered as having diabetes with HbA1c. Our results, based on a large number of surveys from different regions, are consistent with previous smaller studies that compared different diabetes definitions. Diabetes prevalence based on FPG-or-2hOGTT was higher than prevalence based on FPG alone by 18% in an analysis of 13 European cohorts and by 6% in an analysis of 11 Asian cohorts.63, 64 A previous comparison of diabetes prevalence across six studies, including two analysed here, reported that diagnostic sensitivity for HbA1c compared with 2hOGTT ranged from 17% to 78%,67 which is consistent with the results of our analysis. However, this study also found surprisingly low specificities for HbA1c compared with ours.67 Other single-cohort studies also generally reported low but variable sensitivities and high specificities for HbA1c relative to blood glucose. Several studies86, 87, 88, 89 assessed the optimal cutoff for HbA1c in different populations and all reported values lower than 6·5%, which is consistent with our finding that lowering the threshold would increase sensitivity while preserving high specificity. One small study90 examined the effect of anaemia on diagnostic accuracy of HbA1c and reported higher sensitivity (than with FPG) in patients with anaemia, which is consistent with our results.

which is consistent with our finding that lowering the threshold would increase sensitivity while preserving high specificity. One small study90 examined the effect of anaemia on diagnostic accuracy of HbA1c and reported higher sensitivity (than with FPG) in patients with anaemia, which is consistent with our results. Our analysis, which focused on questions that are relevant for population-based surveillance of diabetes and monitoring treatment coverage, has several strengths. We pooled data from a large number of population-based surveys from different world regions, thereby increasing both the precision of our estimates and their generalisability compared with analyses of one or a small number of cohorts. We used consistent eligibility and inclusion criteria, and assessed whether the surveys met these criteria. In particular, we only used surveys that had rigorous protocols for fasting duration and for OGTT. Furthermore, most surveys measured glucose and HbA1c in a laboratory. We also assessed the sources of heterogeneity in how diagnostic criteria compare across surveys, which could not be done in previous analyses because they included few surveys.

ed surveys that had rigorous protocols for fasting duration and for OGTT. Furthermore, most surveys measured glucose and HbA1c in a laboratory. We also assessed the sources of heterogeneity in how diagnostic criteria compare across surveys, which could not be done in previous analyses because they included few surveys. Our results should be interpreted with some limitations in mind. We had few studies from some regions including sub-Saharan Africa, south Asia, the Middle East and north Africa, and central and eastern Europe. We analysed the surveys with consistent methods but surveys might have differed in details such as the exact limit for fasting duration beyond the 6-h limit imposed by us. Because HbA1c measurement has changed over time,91, 92, 93, 94, 95, 96, 97, 98, 99 and to minimise the use of non-standard assays, we did not include any HbA1c data from before the year 2000.71 Despite this exclusion, and the fact that all of our surveys had measured HbA1c in a laboratory, HbA1c measurements can vary between laboratories and instruments,100 about which we did not have complete data. For the same reason, we could not standardise the HbA1c data to account for different assays and instruments used in measurement. Nutritional status—especially iron deficiency—anaemia, malaria and other parasitic diseases, living at high altitudes, and high prevalence of haemoglobinopathies can affect HbA1c,101 but could not be assessed as a source of heterogeneity beyond their effects through mean haemoglobin concentration. Similarly, data for glucose can be affected by unrecorded factors such as inaccurate information about fasting, fluctuations in diet and physical activity in days before measurement, and how samples were handled, including time between drawing blood and laboratory analysis and the type of tube used for collecting and storing blood.

ta for glucose can be affected by unrecorded factors such as inaccurate information about fasting, fluctuations in diet and physical activity in days before measurement, and how samples were handled, including time between drawing blood and laboratory analysis and the type of tube used for collecting and storing blood. Although we assessed the role of geographical region, we did not have data for the ethnic composition of participants in each survey. By their nature, health examination surveys used for population-based surveillance use a single measurement for each participant, whereas diagnosis in a clinical setting might repeat the measurements based on the first test. The use of a single test is affected by within-individual and even within-laboratory variation, and could lead to misclassification of some individuals.99 Finally, we did not have longitudinal follow-up data to assess sensitivity and specificity for diagnosis using one definition (or one cutoff value of HbA1c) compared to another or for development of diabetes complications and sequelae that contribute the bulk of the public health burden of diabetes. Such data are not available in population-based surveys because surveys are typically cross-sectional.

ificity for diagnosis using one definition (or one cutoff value of HbA1c) compared to another or for development of diabetes complications and sequelae that contribute the bulk of the public health burden of diabetes. Such data are not available in population-based surveys because surveys are typically cross-sectional. There is no gold standard definition that captures the phenotypic complexity of diabetes and the risk of its microvascular and macrovascular complications, although 2hOGTT is often treated as the most reliable test.15, 102, 103 In clinical practice, physicians follow an analytical process to diagnose diabetes, in which different sequences of glucose biomarkers are used depending on factors such as a patient's age and symptoms; those with high levels of one biomarker (eg, HbA1c) might be asked to have additional measurements of the same or a different biomarker, and be monitored over time to decide on the best course of treatment. The process might vary from patient to patient to account for their unique characteristics, and might further vary from physician to physician based on available infrastructure and medical resources. In surveillance using population-based surveys, which provides evidence for policies and programmes related to whole populations, repeated measurements are virtually impossible. Therefore, considerations about diabetes definition and diagnosis are different from those of clinical practice, and the emphasis is on comparability of definitions over time and across populations. Our results provide much needed empirical evidence for planning global surveillance of diabetes and coverage of its interventions. Specifically, despite its relative ease of use, using HbA1c alone in health surveys might miss some previously undiagnosed people who would be considered as having diabetes using a glucose-based test, and thus could benefit from lifestyle and treatment interventions. Even so, 2hOGTT is difficult to measure even in a clinical setting, let alone in population-based surveys. Of 493 worldwide population-based diabetes data sources between 1975 and 2014 in the NCD-RisC databases, 448 had measured FPG but only 59 had measured 2hOGTT; 33% of surveys before 1990 had 2hOGTT and only 11% did after 1990. Therefore, a strategy for consistent and comparable surveillance is to use FPG in population-based surveys, be it national or multicountry survey programmes such as the WHO STEPS surveys, and define diabetes based on FPG.

ut only 59 had measured 2hOGTT; 33% of surveys before 1990 had 2hOGTT and only 11% did after 1990. Therefore, a strategy for consistent and comparable surveillance is to use FPG in population-based surveys, be it national or multicountry survey programmes such as the WHO STEPS surveys, and define diabetes based on FPG. Data such as those in figure 2 and table 1 can then be used to relate prevalences based on FPG to those based on FPG-or-2hOGTT. The use of HbA1c in surveillance requires further consideration of how it predicts and helps prevent diabetes complications and sequelae. When HbA1c is used, FPG should ideally also be measured in a subsample of participants to provide information about how the two tests relate. Correspondence to: Prof Majid Ezzati, Imperial College London, London W2 1PG, UK majid.ezzati@imperial.ac.uk Supplementary Material Supplementary appendix Supplementary audio How do different methods for diagnosing diabetes affect estimates of prevalence and how is this likely to affect global estimates of this disease? Acknowledgments This study was funded by the Wellcome Trust and US National Institutes of Health (DK090435). The authors alone are responsible for the views expressed in this Article and they do not necessarily represent the views, decisions, or policies of the institutions with which they are affiliated.

Supplementary audio How do different methods for diagnosing diabetes affect estimates of prevalence and how is this likely to affect global estimates of this disease? Acknowledgments This study was funded by the Wellcome Trust and US National Institutes of Health (DK090435). The authors alone are responsible for the views expressed in this Article and they do not necessarily represent the views, decisions, or policies of the institutions with which they are affiliated. Contributors GD and ME designed the study and oversaw research. Members of the Country and Regional Data Group collected and reanalysed data, and checked pooled data for accuracy of information about their study and other studies in their country. Members of the Pooled Analysis and Writing Group collated data, checked all data sources in consultation with the Country and Regional Data Group, analysed pooled data, and prepared results. GD and ME wrote the first draft of the report with input from other members of Pooled Analysis and Writing Group. Members of Country and Regional Data Group commented on draft report.

ted data, checked all data sources in consultation with the Country and Regional Data Group, analysed pooled data, and prepared results. GD and ME wrote the first draft of the report with input from other members of Pooled Analysis and Writing Group. Members of Country and Regional Data Group commented on draft report. NCD Risk Factor Collaboration (NCD-RisC) Pooled Analysis and Writing (*equal contribution; listed alphabetically)—Goodarz Danaei (Harvard T H Chan School of Public Health, USA)*; Saman Fahimi (Harvard T H Chan School of Public Health, USA)*; Yuan Lu (Harvard T H Chan School of Public Health, USA)*; Bin Zhou (Imperial College London, UK)*; Kaveh Hajifathalian (Harvard T H Chan School of Public Health, USA); Mariachiara Di Cesare (Imperial College London, UK); Wei-Cheng Lo (National Taiwan University, Taiwan); Barbara Reis-Santos (Universidade Federal de Pelotas, Brazil); Melanie J Cowan (World Health Organization, Switzerland); Jonathan E Shaw (Baker IDI Heart and Diabetes Institute, Australia); James Bentham (Imperial College London, UK); John K Lin (University of California San Francisco, USA); Honor Bixby (Imperial College London, UK); Dianna Magliano (Baker IDI Heart and Diabetes Institute, Australia); Pascal Bovet (University of Lausanne, Switzerland; Ministry of Health, Seychelles); J Jaime Miranda (Universidad Peruana Cayetano Heredia, Peru); Young-Ho Khang (Seoul National University, South Korea); Gretchen A Stevens (World Health Organization, Switzerland); Leanne M Riley (World Health Organization, Switzerland); Mohammed K Ali (Emory University, USA); Majid Ezzati (Imperial College London, UK).

J Jaime Miranda (Universidad Peruana Cayetano Heredia, Peru); Young-Ho Khang (Seoul National University, South Korea); Gretchen A Stevens (World Health Organization, Switzerland); Leanne M Riley (World Health Organization, Switzerland); Mohammed K Ali (Emory University, USA); Majid Ezzati (Imperial College London, UK). Country and Regional Data (*equal contribution; listed alphabetically)—Ziad A Abdeen (Al-Quds University, Palestine)*; Khalid Abdul Kadir (Monash University Malaysia, Malaysia)*; Niveen M Abu-Rmeileh (Birzeit University, Palestine)*; Benjamin Acosta-Cazares (Instituto Mexicano del Seguro Social, Mexico)*; Wichai Aekplakorn (Mahidol University, Thailand)*; Carlos A Aguilar-Salinas (Instituto Nacional de Ciencias Médicas y Nutricion, Mexico)*; Alireza Ahmadvand (Tehran University of Medical Sciences, Iran)*; Mohannad Al Nsour (Eastern Mediterranean Public Health Network, Jordan)*; Ala'a Alkerwi (Luxembourg Health Institute, Luxembourg)*; Philippe Amouyel (Lille University and Hospital, France)*; Lars Bo Andersen (University of Southern Denmark, Denmark)*; Sigmund A Anderssen (Norwegian School of Sport Sciences, Norway)*; Dolores S Andrade (Universidad de Cuenca, Ecuador)*; Ranjit Mohan Anjana (Madras Diabetes Research Foundation, India)*; Hajer Aounallah-Skhiri (National Institute of Public Health, Tunisia)*; Tahir Aris (Ministry of Health Malaysia, Malaysia)*; Nimmathota Arlappa (Indian Council of Medical Research, India)*; Dominique Arveiler (Strasbourg University and Hospital, France)*; Felix K Assah (Health of Populations in Transition Research Group, Cameroon)*; Mária Avdicová (Regional Authority of Public Health, Banska Bystrica, Slovakia)*; Nagalla Balakrishna (Indian Council of Medical Research, India)*; Piotr Bandosz (Medical University of Gdansk, Poland)*; Carlo M Barbagallo (University of Palermo, Italy)*; Alberto Barceló (Pan American Health Organization, USA)*; Anwar M Batieha (Jordan University of Science and Technology, Jordan)*; Louise A Baur (University of Sydney, Australia)*; Habiba Ben Romdhane (University Tunis El Manar, Tunisia)*; Antonio Bernabe-Ortiz (Universidad Peruana Cayetano Heredia, Peru)*; Santosh K Bhargava (Sunder Lal Jain Hospital, India)*; Yufang Bi (Shanghai Jiao-Tong University School of Medicine, China)*; Peter Bjerregaard (University of Southern Denmark, Denmark; University of Greenland, Greenland)*; Cecilia Björkelund (University of Gothenburg, Sweden)*; Margaret Blake (NatCen Social Research, UK)*; Anneke Blokstra (National

Jain Hospital, India)*; Yufang Bi (Shanghai Jiao-Tong University School of Medicine, China)*; Peter Bjerregaard (University of Southern Denmark, Denmark; University of Greenland, Greenland)*; Cecilia Björkelund (University of Gothenburg, Sweden)*; Margaret Blake (NatCen Social Research, UK)*; Anneke Blokstra (National Institute for Public Health and the Environment, Netherlands)*; Simona Bo (University of Turin, Italy)*; Bernhard O Boehm (Nanyang Technological University, Singapore)*; Carlos P Boissonnet (Centro de Educación Médica e Investigaciones Clínicas, Argentina)*; Pascal Bovet (University of Lausanne, Switzerland; Ministry of Health, Seychelles)*; Imperia Brajkovich (Universidad Central de Venezuela, Venezuela)*; Juergen Breckenkamp (Bielefeld University, Germany)*; Lizzy M Brewster (University of Amsterdam, Netherlands)*; Garry R Brian (The Fred Hollows Foundation New Zealand, New Zealand)*; Graziella Bruno (University of Turin, Italy)*; Anna Bugge (University of Southern Denmark, Denmark)*; Antonio Cabrera de León (Canarian Health Service, Spain)*; Gunay Can (Istanbul University, Turkey)*; Ana Paula C Cândido (Universidade Federal de Juiz de Fora, Brazil)*; Vincenzo Capuano (Reparto di Cardiologia ed UTIC di Mercato S., Italy)*; Maria J Carvalho (University of Porto, Portugal)*; Felipe F Casanueva (Santiago de Compostela University, Spain)*; Carmelo A Caserta (Associazione Calabrese di Epatologia, Italy)*; Katia Castetbon (French Institute for Health Surveillance, France)*; Snehalatha Chamukuttan (India Diabetes Research Foundation, India)*; Nishi Chaturvedi (University College London, UK)*; Chien-Jen Chen (Academia Sinica, Taiwan)*; Fangfang Chen (Capital Institute of Pediatrics, China)*; Shuohua Chen (Kailuan General Hospital, China)*; Ching-Yu Cheng (Duke-NUS Graduate Medical School, Singapore)*; Angela Chetrit (The Gertner Institute for Epidemiology and Health Policy Research, Israel)*; Shu-Ti Chiou (Ministry of Health and Welfare, Taiwan)*; Yumi Cho (Korea Centers for Disease Control and Prevention, South Korea)*; Jerzy Chudek (Medical University of Silesia, Poland)*; Renata Cifkova (Charles University in Prague, Czech Republic)*; Frank Claessens (Katholieke Universiteit Leuven, Belgium)*; Hans Concin (Agency for Preventive and Social Medicine, Austria)*; Cyrus Cooper (University of Southampton, UK)*; Rachel Cooper (University College London, UK)*; Simona Costanzo (IRCCS Istituto Neurologico Mediterraneo Neuromed, Italy)*; Dominique Cottel (Institut Pasteur de

(Katholieke Universiteit Leuven, Belgium)*; Hans Concin (Agency for Preventive and Social Medicine, Austria)*; Cyrus Cooper (University of Southampton, UK)*; Rachel Cooper (University College London, UK)*; Simona Costanzo (IRCCS Istituto Neurologico Mediterraneo Neuromed, Italy)*; Dominique Cottel (Institut Pasteur de Lille, France)*; Chris Cowell (Westmead University of Sydney, Australia)*; Ana B Crujeiras (CIBERobn, Spain)*; Graziella D'Arrigo (National Research Council, Italy)*; Jean Dallongeville (Institut Pasteur de Lille, France)*; Rachel Dankner (The Gertner Institute for Epidemiology and Health Policy Research, Israel)*; Luc Dauchet (Lille University Hospital, France)*; Giovanni de Gaetano (IRCCS Istituto Neurologico Mediterraneo Neuromed, Italy)*; Stefaan De Henauw (Ghent University, Belgium)*; Mohan Deepa (Madras Diabetes Research Foundation, India)*; Abbas Dehghan (University Medical Center Rotterdam, Netherlands)*; Klodian Dhana (University Medical Center Rotterdam, Netherlands)*; Augusto F Di Castelnuovo (IRCCS Istituto Neurologico Mediterraneo Neuromed, Italy)*; Shirin Djalalinia (Tehran University of Medical Sciences, Iran)*; Kouamelan Doua (Ministère de la Santé et de la Lutte contre le Sida, Côte d'Ivoire)*; Wojciech Drygas (The Cardinal Wyszynski Institute of Cardiology, Poland)*; Yong Du (Robert Koch Institute, Germany)*; Eruke E Egbagbe (University of Benin College of Medical Sciences, Nigeria)*; Robert Eggertsen (University of Gothenburg, Sweden)*; Jalila El Ati (National Institute of Nutrition and Food Technology, Tunisia)*; Roberto Elosua (Institut Hospital del Mar d'Investigacions Mèdiques, Spain)*; Rajiv T Erasmus (University of Stellenbosch, South Africa)*; Cihangir Erem (Karadeniz Technical University, Turkey)*; Gul Ergor (Dokuz Eylul University, Turkey)*; Louise Eriksen (University of Southern Denmark, Denmark)*; Jorge Escobedo-de la Peña (Instituto Mexicano del Seguro Social, Mexico)*; Caroline H Fall (MRC Lifecourse Epidemiology Unit, UK)*; Farshad Farzadfar (Tehran University of Medical Sciences, Iran)*; Francisco J Felix-Redondo (Centro de Salud Villanueva Norte, Spain)*; Trevor S Ferguson (The University of the West Indies, Jamaica)*; Daniel Fernández-Bergés (Hospital Don Benito-Villanueva de la Serena, Spain)*; Marika Ferrari (National Research Institute on Food and Nutrition, Italy)*; Catterina Ferreccio (Pontificia Universidad Católica de Chile, Chile)*; Joseph D Finn (University of Manchester, UK)*; Bernhard Föger (Agency for Preventive an

ca)*; Daniel Fernández-Bergés (Hospital Don Benito-Villanueva de la Serena, Spain)*; Marika Ferrari (National Research Institute on Food and Nutrition, Italy)*; Catterina Ferreccio (Pontificia Universidad Católica de Chile, Chile)*; Joseph D Finn (University of Manchester, UK)*; Bernhard Föger (Agency for Preventive an d Social Medicine, Austria)*; Leng Huat Foo (Universiti Sains Malaysia, Malaysia)*; Heba M Fouad (World Health Organization Regional Office for the Eastern Mediterranean, Egypt)*; Damian K Francis (The University of the West Indies, Jamaica)*; Maria do Carmo Franco (Federal University of São Paulo, Brazil)*; Oscar H Franco (University Medical Center Rotterdam, Netherlands)*; Guillermo Frontera (Hospital Universitario Son Espases, Spain)*; Takuro Furusawa (Kyoto University, Japan)*; Zbigniew Gaciong (Medical University of Warsaw, Poland)*; Andrzej Galbarczyk (Jagiellonian University Medical College, Poland)*; Sarah P Garnett (University of Sydney, Australia)*; Jean-Michel Gaspoz (Geneva University Hospitals, Switzerland)*; Magda Gasull (CIBER en Epidemiología y Salud Pública, Spain)*; Louise Gates (Australian Bureau of Statistics, Australia)*; Johanna M Geleijnse (Wageningen University, Netherlands)*; Anoosheh Ghasemain (Tehran University of Medical Sciences, Iran)*; Simona Giampaoli (Istituto Superiore di Sanita, Italy)*; Francesco Gianfagna (University of Insubria, Italy)*; Jonathan Giovannelli (Lille University Hospital, France)*; Marcela Gonzalez Gross (Universidad Politécnica de Madrid, Spain)*; Juan P González Rivas (The Andes Clinic of Cardio-Metabolic Studies, Venezuela)*; Mariano Bonet Gorbea (National Institute of Hygiene, Epidemiology and Microbiology, Cuba)*; Frederic Gottrand (Université de Lille 2, France)*; Janet F Grant (The University of Adelaide, Australia)*; Tomasz Grodzicki (Jagiellonian University Medical College, Poland)*; Anders Grøntved (University of Southern Denmark, Denmark)*; Grabriella Gruden (University of Turin, Italy)*; Dongfeng Gu (National Center of Cardiovascular Diseases, China)*; Ong Peng Guan (Singapore Eye Research Institute, Singapore)*; Ramiro Guerrero (Universidad Icesi, Colombia)*; Idris Guessous (Geneva University Hospitals, Switzerland)*; Andre L Guimaraes (State University of Montes Claros, Brazil)*; Laura Gutierrez (Institute for Clinical Effectiveness and Health Policy, Argentina)*; Rebecca Hardy (University College London, UK)*; Rachakulla Hari Kumar (Indian Council of Medical Research, India)*; Jiang He (Tulane Un

FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. *The meta-analyses used inverse of variance as survey weights; sensitivity or specificity of either 0% or 100% would make the corresponding variance zero, and therefore the inverse of variance infinite. Figure 2 Prevalence of diabetes defined by FPG-or-2hOGTT versus by FPG only FPG-or-2hOGTT definition was FPG 7·0 mmol/L or more, or 2hOGTT 11·1 mmol/L or more, or history of diabetes or using insulin or oral hypoglycaemic drugs. FPG only definition was FPG 7·0 mmol/L or more, or history of diabetes or using insulin or oral hypoglycaemic drugs. Each point shows one age–sex group in one survey. Table 1 shows the relation summarised as regression coefficients. FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. Figure 3 Prevalence of diabetes defined by HbA1c only versus prevalence defined by (A) FPG only, and (B) FPG-or-2hOGTT HbA1c definition was HbA1c 6·5% or more, or history of diabetes, or using insulin or oral hypoglycaemic drugs. FPG only definition was FPG 7·0 mmol/L or more, or history of diabetes or using insulin or oral hypoglycaemic drugs. FPG-or-2hOGTT definition was FPG 7·0 mmol/L or more, or 2hOGTT 11·1 mmol/L or more, or history of diabetes or using insulin or oral hypoglycaemic drugs. Each point shows one age–sex group in one survey. Table 2 shows the relations summarised as regression coefficients. FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test.

iversity Hospitals, Switzerland)*; Andre L Guimaraes (State University of Montes Claros, Brazil)*; Laura Gutierrez (Institute for Clinical Effectiveness and Health Policy, Argentina)*; Rebecca Hardy (University College London, UK)*; Rachakulla Hari Kumar (Indian Council of Medical Research, India)*; Jiang He (Tulane Un iversity, USA)*; Christin Heidemann (Robert Koch Institute, Germany)*; Ilpo Tapani Hihtaniemi (Imperial College London, UK)*; Sai Yin Ho (University of Hong Kong, China)*; Suzanne C Ho (The Chinese University of Hong Kong, China)*; Albert Hofman (University Medical Center Rotterdam, Netherlands)*; Andrea R V Russo Horimoto (Heart Institute, Brazil)*; Claudia M Hormiga (Fundación Oftalmológica de Santander, Colombia)*; Bernardo L Horta (Universidade Federal de Pelotas, Brazil)*; Leila Houti (University of Oran 1, Algeria)*; Abdullatif S Hussieni (Birzeit University, Palestine)*; Inge Huybrechts (International Agency for Research on Cancer, France)*; Nahla Hwalla (American University of Beirut, Lebanon)*; Licia Iacoviello (IRCCS Istituto Neurologico Mediterraneo Neuromed, Italy)*; Anna G Iannone (Reparto di Cardiologia ed UTIC di Mercato S., Italy)*; Mohsen M Ibrahim (Cairo University, Egypt)*; Nayu Ikeda (National Institute of Health and Nutrition, Japan)*; M Arfan Ikram (University Medical Center Rotterdam, Netherlands)*; Vilma E Irazola (Institute for Clinical Effectiveness and Health Policy, Argentina)*; Muhammad Islam (Aga Khan University, Pakistan)*; Masanori Iwasaki (Niigata University, Japan)*; Jeremy M Jacobs (Hadassah University Medical Center, Israel)*; Tazeen Jafar (Duke-NUS Graduate Medical School, Singapore)*; Grazyna Jasienska (Jagiellonian University Medical College, Poland)*; Chao Qiang Jiang (University of Hong Kong, China)*; Jost B Jonas (Ruprecht-Karls-University of Heidelberg, Germany)*; Pradeep Joshi (World Health Organization Country Office, India)*; Anthony Kafatos (University of Crete, Greece)*; Ofra Kalter-Leibovici (The Gertner Institute for Epidemiology and Health Policy Research, Israel)*; Amir Kasaeian (Tehran University of Medical Sciences, Iran)*; Joanne Katz (Johns Hopkins Bloomberg School of Public Health, USA)*; Prabhdeep Kaur (National Institute of Epidemiology, India)*; Maryam Kavousi (University Medical Center Rotterdam, Netherlands)*; Roya Kelishadi (Isfahan University of Medical Sciences, Iran)*; Andre P Kengne (South African Medical Research Council, South Africa)*; Mathilde Kersting (Research Institute of Child Nutrition,

bhdeep Kaur (National Institute of Epidemiology, India)*; Maryam Kavousi (University Medical Center Rotterdam, Netherlands)*; Roya Kelishadi (Isfahan University of Medical Sciences, Iran)*; Andre P Kengne (South African Medical Research Council, South Africa)*; Mathilde Kersting (Research Institute of Child Nutrition, Germany)*; Yousef Saleh Khader (Jordan University of Science and Technology, Jordan)*; Young-Ho Khang (Seoul National University, South Korea)*; Stefan Kiechl (Medical University Innsbruck, Austria)*; Jeongseon Kim (National Cancer Center, South Korea)*; Yutaka Kiyohara (Kyushu University, Japan)*; Patrick Kolsteren (Institute of Tropical Medicine, Belgium)*; Paul Korrovits (Tartu University Clinics, Estonia)*; Seppo Koskinen (National Institute for Health and Welfare, Finland)*; Wolfgang Kratzer (University Hospital Ulm, Germany)*; Daan Kromhout (Wageningen University, Netherlands)*; Krzysztof Kula (Medical University of Łodz, Poland)*; Pawel Kurjata (The Cardinal Wyszynski Institute of Cardiology, Poland)*; Catherine Kyobutungi (African Population and Health Research Center, Kenya)*; Carl Lachat (Ghent University, Belgium)*; Youcef Laid (National Institute of Public Health, Algeria)*; Tai Hing Lam (University of Hong Kong, China)*; Orlando Landrove (Ministerio de Salud Pública, Cuba)*; Vera Lanska (Institute for Clinical and Experimental Medicine Prague, Czech Republic)*; Georg Lappas (Sahlgrenska Academy, Sweden)*; Avula Laxmaiah (Indian Council of Medical Research, India)*; Catherine Leclercq (Food and Agriculture Organization, Italy)*; Jeannette Lee (National University of Singapore, Singapore)*; Jeonghee Lee (National Cancer Center, South Korea)*; Terho Lehtimäki (Tampere University Hospital, Finland)*; Rampal Lekhraj (Universiti Putra Malaysia, Malaysia)*; Luz M León-Muñoz (Universidad Autónoma de Madrid, Spain)*; Yanping Li (Harvard T H Chan School of Public Health, USA)*; Wei-Yen Lim (National University of Singapore, Singapore)*; M Fernanda Lima-Costa (Oswaldo Cruz Foundation Rene Rachou Research Institute, Brazil)*; Hsien-Ho Lin (National Taiwan University, Taiwan)*; Xu Lin (University of Chinese Academy of Sciences, China)*; Lauren Lissner (University of Gothenburg, Sweden)*; Roberto Lorbeer (University Medicine Greifswald, Germany)*; José Eugenio Lozano (Consejería de Sanidad Junta de Castilla y León, Spain)*; Annamari Lundqvist (National Institute for Health and Welfare, Finland)*; Per Lytsy (University of Uppsala, Sweden)*; Guansheng Ma (Peking Un

ner (University of Gothenburg, Sweden)*; Roberto Lorbeer (University Medicine Greifswald, Germany)*; José Eugenio Lozano (Consejería de Sanidad Junta de Castilla y León, Spain)*; Annamari Lundqvist (National Institute for Health and Welfare, Finland)*; Per Lytsy (University of Uppsala, Sweden)*; Guansheng Ma (Peking Un iversity, China)*; George L L Machado-Coelho (Universidade Federal de Ouro Preto, Brazil)*; Suka Machi (The Jikei University School of Medicine, Japan)*; Stefania Maggi (National Research Council, Italy)*; Dianna Magliano (Baker IDI Heart and Diabetes Institute, Australia)*; Marcia Makdisse (Hospital Israelita Albert Einstein, Brazil)*; Kodavanti Mallikharjuna Rao (Indian Council of Medical Research, India)*; Yannis Manios (Harokopio University of Athens, Greece)*; Enzo Manzato (University of Padova, Italy)*; Paula Margozzini (Pontificia Universidad Católica de Chile, Chile)*; Pedro Marques-Vidal (Lausanne University Hospital, Switzerland)*; Reynaldo Martorell (Emory University, USA)*; Shariq R Masoodi (Sher-i-Kashmir Institute of Medical Sciences, India)*; Tandi E Matsha (Cape Peninsula University of Technology, South Africa)*; Jean Claude N Mbanya (University of Yaoundé 1, Cameroon)*; Shelly R McFarlane (The University of the West Indies, Jamaica)*; Stephen T McGarvey (Brown University, USA)*; Stela McLachlan (University of Edinburgh, UK)*; Breige A McNulty (University College Dublin, Ireland)*; Sounnia Mediene-Benchekor (University of Oran 1, Algeria)*; Aline Meirhaeghe (INSERM, France)*; Ana Maria B Menezes (Universidade Federal de Pelotas, Brazil)*; Shahin Merat (Tehran University of Medical Sciences, Iran)*; Indrapal I Meshram (Indian Council of Medical Research, India)*; Jie Mi (Capital Institute of Pediatrics, China)*; Juan Francisco Miquel (Pontificia Universidad Católica de Chile, Chile)*; J Jaime Miranda (Universidad Peruana Cayetano Heredia, Peru)*; Mostafa K Mohamed (Ain Shams University, Egypt)*; Kazem Mohammad (Tehran University of Medical Sciences, Iran)*; Viswanathan Mohan (Madras Diabetes Research Foundation, India)*; Muhammad Fadhli Mohd Yusoff (Ministry of Health Malaysia, Malaysia)*; Niels C Møller (University of Southern Denmark, Denmark)*; Denes Molnar (University of Pécs, Hungary)*; Charles K Mondo (Mulago Hospital, Uganda)*; Luis A Moreno (Universidad de Zaragoza, Spain)*; Karen Morgan (PU-RCSI School of Medicine, Malaysia)*; George Moschonis (Harokopio University of Athens, Greece)*; Malgorzata Mossakowska (International Institute of Mo

rk, Denmark)*; Denes Molnar (University of Pécs, Hungary)*; Charles K Mondo (Mulago Hospital, Uganda)*; Luis A Moreno (Universidad de Zaragoza, Spain)*; Karen Morgan (PU-RCSI School of Medicine, Malaysia)*; George Moschonis (Harokopio University of Athens, Greece)*; Malgorzata Mossakowska (International Institute of Mo lecular and Cell Biology, Poland)*; Aya Mostafa (Ain Shams University, Egypt)*; Jorge Mota (University of Porto, Portugal)*; Maria L Muiesan (University of Brescia, Italy)*; Martina Müller-Nurasyid (Helmholtz Zentrum München, Germany)*; Jaakko Mursu (University of Eastern Finland, Finland)*; Gabriele Nagel (University of Ulm, Germany)*; Jana Námešná (Regional Authority of Public Health, Banska Bystrica, Slovakia)*; Ei Ei K Nang (National University of Singapore, Singapore)*; Vinay B Nangia (Suraj Eye Institute, India)*; Eva Maria Navarrete-Muñoz (CIBER de Epidemiología y Salud Pública, Spain)*; Ndeye Coumba Ndiaye (INSERM, France)*; Flavio Nervi (Pontificia Universidad Católica de Chile, Chile)*; Nguyen D Nguyen (University of Pharmacy and Medicine of Ho Chi Minh City, Vietnam)*; Ramfis E Nieto-Martínez (Universidad Centro-Occidental Lisandro Alvarado, Venezuela)*; Guang Ning (Shanghai Jiao-Tong University School of Medicine, China)*; Toshiharu Ninomiya (Kyushu University, Japan)*; Marianna Noale (National Research Council, Italy)*; Davide Noto (University of Palermo, Italy)*; Angélica M Ochoa-Avilés (Universidad de Cuenca, Ecuador)*; Kyungwon Oh (Korea Centers for Disease Control and Prevention, South Korea)*; Altan Onat (Istanbul University, Turkey)*; Clive Osmond (MRC Lifecourse Epidemiology Unit, UK)*; Johanna A Otero (Fundación Oftalmológica de Santander, Colombia)*; Luigi Palmieri (Istituto Superiore di Sanita, Italy)*; Songhomitra Panda-Jonas (Ruprecht-Karls-University of Heidelberg, Germany)*; Francesco Panza (Unversity of Bari, Italy)*; Mahboubeh Parsaeian (Tehran University of Medical Sciences, Iran)*; Sergio Viana Peixoto (Oswaldo Cruz Foundation Rene Rachou Research Institute, Brazil)*; Alexandre C Pereira (Heart Institute, Brazil)*; Annette Peters (Helmholtz Zentrum München, Germany)*; Niloofar Peykari (Tehran University of Medical Sciences, Iran)*; Aida Pilav (Federal Ministry of Health, Bosnia and Herzegovina)*; Freda Pitakaka (University of New South Wales, Australia)*; Aleksandra Piwonska (The Cardinal Wyszynski Institute of Cardiology, Poland)*; Jerzy Piwonski (The Cardinal Wyszynski Institute of Cardiology, Poland)*; Pedro Plans-Rubió (Public

ences, Iran)*; Aida Pilav (Federal Ministry of Health, Bosnia and Herzegovina)*; Freda Pitakaka (University of New South Wales, Australia)*; Aleksandra Piwonska (The Cardinal Wyszynski Institute of Cardiology, Poland)*; Jerzy Piwonski (The Cardinal Wyszynski Institute of Cardiology, Poland)*; Pedro Plans-Rubió (Public Health Agency of Catalonia, Spain)*; Miquel Porta (Hospital del Mar Medical Research Institute-IMIM, Spain)*; Marileen L P Portegies (University Medical Center Rotterdam, Netherlands)*; Hossein Poustchi (Tehran University of Medical Sciences, Iran)*; Rajendra Pradeepa (Madras Diabetes Research Foundation, India)*; Jacqueline F Price (University of Edinburgh, UK)*; Margus Punab (Tartu University Clinics, Estonia)*; Radwan F Qasrawi (Al-Quds University, Palestine)*; Mostafa Qorbani (Alborz University of Medical Sciences, Iran)*; Olli Raitakari (University of Turku, Finland)*; Sudha Ramachandra Rao (National Institute of Epidemiology, India)*; Ambady Ramachandran (India Diabetes Research Foundation, India)*; Rafel Ramos (Institut Universitari d'Investigació en Atenció Primària Jordi Gol, Spain)*; Sanjay Rampal (University of Malaya, Malaysia)*; Wolfgang Rathmann (German Diabetes Center, Germany)*; Josep Redon (University of Valencia, Spain)*; Paul Ferdinand M Reganit (University of the Philippines, Philippines)*; Fernando Rigo (Health Center San Agustin, Spain)*; Sian M Robinson (University of Southampton, UK)*; Cynthia Robitaille (Public Health Agency of Canada, Canada)*; Laura A Rodríguez (Universidad Industrial de Santander, Colombia)*; Fernando Rodríguez-Artalejo (Universidad Autónoma de Madrid, Spain)*; María del Cristo Rodriguez-Perez (Canarian Health Service, Spain)*; Rosalba Rojas-Martinez (Instituto Nacional de Salud Pública, Mexico)*; Dora Romaguera (Centre for Research in Environmental Epidemiology, Spain)*; Annika Rosengren (Sahlgrenska Academy, Sweden)*; Adolfo Rubinstein (Institute for Clinical Effectiveness and Health Policy, Argentina)*; Ornelas Rui (University of Madeira, Portugal)*; Blanca Sandra Ruiz-Betancourt (Instituto Mexicano del Seguro Social, Mexico)*; Marcin Rutkowski (Medical University of Gdansk, Poland)*; Charumathi Sabanayagam (Singapore Eye Research Institute, Singapore)*; Harshpal S Sachdev (Sitaram Bhartia Institute of Science and Research, India)*; Olfa Saidi (University Tunis El Manar, Tunisia)*; Sibel Sakarya (Marmara University, Turkey)*; Benoit Salanave (French Institute for Health Surveillance, France)*; Jukka T Salonen (Univ

ayagam (Singapore Eye Research Institute, Singapore)*; Harshpal S Sachdev (Sitaram Bhartia Institute of Science and Research, India)*; Olfa Saidi (University Tunis El Manar, Tunisia)*; Sibel Sakarya (Marmara University, Turkey)*; Benoit Salanave (French Institute for Health Surveillance, France)*; Jukka T Salonen (Univ ersity of Helsinki, Finland)*; Massimo Salvetti (University of Brescia, Italy)*; Jose Sánchez-Abanto (National Institute of Health, Peru)*; Renata Nunes dos Santos (University of São Paulo, Brazil)*; Rute Santos (University of Porto, Portugal)*; Luis B Sardinha (Universidade de Lisboa, Portugal)*; Marcia Scazufca (University of São Paulo, Brazil)*; Herman Schargrodsky (Hospital Italiano de Buenos Aires, Argentina)*; Christa Scheidt-Nave (Robert Koch Institute, Germany)*; Jonathan E Shaw (Baker IDI Heart and Diabetes Institute, Australia)*; Kenji Shibuya (The University of Tokyo, Japan)*; Youchan Shin (Singapore Eye Research Institute, Singapore)*; Rahman Shiri (Finnish Institute of Occupational Health, Finland)*; Rosalynn Siantar (Singapore Eye Research Institute, Singapore)*; Abla M Sibai (American University of Beirut, Lebanon)*; Mary Simon (India Diabetes Research Foundation, India)*; Judith Simons (St Vincent's Hospital, Australia)*; Leon A Simons (University of New South Wales, Australia)*; Michael Sjostrom (Karolinska Institute, Sweden)*; Jolanta Slowikowska-Hilczer (Medical University of Łodz, Poland)*; Przemyslaw Slusarczyk (International Institute of Molecular and Cell Biology, Poland)*; Liam Smeeth (London School of Hygiene & Tropical Medicine, UK)*; Marieke B Snijder (University of Amsterdam, Netherlands)*; Vincenzo Solfrizzi (University of Bari, Italy)*; Emily Sonestedt (Lund University, Sweden)*; Aicha Soumare (University of Bordeaux, France)*; Jan A Staessen (University of Leuven, Belgium)*; Jostein Steene-Johannessen (Norwegian School of Sport Sciences, Norway)*; Peter Stehle (Bonn University, Germany)*; Aryeh D Stein (Emory University, USA)*; Jochanan Stessman (Hadassah University Medical Center, Israel)*; Doris Stöckl (Helmholtz Zentrum München, Germany)*; Jakub Stokwiszewski (National Institute of Hygiene, Poland)*; Maria Wany Strufaldi (Federal University of São Paulo, Brazil)*; Chien-An Sun (Fu Jen Catholic University, Taiwan)*; Johan Sundström (Uppsala University, Sweden)*; Paibul Suriyawongpaisal (Mahidol University, Thailand)*; Rody G Sy (University of the Philippines, Philippines)*; E Shyong Tai (National University of Singapore, Singapor

ufaldi (Federal University of São Paulo, Brazil)*; Chien-An Sun (Fu Jen Catholic University, Taiwan)*; Johan Sundström (Uppsala University, Sweden)*; Paibul Suriyawongpaisal (Mahidol University, Thailand)*; Rody G Sy (University of the Philippines, Philippines)*; E Shyong Tai (National University of Singapore, Singapor e)*; Mohammed Tarawneh (Ministry of Health, Jordan)*; Carolina B Tarqui-Mamani (National Institute of Health, Peru)*; Lutgarde Thijs (University of Leuven, Belgium)*; Janne S Tolstrup (University of Southern Denmark, Denmark)*; Murat Topbas (Karadeniz Technical University, Turkey)*; Maties Torrent (Area de Salut de Menorca, Spain)*; Pierre Traissac (Institut de Recherche pour le Développement, France)*; Oanh T H Trinh (University of Pharmacy and Medicine of Ho Chi Minh City, Vietnam)*; Marshall K Tulloch-Reid (The University of the West Indies, Jamaica)*; Tomi-Pekka Tuomainen (University of Eastern Finland, Finland)*; Maria L Turley (Ministry of Health, New Zealand)*; Christophe Tzourio (University of Bordeaux, France)*; Peter Ueda (Harvard T H Chan School of Public Health, USA)*; Flora M Ukoli (Meharry Medical College, USA)*; Hanno Ulmer (Medical University of Innsbruck, Austria)*; Gonzalo Valdivia (Pontificia Universidad Católica de Chile, Chile)*; Irene G M van Valkengoed (Academic Medical Center of University of Amsterdam, Netherlands)*; Dirk Vanderschueren (Katholieke Universiteit Leuven, Belgium)*; Diego Vanuzzo (Centro di Prevenzione Cardiovascolare, Italy)*; Tomas Vega (Consejería de Sanidad, Junta de Castilla y León, Spain)*; Gustavo Velasquez-Melendez (Universidade Federal de Minas Gerais, Brazil)*; Giovanni Veronesi (University of Insubria, Italy)*; Monique Verschuren (National Institute for Public Health and the Environment, Netherlands)*; Jesus Vioque (Universidad Miguel Hernandez, Spain)*; Jyrki Virtanen (University of Eastern Finland, Finland)*; Sophie Visvikis-Siest (INSERM, France)*; Bharathi Viswanathan (Ministry of Health, Seychelles)*; Peter Vollenweider (Lausanne University Hospital, Switzerland)*; Sari Voutilainen (University of Eastern Finland, Finland)*; Alisha N Wade (University of the Witwatersrand, South Africa)*; Aline Wagner (University of Strasbourg, France)*; Janette Walton (University College Cork, Ireland)*; Wan Nazaimoon Wan Mohamud (Institute for Medical Research, Malaysia)*; Ming-Dong Wang (Public Health Agency of Canada, Canada)*; Ya Xing Wang (Beijing Tongren Hospital, China)*; S Goya Wannamethee (University College London,

agner (University of Strasbourg, France)*; Janette Walton (University College Cork, Ireland)*; Wan Nazaimoon Wan Mohamud (Institute for Medical Research, Malaysia)*; Ming-Dong Wang (Public Health Agency of Canada, Canada)*; Ya Xing Wang (Beijing Tongren Hospital, China)*; S Goya Wannamethee (University College London, UK)*; Deepa Weerasekera (Ministry of Health, New Zealand)*; Peter H Whincup (St George's, University of London, UK)*; Kurt Widhalm (Medical University of Vienna, Austria)*; Andrzej Wiecek (Medical University of Silesia, Poland)*; Rainford J Wilks (The University of the West Indies, Jamaica)*; Johann Willeit (Medical University Innsbruck, Austria)*; Bogdan Wojtyniak (National Institute of Hygiene, Poland)*; Tien Yin Wong (Duke-NUS Graduate Medical School, Singapore)*; Jean Woo (The Chinese University of Hong Kong, China)*; Mark Woodward (University of Sydney, Australia; University of Oxford, UK)*; Aleksander Giwercman Wu (Lund University, Sweden)*; Frederick C Wu (University of Manchester, UK)*; Shou Ling Wu (Kailuan General Hospital, China)*; Haiquan Xu (Institute of Food and Nutrition Development of Ministry of Agriculture, China)*; Xiaoguang Yang (Chinese Center for Disease Control and Prevention, China)*; Xingwang Ye (University of Chinese Academy of Sciences, China)*; Akihiro Yoshihara (Niigata University, Japan)*; Novie O Younger-Coleman (The University of the West Indies, Jamaica)*; Sabina Zambon (University of Padova, Italy)*; Abdul Hamid Zargar (Center for Diabetes and Endocrine Care, India)*; Tomasz Zdrojewski (Medical University of Gdansk, Poland)*; Wenhua Zhao (Chinese Center for Disease Control and Prevention, China)*; Yingfeng Zheng (Singapore Eye Research Institute, Singapore)*.

Jamaica)*; Sabina Zambon (University of Padova, Italy)*; Abdul Hamid Zargar (Center for Diabetes and Endocrine Care, India)*; Tomasz Zdrojewski (Medical University of Gdansk, Poland)*; Wenhua Zhao (Chinese Center for Disease Control and Prevention, China)*; Yingfeng Zheng (Singapore Eye Research Institute, Singapore)*. Declaration of interests JJM has received funding from Medtronics Foundation outside the submitted work. DM has received grants to her institution from Novartis Pharmaceutical, Novo Nordisk Pharmaceutical, Pharmacia and Upjohn, Pfizer, Sanofi Synthelabo, and Servier Laboratories. JES has received grants to his institution from Abbott, Alphapharm, AstraZeneca, Aventis Pharmaceutical, Eli Lilly, GlaxoSmithKline, Janssen-Cilag, Merck Lipha, Merck Sharp & Dohme, Novartis Pharmaceutical, Novo Nordisk Pharmaceutical, Pharmacia and Upjohn, Pfizer, Sanofi Synthelabo, and Servier Laboratories. All other Pooled Analysis and Writing Group members report no competing interests. Figure 1 Study and data inclusion FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. *The meta-analyses used inverse of variance as survey weights; sensitivity or specificity of either 0% or 100% would make the corresponding variance zero, and therefore the inverse of variance infinite. Figure 2 Prevalence of diabetes defined by FPG-or-2hOGTT versus by FPG only

inition was FPG 7·0 mmol/L or more, or 2hOGTT 11·1 mmol/L or more, or history of diabetes or using insulin or oral hypoglycaemic drugs. Each point shows one age–sex group in one survey. Table 2 shows the relations summarised as regression coefficients. FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. Table 1 Regression coefficients for the relation between probit-transformed prevalence of diabetes based on FPG-or-2hOGTT versus diabetes based on FPG only Coefficient (95% CI) p value Univariate R2* Semipartial R2† Intercept 0·135 (−0·020 to 0·290) 0·0872 NA NA Probit-transformed prevalence of diabetes based on FPG 0·903 (0·880 to 0·927) <0·0001 0·963 0·368 Mean age of age–sex group (per 10 years older) 0·048 (0·039 to 0·056) <0·0001 0·444 0·008 Study midyear (per one more recent year since 1976) −0·001 (−0·002 to 0·000) 0·1643 0·003 <0·001 Natural logarithm of per person gross domestic product −0·033 (−0·046 to −0·019) <0·0001 0·004 0·001 Mean BMI 0·000 (−0·004 to 0·004) 0·9057 0·092 <0·001 Study representativeness .. .. 0·021 0·001 National Reference .. .. .. Subnational −0·031 (−0·070 to 0·008) 0·1141 .. .. Community −0·070 (−0·101 to −0·039) <0·0001 .. .. FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. * Calculated by regressing against each independent variable alone; equals the square of the correlation coefficient. † Shows how much R2 decreases if that independent variable is removed from the full model; the overall R2 for the model was 0·973.

Coefficient (95% CI) p value Univariate R2* Semipartial R2† Intercept 0·135 (−0·020 to 0·290) 0·0872 NA NA Probit-transformed prevalence of diabetes based on FPG 0·903 (0·880 to 0·927) <0·0001 0·963 0·368 Mean age of age–sex group (per 10 years older) 0·048 (0·039 to 0·056) <0·0001 0·444 0·008 Study midyear (per one more recent year since 1976) −0·001 (−0·002 to 0·000) 0·1643 0·003 <0·001 Natural logarithm of per person gross domestic product −0·033 (−0·046 to −0·019) <0·0001 0·004 0·001 Mean BMI 0·000 (−0·004 to 0·004) 0·9057 0·092 <0·001 Study representativeness .. .. 0·021 0·001 National Reference .. .. .. Subnational −0·031 (−0·070 to 0·008) 0·1141 .. .. Community −0·070 (−0·101 to −0·039) <0·0001 .. .. FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. * Calculated by regressing against each independent variable alone; equals the square of the correlation coefficient. † Shows how much R2 decreases if that independent variable is removed from the full model; the overall R2 for the model was 0·973. Table 2 Regression coefficients for the association between probit-transformed prevalence of diabetes based on HbA1c and probit-transformed prevalence based on FPG

* Calculated by regressing against each independent variable alone; equals the square of the correlation coefficient. † Shows how much R2 decreases if that independent variable is removed from the full model; the overall R2 for the model was 0·973. Table 2 Regression coefficients for the association between probit-transformed prevalence of diabetes based on HbA1c and probit-transformed prevalence based on FPG Coefficient (95% CI) p value* Univariate R2† Semipartial R2‡ Intercept −1·761 (−2·229 to −1·266) <0·0001 NA NA Probit-transformed prevalence of diabetes based on FPG 0·799 (0·763 to 0·835) <0·0001 0·915 0·075 Mean age of age–sex group (per 10 years older) 0·052 (0·042 to 0·062) <0·0001 0·601 0·011 Study midyear (per one more recent year since 2000) 0·012 (0·009 to 0·015) <0·0001 0·014 0·006 Natural logarithm of per person gross domestic product 0·076 (0·035 to 0·114) 0·0001 0·052 0·003 Mean BMI 0·018 (0·010 to 0·027) <0·0001 0·022 0·002 Study representativeness .. .. 0·013 0·004 National Reference .. .. .. Subnational −0·004 (−0·047 to 0·040) 0·8758 .. .. Community 0·090 (0·060 to 0·119) <0·0001 .. .. The appendix shows regional random effects. FPG=fasting plasma glucose. * p values using likelihood ratio test, which compares the likelihood of the models with and without the variable of interest.78 † Calculated by regressing against each independent variable alone, without the regional random effect; equals the square of the correlation coefficient.

Coefficient (95% CI) p value* Univariate R2† Semipartial R2‡ Intercept −1·761 (−2·229 to −1·266) <0·0001 NA NA Probit-transformed prevalence of diabetes based on FPG 0·799 (0·763 to 0·835) <0·0001 0·915 0·075 Mean age of age–sex group (per 10 years older) 0·052 (0·042 to 0·062) <0·0001 0·601 0·011 Study midyear (per one more recent year since 2000) 0·012 (0·009 to 0·015) <0·0001 0·014 0·006 Natural logarithm of per person gross domestic product 0·076 (0·035 to 0·114) 0·0001 0·052 0·003 Mean BMI 0·018 (0·010 to 0·027) <0·0001 0·022 0·002 Study representativeness .. .. 0·013 0·004 National Reference .. .. .. Subnational −0·004 (−0·047 to 0·040) 0·8758 .. .. Community 0·090 (0·060 to 0·119) <0·0001 .. .. The appendix shows regional random effects. FPG=fasting plasma glucose. * p values using likelihood ratio test, which compares the likelihood of the models with and without the variable of interest.78 † Calculated by regressing against each independent variable alone, without the regional random effect; equals the square of the correlation coefficient. ‡ Is the decrease of R2 if one of the independent variables is removed from the full model; however, traditional R2 is not clearly defined for mixed-effect models, we have used the conditional R2 that describes the proportion of variance explained by both fixed and random factors.79 The overall conditional R2 for the model was 0·949. Table 3 Pooled sensitivity and specificity of diabetes diagnosis using different definitions among participants without diagnosed diabetes

‡ Is the decrease of R2 if one of the independent variables is removed from the full model; however, traditional R2 is not clearly defined for mixed-effect models, we have used the conditional R2 that describes the proportion of variance explained by both fixed and random factors.79 The overall conditional R2 for the model was 0·949. Table 3 Pooled sensitivity and specificity of diabetes diagnosis using different definitions among participants without diagnosed diabetes Number of surveys Sensitivity Specificity (%; 95% CI) I2 (%; 95% CI) I2 HbA1cvs FPG 27 52·82 (51·33–54·30) 97·6% 99·74 (99·71–99·78) 98·2% HbA1cvs 2hOGTT 9 37·16 (35·05–39·28) 97·6% 99·84 (99·79–99·89) 97·3% HbA1cvs FPG-or-2hOGTT 9 30·46 (28·66–32·25) 97·9% 99·69 (99·63–99·76) 98·0% FPG vs 2hOGTT 33 54·42 (53·26–55·57) 96·9% 98·90 (98·83–98·97) 94·4% The appendix shows detailed results of these meta-analyses. Diabetes was defined as HbA1c ≥6·5%, FPG ≥7·0 mmol/L, and 2hOGTT ≥11·1 mmol/L. FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. Table 4 Univariate metaregression coefficients for sensitivity of HbA1c versus FPG in participants without diagnosed diabetes

Number of surveys Sensitivity Specificity (%; 95% CI) I2 (%; 95% CI) I2 HbA1cvs FPG 27 52·82 (51·33–54·30) 97·6% 99·74 (99·71–99·78) 98·2% HbA1cvs 2hOGTT 9 37·16 (35·05–39·28) 97·6% 99·84 (99·79–99·89) 97·3% HbA1cvs FPG-or-2hOGTT 9 30·46 (28·66–32·25) 97·9% 99·69 (99·63–99·76) 98·0% FPG vs 2hOGTT 33 54·42 (53·26–55·57) 96·9% 98·90 (98·83–98·97) 94·4% The appendix shows detailed results of these meta-analyses. Diabetes was defined as HbA1c ≥6·5%, FPG ≥7·0 mmol/L, and 2hOGTT ≥11·1 mmol/L. FPG=fasting plasma glucose. 2hOGTT=2-h oral glucose tolerance test. Table 4 Univariate metaregression coefficients for sensitivity of HbA1c versus FPG in participants without diagnosed diabetes Mean difference in sensitivity (percentage points; 95% CI) p value Mean age (per 10 years older) −4·1 (−12·7 to 4·5) 0·3361 Percent male participants (per 10% more male) 4·6 (−9·0 to 18·2) 0·4901 Study midyear (per one more recent year) 1·2 (−0·9 to 3·2) 0·2566 Region .. 0·2097 High-income western countries Reference group .. East, south, and southeast Asia 21·0 (−0·3 to 42·2) .. Latin America and the Caribbean 8·5 (−17·9 to 34·9) .. Sub-Saharan Africa 17·6 (−14·1 to 49·2) .. Study representativeness .. 0·0915 National Reference group .. Subnational 1·7 (−28·6 to 31·9) .. Community 21·4 (2·1 to 40·8) .. Prevalence of undiagnosed diabetes (percentage point higher undiagnosed diabetes) −0·7 (−4·0 to 2·6) 0·6780 Sample size (per 1000 participants without diagnosed diabetes) −1·6 (−4·6 to 1·4) 0·2730 Natural logarithm of per person gross domestic product −6·5 (−17·6 to 4·6) 0·2410 Mean haemoglobin (per g/L)* −2·0 (−4·1 to 0·2) 0·0677 We used a HbA1c definition of 6·5% or more and a FPG definition of 7·0 mmol/L or more. FPG=fasting plasma glucose.

00 participants without diagnosed diabetes) −1·6 (−4·6 to 1·4) 0·2730 Natural logarithm of per person gross domestic product −6·5 (−17·6 to 4·6) 0·2410 Mean haemoglobin (per g/L)* −2·0 (−4·1 to 0·2) 0·0677 We used a HbA1c definition of 6·5% or more and a FPG definition of 7·0 mmol/L or more. FPG=fasting plasma glucose. * Reliable mean haemoglobin data were available only for women of child-bearing age.82 The national mean for each country-year was used for both men and women; restricting the analysis to women led to similar results, with a mean difference of −2·1 (−4·5 to 0·3, p=0·0929).

Introduction The optimum diuretic for hypertension remains uncertain. Disparity has been growing between the drugs and doses proven to reduce risk of stroke, myocardial infarction, and heart failure, and those recommended by guidelines.1, 2 This move away from recommendation of diuretics in guidelines was driven by an awareness that thiazide and thiazide-like diuretics can increase risk of developing type 2 diabetes.3, 4, 5, 6, 7 The risk seems linked to potassium depletion, and might be avoided by use of potassium-sparing diuretics,3, 4, 8 which are conventionally thought to be the weakest class of diuretic because most filtered sodium is reabsorbed upstream of their site of action in the nephron. But potassium-sparing diuretics target a common site of sodium retention in hypertension, and might be essential in the prevention of compensatory responses to the more proximally acting thiazide and loop diuretics.9 Thus, the hypothesis arose for the present study that an adequate dose of potassium-sparing diuretic would have opposite effects on potassium and glucose to those of a thiazide diuretic, but would have similar or additional effects on blood pressure when the two were compared or combined. Panel Research in context Evidence before this study

Introduction The optimum diuretic for hypertension remains uncertain. Disparity has been growing between the drugs and doses proven to reduce risk of stroke, myocardial infarction, and heart failure, and those recommended by guidelines.1, 2 This move away from recommendation of diuretics in guidelines was driven by an awareness that thiazide and thiazide-like diuretics can increase risk of developing type 2 diabetes.3, 4, 5, 6, 7 The risk seems linked to potassium depletion, and might be avoided by use of potassium-sparing diuretics,3, 4, 8 which are conventionally thought to be the weakest class of diuretic because most filtered sodium is reabsorbed upstream of their site of action in the nephron. But potassium-sparing diuretics target a common site of sodium retention in hypertension, and might be essential in the prevention of compensatory responses to the more proximally acting thiazide and loop diuretics.9 Thus, the hypothesis arose for the present study that an adequate dose of potassium-sparing diuretic would have opposite effects on potassium and glucose to those of a thiazide diuretic, but would have similar or additional effects on blood pressure when the two were compared or combined. Panel Research in context Evidence before this study We searched MEDLINE and Ovid with the terms “thiazide diuretic”, “potassium”, and “glucose tolerance” under the medical subject headings “diabetes” and “hypertension” for observational studies or clinical trials published in English of diuretic use, diabetes, and glucose tolerance in hypertension. We did our last search on July 23, 2015. A network meta-analysis by Elliott and Meyer of incident diabetes in 22 clinical trials of antihypertensive drugs involving 143 153 participants showed that placebo groups had a lower odds ratio of developing diabetes (0·77, 95% CI 0·63–0·94) when compared with thiazide-assigned groups. In a retrospective meta-analysis by Zillich and colleagues, hypokalaemia and hyperglycaemia were significantly associated in patients given thiazide diuretics. There was an average reduction in serum potassium of 0·23 mmol/L and an increase in glucose of 3·26 mg/dL in studies in which potassium supplements or potassium-sparing drugs were used. In studies in which potassium supplements or potassium-sparing drugs were not used, the average reduction in serum potassium concentrations was 0·37 mmol/L, and the increase in serum glucose was 6·01 mg/dL (p<0·03). A quantitative review by the National Heart, Lung, and Blood Institute in 2008 cast some doubt on the strength of the observational studies that first led to recognition of the thiazide association with diabetes, and investigators noted that the strongest evidence for an adverse effect of thiazides was in prospective outcome comparisons with placebo. The conclusion was that prospective studies should be done to investigate the hypothesis that hypokalaemia is the mediator of thiazide-induced dysglycaemia. Since 2008, diuretic doses and use have fallen because of concerns about metabolic consequences.

of thiazides was in prospective outcome comparisons with placebo. The conclusion was that prospective studies should be done to investigate the hypothesis that hypokalaemia is the mediator of thiazide-induced dysglycaemia. Since 2008, diuretic doses and use have fallen because of concerns about metabolic consequences. Added value of this study In our study we compared a potassium-sparing diuretic, a potassium-losing diuretic, and a combination of the two during 24 weeks in 441 patients who were either previously untreated or taking angiotensin-converting enzyme inhibitors, angiotensin-receptor blockers, or calcium-channel blockers. Our results supported our hypothesis that thiazide-induced glucose intolerance, as shown by 2 h glucose concentrations in oral glucose tolerance tests, would not occur in the absence of potassium depletion. Our results do not completely prove that potassium depletion causes the effect of thiazide on glucose tolerance, but show that it is possible to potentiate the benefit of two classes of diuretic on blood pressure while cancelling out undesired effects of thiazides on glucose and potassium concentrations. Implications of all the available evidence

In our study we compared a potassium-sparing diuretic, a potassium-losing diuretic, and a combination of the two during 24 weeks in 441 patients who were either previously untreated or taking angiotensin-converting enzyme inhibitors, angiotensin-receptor blockers, or calcium-channel blockers. Our results supported our hypothesis that thiazide-induced glucose intolerance, as shown by 2 h glucose concentrations in oral glucose tolerance tests, would not occur in the absence of potassium depletion. Our results do not completely prove that potassium depletion causes the effect of thiazide on glucose tolerance, but show that it is possible to potentiate the benefit of two classes of diuretic on blood pressure while cancelling out undesired effects of thiazides on glucose and potassium concentrations. Implications of all the available evidence At an adequate dose (10–20 mg), amiloride is as efficacious as 25–50 mg hydrochlorothiazide, and is not associated with undesirable metabolic consequences. The combination of amiloride and hydrochlorothiazide was already known, from the Medical Research Council's Elderly trial and INSIGHT, to be more efficacious than comparator drugs at preventing some complications of hypertension; however, the 2·5–5 mg dose of amiloride given in these studies was inadequate to prevent hydrochlorothiazide-associated hypokalaemia or diabetes. We propose that the amiloride–hydrochlorothiazide combination tested in this study should be considered in patients taking either an angiotensin-converting enzyme inhibitor, an angiotensin-receptor blocker, or a calcium-channel blocker as a first-line treatment for hypertension.

ide-associated hypokalaemia or diabetes. We propose that the amiloride–hydrochlorothiazide combination tested in this study should be considered in patients taking either an angiotensin-converting enzyme inhibitor, an angiotensin-receptor blocker, or a calcium-channel blocker as a first-line treatment for hypertension. The prevailing view in the 1990s and 2000s was that, at low doses, thiazides did not cause metabolic consequences but remained maximally efficacious at lowering blood pressure.10 Such a view was initially supported by under-powered comparisons of doses, in which no difference in effects on blood pressure were reported.11, 12 However, an apparent dose–response correlation for blood pressure was rediscovered during treatment titration for an outcome comparison of diuretics with calcium-channel blockers, and this relation was confirmed by a formal, crossover comparison of doses in the Spironolactone, Amiloride, Losartan, Thiazide (SALT) study.13, 14 The results of SALT also showed the potential value of potassium-sparing diuretics as an alternative to high-dose thiazides. However, in practice, concerns about hyperkalaemia limit the use of potassium-sparing diuretics, especially in an era when most patients are receiving a renin–angiotensin system (RAS) blocker. Additionally, without proof that potassium-sparing diuretics are not associated with glucose intolerance, or indeed that they prevent major complications of hypertension, drugs such as spironolactone have not been included as options before stage 4 hypertension in most national guidelines, still less a replacement for thiazides as first-line treatment.11, 12, 15, 16

ng diuretics are not associated with glucose intolerance, or indeed that they prevent major complications of hypertension, drugs such as spironolactone have not been included as options before stage 4 hypertension in most national guidelines, still less a replacement for thiazides as first-line treatment.11, 12, 15, 16 In combination with a thiazide diuretic, however, amiloride has been used for many years, and in two studies13, 17 of morbidity and mortality, primary or secondary outcomes were significantly better with an amiloride–hydrochlorothiazide combination than with a calcium-channel blocker or a β blocker. But the low dose of amiloride in the widely available fixed-dose combination used in these studies is insufficient to prevent hypokalaemia. Furthermore, the overall dose of amiloride–hydrochlorothiazide needed to achieve the same reduction in blood pressure produced by the use of a calcium-channel blocker resulted in a 25% excess of new-onset diabetes in a previous trial.13

xed-dose combination used in these studies is insufficient to prevent hypokalaemia. Furthermore, the overall dose of amiloride–hydrochlorothiazide needed to achieve the same reduction in blood pressure produced by the use of a calcium-channel blocker resulted in a 25% excess of new-onset diabetes in a previous trial.13 In the Study of Trandolapril/verapamil SR [sustained release] and Insulin Resistance (STAR),18 200 obese patients were randomly assigned to receive either an angiotensin-converting enzyme (ACE) inhibitor plus a calcium-channel blocker or an angiotensin-receptor blocker (ARB) plus a low-dose thiazide diuretic. After 12 weeks, the ARB–diuretic group had significantly higher 2 h glucose concentrations on an oral glucose tolerance test (OGTT). Furthermore, pilot studies19, 20 showed that high doses of amiloride might be safely used in patients taking a RAS blocker, and have neutral or beneficial effects on glucose tolerance. These studies informed the duration and size of our three-way, randomised, parallel-group study, the British Hypertension Society's Prevention and Treatment of Hypertension with Algorithm-based Therapy (PATHWAY) 3 study, in which we compared the effects of hydrochlorothiazide with those of amiloride, either alone or in combination, on glucose tolerance and blood pressure.

ree-way, randomised, parallel-group study, the British Hypertension Society's Prevention and Treatment of Hypertension with Algorithm-based Therapy (PATHWAY) 3 study, in which we compared the effects of hydrochlorothiazide with those of amiloride, either alone or in combination, on glucose tolerance and blood pressure. Methods Study design and participants PATHWAY-3 was a 24 week, parallel-group, randomised, double-blind, phase 4 trial done by the British Hypertension Society Research Network of Investigators at 11 secondary care and two primary care centres in the UK. The study design and rationale have been published previously (appendix).21

udy design and participants PATHWAY-3 was a 24 week, parallel-group, randomised, double-blind, phase 4 trial done by the British Hypertension Society Research Network of Investigators at 11 secondary care and two primary care centres in the UK. The study design and rationale have been published previously (appendix).21 Eligible patients were aged 18–80 years and had clinic systolic blood pressure of 140 mmHg or higher, home systolic blood pressure of 130 mm Hg or higher, an indication for diuretic treatment, such as high systolic pressure despite treatment with an RAS blocker, and at least one component of the metabolic syndrome in addition to hypertension. Any permutation of ACE inhibitors, ARBs, β blockers, calcium-channel blockers, and direct renin inhibitors was permitted as background treatment, which could be changed at the patient's screening visit for inclusion, but not thereafter. Patients were permitted to take drugs for other disorders, with some specific exceptions (appendix). Diuretic treatment at screening was permitted if it could be discontinued during the 1 month run-in period and replaced with the randomly allocated diuretic. Patients who had not previously been treated for hypertension were eligible for inclusion provided that they were older than 55 years or black, or had a plasma renin concentration of less than 12 mU/L, or any combination thereof. Key exclusion criteria were diabetes diagnosed before enrolment, estimated glomerular filtration rate of less than 45 mL/min per 1·73m2, and plasma potassium concentrations outside normal ranges (appendix). A full list of inclusion and exclusion criteria is provided in the appendix. All patients gave informed written consent. The protocol was approved by Cambridge South Ethics Committee.

ated glomerular filtration rate of less than 45 mL/min per 1·73m2, and plasma potassium concentrations outside normal ranges (appendix). A full list of inclusion and exclusion criteria is provided in the appendix. All patients gave informed written consent. The protocol was approved by Cambridge South Ethics Committee. Randomisation and masking After a month's placebo run-in, during which patients were masked to treatment, enrolled patients were randomly assigned (1:1:1) to 10 mg amiloride hydrochloride force-titrated to 20 mg, 25 mg hydrochlorothiazide force-titrated to 50 mg, or a combination of 5 mg amiloride plus 12·5 mg hydrochlorothiazide force-titrated to 10/25 mg orally daily. Doses were doubled in a single step after 12 weeks of treatment. Treatments were allocated according to randomly permuted blocks of size six with each treatment group appearing twice in each block. Randomisation was achieved via a computer-generated list of pseudo-random numbers in the Robertson Centre for Biostatistics (University of Glasgow, UK), which was the data management centre for the study. Complete blocks of treatment packs were allocated to study sites, thus ensuring balanced allocation of treatments within sites and over time. Treatment allocation was done by study nurses, who accessed a web-based randomisation system on a server in the Robertson Centre for Biostatistics.

nagement centre for the study. Complete blocks of treatment packs were allocated to study sites, thus ensuring balanced allocation of treatments within sites and over time. Treatment allocation was done by study nurses, who accessed a web-based randomisation system on a server in the Robertson Centre for Biostatistics. All trial drugs were packed in identical containers by Alan Wong and colleagues at the Royal Free Hospital Pharmacy, London, UK, and labelled only by subject number and study phase. Investigators, laboratory staff, and patients were masked to the identity of drugs (although the individual drugs, amiloride and hydrochlorothiazide, had a different appearance), and to their sequence allocation. All other antihypertensive drugs already being taken at the time of randomisation were continued unchanged and open-label. Procedures After randomisation, patients took the initial doses of their assigned drug for 12 weeks (phase 1). The doses were then doubled, and treatment continued for a further 12 weeks (phase 2). An OGTT was done at baseline, 12 weeks, and 24 weeks. Blood was taken at 0, 30, 60, and 120 min after administration of a 75 g glucose drink; glucose was tested at all timepoints and insulin at 0 and 30 min.

g for 12 weeks (phase 1). The doses were then doubled, and treatment continued for a further 12 weeks (phase 2). An OGTT was done at baseline, 12 weeks, and 24 weeks. Blood was taken at 0, 30, 60, and 120 min after administration of a 75 g glucose drink; glucose was tested at all timepoints and insulin at 0 and 30 min. Other measurements were done at the same timepoints as the OGTTs. Seated home and clinic blood pressure were measured for each patient with their allocated, approved, automated blood pressure monitor (WatchBP Home, Microlife; Clearwater, FL, USA) for the duration of the trial. Patients took home blood pressure readings in the morning and the evening in triplicate on 4 consecutive days before each OGTT. Participants were instructed by specialist nurses in the use of the monitor, and training was reinforced at each clinic visit at which seated clinic blood pressure was measured—the mean of the last two measurements (three were taken). Before each OGTT, patients were also weighed, and had blood drawn for measurement of renin concentrations (analysed by the Diasorin Liaison automated chemiluminescent immunoassay for direct renin mass; Gerenzano, Italy22) and other biochemical parameters. Outcomes The primary outcome was the change from baseline in plasma glucose concentrations 2 h after oral administration of a 75 g glucose drink at 12 and 24 weeks.

Other measurements were done at the same timepoints as the OGTTs. Seated home and clinic blood pressure were measured for each patient with their allocated, approved, automated blood pressure monitor (WatchBP Home, Microlife; Clearwater, FL, USA) for the duration of the trial. Patients took home blood pressure readings in the morning and the evening in triplicate on 4 consecutive days before each OGTT. Participants were instructed by specialist nurses in the use of the monitor, and training was reinforced at each clinic visit at which seated clinic blood pressure was measured—the mean of the last two measurements (three were taken). Before each OGTT, patients were also weighed, and had blood drawn for measurement of renin concentrations (analysed by the Diasorin Liaison automated chemiluminescent immunoassay for direct renin mass; Gerenzano, Italy22) and other biochemical parameters. Outcomes The primary outcome was the change from baseline in plasma glucose concentrations 2 h after oral administration of a 75 g glucose drink at 12 and 24 weeks. The main secondary outcome was the change in home systolic blood pressure from baseline at at 12 and 24 weeks. Other secondary endpoints were changes in clinic systolic blood pressure, weight, electrolytes and calcium, uric acid, renin (as measure of natriuresis), HbA1c, insulin (at 0 and 30 min during OGTT), area under the curve for glucose during OGTT, and lipid profile. Development of diabetes was defined by fasting glucose concentrations ≥7 mmol/L or 2 h glucose concentrations ≥11·1 mmol/L or HbA1c ≥6·5% (47·5 mmol/mol).

calcium, uric acid, renin (as measure of natriuresis), HbA1c, insulin (at 0 and 30 min during OGTT), area under the curve for glucose during OGTT, and lipid profile. Development of diabetes was defined by fasting glucose concentrations ≥7 mmol/L or 2 h glucose concentrations ≥11·1 mmol/L or HbA1c ≥6·5% (47·5 mmol/mol). Adverse events were recorded in free text at each visit, and coded by the data management centre on the basis of the medical dictionary for regulatory activities. Serious adverse events were documented and reported to the chief investigator and regulatory authorities, in accordance with local and national requirements. Statistical analysis Based on at least 80% power to detect a mean difference in 2 h glucose between two treatment groups of 1 mmol/L (SD 2·2) with two-sample t tests with a 1% significance level, 414 patients were needed. 1 mmol/L was the observed difference in 2 h glucose concentrations in the largest previous trial of glucose intolerance caused by hydrochlorothiazide.18

tect a mean difference in 2 h glucose between two treatment groups of 1 mmol/L (SD 2·2) with two-sample t tests with a 1% significance level, 414 patients were needed. 1 mmol/L was the observed difference in 2 h glucose concentrations in the largest previous trial of glucose intolerance caused by hydrochlorothiazide.18 To test both the mechanistic hypothesis—ie, that a potassium-sparing diuretic (amiloride) is better than a potassium-losing diuretic (hydrochlorothiazide) in terms of effect on glucose tolerance—and that their combination is a practical alternative to hydrochlorothiazide monotherapy, the study had two hierarchical primary endpoints: we first compared 2 h glucose in the hydrochlorothiazide group with that of the amiloride group. If the difference between the groups was significant, we then compared 2 h glucose in the hydrochlorothiazide group with the combination amiloride–hydrochlorothiazide group. Hierarchical analysis prevents loss of power, because the prespecified first test needs to be positive before the next can be examined.

ide group. If the difference between the groups was significant, we then compared 2 h glucose in the hydrochlorothiazide group with the combination amiloride–hydrochlorothiazide group. Hierarchical analysis prevents loss of power, because the prespecified first test needs to be positive before the next can be examined. We tested for differences between groups for primary and secondary endpoints by using mixed-effect models to analyse continuous variables, with unstructured covariances for repeated measures within a patient, and adjustments for prespecified baseline covariates (sex, age, height, weight, smoking history, and the baseline value of the outcome being analysed). We estimated least-squares means for each treatment from these models, which are averaged from measurements at 12 and 24 weeks unless stated otherwise. We used similar models to assess baseline measurements that predicted response in 2 h glucose and home systolic blood pressure. We used logistic models to compare the proportion of patients who achieved target systolic blood pressure (defined as ≤140 mm Hg) at 24 weeks between the treatment groups and to compare the proportion of patients who developed diabetes by the end of the study between the three treatment groups. We used Fisher's exact test for comparisons of adverse events between groups.

patients who achieved target systolic blood pressure (defined as ≤140 mm Hg) at 24 weeks between the treatment groups and to compare the proportion of patients who developed diabetes by the end of the study between the three treatment groups. We used Fisher's exact test for comparisons of adverse events between groups. We used SAS (version 9.3) for our data analyses for the modified intention-to-treat population, which included all randomly assigned participants except for those with no primary outcome data from any follow-up visits. We included other participants for whom data were missing, and assumed that data were missing at random (ie, its absence was unrelated to the unobserved value). We did sensitivity analyses in the per-protocol population, which included participants who completed all follow-up visits with no major protocol deviation (adjudicated before breaking the study masking). The safety population, in whom the rate of adverse events and withdrawals was determined, was all randomised patients. There was no data monitoring board. This trial is registered with Clinicaltrials.gov, number NCT00797862 and the MHRA, Eudract number 2009-010068-41. Role of the funding source The funders of the study had no role in study design (other than through the peer-review process), the collection, analysis, or interpretation of the data, or the writing of the report. The corresponding author had full access to all the data and the final responsibility to submit for publication.

of the funding source The funders of the study had no role in study design (other than through the peer-review process), the collection, analysis, or interpretation of the data, or the writing of the report. The corresponding author had full access to all the data and the final responsibility to submit for publication. Results Between Nov 18, 2009, and Dec 15, 2014 we screened 663 patients and randomly assigned 441. 145 patients were assigned to the amiloride group, 146 to the hydrochlorothiazide group, and 150 to the amiloride–hydrochlorothiazide combination group (figure 1). The modified intention-to-treat analysis included 132 patients in the amiloride group, 134 patients in the hydrochlorothiazide group, and 133 patients in the amiloride plus hydrochlorothiazide group (figure 1; appendix). Table 1 shows baseline characteristics of the modified intention-to-treat population. The commonest indication for diuretic therapy was blood pressure uncontrolled by an ACE inhibitor or ARB, with more than 85% of participants in each group already taking these drugs. The commonest other component of the metabolic syndrome was central obesity, which was present in 99% of subjects.

n-to-treat population. The commonest indication for diuretic therapy was blood pressure uncontrolled by an ACE inhibitor or ARB, with more than 85% of participants in each group already taking these drugs. The commonest other component of the metabolic syndrome was central obesity, which was present in 99% of subjects. The mean change from baseline in plasma glucose concentration at the 2 h timepoint during OGTT was −0·55 mmol/L (95% CI −0·96 to −0·14; p=0·0093) in the amiloride group versus the hydrochlorothiazide group when 12 week and 24 week measurements were averaged (table 2). Because this difference was significant, we then examined the other hierarchical primary endpoint: mean change from baseline in 2 h plasma glucose concentration in the combination group was significantly lower than that in the thiazide group (−0·42 [–0·84 to −0·004; p=0·048).

week measurements were averaged (table 2). Because this difference was significant, we then examined the other hierarchical primary endpoint: mean change from baseline in 2 h plasma glucose concentration in the combination group was significantly lower than that in the thiazide group (−0·42 [–0·84 to −0·004; p=0·048). Differences in 2 h glucose concentrations between the hydrochlorothiazide group and the other groups increased with time and dose (figure 2), and similar differences were noted in the per-protocol population (appendix). The mean change in home systolic blood pressure averaged over 24 weeks was −12·9 mm Hg (95% CI −14·7 to —11·2) in the amiloride group, −12·2 mm Hg (−13·9 to −10·5) in the hydrochlorothiazide group, and −15·6 mm Hg (−17·3 to −13·8) in the combination group. The fall in home systolic blood pressure was 3·4 mm Hg (0·9 to 5·8) greater in the combination than in the hydrochlorothiazide group (figure 2B; p=0·0068), averaged over 24 weeks. Mean changes in clinic systolic blood pressure during 24 weeks did not differ significantly between the amiloride group (change from baseline −16·8 mm Hg [–18·8 to −14·8]) and the hydrochlorothiazide group (−16·5 mm Hg [–18·4 to −14·5]). In the combination group, systolic blood pressure fell by −20·4 mm Hg (−22·4 to −18·4), which was a significantly greater reduction compared with the hydrochlorothiazide group (between group difference −3·9 mm Hg [–6·7 to −1·1], p=0·0064). Blood pressures at each visit are shown in table 3. Blood pressure was more likely to be controlled by the combination than by hydrochlorothiazide monotherapy (odds ratio 1·77 [1·05 to 2·96], p=0·031; appendix). The mean rise in renin concentrations was 1·69 (1·36 to 2·01, p<0·0001) times higher at 12 and 24 weeks (average across the two timepoints) in the combination group than in the hydrochlorothiazide group (figure 3A). Significant differences were also noted in other markers of natriuresis, such as urea and creatinine concentrations, but not bodyweight (table 3).

1·69 (1·36 to 2·01, p<0·0001) times higher at 12 and 24 weeks (average across the two timepoints) in the combination group than in the hydrochlorothiazide group (figure 3A). Significant differences were also noted in other markers of natriuresis, such as urea and creatinine concentrations, but not bodyweight (table 3). Plasma potassium was unchanged at 24 weeks in the combination group, whereas a significant dose-dependent rise in concentration was noted in the amiloride group (by 0·63 mmol/L [0·56 to 0·70]; p<0·0001 at 24 weeks) and a significant fall was recorded in the hydrochlorothiazide group (by −0·27 mmol/L [–0·34 to −0·20]; p<0·0001 at 24 weeks; figure 3B). Plasma concentrations were significantly higher at 24 weeks in both the amiloride and combination groups than in the hydrochlorothiazide group (p<0·0001 for both when adjusted for baseline covariates; figure 3B).

orded in the hydrochlorothiazide group (by −0·27 mmol/L [–0·34 to −0·20]; p<0·0001 at 24 weeks; figure 3B). Plasma concentrations were significantly higher at 24 weeks in both the amiloride and combination groups than in the hydrochlorothiazide group (p<0·0001 for both when adjusted for baseline covariates; figure 3B). Uric acid concentrations rose in the hydrochlorothiazide group but were unchanged in the amiloride group (p<0·0001 for difference between groups; table 3, figure 3C); uric acid concentrations did not differ significantly between the hydrochlorothiazide and combination groups. Area under the curve during OGTT was significantly smaller in the amiloride group than in the hydrochlorothiazide group, but did not differ significantly between the combination and hydrochlorothiazide groups (table 3). We noted no significant differences between groups in fasting glucose, insulin (at 0 and 30 min), HbA1c, or homoeostatic model assessment indices of insulin resistance and secretion (table 3), except for a difference between the hydrochlorothiazide and combination groups for insulin at 30 min. However, we noted a mean numerical increase across all groups in HbA1c at both 12 weeks (0·086% [0·033–0·139], p=0·0006) and 24 weeks (0·126% [0·082–0·170], p<0·0001). 11 patients in the amiloride (11·8% adjusted for baseline covariates [95% CI 6·3–21·1]), nine (8·0% [3·7–16·2]) in the combination, and 13 (12·6% [6·5–23·0]) in the hydrochlorothiazide group developed diabetes during the study. Odds ratios for developing diabetes compared with the hydrochlorothiazide group were 0·65 (0·25–1·69) for the combination group and 1·07 (0·43–2·64) for the amiloride group.

21·1]), nine (8·0% [3·7–16·2]) in the combination, and 13 (12·6% [6·5–23·0]) in the hydrochlorothiazide group developed diabetes during the study. Odds ratios for developing diabetes compared with the hydrochlorothiazide group were 0·65 (0·25–1·69) for the combination group and 1·07 (0·43–2·64) for the amiloride group. Predictors of the glucose and blood pressure responses to study drugs are shown in the appendix. For change from baseline in 2 h glucose concentrations the main predictors were baseline fasting or 2 h glucose concentrations. For blood pressure, the main predictors were baseline blood pressure and plasma renin concentration. Blood glucose at the end of the study was weakly and inversely correlated with serum potassium concentrations (N=314, r2=0·02, p=0·010; no differences between treatments [p=0·60]; appendix). All drugs were well tolerated. 13 serious adverse events were reported. All adverse events are listed in the appendix. Only dizziness and muscle spasms were recorded in nine or more participants in each group; frequency of these events did not differ significantly between groups (table 4). Hyperkalaemia was reported in ten patients receiving amiloride alone or with hydrochlorothiazide (table 4). However the highest recorded potassium concentration was 5·8 mmol/L (in the amiloride group), and most values were between 5·1 mmol/L and 5·3 mmol/L (data not shown) in patients at a site where 5·0 was the upper limit of the normal laboratory range (appendix).

ving amiloride alone or with hydrochlorothiazide (table 4). However the highest recorded potassium concentration was 5·8 mmol/L (in the amiloride group), and most values were between 5·1 mmol/L and 5·3 mmol/L (data not shown) in patients at a site where 5·0 was the upper limit of the normal laboratory range (appendix). Discussion Diuretics have been used as the control drug in many large hypertension studies, but have rarely in the past 10 years been the main target of interest or studied in maximally efficacious doses.23, 24 Our study provides answers to several questions about diuretics that had not been previously investigated or resolved. We showed that, after 24 weeks of treatment, a potassium-sparing diuretic reduces blood pressure as efficaciously as high-dose thiazide without inducing adverse effects on blood glucose concentrations. Furthermore, a combination of half the conventional doses of amiloride and hydrochlorothiazide was not associated with increased 2 h glucose concentrations compared with hydrochlorothiazide treatment alone but produced significantly larger reductions in blood pressure than full doses of either diuretic given alone. Amiloride monotherapy did not cause clinically significant hyperkalaemia, and the amiloride–hydrochlorothiazide combination did not significantly affect potassium concentrations.

chlorothiazide treatment alone but produced significantly larger reductions in blood pressure than full doses of either diuretic given alone. Amiloride monotherapy did not cause clinically significant hyperkalaemia, and the amiloride–hydrochlorothiazide combination did not significantly affect potassium concentrations. Hitherto, the mechanism and prospects for prevention of thiazide-induced glucose intolerance were uncertain; the role of potassium in this problem was also unclear. A National Heart, Lung, and Blood Institute working party in 2008 identified potassium as “perhaps the most attractive variable” in developing a hypothesis for the mechanism of the thiazide response, and called for studies of potassium-sparing diuretics, among others.6 Amiloride has been licensed for hypertension for almost as long as hydrochlorothiazide, but has rarely been used or studied in doses that lower blood pressure as effectively as high-dose thiazides or other diuretic classes.25 That matched doses of thiazides and potassium-sparing diuretics, with similar efficacy on blood pressure, could neutralise the undesirable effects of each class while synergising to enhance reduction of blood pressure was an attractive hypothesis, but there were many unknowns, such as whether amiloride—in the context of blockade of the RAS in most patients—could be safely used at a dose large enough to match the blood pressure reduction of hydrochlorothiazide without causing hazardous electrolyte abnormalities.

ion of blood pressure was an attractive hypothesis, but there were many unknowns, such as whether amiloride—in the context of blockade of the RAS in most patients—could be safely used at a dose large enough to match the blood pressure reduction of hydrochlorothiazide without causing hazardous electrolyte abnormalities. Glucose concentration at 2 h in the OGTT is the best single measure for prediction of the long-term development of diabetes,26, 27, 28 and is also a strong predictor of cardiovascular morbidity.29 Sequential OGTTs offered the possibility of testing the hypothesis that prevention of potassium depletion would protect against thiazide-induced glucose intolerance. Although some of the difference in glucose profiles between the groups was due to a progressive increase in glucose intolerance in the hydrochlorothiazide group, glucose concentrations fell significantly in the amiloride group, at least when 12 week and 24 week data were compared with baseline. The importance of even minor degrees of glucose intolerance has been long evident from the Whitehall study,30 which showed that, during a period of 7·5 years, mortality from coronary heart disease doubled in participants with a 2 h glucose concentration greater than 5·3 mmol/L compared with patients with 2 h concentations of less than 5·3 mmol/L.

degrees of glucose intolerance has been long evident from the Whitehall study,30 which showed that, during a period of 7·5 years, mortality from coronary heart disease doubled in participants with a 2 h glucose concentration greater than 5·3 mmol/L compared with patients with 2 h concentations of less than 5·3 mmol/L. That amiloride and hydrochlorothiazide have opposite effects on glucose tolerance is consistent with potassium depletion being the cause of the rise in blood glucose concentration in patients taking thiazide diuretics. The poor correlation in our study between plasma glucose and potassium might be related to how poorly plasma electrolytes reflect overall electrolyte balance—plasma potassium concentrations are often normal in primary aldosteronism, for example.31 However, we included an intermediate, combination group in the study, both to confirm the role of potassium and to investigate a treatment that could be implemented in practice. The combination group necessitated selection of doses of each drug that, unlike the available fixed-dose combinations (in which the dose of amiloride is only a tenth that of hydrochlorothiazide), were predicted to neutralise changes in potassium. If the effects on blood glucose concentrations in the combination group were, as we noted with potassium concentrations, halfway between those in the two monotherapy groups, an even larger study than PATHWAY-3 would have been needed. But our additional predictions were that blood pressure would trump other effects on blood glucose, and that combining half doses of two diuretics with different targets in the kidney would have synergistic effects on sodium excretion, hence leading to reduction in blood pressure.

than PATHWAY-3 would have been needed. But our additional predictions were that blood pressure would trump other effects on blood glucose, and that combining half doses of two diuretics with different targets in the kidney would have synergistic effects on sodium excretion, hence leading to reduction in blood pressure. Our prediction of natriuretic synergism seems to be confirmed by the significantly greater reduction in home and clinic systolic blood pressure in the combination group than in the hydrochlorothiazide group. The blood pressure reduction in the combination group was associated with a near-doubling of the reactive rise in renin concentrations compared with those in the other two groups. This rise in renin concentrations is not clinically important in patients taking RAS blockers, but is a sensitive measure of natriuresis.32 Results of cross-sectional studies33, 34 have suggested the importance of blood pressure control in the prevention of glucose intolerance, which might be underestimated in comparisons of different antihypertensive drug classes. In studies of single antihypertensive drugs versus placebo, the reduced incidence of diabetes might be ascribed to the specific class effect rather than reductions in blood pressure reduction.35 The results of our pilot crossover studies showed that drugs that impair glucose tolerance when used alone (ie, β blockers and thiazide diuretics) have a neutral effect on glucose tolerance when combined with each other, and that the more favourable effects of the combination (compared with monotherapy) on glucose tolerance is associated with superior blood pressure reduction.19 Therefore the more efficacious blood pressure reduction in the amiloride–hydrochlorothiazide group compared with either drug alone is probably what underpins the success of the combination in avoiding induction of glucose intolerance.

therapy) on glucose tolerance is associated with superior blood pressure reduction.19 Therefore the more efficacious blood pressure reduction in the amiloride–hydrochlorothiazide group compared with either drug alone is probably what underpins the success of the combination in avoiding induction of glucose intolerance. The main limitation of our study is the short duration. 24 weeks was long enough to test the hypotheses that drugs with opposite effects on potassium concentrations would have opposite effects on 2 h glucose concentrations, and that diuretics with different sites of action would have synergistic effects on sodium loss, leading to blood pressure reduction, leading to glucose tolerance. However, not all patients who had a rise in 2 h glucose concentration in our study will progress to diabetes, and our results cannot show that longer-term usage of potassium-sparing diuretics will prevent an increase in incidence of diabetes as that which occurs with thiazide diuretics. Furthermore, 24 weeks' exposure is not sufficient to provide complete assurance of long-term safety when amiloride is added to an RAS blocker—intermittent monitoring of electrolytes and estimated glomerular filtration rates will be necessary in practice.

cidence of diabetes as that which occurs with thiazide diuretics. Furthermore, 24 weeks' exposure is not sufficient to provide complete assurance of long-term safety when amiloride is added to an RAS blocker—intermittent monitoring of electrolytes and estimated glomerular filtration rates will be necessary in practice. The short duration of the study also limits our ability to explain why potassium-sparing and potassium-losing diuretics have opposing effects on blood glucose. On several measures—insulin concentrations, HbA1c, and the calculated homoeostatic model assessments of insulin secretion and resistance—we found no differences between groups. Good control of blood pressure and the high prevalence of ACE inhibitors or ARBs in our study population probably mitigated the rise in 2 h blood glucose in the hydrochlorothiazide group, despite the high proportion of patients with central adiposity.33, 34, 35 But results in the amiloride group—in which there were seemingly contrary trends for HbA1c and 2 h glucose concentrations—show previously noted limitations of using HbA1c as a surrogate of glucose intolerance.36 Thus, the adverse effect of spironolactone on glucose tolerance, which was imputed from a similar small rise in HbA1c, could possibly be an artifact.37 The mechanism to explain why HbA1c does not accurately reflect glucose intolerance in this setting can only be speculated, but perhaps redistribution and reduction of blood supply by diuretics affects red cell disposal and hence the turnover of HbA1c. A final limitation is that we enrolled patients with an increased likelihood of developing glucose intolerance. The beneficial effects of amiloride might be less certain or applicable in thinner patients.

edistribution and reduction of blood supply by diuretics affects red cell disposal and hence the turnover of HbA1c. A final limitation is that we enrolled patients with an increased likelihood of developing glucose intolerance. The beneficial effects of amiloride might be less certain or applicable in thinner patients. Diuretics are the oldest among commonly used antihypertensive drugs, and their target (sodium retention) is one of the few universally agreed contributors to the pathogenesis of hypertension.1 Yet they have slipped in priority in some guidelines, partly because of trials in which suboptimum doses of diuretics were compared with optimum doses of other classes, and partly because of thiazide-induced diabetes. Even if thiazide-induced diabetes does not carry the same cardiovascular risks as spontaneous diabetes, diuretics cease to be cost effective when extra clinic visits and treatments for diabetes are factored into analyses.38 On the basis of our results from PATHWAY-3, we recommend that the combination of amiloride and hydrochlorothiazide, in doses equipotent for blood pressure reduction, becomes the first-choice diuretic in patients in whom adequate diuretic has not yet been prescribed. Our results suggest that this drug combination will confer the proven long-term benefits of hydrochlorothiazide without the possible downside of glucose intolerance. Low doses of amiloride with hydrochlorothiazide in the INSIGHT trial13 were as efficacious as nifedipine in the prevention of stroke and myocardial infarction, and significantly more efficacious in the prevention of heart failure. In the Medical Research Council's Elderly trial,17 the combination was significantly superior to atenolol in all cardiovascular endpoints.17 On the basis of our findings, the combination of hydrochlorothiazide with four times as much amiloride as was used in these studies can be predicted to increase its antihypertensive efficacy and counter the risks and costs of hypokalaemia and glucose intolerance. The efficacy of potassium-sparing diuretics revealed by PATHWAY-3, and the parallel PATHWAY-2 study of spironolactone in patients with resistant hypertension,39 warrants investigation of their long-term benefits in hypertension.

ertensive efficacy and counter the risks and costs of hypokalaemia and glucose intolerance. The efficacy of potassium-sparing diuretics revealed by PATHWAY-3, and the parallel PATHWAY-2 study of spironolactone in patients with resistant hypertension,39 warrants investigation of their long-term benefits in hypertension. Supplementary Material Supplementary appendix Acknowledgments The study was funded by a special project grant from the British Heart Foundation (number SP/08/002), and support from the National Institute of Health Research (NIHR) Comprehensive Local Research Networks. We are grateful to Sir Stephen O'Rahilly for discussion of the metabolic results. We thank Alan Wong and colleagues and the Royal Free Hospital Pharmacy for package and distribution of masked drugs. BW, PS, MJC, and MJB are NIHR senior investigators, and are supported by, respectively, the NIHR University College London/University College London Hospitals Biomedical Research Centre, the Biomedical Research Centre award to Imperial College Healthcare NHS Trust, the NIHR Cardiovascular Biomedical Research Unit at St Bartholomew's Hospital, London, and the NIHR Biomedical Research Centre award to Cambridge University Hopsitals NHS Trust.

University College London Hospitals Biomedical Research Centre, the Biomedical Research Centre award to Imperial College Healthcare NHS Trust, the NIHR Cardiovascular Biomedical Research Unit at St Bartholomew's Hospital, London, and the NIHR Biomedical Research Centre award to Cambridge University Hopsitals NHS Trust. Contributors MJB, BW, DJW, MJC, IF, GM, PS, and TMM designed the trial and were coapplicants to the British Heart Foundation for funding of the PATHWAY programme, which was led by MJB. MJB, BW, DJW, MJC, JKC, IF, GM, PS, ISM, SP, and TMM were the trial's steering committee. MJB and JS drafted the protocol, with help from BW and TMM. JS was trial coordinator. SVM and IF advised on statistical analysis. MJB, BW, and TMM were the's trial executive committee. MJB, BW, SVM, and TMM wrote the manuscript, with assistance from DJW. All authors approved the final draft.

trial's steering committee. MJB and JS drafted the protocol, with help from BW and TMM. JS was trial coordinator. SVM and IF advised on statistical analysis. MJB, BW, and TMM were the's trial executive committee. MJB, BW, SVM, and TMM wrote the manuscript, with assistance from DJW. All authors approved the final draft. Declaration of interests MJB has received honoraria from Novartis. BW has received honoraria for lectures on hypertension from Novartis, Boehringer Ingelheim, Servier, Daiichi Sankyo and Pfizer. DJW has received funding for membership of Independent Data Monitoring Committees for Abbvie in relation to clinical trials in diabetic nephropathy and is president-elect of the British Pharmacological Society and a board member of Medicines and Healthcare Products Regulatory Agency. MJC received honoraria from Medtronic and Quantum Genomics during this trial that were unrelated to the work. JKC is vice-president of the Artery Society. GM has received honoraria from Novartis. TMM is chief investigator on two investigator-initiated, industry-funded, university-sponsored cardiovascular outcome studies (funded by Pfizer and Menarini, IPSEN, and Teijin), neither of which focuses on blood pressure. He has provided consultancy or received honoraria for speaking from Novartis, Takeda, Daiichi Sankyo, Shire, and Astellus, and is president of the British Hypertension Society. His research unit also does studies funded by Novartis and Amgen. ISM has received research grants from Amgen, Menarini, and Novartis, and personal fees from MSD, unrelated to the present study. SVM, IF, JS, SP, and PS declare no competing interests.

, and Astellus, and is president of the British Hypertension Society. His research unit also does studies funded by Novartis and Amgen. ISM has received research grants from Amgen, Menarini, and Novartis, and personal fees from MSD, unrelated to the present study. SVM, IF, JS, SP, and PS declare no competing interests. Figure 1 Trial profile Patients who underwent at least one follow-up observation for the primary endpoint after randomisation were included in the modified intention-to-treat analyses. The appendix lists reasons for non-randomisation, dropout between randomisation and modified intention-to-treat populations, and “other” reasons for discontinuation of randomly assigned treatment. Reasons for exclusion from the per-protocol cohort are not mutually exclusive. *Some discontinuations were also protocol deviations, so the differences between modified intention-to-treat and per-protocol populations are less than sum of protocol deviations and discontinuations. Figure 2 Changes in 2 h blood glucose concentrations (A), home systolic blood pressure (B), and clinic systolic blood pressure (C)

Patients who underwent at least one follow-up observation for the primary endpoint after randomisation were included in the modified intention-to-treat analyses. The appendix lists reasons for non-randomisation, dropout between randomisation and modified intention-to-treat populations, and “other” reasons for discontinuation of randomly assigned treatment. Reasons for exclusion from the per-protocol cohort are not mutually exclusive. *Some discontinuations were also protocol deviations, so the differences between modified intention-to-treat and per-protocol populations are less than sum of protocol deviations and discontinuations. Figure 2 Changes in 2 h blood glucose concentrations (A), home systolic blood pressure (B), and clinic systolic blood pressure (C) Data are adjusted means; error bars show 95% CIs. For (A), p=0·0026 for the comparison between the amiloride and hydrochlorothiazide groups and 0·039 for comparisons between the combination and hydrochlorothiazide groups at 24 weeks, in a model adjusting for baseline covariates. For (B), averaged across 12 weeks and 24 weeks, the fall in home blood pressure was significantly greater in the combination group than in the hydrochlorothiazide group (p=0·0068). For (C), averaged across 12 weeks and 24 weeks, the fall in clinic blood pressure was significantly greater in the combination group than in the hydrochlorothiazide group (p=0·0064). Figure 3 Changes in plasma renin (A), serum potassium (B), and serum uric acid (C) concentrations

Data are adjusted means; error bars show 95% CIs. For (A), p=0·0026 for the comparison between the amiloride and hydrochlorothiazide groups and 0·039 for comparisons between the combination and hydrochlorothiazide groups at 24 weeks, in a model adjusting for baseline covariates. For (B), averaged across 12 weeks and 24 weeks, the fall in home blood pressure was significantly greater in the combination group than in the hydrochlorothiazide group (p=0·0068). For (C), averaged across 12 weeks and 24 weeks, the fall in clinic blood pressure was significantly greater in the combination group than in the hydrochlorothiazide group (p=0·0064). Figure 3 Changes in plasma renin (A), serum potassium (B), and serum uric acid (C) concentrations Data for (A) are log-transformed changes from baseline; data for (B) and (C) are adjusted means. Error bars show 95% CIs. For (A), p=0·0002 for the comparison between the combination and hydrochlorothiazide groups at 24 weeks, from a model adjusting for baseline covariates. For (B), p<0·0001 for the comparison between the amiloride and the hydrochlorothiazide groups and between the combination and hydrochlorothiazide groups at 24 weeks, from a model adjusting for baseline covariates. For (C), p<0·0001 for the comparison between the amiloride and hydrochlorothiazide groups at 24 weeks, from a model adjusting for baseline covariates. Table 1 Baseline characteristics in the modified intention-to-treat population

Data for (A) are log-transformed changes from baseline; data for (B) and (C) are adjusted means. Error bars show 95% CIs. For (A), p=0·0002 for the comparison between the combination and hydrochlorothiazide groups at 24 weeks, from a model adjusting for baseline covariates. For (B), p<0·0001 for the comparison between the amiloride and the hydrochlorothiazide groups and between the combination and hydrochlorothiazide groups at 24 weeks, from a model adjusting for baseline covariates. For (C), p<0·0001 for the comparison between the amiloride and hydrochlorothiazide groups at 24 weeks, from a model adjusting for baseline covariates. Table 1 Baseline characteristics in the modified intention-to-treat population Amiloride (n=132) Amiloride plus hydrochlorothiazide (n=133) Hydrochlorothiazide (n=134) Age (years) 62·1 (10·4) 61·5 (10·2) 62·8 (9·9) Female sex 52 (39·4%) 63 (47·4%) 47 (35·1%) Weight (kg) 89·3 (16·7) 88·8 (16·7) 88·2 (17·1) BMI (kg/m2) 31·4 (7·6) 31·0 (4·7) 30·6 (5·1) Number of current smokers 10 (7·6%) 12 (9·0%) 15 (11·2%) Blood pressure (mm Hg) Clinic systolic 153·8 (11·4) 156·2 (12·4) 154·4 (11·7) Clinic diastolic 91·3 (9·7) 91·2 (9·4) 90·0 (10·2) Home systolic 149·3 (12·4) 150·6 (11·4) 148·8 (10·9) Home diastolic 86·9 (9·8) 86·6 (8·9) 85·1 (9·6) Previously untreated 7 (5·3%) 12 (9·0%) 11 (8·2%) Receiving ACE inhibitor or ARB 119 (90·2%) 115 (86·5%) 117(87·3%) Receiving β blocker 18 (13·6%) 24 (18·0%) 23 (17·2%) Receiving calcium-channel blocker 56 (42·4%) 57 (42·9%) 56 (41·8%) Number of drugs (if treated) 1·5 (0·7) 1·5 (0·7) 1·6 (0·7) Central obesity* 129 (97·7%) 133 (100·0%) 132 (98·5%) Serum potassium (mmol/L) 4·1 (0·4) 4·2 (0·3) 4·2 (0·4) 2 hour glucose during OGTT (mmol/L) 7·2 (2·3) 7·2 (2·1) 6·9 (2·4) Impaired glucose tolerance† 44 (33·3%) 45 (33·8%) 42 (31·3%) Data are mean (SD) or n (%). ACE=angiotensin-converting enzyme. ARB=angiotensin-receptor blocker. OGTT=oral glucose tolerance test.

98·5%) Serum potassium (mmol/L) 4·1 (0·4) 4·2 (0·3) 4·2 (0·4) 2 hour glucose during OGTT (mmol/L) 7·2 (2·3) 7·2 (2·1) 6·9 (2·4) Impaired glucose tolerance† 44 (33·3%) 45 (33·8%) 42 (31·3%) Data are mean (SD) or n (%). ACE=angiotensin-converting enzyme. ARB=angiotensin-receptor blocker. OGTT=oral glucose tolerance test. * Central obesity is defined as a waist circumference of greater than 94 cm in men and greater than 80 cm in women. † Impaired glucose tolerance is defined as a 2 h glucose concentration of greater than 7·8 mmol/L. Table 2 Changes from baseline in 2 h glucose concentrations during oral glucose tolerance tests in the modified intention-to-treat population Amiloride (n=132) Amiloride plus hydrochlorothiazide (n=133) Hydrochlorothiazide (n=134) Mean change from baseline (mmol/L) −0·35 (−0·69 to −0·01) −0·22 (−0·56 to 0·11) 0·20 (−0·12 to 0·51) Difference from hydrochlorothiazide (mmol/L) −0·55 (−0·96 to −0·14; p=0·0093) −0·42 (−0·84 to 0·00; p=0·048) .. Data in parentheses are 95% CIs. Mean change from baseline was calculated on the basis of data at 12 weeks (low-dose treatment) and 24 weeks (high-dose treatment). Least squares estimates were adjusted for prespecified baseline covariates in a mixed-effects model; p values are for comparisons with hydroclorothiazide. Table 3 Changes in blood pressure and biochemical parameters in the modified intention-to-treat population

Amiloride (n=132) Amiloride plus hydrochlorothiazide (n=133) Hydrochlorothiazide (n=134) Mean change from baseline (mmol/L) −0·35 (−0·69 to −0·01) −0·22 (−0·56 to 0·11) 0·20 (−0·12 to 0·51) Difference from hydrochlorothiazide (mmol/L) −0·55 (−0·96 to −0·14; p=0·0093) −0·42 (−0·84 to 0·00; p=0·048) .. Data in parentheses are 95% CIs. Mean change from baseline was calculated on the basis of data at 12 weeks (low-dose treatment) and 24 weeks (high-dose treatment). Least squares estimates were adjusted for prespecified baseline covariates in a mixed-effects model; p values are for comparisons with hydroclorothiazide. Table 3 Changes in blood pressure and biochemical parameters in the modified intention-to-treat population Unadjusted treatment means (95% CI) Adjusted treatment differences (95% CI; p) Amiloride 10–20 mg Combination amiloride/hydrochlorothiazide 5 mg/12·5 mg–10 mg/25 mg Hydrochlorothiazide 25–50 mg Amiloride vs hydrochlorothiazide Combination vs hydrochlorothiazide Home systolic blood pressure (mm Hg) Baseline 149·3 (147·2 to 151·5) 150·6 (148·6 to 152·6) 148·8 (146·9 to 150·6) .. .. 12 weeks 138·3 (136·0 to 140·5) 136·1 (133·8 to 138·3) 138·5 (136·6 to 140·5) −0·5 (−3·2 to 2·1; p=0·70) −3·2 (−5·8 to −0·5; p=0·020) 24 weeks 134·4 (132·3 to 136·5) 132·3 (130·0 to 134·6) 135·0 (133·0 to 137·0) −1·0 (−3·7 to 1·7; p=0·49) −3·5 (−6·2 to −0·8; p=0·011) Clinic systolic blood pressure (mm Hg) Baseline 153·8 (151·9 to 155·8) 156·2 (154·0 to 158·3) 154·4 (152·4 to 156·4) .. .. 12 weeks 140·8 (138·5 to 143·1) 136·7 (134·2 to 139·3) 140·3 (137·7 to 142·9) 0·3 (−2·9 to 3·5; p=0·84) −3·9 (−7·1 to −0·7; p=0·016) 24 weeks 135·4 (132·9 to 137·9) 133·4 (131·0 to 135·8) 135·8 (133·3 to 138·2) −1·0 (−4·3 to 2·3; p=0·55) −4·0 (−7·2 to −0·7; p=0·018) Bodyweight (kg) Baseline 89·3 (86·4 to 92·1) 88·8 (85·9 to 91·7) 88·2 (85·3 to 91·1) .. .. 12 weeks 90·5 (87·6 to 93·5) 89·1 (86·2 to 92·1) 88·9 (85·9 to 91·8) 0·2 (−0·5 to 0·8; p=0·60) −0·2 (−0·8 to 0·5; p=0·64) 24 weeks 89·1 (85·9 to 92·2) 89·8 (86·4 to 93·2) 87·2 (84·3 to 90·0) −0·2 (−0·8 to 0·5; p=0·63) −0·1 (−0·8 to 0·6; p=0·82) Renin (mU/L, log base 10) Baseline 1·20 (1·09 to 1·31) 1·14 (1·03 to 1·24) 1·20 (1·11 to 1·30) .. .. 12 weeks 1·67 (1·53 to 1·80) 1·75 (1·61 to 1·89) 1·62 (1·50 to 1·74) 0·03 (−0·10 to 0·16; p=0·63) 0·21 (0·08 to 0·34; p=0·0018) 24 weeks 1·83 (1·70 to 1·95) 1·95 (1·82 to 2·09) 1·74 (1·62 to 1·86) 0·06 (−0·07 to 0·20; p=0·34) 0·25 0·12 to 0·39; p=0·0002) Sodium (mmol/L) Baseline 139·7 (139·3 to 140·2) 139·9 (139·5 to 140·4) 139·7 (139·2 to 140·2) .. 12 weeks 138·2 (137·8 to 138·6) 138·1 (137·7 to 138·5) 138·9 (138·5 to 139·3) −0·7 (−1·3 to −0·1; p=0·03) −1·1 (−1·7 to −0·4; p=0·0010) 24 weeks 138·1 (137·7 to 138·5) 138·0 (137·5 to 138·4) 138·6 (138·1 to 139·0) −0·6 (−1·2 to 0·1; p=0·097) −0·9 (−1·6 to −0·2; p=0·0075) Potassium (mmol/L) Baseline 4·09 (4·01 to 4·16) 4·16 (4·10 to 4·22) 4·21 (4·14 to 4·28) .. ..

to 138·5) 138·9 (138·5 to 139·3) −0·7 (−1·3 to −0·1; p=0·03) −1·1 (−1·7 to −0·4; p=0·0010) 24 weeks 138·1 (137·7 to 138·5) 138·0 (137·5 to 138·4) 138·6 (138·1 to 139·0) −0·6 (−1·2 to 0·1; p=0·097) −0·9 (−1·6 to −0·2; p=0·0075) Potassium (mmol/L) Baseline 4·09 (4·01 to 4·16) 4·16 (4·10 to 4·22) 4·21 (4·14 to 4·28) .. .. 12 weeks 4·55 (4·50 to 4·61) 4·31 (4·26 to 4·36) 3·97 (3·92 to 4·03) 0·68 (0·60 to 0·76; p<0·0001) 0·42 (0·34 to 0·50; p<0·0001) 24 weeks 4·61 (4·56 to 4·66) 4·17 (4·12 to 4·23) 3·79 (3·74 to 3·84) 0·90 (0·81 to 0·98; p<0·0001) 0·48 (0·39 to 0·56; p<0·0001) Urea (mmol/L) Baseline 5·36 (5·12 to 5·60) 5·15 (4·92 to 5·38) 5·40 (5·15 to 5·65) .. .. 12 weeks 5·90 (5·68 to 6·12) 6·08 (5·85 to 6·32) 6·03 (5·82 to 6·25) −0·02 (−0·31 to 0·28; p=0·91) 0·34 (0·05 to 0·64; p=0·024) 24 weeks 5·82 (5·59 to 6·06) 6·21 (5·98 to 6·44) 6·13 (5·87 to 6·39) −0·17 (−0·48 to 0·13; p=0·26) 0·45 (0·14 to 0·76; p=0·0042) Creatinine (μmol/L) Baseline 77·4 (74·4 to 80·4) 76·6 (73·3 to 79·9) 76·7 (73·8 to 79·6) .. .. 12 weeks 79·8 (77·5 to 82·2) 80·1 (77·6 to 82·7) 79·3 (76·9 to 81·6) 1·1 (−1·5 to 3·7; p=0·40) 2·3 (−0·4 to 4·9; p=0·089) 24 weeks 79·7 (77·2 to 82·3) 80·3 (77·7 to 83·0) 78·1 (75·6 to 80·6) 2·9 (0·2 to 5·6; p=0·03 3·8 (1·1 to 6·5; p=0·0057) Uric acid (μmol/L) Baseline 354 (340 to 369) 349 (332 to 365) 342 (324 to 359) .. .. 12 weeks 355 (340 to 371) 365 (347 to 383) 375 (358 to 392) −32 (−49 to −15; p=0·0003) −7 (−23 to 10; p=0·43) 24 weeks 351 (335 to 368) 380 (362 to 397) 392 (374 to 410) −50 (−67 to −33; p<0·0001) −3 (−20 to 14; p=0·76) Total cholesterol (mmol/L) Baseline 1·40 (1·27 to 1·52) 1·53 (1·37 to 1·69) 1·34 (1·23 to 1·46) .. .. 12 weeks 1·51 (1·34 to 1·68) 1·59 (1·42 to 1·76) 1·44 (1·30 to 1·58) 0·11 (−0·06 to 0·27; p=0·21) 0·10 (−0·07 to 0·26; p=0·25) 24 weeks 1·59 (1·40 to 1·77) 1·49 (1·35 to 1·63) 1·54 (1·39 to 1·69) 0·01 (−0·16 to 0·18; p=0·93) −0·16 (−0·33 to 0·01; p=0·057) Triglycerides (mmol/L) Baseline 1·40 (1·27 to 1·52) 1·53 (1·37 to 1·69) 1·34 (1·23 to 1·46) .. .. 12 weeks 1·51 (1·34 to 1·68) 1·59 (1·42 to 1·76) 1·44 (1·30 to 1·58) 0·11 (−0·06 to 0·27; p=0·21) 0·10 (−0·07 to 0·26; p=0·25) 24 weeks 1·59 (1·40 to 1·77) 1·49 (1·35 to 1·63) 1·54 (1·39 to 1·69) 0·01 (−0·16 to 0·18; p=0·93) −0·16 (−0·33 to 0·01; p=0·057) LDL cholesterol (mmol/L) Baseline 2·88 (2·70 to 3·05) 2·94 (2·78 to 3·10) 2·95 (2·79 to 3·11) .. ..

1·42 to 1·76) 1·44 (1·30 to 1·58) 0·11 (−0·06 to 0·27; p=0·21) 0·10 (−0·07 to 0·26; p=0·25) 24 weeks 1·59 (1·40 to 1·77) 1·49 (1·35 to 1·63) 1·54 (1·39 to 1·69) 0·01 (−0·16 to 0·18; p=0·93) −0·16 (−0·33 to 0·01; p=0·057) LDL cholesterol (mmol/L) Baseline 2·88 (2·70 to 3·05) 2·94 (2·78 to 3·10) 2·95 (2·79 to 3·11) .. .. 12 weeks 2·87 (2·71 to 3·03) 2·92 (2·75 to 3·09) 2·87 (2·72 to 3·03) −0·00 (−0·16 to 0·16; p=0·96) 0·07 (−0·09 to 0·23; p=0·3605) 24 weeks 3·05 (2·86 to 3·24) 2·95 (2·77 to 3·13) 2·96 (2·80 to 3·12) 0·10 (−0·06 to 0·26; p=0·22) 0·04 (−0·12 to 0·21; p=0·5849) HDL cholesterol (mmol/L) Baseline 1·27 (1·17 to 1·36) 1·33 (1·23 to 1·43) 1·35 (1·24 to 1·45) .. 12 weeks 1·17 (1·05 to 1·28) 1·34 (1·24 to 1·44) 1·35 (1·24 to 1·46) −0·18 (−0·34 to −0·03; p=0·018) −0·00 (−0·15 to 0·15; p=1·00) 24 weeks 1·20 (1·08 to 1·32) 1·32 (1·20 to 1·44) 1·35 (1·25 to 1·45) −0·13 (−0·28 to 0·02; p=0·09) −0·02 (−0·17 to 0·14; p=0·85) Fasting glucose (mmol/L) Baseline 5·20 (5·07 to 5·32) 5·13 (5·04 to 5·23) 5·30 (5·13 to 5·46) .. .. 12 weeks 5·30 (5·13 to 5·47) 5·26 (5·15 to 5·38) 5·38 (5·25 to 5·52) −0·00 (−0·16 to 0·16; p=0·98) −0·01 (−0·17 to 0·14; p=0·86) 24 weeks 5·23 (5·08 to 5·37) 5·27 (5·15 to 5·39) 5·39 (5·24 to 5·54) −0·08 (−0·23 to 0·08; p=0·36) −0·04 (−0·20 to 0·12; p=0·61) 2 h glucose (mmol/L) Baseline 7·35 (6·87 to 7·82) 7·21 (6·84 to 7·59) 7·03 (6·62 to 7·44) .. .. 12 weeks 6·97 (6·59 to 7·35) 7·03 (6·66 to 7·41) 7·33 (6·98 to 7·69) −0·36 (−0·84 to 0·11; p=0·13) −0·30 (−0·78 to 0·18; p=0·21) 24 weeks 6·73 (6·35 to 7·10) 6·92 (6·55 to 7·30) 7·46 (7·11 to 7·81) −0·74 (−1·21 to −0·26; p=0·0023) −0·54 (−1·01 to −0·06; p=0·026) Fasting insulin (pmol/L) Baseline 80·1 (65·1 to 95·0) 86·2 (60·8 to 111·5) 95·6 (63·7 to 127·5) .. .. 12 weeks 85·1 (74·4 to 95·7) 121·1 (81·4 to 160·7) 101·9 (81·1 to 122·6) −17·0 (−46·1 to 12·0; p=0·25) −0·9 (−29·6 to 27·7; p=0·95) 24 weeks 88·7 (76·3 to 101·1) 107·7 (78·8 to 136·6) 102·1 (75·6 to 128·6) −17·7 (−47·5 to 12·1; p=0·24) 4·9 (−24·6 to 34·5; p=0·74) 30 min insulin (pmol/L) Baseline 487 (368 to 605) 427 (379 to 474) 445 (378 to 512) .. .. 12 weeks 547 (448 to 645) 616 (484 to 748) 499 (431 to 568) 14 (−72 to 99; p=0·75) 68 (−14 to 150; p=0·10) 24 weeks 584 (488 to 680) 531 (437 to 626) 465 (397 to 533) 51 (−34 to 136; p=0·24) 88 (2 to 174; p=0·04) Homoeostatic model assessment—insulin resistance Baseline 2·70 (2·18 to 3·23) 2·93 (2·00 to 3·86) 3·29 (2·20 to 4·38) .. ..

8 to 645) 616 (484 to 748) 499 (431 to 568) 14 (−72 to 99; p=0·75) 68 (−14 to 150; p=0·10) 24 weeks 584 (488 to 680) 531 (437 to 626) 465 (397 to 533) 51 (−34 to 136; p=0·24) 88 (2 to 174; p=0·04) Homoeostatic model assessment—insulin resistance Baseline 2·70 (2·18 to 3·23) 2·93 (2·00 to 3·86) 3·29 (2·20 to 4·38) .. .. 12 weeks 2·90 (2·49 to 3·31) 4·34 (2·80 to 5·87) 3·57 (2·85 to 4·29) −0·65 (−1·84 to 0·55; p=0·29) 0·02 (−1·14 to 1·19; p=0·97) 24 weeks 3·01 (2·57 to 3·44) 3·88 (2·54 to 5·22) 3·57 (2·60 to 4·53) −0·84 (−2·05 to 0·37; p=0·17) 0·17 (−1·03 to 1·37; p=0·78) Homoeostatic model assessment—β cells Baseline 122·1 (98·6 to 145·7) 134·2 (103·1 to 165·2) 138·3 (93·1 to 183·4) .. .. 12 weeks 55·6 (43·8 to 67·4) 77·9 (54·0 to 101·8) 62·8 (47·7 to 77·9) −9·8 (−33·8 to 14·3; p=0·42) 13·8 (−10·2 to 37·8; p=0·26) 24 weeks 63·0 (47·1 to 79·0) 71·0 (54·4 to 87·5) 71·4 (51·0 to 91·8) −11·6 (−36·8 to 13·5; p=0·36) −2·9 (−28·0 to 22·3; p=0·82) HbA1c(%) Baseline 5·73 (5·63 to 5·83) 5·63 (5·55 to 5·70) 5·65 (5·57 to 5·74) .. .. 12 weeks 5·81 (5·70 to 5·91) 5·70 (5·62 to 5·78) 5·68 (5·54 to 5·82) 0·05 (−0·07 to 0·18; p=0·39) 0·05 (−0·07 to 0·17; p=0·38) 24 weeks 5·76 (5·64 to 5·88) 5·70 (5·60 to 5·81) 5·75 (5·65 to 5·85) −0·04 (−0·16 to 0·09; p=0·56) 0·00 (−0·12 to 0·13; p=0·95) Calcium (mmol/L) Baseline 2·34 (2·32 to 2·35) 2·33 (2·32 to 2·35) 2·32 (2·30 to 2·33) .. .. 12 weeks 2·30 (2·28 to 2·32) 2·31 (2·29 to 2·33) 2·29 (2·27 to 2·31) −0·01 (−0·04 to 0·02; p=0·62) 0·02 (−0·01 to 0·05; p=0·22) 24 weeks 2·31 (2·28 to 2·34) 2·32 (2·29 to 2·34) 2·31 (2·28 to 2·34) −0·01 (−0·04 to 0·02; p=0·46) 0·01 (−0·02 to 0·04; p=0·46) Change in area under curve during oral glucose tolerance test Baseline 987 (947 to 1027) 962 (931 to 992) 975 (941 to 1010) .. .. 12 weeks 951 (908 to 994) 965 (925 to 1005) 1005 (965 to 1046) −56 (−98 to −14; p=0·0087) −29 (−70 to 12; p=0·17) 24 weeks 930 (891 to 968) 975 (932 to 1018) 990 (947 to 1032) −60 (−103 to −18; p=0·0052) −20 (−63 to 23; p=0·37) Comparisons are adjusted for baseline values. The slight differences for 2 h glucose at week 24 between this table and table 2 arise because missing values were imputed in the calculations used for this table whereas they were omitted in table 2. To convert HbA1c to mmol/mol, multiply the percentages by 10·93 and subtract 23·5. To convert renin to pmol/L, multiply concentration in mU/L by 1·56.

glucose at week 24 between this table and table 2 arise because missing values were imputed in the calculations used for this table whereas they were omitted in table 2. To convert HbA1c to mmol/mol, multiply the percentages by 10·93 and subtract 23·5. To convert renin to pmol/L, multiply concentration in mU/L by 1·56. Table 4 Adverse events and withdrawals in the modified intention-to-treat population Amiloride (n=145) Amiloride plus hydrochlorothiazide (n=150) Hydrochlorothiazide (n=146) Amiloridevshydrochlorothizaide Combinationvshydrochlorothiazide Withdrawals 17 (11·7%) 16 (10·7%) 10 (6·8%) 0·16 0·31 Serious adverse events 7 (4·8%) 4 (2·7%) 2 (1·4%) 0·10 0·68 Any adverse event 97 (66·9%) 92 (61·3%) 95 (65·1%) 0·80 0·55 Dizziness 9 (6·2%) 15 (10·0%) 16 (11·0%) 0·21 0·85 Muscle spasms 12 (8·3%) 14 (9·3%) 10 (6·8%) 0·66 0·52 Hyperkalaemia 7 (4·8%) 3 (2·0%) 0 0·0071 0·25 Data are n (%) unless otherwise specified. Fisher's exact test was used to calculate p values.

Introduction Cardiovascular disease guidelines recommend strategies that predict and prevent composite endpoints for coronary heart disease and stroke.1, 2, 3, 4 A rationale for this combined approach is to enhance efficiency of cardiovascular disease screening by capitalising on shared risk factors and preventive interventions, even though coronary heart disease and stroke are aetiologically distinct. Such a rationale could be extended to heart failure. The age-specific incidence of heart failure is increasing; it is a common initial presentation of cardiovascular disease.5 Furthermore, statins and antihypertensive treatments might, in addition to their benefits for primary prevention of coronary heart disease and stroke, be effective at reducing the risk of new-onset heart failure.6, 7 Practical advantages of a strategy that integrates heart failure prediction into cardiovascular disease risk assessment could exist since coronary heart disease and stroke risk assessment is already widespread, whereas primary prevention of heart failure is not addressed by current guidelines.8, 9

et heart failure.6, 7 Practical advantages of a strategy that integrates heart failure prediction into cardiovascular disease risk assessment could exist since coronary heart disease and stroke risk assessment is already widespread, whereas primary prevention of heart failure is not addressed by current guidelines.8, 9 One approach that could enable such an integrated strategy is measurement of soluble natriuretic peptides. These molecules play important roles in regulation of blood pressure, blood volume, and sodium balance.10 Assessment of circulating B-type natriuretic peptide concentration and its more stable by-product N-terminal-pro-B-type natriuretic peptide (NT-proBNP) is recommended by guidelines for diagnosis and management of patients with heart failure.8, 9 As natriuretic peptides are markers of vascular remodelling, their measurement could also serve as an adjunct in prediction of first-ever coronary heart disease and stroke outcomes.11 However, to what extent assessment of natriuretic peptides can predict first-onset heart failure outcomes or improve prediction of coronary heart disease and stroke in people without known cardiovascular disease is uncertain.12, 13, 14, 15, 16 To address these questions, we established the Natriuretic Peptides Studies Collaboration, an international consortium of individual-participant data from individuals without a history of cardiovascular disease at baseline. Research in context Evidence before this study

One approach that could enable such an integrated strategy is measurement of soluble natriuretic peptides. These molecules play important roles in regulation of blood pressure, blood volume, and sodium balance.10 Assessment of circulating B-type natriuretic peptide concentration and its more stable by-product N-terminal-pro-B-type natriuretic peptide (NT-proBNP) is recommended by guidelines for diagnosis and management of patients with heart failure.8, 9 As natriuretic peptides are markers of vascular remodelling, their measurement could also serve as an adjunct in prediction of first-ever coronary heart disease and stroke outcomes.11 However, to what extent assessment of natriuretic peptides can predict first-onset heart failure outcomes or improve prediction of coronary heart disease and stroke in people without known cardiovascular disease is uncertain.12, 13, 14, 15, 16 To address these questions, we established the Natriuretic Peptides Studies Collaboration, an international consortium of individual-participant data from individuals without a history of cardiovascular disease at baseline. Research in context Evidence before this study We hypothesised that integrated cardiovascular disease risk assessment strategies could be extended to primary prevention of heart failure through measurement of N-terminal-pro-B-type natriuretic peptide (NT-proBNP) concentration. In a systematic review of the published literature (searches of PubMed, Scientific Citation Index Expanded, and Embase for relevant articles published up to Sept 4, 2014, using search terms related to natriuretic peptide family members and the primary outcomes, with no language restrictions), we identified 33 relevant prospective studies of natriuretic peptides and incident coronary heart disease, stroke, or heart failure outcomes. We attempted a synthesis of these results in a previous literature-based review, but we found that using published results was insufficiently powered or detailed or both to enable reliable assessment of whether or not NT-proBNP concentration measurement could augment cardiovascular disease risk assessment for coronary heart disease and stroke, and investigators of only few population-based prospective studies reported on associations between NT-proBNP concentration and first-onset heart failure.

reliable assessment of whether or not NT-proBNP concentration measurement could augment cardiovascular disease risk assessment for coronary heart disease and stroke, and investigators of only few population-based prospective studies reported on associations between NT-proBNP concentration and first-onset heart failure. Added value of this study The Natriuretic Peptides Studies Collaboration involved new NT-proBNP concentration measurements in eight prospective studies as well as collation and harmonisation of individual-participant data from a further 32 relevant prospective cohorts identified by an updated systematic review. This effort enabled a detailed and standardised analysis of primary data for 95 617 participants without a history of cardiovascular disease recruited into 40 prospective studies in 12 different countries. The key added value of this collaboration is its ability to derive valid and new insights by combination of individual-participant data, information about various established and emerging risk factors, extended follow-up, breadth of cardiovascular disease outcomes recorded (eg, fatal and non-fatal heart failure, coronary heart disease, and stroke), study of several different measures of predictive ability, and generalisability to several high-income industrialised countries. Implications of all the available evidence

The Natriuretic Peptides Studies Collaboration involved new NT-proBNP concentration measurements in eight prospective studies as well as collation and harmonisation of individual-participant data from a further 32 relevant prospective cohorts identified by an updated systematic review. This effort enabled a detailed and standardised analysis of primary data for 95 617 participants without a history of cardiovascular disease recruited into 40 prospective studies in 12 different countries. The key added value of this collaboration is its ability to derive valid and new insights by combination of individual-participant data, information about various established and emerging risk factors, extended follow-up, breadth of cardiovascular disease outcomes recorded (eg, fatal and non-fatal heart failure, coronary heart disease, and stroke), study of several different measures of predictive ability, and generalisability to several high-income industrialised countries. Implications of all the available evidence We found that NT-proBNP concentration assessment strongly predicted first-onset heart failure and augmented coronary heart disease and stroke prediction. The incremental predictive ability of NT-proBNP concentration for coronary heart disease and stroke was moderate, but still greater than were those for HDL cholesterol or C-reactive protein concentrations. Our results have suggested that NT-proBNP concentration assessment could serve as a multipurpose biomarker in new approaches that integrate heart failure into cardiovascular disease primary prevention.

e and stroke was moderate, but still greater than were those for HDL cholesterol or C-reactive protein concentrations. Our results have suggested that NT-proBNP concentration assessment could serve as a multipurpose biomarker in new approaches that integrate heart failure into cardiovascular disease primary prevention. Methods Data sources Using two complementary approaches, we generated, collated, and harmonised individual-participant-level data from relevant prospective cohorts. First, de-novo NT-proBNP concentration measurements of stored samples were done by technicians masked to case-control status for some studies using the Elecsys2010 electrochemiluminescence method (proBNP Generation II; Roche, Burgess Hill, UK; appendix p 4). Second, we sought individual-participant data from relevant prospective studies identified through systematic searches of the published literature (PubMed, Scientific Citation Index Expanded, and Embase) for articles published up to Sept 4, 2014, using search terms related to natriuretic peptide family members and the primary outcomes, with no language restrictions (appendix p 7). We also scanned reference lists of identified articles for additional relevant studies. Studies were eligible if they had assayed NT-proBNP or B-type natriuretic peptide (BNP) concentration; recorded baseline information about age, sex, smoking status, systolic blood pressure, history of diabetes, and total and HDL cholesterol concentration (conventional risk factors); included participants without a known history of cardiovascular disease (ie, coronary heart disease, stroke, transient ischaemic attack, peripheral vascular disease, cardiovascular surgery, pulmonary heart disease, atrial fibrillation, or heart failure) at entry into the study; and recorded cause-specific deaths or major cardiovascular morbidity (non-fatal myocardial infarction, stroke, or heart failure) using well defined criteria over at least 1 year of follow-up.

scular disease, cardiovascular surgery, pulmonary heart disease, atrial fibrillation, or heart failure) at entry into the study; and recorded cause-specific deaths or major cardiovascular morbidity (non-fatal myocardial infarction, stroke, or heart failure) using well defined criteria over at least 1 year of follow-up. The appendix (p 4) provides details of the methods used to collect and harmonise data. Contributing studies classified deaths according to the primary cause (or, in its absence, the underlying cause) on the basis of International Classification of Diseases coding, revisions 8–10, to at least three digits, or according to study-specific classification systems. We based ascertainment of fatal outcomes on death certificates, supplemented in 26 cohorts by additional data, and of non-fatal outcomes on WHO (or similar) criteria for myocardial infarction and on clinical and imaging features for stroke and heart failure (appendix p 18). This Article follows Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Individual Patient Data reporting (appendix pp 8–11).17 The study was designed and done by the Natriuretic Peptides Studies Collaboration's independent coordinating centre and approved by the Cambridgeshire Ethics Review Committee.

s Article follows Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Individual Patient Data reporting (appendix pp 8–11).17 The study was designed and done by the Natriuretic Peptides Studies Collaboration's independent coordinating centre and approved by the Cambridgeshire Ethics Review Committee. Data analysis The analysis involved three inter-related components. First, we characterised cross-sectional associations of NT-proBNP concentration with established and emerging risk factors. Second, we assessed associations of NT-proBNP concentration with first-onset coronary heart disease, stroke, and heart failure, singly and in combination. Third, we quantified the incremental predictive value of assessment of NT-proBNP concentration in addition to conventional risk factors for major cardiovascular disease outcomes.

sessed associations of NT-proBNP concentration with first-onset coronary heart disease, stroke, and heart failure, singly and in combination. Third, we quantified the incremental predictive value of assessment of NT-proBNP concentration in addition to conventional risk factors for major cardiovascular disease outcomes. We focused the principal analyses on NT-proBNP concentration data because NT-proBNP is a more stable analyte than is BNP and encompassed more than 95% of the data in the collaboration (reserving the sparse BNP data for supplementary analyses). We defined two primary outcomes: 1) a combination of coronary heart disease (defined as fatal or non-fatal myocardial infarction) and stroke and 2) a combination of coronary heart disease, stroke, and heart failure. Participants contributed only the first cardiovascular disease outcome (whether non-fatal or fatal) recorded during follow-up (ie, we did not include deaths preceded by non-fatal cardiovascular disease events). Secondary outcomes were the component cardiovascular disease outcomes (ie, coronary heart disease, stroke, and heart failure) and the aggregate of death due to additional cardiovascular disease outcomes (ie, cardiac arrhythmia, hypertensive disease, pulmonary embolism, complications and ill defined descriptions of heart disease, sudden death, aortic aneurysms, and peripheral vascular disease). We censored outcomes if a participant was lost to follow-up, died from causes other than cardiovascular diseases, or reached the end of the follow-up period.

ensive disease, pulmonary embolism, complications and ill defined descriptions of heart disease, sudden death, aortic aneurysms, and peripheral vascular disease). We censored outcomes if a participant was lost to follow-up, died from causes other than cardiovascular diseases, or reached the end of the follow-up period. We calculated hazard ratios from prospective studies with Cox proportional hazard regression models, stratified by sex, using time-on-study as a timescale. We assessed the proportional hazards assumption, which was satisfied, as previously described.18 Analyses of case-cohort data involved Prentice weights and robust SEs.19 We calculated odds ratios from nested case-control studies using logistic regression models. We assumed hazard and odds ratios to represent the same relative risk, collectively describing them as risk ratios. We calculated risk ratios for a comparison of individuals in the top third with those in the bottom third of baseline NT-proBNP values using a two-stage approach, with estimates calculated separately within each study before pooling across studies with multivariate random-effects meta-analysis.18 To characterise shapes of associations, we calculated pooled risk ratios within overall tenths of NT-proBNP concentration and plotted them against the pooled geometric mean of NT-proBNP concentration within each tenth. We adjusted risk ratios for baseline levels of conventional risk factors. We investigated effect modification by study-level and individual characteristics with meta-regression and formal tests of interaction.18 We assessed between-study heterogeneity with the I2 statistic.20

NT-proBNP concentration within each tenth. We adjusted risk ratios for baseline levels of conventional risk factors. We investigated effect modification by study-level and individual characteristics with meta-regression and formal tests of interaction.18 We assessed between-study heterogeneity with the I2 statistic.20 We developed cardiovascular disease risk prediction models containing information about conventional risk factors with or without NT-proBNP concentration only in cohort and case-cohort studies and quantified improvements in predictive ability using measures of risk discrimination and reclassification.21, 22 We calculated C-indices and C-index changes within each study before pooling results weighted by the number of outcomes contributed. We calculated measures of risk reclassification (ie, integrated discrimination improvement and categorical and continuous net reclassification improvement) using data from studies in which both fatal and non-fatal events had been recorded.21 We examined categorical net reclassification of participants across predicted 10 year risk categories using cutoffs defined by the American College of Cardiology (ACC) and American Heart Association (AHA) 2013 (ie, <5%, 5% to <7·5%, and ≥7·5%),1 National Institute of Health and Care Excellence 2014,4 American College of Cardiology Foundation and American Heart Association 2010,3 and European Society of Cardiology 2016 guidelines.2 We log-transformed NT-proBNP concentration and modelled it using both linear and quadratic terms (with similar approaches used for the analysis of HDL cholesterol and C-reactive protein [CRP] concentration). We did analyses using Stata software, version 12.1. All p values are two sided. The appendix (pp 5–6) provides further details of the analytical methods used.

modelled it using both linear and quadratic terms (with similar approaches used for the analysis of HDL cholesterol and C-reactive protein [CRP] concentration). We did analyses using Stata software, version 12.1. All p values are two sided. The appendix (pp 5–6) provides further details of the analytical methods used. 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. PWi, JD, and EDA had full access to all the data in the study and had final responsibility for the decision to submit for publication.

modelled it using both linear and quadratic terms (with similar approaches used for the analysis of HDL cholesterol and C-reactive protein [CRP] concentration). We did analyses using Stata software, version 12.1. All p values are two sided. The appendix (pp 5–6) provides further details of the analytical methods used. 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. PWi, JD, and EDA had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Measurement of stored samples from 7129 participants (including 1173 incident cardiovascular disease cases) was done for eight prospective studies (the Reykjavik Offspring Study,23 the Northern Sweden Health and Disease Study,24 the Bruneck Study,25 and five cohorts contributing to the DAN-MONICA study;26 appendix p 3). We sought individual-participant data from 33 relevant prospective studies. Only one potentially relevant study27 (comprising <3% of the cardiovascular disease outcomes) was unable to contribute data, yielding a total of 40 contributing prospective studies from 12 countries (of which 30 had been analysed as cohort studies, eight as case-cohort studies, and two as nested case-control studies) and 95 617 participants without a history of cardiovascular disease. Details of the 40 contributing studies are provided in the appendix (pp 12, 17–20).12, 13, 14, 15, 16, 23, 24, 25, 26, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51

d two as nested case-control studies) and 95 617 participants without a history of cardiovascular disease. Details of the 40 contributing studies are provided in the appendix (pp 12, 17–20).12, 13, 14, 15, 16, 23, 24, 25, 26, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51 48 528 (51%) of participants were women and 61 451 (64%) were from Europe, and mean age at baseline was 61 years (SD 10). Median NT-proBNP concentration was 64 pg/mL (IQR 30–135; appendix pp 19–20, 27). NT-proBNP concentrations were approximately linearly associated with BNP concentrations across the range of values (appendix p 28). NT-proBNP and BNP concentrations increased with age and were higher in women, but were only weakly associated with several other characteristics, including ethnicity, history of hypertension, use of antihypertensive medication, systolic blood pressure, total and HDL cholesterol concentration, and estimated glomerular filtration rate (appendix pp 21, 22, 29).

ed with age and were higher in women, but were only weakly associated with several other characteristics, including ethnicity, history of hypertension, use of antihypertensive medication, systolic blood pressure, total and HDL cholesterol concentration, and estimated glomerular filtration rate (appendix pp 21, 22, 29). During 809 525 person-years at risk (median follow-up 7·8 years [IQR 5·2–11·8]), 5500 coronary heart disease, 4002 stroke, and 2212 heart failure outcomes occurred. NT-proBNP concentration was non-linearly associated with the risk of each of these diseases (figure 1). Risk ratios (top third vs bottom third of NT-proBNP concentration) adjusted for conventional risk factors were 1·76 (95% CI 1·56–1·98) for the combination of coronary heart disease and stroke; 2·00 (1·77–2·26) for the combination of coronary heart disease, stroke, and heart failure; 1·67 (1·45–1·93) for coronary heart disease; 1·81 (1·58–2·07) for stroke; 3·45 (2·66–4·46) for heart failure; and 3·11 (2·34–4·15) for cardiovascular disease deaths due to additional causes (figure 2; appendix p 30). Risk ratios were somewhat higher for fatal than for non-fatal coronary heart disease (p<0·0001), but similar for ischaemic and haemorrhagic stroke (p=0·44). In the same participants, corresponding risk ratios with lower HDL cholesterol concentration were 1·61 (1·45–1·78) for the combination of coronary heart disease and stroke and 1·47 (1·31–1·66) for the combination of coronary heart disease, stroke, and heart failure.

milar for ischaemic and haemorrhagic stroke (p=0·44). In the same participants, corresponding risk ratios with lower HDL cholesterol concentration were 1·61 (1·45–1·78) for the combination of coronary heart disease and stroke and 1·47 (1·31–1·66) for the combination of coronary heart disease, stroke, and heart failure. Risk ratios for NT-proBNP concentration did not materially change with further adjustment for body-mass index or estimated glomerular filtration rate, but they reduced somewhat with adjustment for CRP concentration (appendix p 23). Risk ratios for heart failure were higher in men than in women (4·25 vs 2·44; p<0·0001), in participants with a low body-mass index than a high body-mass index (3·61 vs 2·76; p=0·0004), and in studies that had stored samples for 10 years or fewer before analysis than longer than 10 years (6·20 vs 2·68; p=0·0018; appendix p 31–32). Otherwise, risk ratios did not vary substantially with levels of conventional risk factors or in other clinically relevant subgroups (appendix pp 31–32). We observed qualitatively similar findings in analyses that defined thirds separately for men and women, excluded people with high baseline concentrations of NT-proBNP, excluded the initial 5 years of follow-up, and were restricted to studies recording both fatal and non-fatal outcomes (appendix p 33). Similar findings were also noted in analyses that compared studies grouped by NT-proBNP concentration assay type or generation (appendix p 32), compared studies with different lengths of follow-up (appendix p 34), used per one SD higher log NT-proBNP concentration (appendix p 24), and focused on fatal outcomes only (appendix p 35). In analyses of 15 909 participants for coronary heart disease and stroke and 12 202 participants for heart failure from seven studies with available information about BNP concentration,28, 29, 30, 31, 32, 33, 34 risk ratios for coronary heart disease, stroke, and heart failure observed with BNP concentration were weaker than were those observed with NT-proBNP concentration (appendix p 25). We noted moderate heterogeneity of risk ratios across studies (appendix p 30). I2 values were 45% for coronary heart disease, 23% for stroke, and 54% for heart failure.

t disease, stroke, and heart failure observed with BNP concentration were weaker than were those observed with NT-proBNP concentration (appendix p 25). We noted moderate heterogeneity of risk ratios across studies (appendix p 30). I2 values were 45% for coronary heart disease, 23% for stroke, and 54% for heart failure. After addition of NT-proBNP concentration to a model with conventional risk factors only, the C-index increased by 0·012 (95% CI 0·010–0·014) for the combination of coronary heart disease and stroke; 0·019 (0·016–0·022) for the combination of coronary heart disease, stroke, and heart failure; 0·012 (0·009–0·015) for coronary heart disease; 0·011 (0·008–0·015) for stroke; and 0·038 (0·030–0·045) for heart failure (figure 3; appendix pp 36–37). Overall net reclassification improvements for NT-proBNP concentration across predicted 10 year risk categories defined by the 2013 ACC and AHA guidelines1 were 0·027 (0·019–0·036) for the combination of coronary heart disease and stroke and 0·028 (0·019–0·038) for the combination of coronary heart disease, stroke, and heart failure (table). Continuous net reclassification improvements were 0·154 (0·111–0·198) for the combination of coronary heart disease and stroke and 0·198 (0·162–0·234) for the combination of coronary heart disease, stroke, and heart failure, and integrated discrimination improvements were 0·013 (0·011–0·015) for the combination of coronary heart disease and stroke and 0·030 (0·026–0·033) for the combination of coronary heart disease, stroke, and heart failure (appendix p 26). Incremental risk prediction afforded by NT-proBNP concentration assessment was greater than that afforded by HDL cholesterol or CRP concentration assessment (figure 3, figure 4, table). NT-proBNP and CRP concentration provided essentially non-overlapping incremental risk discrimination (figure 4).

lure (appendix p 26). Incremental risk prediction afforded by NT-proBNP concentration assessment was greater than that afforded by HDL cholesterol or CRP concentration assessment (figure 3, figure 4, table). NT-proBNP and CRP concentration provided essentially non-overlapping incremental risk discrimination (figure 4). In further analyses that involved the combination of coronary heart disease, stroke, and heart failure as the outcome, improvements in C-index with NT-proBNP concentration assessment were possibly greater among older individuals and people with a history of diabetes, who used antihypertensives, who had a higher systolic blood pressure, and who had a lower total cholesterol concentration (appendix p 38). However, we did not adjust these exploratory analyses for multiple comparisons. In further sensitivity analyses, we found that C-index improvements were similar when the base model additionally included information about ethnicity and antihypertensive treatment (appendix p 39), but somewhat smaller in analyses that excluded people with high baseline concentrations of NT-proBNP or modelled NT-proBNP concentration using a prespecified cutoff value rather than continuous values (appendix p 40). Net reclassification improvements were similar or larger than were those in the main analysis when analysis involved cutoffs for clinical risk categories defined by guidelines other than the 2013 ACC and AHA guidelines1 (appendix p 26).

on using a prespecified cutoff value rather than continuous values (appendix p 40). Net reclassification improvements were similar or larger than were those in the main analysis when analysis involved cutoffs for clinical risk categories defined by guidelines other than the 2013 ACC and AHA guidelines1 (appendix p 26). Discussion In this study, we found that NT-proBNP concentration assessment strongly predicted first-onset heart failure and augmented coronary heart disease and stroke prediction, suggesting that NT-proBNP concentration assessment could serve as a multipurpose biomarker in new approaches that integrate heart failure into cardiovascular disease primary prevention. A key observation was our study's demonstration of graded associations between NT-proBNP concentration and the incidence of coronary heart disease, stroke, and heart failure. The continuous nature of these associations suggests that NT-proBNP concentration measurement is potentially suitable for population-level risk assessment. We also made the surprising observation that NT-proBNP concentration predicts stroke at least as strongly as it does coronary heart disease, by contrast with the idea that NT-proBNP concentration is predominantly a coronary biomarker. The stroke associations that we noted could partly be explained by associations previously reported between NT-proBNP concentration and stroke risk factors (eg, left ventricular hypertrophy and atrial fibrillation),15, 52 but further work is needed to elucidate the common pathobiology for coronary heart disease, stroke, and heart failure reflected by preceding NT-proBNP concentration. Furthermore, we found that NT-proBNP concentration predicted deaths due to additional cardiovascular causes, such as cardiac arrhythmia and sudden death.53 Collectively, these results encourage evaluation of NT-proBNP concentration for prediction of an even wider range of cardiovascular disease outcomes than that we studied.

rmore, we found that NT-proBNP concentration predicted deaths due to additional cardiovascular causes, such as cardiac arrhythmia and sudden death.53 Collectively, these results encourage evaluation of NT-proBNP concentration for prediction of an even wider range of cardiovascular disease outcomes than that we studied. Our conclusions on the incremental predictive ability of assessment of NT-proBNP concentration were strengthened by broadly concordant results when we studied varying cardiovascular disease outcomes and used different measures of risk discrimination and reclassification. Importantly, the modest improvements that we observed in risk reclassification with NT-proBNP concentration assessment applied similarly across the absolute risk thresholds used in different clinical guidelines.1, 2, 3, 4 In particular, NT-proBNP concentration assessment improved the specificity of risk prediction by appropriately downclassifying the clinical risk category of many individuals who did not go on to develop cardiovascular disease outcomes. Hence, addition of NT-proBNP concentration measurement to cardiovascular disease risk assessment could improve targeting of preventive treatments (such as statins) and allocation of resources for detailed screening (such as comprehensive tests for heart failure at specialised cardiology clinics), as exemplified by previous natriuretic peptide-guided trials in patients with diabetes54 or heart failure.55, 56 Data from future studies are needed to establish the cost-effectiveness and feasibility of NT-proBNP concentration screening for prediction of first composite cardiovascular disease outcomes, analogous with previous work on left ventricular systolic dysfunction.57, 58, 59

atients with diabetes54 or heart failure.55, 56 Data from future studies are needed to establish the cost-effectiveness and feasibility of NT-proBNP concentration screening for prediction of first composite cardiovascular disease outcomes, analogous with previous work on left ventricular systolic dysfunction.57, 58, 59 To provide clinical context, we compared incremental improvements afforded by NT-proBNP concentration assessment with those afforded by HDL cholesterol, a widely used biomarker in cardiovascular disease risk assessment (this comparison is additionally relevant because HDL cholesterol concentration, like NT-proBNP concentration, is a biomarker of unknown relevance to the cause of cardiovascular disease60). We found that improvements in risk discrimination with NT-proBNP concentration were greater than those provided by HDL cholesterol, even though our evaluation was skewed in favour of HDL cholesterol concentration since we added HDL cholesterol concentration only to other conventional risk factors (omitting NT-proBNP concentration), whereas we added NT-proBNP concentration to conventional risk factors, including HDL cholesterol concentration. Furthermore, in a head-to-head comparison, we found that the improvement in risk discrimination with NT-proBNP concentration was about three times greater than was the improvement in risk discrimination using CRP concentration. The idea that NT-proBNP concentration captures information about non-traditional cardiovascular disease pathways61, 62 was supported by our observation that NT-proBNP concentration was uncorrelated or weakly correlated with the established and emerging risk factors that we studied.

scrimination using CRP concentration. The idea that NT-proBNP concentration captures information about non-traditional cardiovascular disease pathways61, 62 was supported by our observation that NT-proBNP concentration was uncorrelated or weakly correlated with the established and emerging risk factors that we studied. Our study had major strengths. Because of its considerable statistical power, we could provide precise estimates, even for analyses that involved categorisation of NT-proBNP concentrations. More than 90% of the NT-proBNP concentration data in our analysis were generated with use of a common gold-standard assay. We recorded information about the incidence of various cardiovascular disease outcomes using well validated endpoint definitions. We centrally analysed individual-participant data, which were harmonised from prospective studies with extended follow-up, enabling time-to-event analyses, exclusion of people with a baseline history of cardiovascular disease (including heart failure), and adoption of a uniform approach to statistical analyses. To enhance validity further, we restricted analyses to people with complete information about a set of relevant risk factors. Our primary analysis excluded participants with a reported baseline history of heart failure and, moreover, the findings were robust to exclusion of participants with high baseline NT-proBNP concentrations. The generalisability of our findings was enhanced by inclusion of data from 12 countries and by the robustness of results to various sensitivity analyses.

s with a reported baseline history of heart failure and, moreover, the findings were robust to exclusion of participants with high baseline NT-proBNP concentrations. The generalisability of our findings was enhanced by inclusion of data from 12 countries and by the robustness of results to various sensitivity analyses. Our study had potential limitations. Misclassification of heart failure outcomes could have led to underestimation of associations between NT-proBNP concentration and heart failure risk and, conversely, overestimation of associations with non-heart failure outcomes. Most of our data were derived from people of European continental ancestry. We could not compare the performance of NT-proBNP concentration with cardiac troponin, coronary calcium scoring, or other biomarkers apart from HDL cholesterol and CRP concentrations. We conclude that assessment of NT-proBNP concentration could serve as a multipurpose biomarker in new approaches that integrate heart failure into primary prevention of cardiovascular diseases. Supplementary Material Supplementary appendix Contributors PWi, VG, NS, JD, and EDA conceived the idea for this study. PWi, JD, and EDA drafted the report. PWi did literature searches and analysed data. All investigators shared data and had opportunities to interpret results and critically revise the report. All members of the writing committee provided critical revisions. All members of the coordinating centre collected, harmonised, analysed, and interpreted data. The data management team collected and harmonised data.

l investigators shared data and had opportunities to interpret results and critically revise the report. All members of the writing committee provided critical revisions. All members of the coordinating centre collected, harmonised, analysed, and interpreted data. The data management team collected and harmonised data. Writing committee Peter Willeit (University of Cambridge, Cambridge, UK, and Medical University Innsbruck, Innsbruck, Austria); Stephen Kaptoge* (University of Cambridge, Cambridge, UK); Paul Welsh* (University of Glasgow, Glasgow, UK); Adam S Butterworth* (University of Cambridge, Cambridge, UK); Rajiv Chowdhury (University of Cambridge, Cambridge, UK); Sarah A Spackman (University of Cambridge, Cambridge, UK); Lisa Pennells (University of Cambridge, Cambridge, UK); Pei Gao (University of Cambridge, Cambridge, UK, and Peking University, Beijing, China); Stephen Burgess (University of Cambridge, Cambridge, UK); Daniel F Freitag (University of Cambridge, Cambridge, UK); Michael Sweeting (University of Cambridge, Cambridge, UK); Angela M Wood (University of Cambridge, Cambridge, UK); Nancy R Cook (Brigham and Women's Hospital and Harvard Medical School, Boston, MA, USA); Suzanne Judd (University of Alabama at Birmingham, Birmingham, AL, USA); Stella Trompet (Leiden University Medical Centre, Leiden, Netherlands); Vijay Nambi (Michael E DeBakey Baylor College of Medicine and Houston Methodist DeBakey Heart and Vascular Center, Houston, TX, USA); Michael Hecht Olsen (Odense University Hospital, Odense, Denmark); Brendan M Everett (Brigham and Women's Hospital and Harvard Medical School, Boston, MA, USA); Frank Kee (Queen's University Belfast, Belfast, UK); Johan Ärnlöv (Uppsala University, Uppsala, Sweden); Veikko Salomaa (National Institute for Health and Welfare, Helsinki, Finland); Daniel Levy (National Institutes of Health, Bethesda, MD, USA); Jussi Kauhanen (University of Eastern Finland, Kuopio, Finland); Jari A Laukkanen (University of Eastern Finland, Kuopio, Finland); Maryam Kavousi (Erasmus Medical Center, Rotterdam, Netherlands); Toshiharu Ninomiya (Kyushu University, Fukuoka, Japan); Juan-Pablo Casas (Farr Institute of Health Informatics, University College London, London, UK); Lori B Daniels (University of California San Diego, San Diego, CA, USA); Lars Lind (Uppsala University, Uppsala, Sweden); Caroline N Kistorp (University of Copenhagen and Herlev Hospital, Copenhagen, Denmark); Jens Rosenberg (Copenhagen University Hospital Glostrup, Glostrup, Denm

niversity College London, London, UK); Lori B Daniels (University of California San Diego, San Diego, CA, USA); Lars Lind (Uppsala University, Uppsala, Sweden); Caroline N Kistorp (University of Copenhagen and Herlev Hospital, Copenhagen, Denmark); Jens Rosenberg (Copenhagen University Hospital Glostrup, Glostrup, Denm ark); Thomas Mueller (Konventhospital Barmherzige Brüder, Linz, Austria); Speranza Rubattu (University Sapienza of Rome, Rome, Italy, and Istituto di Ricovero e Cura a Carattere Scientifico Neuromed, Pozzilli, Italy); Demosthenes B Panagiotakos (Harokopio University of Athens, Athens, Greece); Oscar H Franco (Erasmus Medical Center, Rotterdam, Netherlands); James A de Lemos (University of Texas Southwestern Medical School, Dallas, TX, USA); Andreas Luchner (Universitätsklinikum Regensburg, Regensburg, Germany, and Klinikum St Marien, Amberg, Germany); Jorge R Kizer (Albert Einstein College of Medicine, Bronx, NY, USA); Stefan Kiechl (Medical University Innsbruck, Innsbruck, Austria); Jukka T Salonen (Metabolic Analytical Services, Helsinki, Finland); S Goya Wannamethee (University College London, London, UK); Rudolf A de Boer (University of Groningen, Groningen, Netherlands); Børge G Nordestgaard (University of Copenhagen, Copenhagen, Denmark); Jonas Andersson (Umeå University, Umeå, Sweden); Torben Jørgensen (Research Centre for Prevention and Health, Glostrup, Denmark); Olle Melander (Malmö University Hospital, Malmö, Sweden); Christie M Ballantyne (Baylor College of Medicine and Houston Methodist DeBakey Heart and Vascular Center, Houston, TX, USA); Christopher DeFilippi (University of Maryland School of Medicine, Baltimore, MD, USA); Paul M Ridker (Brigham and Women's Hospital and Harvard Medical School, Boston, MA, USA); Mary Cushman (University of Vermont, Burlington, VT, USA); Wayne D Rosamond (University of North Carolina at Chapel Hill, Chapel Hill, NC, USA); Simon G Thompson (University of Cambridge, Cambridge, UK); Vilmundur Gudnason† (Icelandic Heart Association, Kopavogur, Iceland and University of Iceland, Reykjavik, Iceland); Naveed Sattar† (University of Glasgow, Glasgow, UK); John Danesh† (University of Cambridge, Cambridge, UK); Emanuele Di Angelantonio† (University of Cambridge, Cambridge, UK).

Cambridge, Cambridge, UK); Vilmundur Gudnason† (Icelandic Heart Association, Kopavogur, Iceland and University of Iceland, Reykjavik, Iceland); Naveed Sattar† (University of Glasgow, Glasgow, UK); John Danesh† (University of Cambridge, Cambridge, UK); Emanuele Di Angelantonio† (University of Cambridge, Cambridge, UK). *Contributed equally. †Contributed equally.

Cambridge, Cambridge, UK); Vilmundur Gudnason† (Icelandic Heart Association, Kopavogur, Iceland and University of Iceland, Reykjavik, Iceland); Naveed Sattar† (University of Glasgow, Glasgow, UK); John Danesh† (University of Cambridge, Cambridge, UK); Emanuele Di Angelantonio† (University of Cambridge, Cambridge, UK). *Contributed equally. †Contributed equally. Natriuretic Peptides Studies Collaboration ARIC: Vijay Nambi, Christie M Ballantyne, Ron C Hoogeveen, Sunil K Agarwal; ATTICA: Demosthenes B Panagiotakos; BRHS: S Goya Wannamethee, Peter H Whincup; BRUN: Stefan Kiechl, Johann Willeit, Georg Schett, Peter Santer, Peter Willeit; BWHHS: Juan-Pablo Casas, Debbie A Lawlor; CHS: Christopher DeFilippi, Richard A Kronmal, Bruce M Psaty, Mary Cushman; COPEN: Børge G Nordestgaard; DANMON: Michael Hecht Olsen, Torben Jørgensen; DHS: James A de Lemos, Darren K McGuire, Sandeep R Das, Mark H Drazner; FINRISK97: Veikko Salomaa, Erkki Vartiainen, Kennet Harald, Tanja Zeller; FRAMOFF: Daniel Levy; HISAYAMA: Toshiharu Ninomiya, Jun Hata, Yutaka Kiyohara; KIHD: Jussi Kauhanen, Jukka T Salonen, Jari A Laukkanen, Tomi-Pekka Tuomainen, Heikki Ruskoaho; KISTORP: Caroline N Kistorp, Ilan Raymond; LIFE: Michael Hecht Olsen; LIPAD: Thomas Mueller, Benjamin Dieplinger, Meinhard Haltmayer; MDCS: Olle Melander; MESA: Mary Cushman, Susan R Heckbert, Bruce M Psaty, João A Lima; MONICA/KORA3: Andreas Luchner, Klaus Stark, Iris M Heid, Annette Peters; NSHDS: Jonas Andersson, Jan-Håkan Jansson, Patrik Wennberg; OHS: Speranza Rubattu, Massimo Volpe, Pasquale Strazzullo; PIVUS: Lars Lind, Per Venge, Bertil Lindahl; PREVEND: Rudolf A de Boer, Stephan J L Bakker, Ron T Gansevoort; PRIME: Frank Kee, Alun Evans, John W G Yarnell; PROSPER: Stella Trompet, J Wouter Jukema, David J Stott, Anton J M de Craen; PTLBNP: Jens Rosenberg, Per R Hildebrandt, Finn Gustafsson, Morten Schou; RANCHO: Lori B Daniels, Elizabeth Barrett-Connor, Alan S Maisel; REGARDS: Mary Cushman, Suzanne Judd, Monika M Safford, Virginia J Howard, Neil A Zakai; REYKOFF: Vilmundur Gudnason, Thor Aspelund, Gudny Eiriksdottir, Uggi Agnarsson, Margret B Andresdottir; RS-I: Maryam Kavousi, Albert Hofman, Symen Ligthart, Anton H van den Meiracker; RS-II: Oscar H Franco, Abbas Dehghan, Frank J A van Rooij, M Arfan Ikram; SHS: Jorge R Kizer, Lyle G Best, Richard B Devereux, Jason G Umans; ULSAM: Johan Ärnlöv, Björn Zethelius, Lars Lannfelt, Vilmantas Giedraitis; WHIOS: Nancy R Cook, JoAnn E Manson, Brendan M Everett, Paul M Ridker; WHS: Brendan M Everett, Nancy R Cook, Paul M Ridker, Aruna D

Dehghan, Frank J A van Rooij, M Arfan Ikram; SHS: Jorge R Kizer, Lyle G Best, Richard B Devereux, Jason G Umans; ULSAM: Johan Ärnlöv, Björn Zethelius, Lars Lannfelt, Vilmantas Giedraitis; WHIOS: Nancy R Cook, JoAnn E Manson, Brendan M Everett, Paul M Ridker; WHS: Brendan M Everett, Nancy R Cook, Paul M Ridker, Aruna D Pradhan; WOSCOPS: Naveed Sattar, Ian Ford, Chris J Packard, Paul Welsh. Data management team Sarah A Spackman, Thomas Bolton, Matthew Walker. Coordinating centre Narinder Bansal, Thomas Bolton, Stephen Burgess, Adam S Butterworth, Rajiv Chowdhury, Emanuele Di Angelantonio, Daniel F Freitag, Pei Gao, Eric Harshfield, Stephen Kaptoge, Linda M O'Keeffe, Lisa Pennells, Anna Ramond, Sarah A Spackman, Michael Sweeting, Simon G Thompson, Matthew Walker, Peter Willeit, Angela M Wood, John Danesh.

nsal, Thomas Bolton, Stephen Burgess, Adam S Butterworth, Rajiv Chowdhury, Emanuele Di Angelantonio, Daniel F Freitag, Pei Gao, Eric Harshfield, Stephen Kaptoge, Linda M O'Keeffe, Lisa Pennells, Anna Ramond, Sarah A Spackman, Michael Sweeting, Simon G Thompson, Matthew Walker, Peter Willeit, Angela M Wood, John Danesh. Declaration of interests PWi was supported by a PhD studentship from the British Heart Foundation (FS/10/037/28413) and an Erwin Schrödinger fellowship from the Austrian Science Fund (J 3679-B13). PWe was supported by a British Heart Foundation fellowship (FS/12/62/29889). NRC received grants from the National Heart, Lung, and Blood Institute. VN received grants from the National Institute of Health. BME received investigator-initiated awards from Roche Diagnostics. JR received grants from the Danish Heart Foundation. JAdL received grants from Roche Diagnostics. AL and CMB received personal fees from Roche Diagnostics. SGT received grants from the UK Medical Research Council and British Heart Foundation. JD received research funding from the British Heart Foundation, the National Institute for Health Reasearch Cambridge Comprehensive Biomedical Research Centre, the Bupa Foundation, diaDexus, the European Research Council, the European Union, the Evelyn Trust, the Fogarty International Centre, GlaxoSmithKline, Merck, the National Heart, Lung, and Blood Institute, the National Institute for Health Research, the National Institute of Neurological Disorders and Stroke, NHS Blood and Transplant, Novartis, Pfizer, the UK Medical Research Council, and the Wellcome Trust. EDA received research funding from the UK Medical Research Council, the British Heart Foundation, the National Institute of Health Research, NHS Blood and Transplant, and the European Commission Framework Programme during the conduct of the study. All other members of the writing committee declare no competing interests.

atement. 18 (62%) expressed fewer concerns about their glycaemic control while using the closed-loop system. 14 (48%) participants reported improved sleep during the closed-loop period. 23 (88%) of 26 participants reported feeling safe while using the closed-loop system, and 26 (96%) of 27 would recommend it to others. No serious adverse events, episodes of severe hypoglycaemia, or episodes of hyperglycaemia with ketosis were reported. Skin irritations related to sensor use occurred in four participants. Two participants had mild respiratory tract infections (one during the run-in period and one during the control period). One participant had cystitis during the closed-loop period and one reported allergic rhinoconjunctivitis during the control period. All reported adverse events were resolved without sequelae.

A received research funding from the UK Medical Research Council, the British Heart Foundation, the National Institute of Health Research, NHS Blood and Transplant, and the European Commission Framework Programme during the conduct of the study. All other members of the writing committee declare no competing interests. Acknowledgments The work of the coordinating centre was funded by the UK Medical Research Council (G0800270), British Heart Foundation (SP/09/002), British Heart Foundation Cambridge Cardiovascular Centre of Excellence, National Institute for Health Research Cambridge Biomedical Research Centre, European Research Council, and European Commission Framework Programme 7 (HEALTH-F2–2012–279233). The collaboration's website has compiled a list provided by investigators of some of the funders of the component studies in this analysis. Laboratory measurements were supported by a grant from the Evelyn Trust. Roche donated N-terminal-pro-B-type natriuretic peptide reagents. Figure 1 Associations of NT-proBNP and HDL-C concentrations with first-onset coronary heart disease, stroke, and heart failure

Acknowledgments The work of the coordinating centre was funded by the UK Medical Research Council (G0800270), British Heart Foundation (SP/09/002), British Heart Foundation Cambridge Cardiovascular Centre of Excellence, National Institute for Health Research Cambridge Biomedical Research Centre, European Research Council, and European Commission Framework Programme 7 (HEALTH-F2–2012–279233). The collaboration's website has compiled a list provided by investigators of some of the funders of the component studies in this analysis. Laboratory measurements were supported by a grant from the Evelyn Trust. Roche donated N-terminal-pro-B-type natriuretic peptide reagents. Figure 1 Associations of NT-proBNP and HDL-C concentrations with first-onset coronary heart disease, stroke, and heart failure Risk ratios adjusted for age, smoking status, history of diabetes, systolic blood pressure, and total cholesterol and HDL-C concentration (HDL-C concentration only for NT-proBNP concentration analysis) and stratified by sex. Analyses involved 4716 coronary heart disease outcomes (from 34 cohorts), 3768 stroke outcomes (from 30 cohorts), and 2021 heart failure outcomes (from 16 cohorts). The size of the circles is proportional to the inverse of the variance of the respective estimate. Error bars are 95% CIs, estimated from floated variances. HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. Figure 2 Associations of NT-proBNP and HDL-C concentrations with several incident first-onset cardiovascular outcomes

Risk ratios adjusted for age, smoking status, history of diabetes, systolic blood pressure, and total cholesterol and HDL-C concentration (HDL-C concentration only for NT-proBNP concentration analysis) and stratified by sex. Analyses involved 4716 coronary heart disease outcomes (from 34 cohorts), 3768 stroke outcomes (from 30 cohorts), and 2021 heart failure outcomes (from 16 cohorts). The size of the circles is proportional to the inverse of the variance of the respective estimate. Error bars are 95% CIs, estimated from floated variances. HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. Figure 2 Associations of NT-proBNP and HDL-C concentrations with several incident first-onset cardiovascular outcomes Risk ratios adjusted for age, smoking status, history of diabetes, systolic blood pressure, and total cholesterol and HDL-C concentration (HDL-C concentration only for NT-proBNP concentration analysis) and stratified by sex. HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. *Top versus bottom third of NT-proBNP concentration. †Bottom versus top third of HDL-C concentration. ‡Subsumes deaths due to cardiac arrhythmia, hypertensive disease, pulmonary embolism, complications and ill defined descriptions of heart disease, sudden death, aortic aneurysms, and peripheral vascular disease. Figure 3 Improvement in risk discrimination for first-onset individual and composite cardiovascular disease outcomes by addition of information about NT-proBNP concentration compared with that about HDL-C concentration

Risk ratios adjusted for age, smoking status, history of diabetes, systolic blood pressure, and total cholesterol and HDL-C concentration (HDL-C concentration only for NT-proBNP concentration analysis) and stratified by sex. HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. *Top versus bottom third of NT-proBNP concentration. †Bottom versus top third of HDL-C concentration. ‡Subsumes deaths due to cardiac arrhythmia, hypertensive disease, pulmonary embolism, complications and ill defined descriptions of heart disease, sudden death, aortic aneurysms, and peripheral vascular disease. Figure 3 Improvement in risk discrimination for first-onset individual and composite cardiovascular disease outcomes by addition of information about NT-proBNP concentration compared with that about HDL-C concentration Analyses involved 8323 outcomes for the combination of coronary heart disease plus stroke (from 32 cohorts), 6582 outcomes for the combination of coronary heart disease plus stroke plus heart failure (from 22 cohorts), 4552 coronary heart disease outcomes (from 32 cohorts), 3768 stroke outcomes (from 30 cohorts), and 2021 heart failure outcomes (from 16 cohorts). HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. *The reference model included information about age, sex, smoking, systolic blood pressure, history of diabetes, and concentration of total cholesterol.

8·1] when participants were using usual pump therapy vs 76·2% [6·4] when they used closed-loop; table 2). 24 h sensor glucose and insulin delivery profiles are shown in figure 2. The proportion of time sensor glucose concentration was in the target range seemed unchanged over the 4 week intervention periods (appendix). Day-and-night closed-loop insulin delivery reduced mean glucose concentration by 0·4 mmol/L (0·1–0·7, p=0·0226) compared with usual pump therapy (table 2, figure 3). Compared with the control period, day-and-night closed-loop insulin delivery reduced the proportion of time with glucose concentration below 3·9 mmol/L by 50% (37–59, p<0·0001), below 3·5 mmol/L by 65% (53–74, p<0·0001), below 3·3 mmol/L by 70% (57–78, p<0·0001), and below 2·8 mmol/L by 76% (59–86, p<0·0001), as well as the burden of hypoglycaemia (ie, area under the curve when sensor glucose concentration was less than 3·5 mmol/L) by 73% (59–82, p<0·0001). Closed-loop insulin delivery also reduced the number of nights when glucose concentration was below 3·5 mmol/L for at least 20 min as well as the mean duration of such periods (table 2). Compared with usual pump therapy, closed-loop insulin delivery reduced the proportion of time with glucose concentration above the target range (ie, >10 mmol/L) by 6·9 percentage points (3·5–10·2, p=0·0003), above 13·9 mmol/L by 3·0 percentage points (1·6–4·4, p=0·0002) and above 16·7 mmol/L by 1·2 percentage points (0·6–1·9, p=0·0009; table 2). Moreover, all measures of glycaemic variability were significantly lower in the closed-loop period than in the control period: SD of sensor glucose was 0·5 mmol/L (0·3–0·7) lower (p<0·0001), coefficient of variation of sensor glucose within days was 5·0% (3·0–7·1) lower (p<0·0001), and coefficient of variation of sensor glucose between days was 7·5% (5·3–9·7) lower (p<0·0001; table 2). Total daily insulin was similar between study periods (table 3). Weekly trends in glucose control and insulin delivery are shown in the appendix (p 3).

3768 stroke outcomes (from 30 cohorts), and 2021 heart failure outcomes (from 16 cohorts). HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. *The reference model included information about age, sex, smoking, systolic blood pressure, history of diabetes, and concentration of total cholesterol. Figure 4 Improvement in risk discrimination for first-onset individual and composite cardiovascular outcomes by addition of information about CRP and NT-proBNP concentration to a model with conventional risk factors Analyses involved 7618 outcomes for the combination of coronary heart disease plus stroke (from 28 cohorts), 5492 outcomes for the combination of coronary heart disease plus stroke plus heart failure (from 18 cohorts), 4120 coronary heart disease outcomes (from 27 cohorts), 3487 stroke outcomes (from 26 cohorts), and 1606 heart failure outcomes (from 13 cohorts). CRP=C-reactive protein. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. *The reference model included information about age, sex, smoking, systolic blood pressure, history of diabetes, and concentrations of total and HDL cholesterol. Table Improvement in risk classification for first-onset composite cardiovascular disease outcomes by addition of information about NT-proBNP concentration compared with that about HDL-C

Analyses involved 7618 outcomes for the combination of coronary heart disease plus stroke (from 28 cohorts), 5492 outcomes for the combination of coronary heart disease plus stroke plus heart failure (from 18 cohorts), 4120 coronary heart disease outcomes (from 27 cohorts), 3487 stroke outcomes (from 26 cohorts), and 1606 heart failure outcomes (from 13 cohorts). CRP=C-reactive protein. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. *The reference model included information about age, sex, smoking, systolic blood pressure, history of diabetes, and concentrations of total and HDL cholesterol. Table Improvement in risk classification for first-onset composite cardiovascular disease outcomes by addition of information about NT-proBNP concentration compared with that about HDL-C Non-cases Cases Overall Coronary heart disease plus stroke Conventional risk factors without HDL-C concentration* Reference Reference Reference plus HDL-C concentration 0·001 (−0·003 to 0·004); p=0·70 0·008 (−0·000 to 0·016); p=0·056 0·009 (−0·000 to 0·017); p=0·056 plus HDL-C and NT-proBNP concentration 0·029 (0·025 to 0·032); p<0·0001 −0·001 (−0·009 to 0·007); p=0·79 0·027 (0·019 to 0·036); p<0·0001 Coronary heart disease plus stroke plus heart failure Conventional risk factors without HDL-C concentration* Reference Reference Reference plus HDL-C concentration 0·011 (0·008 to 0·015); p<0·0001 0·006 (−0·001 to 0·013); p=0·10 0·017 (0·009 to 0·025); p<0·0001 plus HDL-C and NT-proBNP concentration 0·036 (0·032 to 0·040); p<0·0001 −0·008 (−0·017 to 0·001); p=0·087 0·028 (0·019 to 0·038); p<0·0001 Data are categorical net reclassification improvement versus preceding model (95% CI); p value. We calculated categorical net reclassification improvement across predicted 10 year cardiovascular disease risk categories defined by the American College of Cardiology and American Heart Association 2013 guidelines.1 Analyses involved 4672 outcomes for the composite outcome of coronary heart disease plus stroke (from 19 cohorts) and 4071 for the composite outcome of coronary heart disease plus stroke plus heart failure (from 16 cohorts). HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide.

n 2013 guidelines.1 Analyses involved 4672 outcomes for the composite outcome of coronary heart disease plus stroke (from 19 cohorts) and 4071 for the composite outcome of coronary heart disease plus stroke plus heart failure (from 16 cohorts). HDL-C=HDL cholesterol. NT-proBNP=N-terminal-pro-B-type natriuretic peptide. * The reference model included information about age, sex, smoking, systolic blood pressure, history of diabetes, and concentration of total cholesterol.

ipants. Two participants had mild respiratory tract infections (one during the run-in period and one during the control period). One participant had cystitis during the closed-loop period and one reported allergic rhinoconjunctivitis during the control period. All reported adverse events were resolved without sequelae. Discussion In this two-centre, open-label, randomised, crossover trial, we showed that, in adults with type 1 diabetes and HbA1c below 7·5%, day-and-night hybrid closed-loop insulin delivery significantly improved overall glucose control while reducing hypoglycaemia progressively by 50–75% at lower glucose thresholds compared with usual insulin pump therapy. Beneficial effects on glycaemic outcomes included increased time spent with glucose concentration in target range (3·9–10·0 mmol/L), reduced time with glucose concentration above and below the target range, and decreased mean glucose concentration and glycaemic variability. The findings of increased time spent in the glucose concentration target range, reduced hypoglycaemia, and decreased glycaemic variability were similarly observed during night-time and daytime periods. These outcomes were achieved without change in total insulin delivery.

lucose concentration and glycaemic variability. The findings of increased time spent in the glucose concentration target range, reduced hypoglycaemia, and decreased glycaemic variability were similarly observed during night-time and daytime periods. These outcomes were achieved without change in total insulin delivery. Hypoglycaemia is associated with increased morbidity and mortality in patients with type 1 diabetes.25 A reduction of at least 30% in risk of hypoglycaemia, as observed in our study, is considered clinically relevant.26 Threshold and predictive low-glucose suspend insulin delivery systems8, 9, 10 cannot step-up insulin delivery and thus do not address the issue of hyperglycaemia. The advantage of a closed-loop system such as ours is the responsive, graduated modulation of insulin delivery, both below and above the pre-set pump regimen. This notion is supported by findings from our study, which showed that reduction in mean glucose concentration was accompanied by significant reduction in all the measured hypoglycaemia parameters. The multiplicity of beneficial outcomes—including increased time with optimum glucose control (in target range) and reduced time below and above target range, which were consistently observed during both night-time and daytime periods—suggests that the benefits of the closed-loop system can be accrued irrespective of the time of day in adults with HbA1c below 7·5%. The control algorithm used in our study had enhanced adaptive features and coped safely with variations in insulin requirements, trading variability in insulin delivery for consistency in glucose concentrations. Several studies have shown that increased glycaemic variability is associated with the burden of hypoglycaemia.27 Glycaemic variability28 and hypoglycaemia25 have both been associated with adverse clinical outcomes. We hypothesise that the significant reductions in glucose variability and hypoglycaemia by closed-loop insulin delivery in our study might have implications for clinical outcomes, although this hypothesis will need confirmation by longer and larger studies.

Introduction So-called western dietary patterns (ie, high in saturated fat, cholesterol, sodium, and added sugars; low in fruits, vegetables, and fibre) increase the risk of obesity and many non-communicable diseases, including diabetes, coronary heart disease, and cancers.1, 2, 3, 4 Overall dietary patterns might be more informative about non-communicable disease risk than individual foods or nutrients.5 Many governments have introduced population-based policies aiming to improve dietary patterns and reduce disease burden. These policies have a common core goal (reflected in the WHO Global Strategy on Diet, Physical Activity and Health6) of decreasing added sugar, sodium, and total fat consumption, and increasing intakes of wholegrain cereals, fruits, vegetables, and fibre. Results from the North Karelia project7 showed that such dietary change can contribute to decreased coronary heart disease mortality at the population level. Research in context Evidence before this study

Introduction So-called western dietary patterns (ie, high in saturated fat, cholesterol, sodium, and added sugars; low in fruits, vegetables, and fibre) increase the risk of obesity and many non-communicable diseases, including diabetes, coronary heart disease, and cancers.1, 2, 3, 4 Overall dietary patterns might be more informative about non-communicable disease risk than individual foods or nutrients.5 Many governments have introduced population-based policies aiming to improve dietary patterns and reduce disease burden. These policies have a common core goal (reflected in the WHO Global Strategy on Diet, Physical Activity and Health6) of decreasing added sugar, sodium, and total fat consumption, and increasing intakes of wholegrain cereals, fruits, vegetables, and fibre. Results from the North Karelia project7 showed that such dietary change can contribute to decreased coronary heart disease mortality at the population level. Research in context Evidence before this study We searched PubMed for studies published in English from database inception to Oct 6, 2016, using the search terms “metabolomics OR metabonomics OR metabolic profiling OR metabolomic OR metabonomic OR metabolite profiling” and “dietary intervention OR dietary intake OR dietary pattern” and excluding non-human studies and review, perspective, opinion, comment, and protocol articles. The search identified 58 studies, of which 45 were related to associations between dietary patterns and metabolite profiles. 27 of these studies were related to consumption of dietary patterns rather than supplementation of a standard diet with a specific food. All 27 studies used self-reported dietary data obtained with instruments such as food frequency questionnaires, dietary recall, and diet diaries, which can be prone to misreporting, with under-reporting biased towards unhealthy foods and over-reporting towards fruits and vegetables. Such inadequate dietary reporting in epidemiological surveys has led to conflicting research findings.

uments such as food frequency questionnaires, dietary recall, and diet diaries, which can be prone to misreporting, with under-reporting biased towards unhealthy foods and over-reporting towards fruits and vegetables. Such inadequate dietary reporting in epidemiological surveys has led to conflicting research findings. We did not find any studies that investigated associations between dietary patterns and metabolic profiles in a controlled crossover clinical trial setting. Established dietary biomarkers such as urinary sodium, potassium, and nitrogen track intake of specific nutrients only, and although a few reports have linked specific metabolite profiles to dietary patterns, these studies used self-reported food intake only. Urine and plasma have been shown to contain individual metabolites that are reflective of self-reported food intake data. In this study, we aimed to develop a new approach to assess dietary patterns using proton nuclear magnetic resonance (1H-NMR) spectroscopic profiling of urine. Added value of this study

We did not find any studies that investigated associations between dietary patterns and metabolic profiles in a controlled crossover clinical trial setting. Established dietary biomarkers such as urinary sodium, potassium, and nitrogen track intake of specific nutrients only, and although a few reports have linked specific metabolite profiles to dietary patterns, these studies used self-reported food intake only. Urine and plasma have been shown to contain individual metabolites that are reflective of self-reported food intake data. In this study, we aimed to develop a new approach to assess dietary patterns using proton nuclear magnetic resonance (1H-NMR) spectroscopic profiling of urine. Added value of this study We developed a metabolic profiling strategy, using NMR spectroscopy of urine samples, that can objectively assess dietary profiles. We have shown that urinary metabolite models, developed in a highly controlled environment, can be used to classify free-living people into consumers of a dietary profile associated with low or high risk of non-communicable disease. Previous studies that related dietary patterns to metabolic profiles relied on self-reported dietary data or were performed in a non-controlled setting (eg, a home environment), which increases the possibility of misreporting and non-compliance. Our study was performed in a controlled environment to guarantee compliance and eliminate misreporting, and we validated our model using both internal controlled clinical trial data and external cohort data. This approach has the potential to be used to monitor dietary patterns objectively in population settings without the risk of making false inferences based on data prone to misreporting or non-compliance, which can confound the true effect of a diet on health.

al controlled clinical trial data and external cohort data. This approach has the potential to be used to monitor dietary patterns objectively in population settings without the risk of making false inferences based on data prone to misreporting or non-compliance, which can confound the true effect of a diet on health. Implications of all the available evidence Unhealthy dietary patterns are major risk factors of multiple common diseases, and many countries have local and national health policies that encourage dietary change. However, existing dietary tools are inadequate for assessing responses in dietary behaviour that result from policy change. Implementation of our metabolic profiling approach and analyses of dietary patterns based on urine metabolite profiles might not only enhance understanding of the relations between diet and health but also offer an objective method for dietary screening of large numbers of people. Our metabolic profiling strategy could be used to obtain objective information on adherence to healthy eating programmes aimed at combating obesity and common diseases.

nly enhance understanding of the relations between diet and health but also offer an objective method for dietary screening of large numbers of people. Our metabolic profiling strategy could be used to obtain objective information on adherence to healthy eating programmes aimed at combating obesity and common diseases. A major limitation of nutritional science is the objective assessment of dietary intake in free-living populations. Monitoring of dietary change in national surveys and large prospective studies relies on self-reported food intake using instruments such as food frequency questionnaires, dietary recall, and diet diaries; the prevalence of misreporting with these tools is estimated at 30–88%.8, 9 Compounding this problem, bias in dietary misreporting (with under-reporting biased towards unhealthy foods and over-reporting towards fruits and vegetables10) contributes to data inaccuracy and misinterpretation.11, 12 Moreover, under-reporting of dietary energy intake is particularly common in obese individuals,8 which is a major concern considering the increasing prevalence of obesity worldwide.13

towards unhealthy foods and over-reporting towards fruits and vegetables10) contributes to data inaccuracy and misinterpretation.11, 12 Moreover, under-reporting of dietary energy intake is particularly common in obese individuals,8 which is a major concern considering the increasing prevalence of obesity worldwide.13 Established dietary biomarkers such as urinary sodium, potassium, and nitrogen track intake of specific nutrients only. To our knowledge, no independent, objective method exists for assessing overall dietary patterns in free-living populations. This limitation has led to conflicting research findings,12 partly because of inadequate dietary reporting in epidemiological surveys.14 Urine and plasma have been shown to contain individual metabolites that are reflective of self-reported food intake data.15, 16, 17, 18 We now propose a new approach to assess dietary patterns using proton nuclear magnetic resonance (1H-NMR) spectroscopic profiling. This technology has potential to simultaneously measure hundreds of metabolites, the concentrations of which are affected by food intake.19 If validated, this approach could enhance understanding of the relation between food consumption and disease risk, a concept embedded in the Precision Medicine Initiative.20 Although a few reports15, 16, 17, 18 have linked metabolite profiles to dietary patterns, these studies used self-reported food intake. To address dietary misreporting by free-living populations,8, 9 we hypothesised that metabolic profiles of urine from volunteers exposed to a range of diets based on WHO dietary guidelines6 for the prevention of non-communicable disease risks (obesity, diabetes, and coronary heart disease), under highly controlled conditions, reflect dietary intake and can be used to model and classify dietary patterns of free-living people from large cross-sectional cohorts, without requiring self-reported food intake. In this study, we aimed to quantify the effect of diverse dietary patterns on urinary metabolic profiles in a highly controlled environment.

eflect dietary intake and can be used to model and classify dietary patterns of free-living people from large cross-sectional cohorts, without requiring self-reported food intake. In this study, we aimed to quantify the effect of diverse dietary patterns on urinary metabolic profiles in a highly controlled environment. Methods Study design and participants In this randomised, controlled, crossover trial, we recruited participants from a database of healthy volunteers at the UK National Institute for Health Research (NIHR)/Wellcome Trust Imperial Clinical Research Facility (CRF). These volunteers had previously been screened at the CRF and had expressed an interest in being contacted regarding future research studies. Volunteers in the database who were aged 21–65 years and had a BMI of 20–35 kg/m2 were contacted with a letter of invitation, and those who responded with an interest in participation were screened, initially by email or telephone and subsequently at the CRF. Potential participants were excluded if they had clinically significant illnesses, if they reported weight loss or gain of 3 kg or more in the preceding 2 months, if they were taking prescription medication, if they were current smokers or substance abusers, or if they presented any abnormalities on physical examination, electrocardiography, or screening blood tests. Women were ineligible if they were pregnant or breastfeeding.

or gain of 3 kg or more in the preceding 2 months, if they were taking prescription medication, if they were current smokers or substance abusers, or if they presented any abnormalities on physical examination, electrocardiography, or screening blood tests. Women were ineligible if they were pregnant or breastfeeding. The study was approved by the London–Brent Research Ethics Committee and done in accordance with the Declaration of Helsinki (13/LO/0078). The study protocol is available in the appendix. All participants provided written informed consent. Randomisation and masking Participants were given one dietary intervention during each inpatient stay. The order of diets was randomly assigned across study visits. Randomisation was done by an investigator who was not directly involved in the study, with the use of opaque, sealed, sequentially numbered envelopes that each contained one of the four dietary interventions in a random order. The envelopes were stored securely, away from the trial site, and opened in sequence by an investigator (ESC) as each participant was enrolled. Although participants and investigators could not be masked from the dietary intervention during the study period, investigators analysing the data were masked from the randomisation order.

were stored securely, away from the trial site, and opened in sequence by an investigator (ESC) as each participant was enrolled. Although participants and investigators could not be masked from the dietary intervention during the study period, investigators analysing the data were masked from the randomisation order. Procedures Participants attended the CRF for a 72 h inpatient period on four occasions, separated by at least 5 days (appendix p 10). We chose 3 days (72 h) for the inpatient period because most food-derived metabolites are absorbed and eliminated in urine within 48 h, as evidenced in numerous studies (including other studies done in our laboratories) of the kinetics of absorption, bioavailability, and elimination of several food metabolites contributing to the urinary metabolome.21 The minimum 5 day gap between dietary intervention periods ensured that any possible carryover was minimised. For example, tartaric acid, a marker of grape consumption, was shown to be cleared from the body within a few hours in an excretion kinetics study.22 Similarly, other metabolites associated with diet, such as proline betaine, creatine, and trimethylamine N-oxide (TMAO), are cleared from the body within a few hours.23, 24

r example, tartaric acid, a marker of grape consumption, was shown to be cleared from the body within a few hours in an excretion kinetics study.22 Similarly, other metabolites associated with diet, such as proline betaine, creatine, and trimethylamine N-oxide (TMAO), are cleared from the body within a few hours.23, 24 The Nutrition and Dietetic Research Group at Imperial College London (London, UK; led by GF) developed four dietary interventions with a stepwise variance in concordance with the WHO healthy eating guidelines6—diet 1 was the most concordant with the guidelines, and diet 4 was the least concordant (table 1; appendix p 5). High energy density is an important driver of the association between poor diets and the risk of obesity and diabetes;25 therefore, the diets had a range of energy densities (table 1; appendix p 5). Participants were asked to consume all the food provided and were allowed to drink water as they wished. The expectation to consume all food provided and not to leave the CRF during each visit was fully explained to potential participants before they provided consent to take part in the study. This adherence was monitored strictly: all food was weighed immediately before being given to the participants, and any uneaten food was weighed. Physical activity was also controlled; participants were allowed to engage in only very light physical activity (no more strenuous than walking from their hospital bed to the toilet).

rence was monitored strictly: all food was weighed immediately before being given to the participants, and any uneaten food was weighed. Physical activity was also controlled; participants were allowed to engage in only very light physical activity (no more strenuous than walking from their hospital bed to the toilet). During each 3 day inpatient period, urine was collected daily over three timed periods: morning collection (0900–1300 h; cumulative sample 1), afternoon collection (1300–1800 h; cumulative sample 2), and evening and overnight collection (1800–0900 h; cumulative sample 3). 24 h urine samples were obtained by pooling these samples (appendix p 10). Urine samples were prepared with a pH 7·4 phosphate buffer for 1H-NMR spectroscopy as described previously.26 We analysed samples at 300 K on a 600 MHz spectrometer (Bruker BioSpin, Karlsruhe, Germany) using a standard one-dimensional pulse sequence with water-presaturation.26 Acquisition parameters are shown in the appendix (pp 2–3).

re prepared with a pH 7·4 phosphate buffer for 1H-NMR spectroscopy as described previously.26 We analysed samples at 300 K on a 600 MHz spectrometer (Bruker BioSpin, Karlsruhe, Germany) using a standard one-dimensional pulse sequence with water-presaturation.26 Acquisition parameters are shown in the appendix (pp 2–3). Statistical analysis To the best of our knowledge, this study was the first-in-human trial of metabolic profiling in a controlled feeding setting; therefore, no formal power calculation could be undertaken. However, to provide a basis for sample size calculation, we used data on urinary proline betaine, which we selected as a representative marker for nutritional intake.23 Heinzmann and colleagues23 suggested that urinary concentration of this metabolite would rise by 50 μmol/L with each incremental rise in fruit intake (ie, pieces of fruit) in the experimental setting. With an SD of 40 μmol/L, assuming a power of 0·95 and an alpha of 0·05 to detect a difference of 50 μmol/L, we estimated that we would need 12 volunteers. Because the protocol required a high amount of volunteer time and involvement (12 inpatient days plus travelling time) and volunteers could withdraw from the study, we requested permission to recruit 30 people, with the aim of having a cohort of roughly 20 people. Of note, in two previous dietary studies27, 28 researchers identified individual biomarkers of food intake after controlled feeding having included fewer than 20 participants in each study. All 19 participants who completed the study were included in the analysis.

, with the aim of having a cohort of roughly 20 people. Of note, in two previous dietary studies27, 28 researchers identified individual biomarkers of food intake after controlled feeding having included fewer than 20 participants in each study. All 19 participants who completed the study were included in the analysis. 1H-NMR spectra (16 000 spectral variables) were manually phased and digitised over the range δ0·5–9·5 and imported into MATLAB (release 2014a). A combination of data-driven29 and experimental structural elucidation techniques and spiking-in of chemical standards was used to aid structural identification of diet-discriminatory metabolites. We used global urinary 1H-NMR spectral profiles representing diets 1 and 4 to generate representative metabolite patterns relating to each diet. Global metabolic profiling entails using methods that aim to measure all metabolites, or as many as possible with the assay, in a sample, as opposed to targeted analysis, in which only specific compounds are measured. Because this study is the first of its kind, we did not know a priori which compounds were of interest; therefore, we used global metabolic profiling to capture as much information as possible rather than limit our information to a set of targeted compounds.

eted analysis, in which only specific compounds are measured. Because this study is the first of its kind, we did not know a priori which compounds were of interest; therefore, we used global metabolic profiling to capture as much information as possible rather than limit our information to a set of targeted compounds. We modelled data with partial least squares discriminant analysis (PLS-DA), using Monte Carlo cross-validation (MCCV) to assess model robustness using a total of 1000 individual models; the data were centred and scaled to account for the repeated-measures design. The mean (Tpred) and variance of each predicted sample were estimated using all MCCV models. We then used this MCCV–PLS-DA model to predict 24 h urinary global profiles of volunteers after 3 days of strict adherence to the intermediate diets (ie, diets 2 and 3) without informing the model whether these urinary profiles belonged to diet 2 or diet 3. Day 3 samples were used for modelling because these were timed to be 48 h after starting the dietary intervention and ensured diet homoeostasis. Data from day 1 and day 2 samples served as internal validation data.

, diets 2 and 3) without informing the model whether these urinary profiles belonged to diet 2 or diet 3. Day 3 samples were used for modelling because these were timed to be 48 h after starting the dietary intervention and ensured diet homoeostasis. Data from day 1 and day 2 samples served as internal validation data. Across the 1000 models the mean prediction (Tpred) of each sample was calculated from all models in which the sample was part of the validation set. A positive Tpred indicates that the urinary metabolic profile of the sample resembles more diet 1 than diet 4, and vice versa for a negative Tpred. The variance of Tpred was estimated from the same predictions as were used for calculating the mean. We then calculated the Kernel density estimate by summing the resulting Gaussian distributions of all samples within each group. A p value was calculated for each variable on the basis of 25 bootstrap resamplings of the training data in each of 1000 models to estimate the variance and the mean coefficient across the 1000 models. Spectral variable importance was assessed with the false discovery rate q value, with a value of 0·01 or less as the cutoff for significance.

d for each variable on the basis of 25 bootstrap resamplings of the training data in each of 1000 models to estimate the variance and the mean coefficient across the 1000 models. Spectral variable importance was assessed with the false discovery rate q value, with a value of 0·01 or less as the cutoff for significance. To assess the ability of our model—based on the 24 h urinary collections—to independently predict healthy eating in a free-living population, we used data from the UK cohort (n=225 from a cohort of 499) of the INTERMAP study as our validation dataset. The INTERMAP study30 investigated dietary and other factors associated with blood pressure in 4680 men and women aged 40–59 years from 17 population samples in four countries (China, Japan, UK, and USA). Dietary intake data were obtained from two consecutive multipass 24 h recalls31 on two occasions that were 3 weeks apart on average. For this analysis, we used the 24 h urine sample data, corresponding to the first two multipass 24 h dietary recalls, from the UK cohort.32 We stratified participants into percentile groups (0 to 10th, 45th to 55th, and 90th to 100th) using the Dietary Approaches to Stop Hypertension (DASH) index (appendix p 2),33 which is a tool used for healthy eating assessment in several countries34 and has been used in INTERMAP.

dietary recalls, from the UK cohort.32 We stratified participants into percentile groups (0 to 10th, 45th to 55th, and 90th to 100th) using the Dietary Approaches to Stop Hypertension (DASH) index (appendix p 2),33 which is a tool used for healthy eating assessment in several countries34 and has been used in INTERMAP. Additionally, to assess the ability of our model to inform a non-UK dataset, we used data from a healthy omnivorous cohort of 66 participants recruited and phenotyped at the University of Copenhagen (Copenhagen, Denmark) for our external validation dataset (appendix p 7). We calculated DASH scores for these participants on the basis of their 4 day dietary records according to the quintiles defined in the appendix (pp 2, 6). In this cohort, metabolite profiling was done on spot urine samples collected after the first morning void following a 10 h overnight fast. Therefore, we mapped these samples to models derived from cumulative sample 1 (morning collection) in our trial. The method used to model spot urine samples is provided in the appendix (pp 4, 14).

metabolite profiling was done on spot urine samples collected after the first morning void following a 10 h overnight fast. Therefore, we mapped these samples to models derived from cumulative sample 1 (morning collection) in our trial. The method used to model spot urine samples is provided in the appendix (pp 4, 14). To account for differences in urine osmolality, we normalised all spectra from our study cohort and the two validation cohorts using Probabilistic Quotient Normalization35 to the median spectrum of diets 1 and 4 combined. This procedure corrects the metabolite concentrations for differences in dilution across samples. Such differences can arise from different intakes of water or liquids between participants (causing differences in metabolite concentrations) and from different amounts of foods consumed (eg, high caloric intakes). Therefore, any effect of these potential confounders is attenuated by the normalisation procedure. We used the Skillings-Mack and Kruskal-Wallis tests, as appropriate, to assess differences among multiple groups, and non-parametric post-hoc (Wilcoxon's signed rank and rank sum) tests to determine pairwise differences. p values from post-hoc tests were adjusted for multiple testing with Hommel's adjustment. More details on the statistical analysis are given in the appendix (pp 3–4). All statistical analyses were done in MATLAB. This trial was registered on the NIHR UK clinical trial gateway and with ISRCTN, number ISRCTN43087333.

We used the Skillings-Mack and Kruskal-Wallis tests, as appropriate, to assess differences among multiple groups, and non-parametric post-hoc (Wilcoxon's signed rank and rank sum) tests to determine pairwise differences. p values from post-hoc tests were adjusted for multiple testing with Hommel's adjustment. More details on the statistical analysis are given in the appendix (pp 3–4). All statistical analyses were done in MATLAB. This trial was registered on the NIHR UK clinical trial gateway and with ISRCTN, number ISRCTN43087333. 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. IG-P, JMP, EH, and GF had full access to all the data in the study, and the corresponding authors (GF and EH) had final responsibility for the decision to submit for publication. Results Of 352 individuals in the database of healthy volunteers at the NIHR/Wellcome Trust Imperial CRF, we contacted 300 who were eligible for the study with a letter of invitation between Aug 13, 2013, and May 18, 2014 (52 were ineligible on the basis of age or BMI). 78 individuals responded to the invitation and, after screening, 20 remained eligible and enrolled into the study (figure 1). Between Oct 2, 2013, and July 29, 2014, 19 participants completed the four inpatient periods and consumed all the food provided; their baseline characteristics are shown in table 2.

age or BMI). 78 individuals responded to the invitation and, after screening, 20 remained eligible and enrolled into the study (figure 1). Between Oct 2, 2013, and July 29, 2014, 19 participants completed the four inpatient periods and consumed all the food provided; their baseline characteristics are shown in table 2. MCCV–PLS-DA models of 24 h urine spectra showed systematic differences between metabolic phenotypes of diets 1 and 4 that were reflected in both the metabolic profile (figure 2) and predicted scores (figure 3). From a total of 486 peaks (appendix p 3) in the mean 1H-NMR spectrum, 19 identified metabolites were present in significantly higher concentrations in urine after consumption of diet 1 than after diet 4, and nine metabolites were present in significantly increased concentrations after consumption of diet 4 (figure 2B; appendix p 8). For example, all 19 individuals had consistent, significant changes in excretion of 28 metabolites (figure 2B), including metabolites with well known dietary associations—hippurate (a urinary marker of fruit and vegetable intake), carnitine (a marker of red meat consumption), and tartrate (a marker of grape intake; figure 2C–E). Substantial between-person variability could be seen in concentrations of hippurate (figure 2C) and carnitine (figure 2D), but the direction of association remained the same. For tartaric acid, between-person variability was also apparent, but because it is a quantitative biomarker of grape consumption22 there was almost no between-person variation after consumption of diet 4, which did not contain any grape-derived products (appendix p 5).

ut the direction of association remained the same. For tartaric acid, between-person variability was also apparent, but because it is a quantitative biomarker of grape consumption22 there was almost no between-person variation after consumption of diet 4, which did not contain any grape-derived products (appendix p 5). Data from the MCCV–PLSA-DA model obtained from the analysis of 24 h urine samples from diets 1 and 4 were used to predict 24 h global urinary metabolic profiles of diets 2 and 3 (figure 3A). We found a significant stepwise increase in predicted scores from diet 4 (negative scores) to diet 1 (positive scores; Skillings-Mack test p=7·21 × 10−9; figure 3B). The urinary metabolic profiles generated from the urine samples obtained after diet 3 clustered next to those obtained after diet 4, and those from diet 2 clustered adjacent to those from diet 1. Additionally, these metabolite patterns were reproducible in 24 h samples obtained from the 19 volunteers on days 1 and 2 of the 72 h inpatient stay (appendix p 11).

from the urine samples obtained after diet 3 clustered next to those obtained after diet 4, and those from diet 2 clustered adjacent to those from diet 1. Additionally, these metabolite patterns were reproducible in 24 h samples obtained from the 19 volunteers on days 1 and 2 of the 72 h inpatient stay (appendix p 11). In our first validation dataset (INTERMAP UK cohort), urinary metabolic profiles of samples from the group with high DASH score (90th to 100th percentile; n=67) clustered near those from the diet 1 samples, whereas metabolic profiles of samples from the group with low DASH score (0 to 10th percentile; n=67) clustered next to those from the diet 4 samples (figure 4A). The urinary metabolic profiles from the group with intermediate DASH scores (45th to 55th percentile; n=91) clustered between the two extreme categories. Although the Kernel density estimate plot showed some overlap between the three groups—which was expected because of dietary misreporting, estimated to be 22·4% from men and 30·9% from women in the INTERMAP UK cohort36—a significant linear association was seen between DASH scores and predicted scores (Kruskal-Wallis test p=5·10 × 10−6; figure 4B). Post-hoc Wilcoxon rank sum test corrected by Hommel's method confirmed that all pairwise comparisons differed significantly (figure 4). In addition to global metabolite profiles, we quantified specific metabolites from the model known to be associated with foods linked to healthy eating—ie, hippurate (fruits and vegetables), 4-hydroxyhippurate (fruits), and S-methyl-L-cysteine-sulfoxide (cruciferous vegetables)—and found that these metabolites were present in significantly higher concentrations in urine samples from INTERMAP participants with high DASH scores than in samples from those with low DASH scores (appendix p 9). However, S-methyl-L-cysteine-sulfoxide (p=0·19) and hippurate (p=0·051) concentrations did not differ significantly between the groups with intermediate DASH scores and high DASH scores; similarly; concentrations of hippurate (p=0·096) and 4-hydroxyhippurate (p=0·15) did not differ significantly between the groups with low DASH scores and intermediate DASH scores. The overall findings indicated that the global urinary metabolic profile measured by 1H-NMR spectroscopy (χ2 24·37, p<0·0001) was a more accurate predictor of dietary patterns than single markers of individual foods consumed (χ2 12·71–21·09, p=0·0017 to p<0·0001).

oups with low DASH scores and intermediate DASH scores. The overall findings indicated that the global urinary metabolic profile measured by 1H-NMR spectroscopy (χ2 24·37, p<0·0001) was a more accurate predictor of dietary patterns than single markers of individual foods consumed (χ2 12·71–21·09, p=0·0017 to p<0·0001). We classified seven INTERMAP participants as metabolic outliers (figure 4B). On detailed examination of their dietary records, one participant (who had low DASH score but positive predicted score) was considered a misreporter because high amounts of proline betaine were found in the urine, but no citrus fruits or other dietary sources of proline betaine were recorded. The urine of another outlier from the group with intermediate DASH scores contained very high amounts of N-methylnicotinate (a vitamin B3 derivative), which contributed greatly to the classification of this sample as close to diet 1. This individual had consumed very high amounts of coffee, which is rich in niacin (vitamin B3), a precursor of N-methylnicotinate, accounting for the high level of urinary excretion. The urine of the remaining five outliers (DASH scores 24, 24, 25, 25, and 30) contained very high amounts of paracetamol, the metabolite signals (specifically sulphate and glucuronide) of which overlap with phenylacetylglutamine signals; the fact that phenylacetylglutamine was associated with diet 4 could help to explain the misclassification.

g five outliers (DASH scores 24, 24, 25, 25, and 30) contained very high amounts of paracetamol, the metabolite signals (specifically sulphate and glucuronide) of which overlap with phenylacetylglutamine signals; the fact that phenylacetylglutamine was associated with diet 4 could help to explain the misclassification. For the external validation cohort (ie, the Danish cohort; n=66), the metabolite profiles aligned with those of diets 1 and 2 (appendix p 12), which was confirmed by the high DASH score of the cohort (median 28·5, range 20–36). The urinary metabolic profiles were again associated with the dietary profiles (p<0·0001). Although our method worked best with 24 h urine models (appendix pp 11, 13), the model created for spot samples from the Danish cohort based on cumulative urine samples also showed good stratification of metabolic phenotypes according to diet. Discussion In this proof-of-principle study, we showed notable differences in urinary metabolic profiles in a controlled feeding condition in which participants consumed four defined diets differing in compliance to the WHO-recommended healthy diet. We then showed, in two independent epidemiological datasets, that these metabolic profile patterns have the potential to classify the dietary intake of free-living individuals. We concluded that this novel application of metabolic phenotyping at the population level has the potential to provide objective measures of adherence to dietary recommendations, without the use of dietary surveys, which are known to be subject to misreporting, incompleteness, and bias.

take of free-living individuals. We concluded that this novel application of metabolic phenotyping at the population level has the potential to provide objective measures of adherence to dietary recommendations, without the use of dietary surveys, which are known to be subject to misreporting, incompleteness, and bias. We showed that urinary concentrations of biomarkers from individual healthy foods—eg, hippurate (a marker of fruit and vegetable consumption), (N-acetyl-)S-methyl-L-cysteine-sulfoxide (cruciferous vegetables), dimethylamine and TMAO (fish), and 1-methylhistidine and 3-methylhistidine (oily fish and chicken)—were significantly higher after consumption of diet 1 than after diet 4, reflecting increased intakes of fruits, cruciferous vegetables, salmon, and chicken that were provided as part of diet 1. TMAO is a cryoprotectant in freshwater and saltwater fish, and urinary concentrations of TMAO are associated with recent fish intake;37 therefore, high urinary TMAO concentrations can be associated with healthy diets that are rich in fish. However, gut bacteria can synthesise TMAO from choline and hence high urinary and plasma TMAO concentrations can also originate from red meat consumption, which is generally associated with adverse health outcomes. Indeed, high concentrations of TMAO in plasma and urine have been associated with cardiovascular and renal disease, respectively.38 Thus, the global pattern of metabolites, which reflects the totality of the diet, is more important in indicating dietary patterns than are individual biomarkers.

adverse health outcomes. Indeed, high concentrations of TMAO in plasma and urine have been associated with cardiovascular and renal disease, respectively.38 Thus, the global pattern of metabolites, which reflects the totality of the diet, is more important in indicating dietary patterns than are individual biomarkers. Although findings from previous studies have shown that metabolic profiling could be used to identify specific dietary biomarkers associated with diet, our new approach allowed differentiation of the dietary interventions on the basis of global urinary metabolic profiles, while also reflecting specific chemicals found in individual foods or beverages. We used the characteristic metabolic profiles of diets 1 and 4 to predict the dietary profiles of individuals consuming diets between these two extremes (ie, diets 2 and 3). A clear separation was seen across the metabolite profiles of the four diets. We also showed significant stepwise differences in metabolite concentrations from diet 1 to diet 4. Changes in individual diet-associated metabolites were consistent across the dietary interventions.

en these two extremes (ie, diets 2 and 3). A clear separation was seen across the metabolite profiles of the four diets. We also showed significant stepwise differences in metabolite concentrations from diet 1 to diet 4. Changes in individual diet-associated metabolites were consistent across the dietary interventions. To assess the feasibility of using metabolic profiling as an objective and unbiased approach to assess dietary patterns in a population setting, we used the INTERMAP UK cohort as a validation dataset to investigate whether free-living individuals with high dietary DASH scores (associated with a reduced risk of non-communicable diseases) and low DASH scores (associated with a high risk of non-communicable diseases33, 34) could be distinguished by use of the metabolite profiles from the controlled feeding experiment. We showed substantial clustering of predicted scores of the metabolic profiles according to the DASH scores, suggesting that dietary patterns could be predicted by interrogating the whole urinary metabolite profile (not just individual metabolites). This approach was based on the concept that differences in dietary intake are reflected in the relative concentrations of many hundreds of urinary metabolites.39 Metabolic analysis revealed that urine from participants with high DASH scores in the INTERMAP cohort contained an increased abundance of biomarkers associated with fruit and vegetable intake—including hippurate, 4-hydroxyhippurate, and S-methyl-L-cysteine-sulfoxide—compared with urine from participants on a diet with low DASH scores. The dietary intervention models were able to predict dietary patterns regardless of the specific dietary components. For example, the citrus fruit marker proline betaine was present in increased abundance in samples from the group with high DASH score (appendix p 9) even though citrus fruits were not provided as part of the dietary interventions. This finding suggests that the derived metabolite profiles are reflective of a wider range of healthy diets than those used in our trial.

ine was present in increased abundance in samples from the group with high DASH score (appendix p 9) even though citrus fruits were not provided as part of the dietary interventions. This finding suggests that the derived metabolite profiles are reflective of a wider range of healthy diets than those used in our trial. Although the energy intakes across the three groups in the INTERMAP cross-sectional study stratified for DASH score did not differ significantly (despite a large spread around the median), the participants with low DASH scores tended to have higher energy intakes than did those with high DASH scores (appendix p 6). However, the potential confounding effects of energy intake and dilution on the urinary metabolome were attenuated by the normalisation procedure applied.

te a large spread around the median), the participants with low DASH scores tended to have higher energy intakes than did those with high DASH scores (appendix p 6). However, the potential confounding effects of energy intake and dilution on the urinary metabolome were attenuated by the normalisation procedure applied. Dietary intake data from free-living participants are subject to dietary misreporting; therefore, this challenge had to be addressed in the validation of our model. The INTERMAP epidemiological data have a track record of low misreporting rates,36 thus allowing us to address this limitation by exploring food intake records of so-called metabolic outliers. Because of potential misreporting issues, we used a second cohort (the Danish cohort) to further validate our model. Since only data from spot urine samples after the first void were available for this cohort, we matched the timing of this spot urine with the morning urine collection in our trial. Our analysis showed that the metabolite profiles of the Danish cohort resembled those of our participants after consumption of diets 1 and 2, and concurred with the dietary analysis showing that this healthy-eating population had a high median DASH score (appendix p 7). These findings suggest that our model is robust across different populations.

metabolite profiles of the Danish cohort resembled those of our participants after consumption of diets 1 and 2, and concurred with the dietary analysis showing that this healthy-eating population had a high median DASH score (appendix p 7). These findings suggest that our model is robust across different populations. Although results from several studies33, 34 have shown that the DASH score is positively associated with health, so far the DASH score has not generally been adopted in public health policies. We based our experimental dietary interventions on internationally accepted healthy eating guidelines.6 Several other dietary scoring methods—such as the alternate healthy eating index40 and Mediterranean41 and alternate Mediterranean scores42—exist, but we chose to apply the DASH index to all the diets (the ones in our trial and the diets consumed by the INTERMAP and Danish cohorts) to provide a common scale in which the range in absolute numbers is not limited. Additionally, the DASH score has been shown to be associated with cardiac risk.43

anean scores42—exist, but we chose to apply the DASH index to all the diets (the ones in our trial and the diets consumed by the INTERMAP and Danish cohorts) to provide a common scale in which the range in absolute numbers is not limited. Additionally, the DASH score has been shown to be associated with cardiac risk.43 It is important to understand the difference in use of spot urine samples compared with 24 h urine samples. Although spot urine samples are commonly used in epidemiological studies, they provide only the snapshot of the urinary metabolome for a very specific sampling time, whereas 24 h samples provide a time-averaged window on metabolism encompassing diurnal variation and other lifestyle-related fluctuations. Global metabolic changes are thus more difficult to assess in spot samples; use of the 24 h model to predict spot samples taken 2 h after lunch, 5 h after lunch, and 2 h after dinner showed that variability was greater in the spot sample predictions (appendix p 13) than in 24 h samples from days 1, 2, and 3 of the trial (appendix p 12). Where possible, use of 24 h urine samples rather than spot samples is advised for measurement of metabolic changes, particularly when excretion kinetics are not known a priori, partly because spot urine samples vary more in dilution. However, where only spot urine samples are available because of study limitations, such as in the case of the Danish cohort, we found that the time-matched cumulative sample obtained between meals predicted the spot samples with more accuracy than did the 24 h model (appendix p 14), which is consistent with previous work showing accurate quantification of grape intake based on both 24 h samples and cumulative samples matching the excretion kinetics window.22 For this reason, we compared the spot samples, taken after the first morning void in the Danish cohort with the model derived from cumulative sample 1 (after breakfast to before lunch) (appendix p 11) rather than the 24 h model, because the sampling time was better matched.

les matching the excretion kinetics window.22 For this reason, we compared the spot samples, taken after the first morning void in the Danish cohort with the model derived from cumulative sample 1 (after breakfast to before lunch) (appendix p 11) rather than the 24 h model, because the sampling time was better matched. Our study has several limitations. The metabolic profiling trial and the models derived from it were based on limited types of foods. However, even with this narrow range of foods, the model clearly classified the diets from the corresponding urinary metabolic profiles. To the best of our knowledge, this study is the first of its kind to use global urinary metabolic profiles to objectively assess dietary patterns. However, we recognise that our approach will benefit from additional testing in a wider range of populations, including those of disease groups and diverse ethnic origins. Additionally, the likely misclassification of individuals consuming very high amounts of paracetamol and coffee in our model could be addressed in future work by using further analytical chemical profiling methods to avoid peak overlap in the urinary spectra. We anticipate that models built from a dataset based on a wider variety of foods would give more robust predictions. In this first-in-human study, we planned to test the hypothesis that adherence to different dietary intake patterns could be assessed from analysis of the urinary metabolome. With this aim in mind, we considered it was important that our controlled feeding study was done in healthy individuals to minimise the risk that metabolite profiles in urine might be confounded by pre-existing disease or concurrent medications, both of which are known to affect the urinary metabolic profile. The performance of the model in populations with non-communicable diseases would be an obvious follow-up. Despite these limitations, our models were predictive of dietary patterns in two well characterised free-living cohorts, thus providing proof of principle that this new approach has value as an objective means to monitor dietary patterns at the population level. Our model showed a clear separation between diets 1 and 4 in the global urinary metabolic profile and individual food-related metabolites, even though energy provided by the diets was not matched to individual estimated requirements.

h has value as an objective means to monitor dietary patterns at the population level. Our model showed a clear separation between diets 1 and 4 in the global urinary metabolic profile and individual food-related metabolites, even though energy provided by the diets was not matched to individual estimated requirements. Existing methods for dietary assessment—eg, dietary diaries (which require coding and data entry), food frequency questionnaires, and dietary recalls—are expensive (our own estimated cost is £60 for the complete analysis, including quality control of a 1 day dietary recall by an experienced nutritionist or dietitian). Accurate reporting assumes knowledge of food ingredients and can involve complex decoding of food product labels. Translating our study method into clinical practice is a cost-effective and time-effective alternative for objective dietary assessment—the cost is roughly £20 per sample for robust analysis by NMR, which takes less than 5 min per sample to run (excluding sample preparation). This metabolic phenotyping strategy circumvents bias caused by misreporting and the fact that dietary scoring methods, such as DASH scores, can be based on arbitrary thresholds. Moreover, similar scores can arise from different dietary patterns. The global metabolic profile, consisting of both food-specific and general food-related metabolites, is an unbiased approach and does not rely on arbitrary cutoffs. Additionally, the global metabolic profile also allows testing of whether metabolites associated with reported foods are found in the urine and therefore the accuracy of dietary records.

onsisting of both food-specific and general food-related metabolites, is an unbiased approach and does not rely on arbitrary cutoffs. Additionally, the global metabolic profile also allows testing of whether metabolites associated with reported foods are found in the urine and therefore the accuracy of dietary records. In conclusion, we showed that urinary metabolic profiles developed in a controlled environment have the potential to be used to assess adherence to dietary patterns in free-living populations without the need to collect dietary data. The extension and application of this strategy at the population level offer a potential step-change for public health nutrition because it provides an objective method to survey dietary intakes. Implementation of this strategy for urine-based dietary pattern analysis might enhance our understanding of the relation between diet and health and might improve clinical nutrition practice by providing health professionals with objective information on adherence to healthy eating guidelines. Supplementary Material Supplementary appendix

In conclusion, we showed that urinary metabolic profiles developed in a controlled environment have the potential to be used to assess adherence to dietary patterns in free-living populations without the need to collect dietary data. The extension and application of this strategy at the population level offer a potential step-change for public health nutrition because it provides an objective method to survey dietary intakes. Implementation of this strategy for urine-based dietary pattern analysis might enhance our understanding of the relation between diet and health and might improve clinical nutrition practice by providing health professionals with objective information on adherence to healthy eating guidelines. Supplementary Material Supplementary appendix Acknowledgments This Article presents independent research funded by the UK National Institute for Health Research (NIHR) and Medical Research Council (MRC). The views expressed are those of the authors and not necessarily those of the UK National Health Service (NHS), the NIHR, or the UK Department of Health. IG-P is supported by a NIHR postgraduate research fellowship (NIHR-PDF-2012-05-456) and a Wellcome Trust Value In People award. GF is supported by an NIHR Senior Investigator award. JCM, JD, MB, GF, and EH are supported by an MRC grant entitled Metabolomics for Monitoring Dietary Exposure (MR/J010308/1). This study was supported by the NIHR/Wellcome Trust Imperial Clinical Research Facility. The Section of Investigative Medicine is funded by grants from the MRC, UK Biotechnology and Biological Sciences Research Council, NIHR, and an Integrative Mammalian Biology (IMB) Capacity Building Award. The INTERMAP and INTERMAP metabonomic study are supported by the US National Heart, Lung, and Blood Institute (R01-HL50490 and R01-HL84228). INTERMAP data collection was also supported by Chicago Health Research Foundation and by national agencies in China, Japan (the Ministry of Education, Science, Sports, and Culture, Grant-in-Aid for Scientific Research (A), number 090357003), and UK. PE acknowledges support from the MRC–Public Health England Centre for Environment & Health, the NIHR Biomedical Research Centre at Imperial College Healthcare NHS Trust and Imperial College London, the NIHR Health Protection Research Unit on Health Impact of Environmental Hazards, and the UK MEDical BIOinformatics partnership—aggregation, integration, visualisation, and analysis of large, complex data (UK MED-BIO), which is supported by the MRC (MR/L01632X/1). PE is an NIHR Senior Investigator. The Novo Nordisk Foundation Center for Basic Metabolic Research is an independent research centre at the University of Copenhagen (Copenhagen, Denmark) partly funded by an unrestricted donation from the Novo Nordisk Foundation. We also thank the MRC–NIHR National Phenome Centre for facilitating this and related work.

. The Novo Nordisk Foundation Center for Basic Metabolic Research is an independent research centre at the University of Copenhagen (Copenhagen, Denmark) partly funded by an unrestricted donation from the Novo Nordisk Foundation. We also thank the MRC–NIHR National Phenome Centre for facilitating this and related work. Contributors IG-P, EH, JCM, JD, and GF designed the study. IG-P and ESC did the clinical trial. IG-P did the analytical research. JMP analysed data and did statistical analyses. IG-P and JMP created the figures. IG-P, JMP, ESC, RG, JKN, JCM, MB, JD, JS, PE, EH, and GF wrote the report. THH, HV, TH, and OP (Danish cohort), and PE and JS (INTERMAP senior investigators) provided data for the free-living population samples. All authors read and revised the report, and approved the final submitted version. GF assumes responsibility for the completeness and accuracy of the data and analyses, and for adherence to the study protocol.

h cohort), and PE and JS (INTERMAP senior investigators) provided data for the free-living population samples. All authors read and revised the report, and approved the final submitted version. GF assumes responsibility for the completeness and accuracy of the data and analyses, and for adherence to the study protocol. Declaration of interests IG-P reports grants from the UK National Institute of Health Research (NIHR) and Wellcome Trust during the conduct of the study; and personal fees from Metabometrix outside the submitted work. MB and JD report grants from the UK Medical Research Council (MRC) during the conduct of the study and have worked on the Cook to Health project (of which Groupe SEB is a collaborator and partly funded by EIT-Health) and the FACET project (of which Abbott, Spain, is a collaborator and partly funded by EIT-Health) outside the submitted work. JS and PE report grants from the US National Institutes of Health during the conduct of the study. JKN is non-executive Director of Metabometrix and reports grants from Nestlé outside the submitted work. JCM reports grants from MRC during the conduct of the study. EH reports grants from NIHR and MRC during the conduct of the study; and is non-executive Director of Metabometrix outside the submitted work. GF reports grants from NIHR and MRC during the conduct of the study; is lead for the Imperial Nestlé Collaboration outside the submitted work; and reports personal fees from Unilever outside the submitted work. All other authors declare no competing interests. Figure 1 Trial profile

Declaration of interests IG-P reports grants from the UK National Institute of Health Research (NIHR) and Wellcome Trust during the conduct of the study; and personal fees from Metabometrix outside the submitted work. MB and JD report grants from the UK Medical Research Council (MRC) during the conduct of the study and have worked on the Cook to Health project (of which Groupe SEB is a collaborator and partly funded by EIT-Health) and the FACET project (of which Abbott, Spain, is a collaborator and partly funded by EIT-Health) outside the submitted work. JS and PE report grants from the US National Institutes of Health during the conduct of the study. JKN is non-executive Director of Metabometrix and reports grants from Nestlé outside the submitted work. JCM reports grants from MRC during the conduct of the study. EH reports grants from NIHR and MRC during the conduct of the study; and is non-executive Director of Metabometrix outside the submitted work. GF reports grants from NIHR and MRC during the conduct of the study; is lead for the Imperial Nestlé Collaboration outside the submitted work; and reports personal fees from Unilever outside the submitted work. All other authors declare no competing interests. Figure 1 Trial profile Figure 2 Associations of urinary metabolites with diets 1 and 4 in 19 participants

Declaration of interests IG-P reports grants from the UK National Institute of Health Research (NIHR) and Wellcome Trust during the conduct of the study; and personal fees from Metabometrix outside the submitted work. MB and JD report grants from the UK Medical Research Council (MRC) during the conduct of the study and have worked on the Cook to Health project (of which Groupe SEB is a collaborator and partly funded by EIT-Health) and the FACET project (of which Abbott, Spain, is a collaborator and partly funded by EIT-Health) outside the submitted work. JS and PE report grants from the US National Institutes of Health during the conduct of the study. JKN is non-executive Director of Metabometrix and reports grants from Nestlé outside the submitted work. JCM reports grants from MRC during the conduct of the study. EH reports grants from NIHR and MRC during the conduct of the study; and is non-executive Director of Metabometrix outside the submitted work. GF reports grants from NIHR and MRC during the conduct of the study; is lead for the Imperial Nestlé Collaboration outside the submitted work; and reports personal fees from Unilever outside the submitted work. All other authors declare no competing interests. Figure 1 Trial profile Figure 2 Associations of urinary metabolites with diets 1 and 4 in 19 participants Data from the third 24 h urine collection are shown; data from the first and second 24 h urine collection are shown in the appendix (p 11). (A) Mean 600 MHz 1H-NMR spectrum of the 19 participants. (B) Manhattan plot showing −log10(q) × sign of regression coefficient (β) of the MCCV–PLS-DA model for the 16 000 spectral variables. A p value was calculated for each variable on the basis of 25 bootstrap resamplings of the training data in each of 1000 models to estimate the variance. Red peaks represent the 19 metabolites excreted in higher amounts after diet 1 and blue peaks represent the nine metabolites excreted in higher amounts after diet 4. The two horizontal lines indicate the cutoffs for the false discovery rate on the log10 scale. (C–E) Metabolite concentrations for 19 participants after following diet 1 and diet 4 for (C) hippurate (a urinary marker of fruit and vegetable consumption; number 24 in part A), (D) carnitine (a marker of red meat consumption; number 11 in part A), and (E) tartrate (a marker of grape intake; number 25 in part A). 1H-NMR=proton nuclear magnetic resonance. AU=arbitrary unit. MCCV=Monte Carlo cross-validation. PLS-DA=partial least squares discriminant analysis.

consumption; number 24 in part A), (D) carnitine (a marker of red meat consumption; number 11 in part A), and (E) tartrate (a marker of grape intake; number 25 in part A). 1H-NMR=proton nuclear magnetic resonance. AU=arbitrary unit. MCCV=Monte Carlo cross-validation. PLS-DA=partial least squares discriminant analysis. Figure 3 The MCCV–PLS-DA model of metabolic patterns of the four diets for 19 participants Data from the third 24 h urine collection. (A) Kernel density estimate of the predicted scores (Tpred) for the four diets. (B) Mean predicted score for individuals' spectra after following the diets. (C) Tpred of the four diets. Box and whiskers plots indicate median with 25th and 75th percentiles (boxes), interval endpoints (notches of boxes), and 1·5 × IQR above or below the 75th and 25th percentiles (whiskers); points are outliers. Post-hoc Wilcoxon's signed rank test for pairwise differences (adjusted for multiple testing with Hommel's method) gave the following p values: diet 1 vs diet 2 p=6·71 × 10−4; diet 2 vs diet 3 p=5·04 × 10−4; diet 3 vs diet 4 p=1·96 × 10−1; diet 1 vs diet 3 p=3·05 × 10−5; diet 2 vs diet 4 p=4·58 × 10−5; diet 1 vs diet 4 p=3·05 × 10−5. MCCV=Monte Carlo cross-validation. PLS-DA=partial least squares discriminant analysis. Figure 4 Applicability of our model to predict adherence to diverse diets in the INTERMAP UK cohort

Data from the third 24 h urine collection. (A) Kernel density estimate of the predicted scores (Tpred) for the four diets. (B) Mean predicted score for individuals' spectra after following the diets. (C) Tpred of the four diets. Box and whiskers plots indicate median with 25th and 75th percentiles (boxes), interval endpoints (notches of boxes), and 1·5 × IQR above or below the 75th and 25th percentiles (whiskers); points are outliers. Post-hoc Wilcoxon's signed rank test for pairwise differences (adjusted for multiple testing with Hommel's method) gave the following p values: diet 1 vs diet 2 p=6·71 × 10−4; diet 2 vs diet 3 p=5·04 × 10−4; diet 3 vs diet 4 p=1·96 × 10−1; diet 1 vs diet 3 p=3·05 × 10−5; diet 2 vs diet 4 p=4·58 × 10−5; diet 1 vs diet 4 p=3·05 × 10−5. MCCV=Monte Carlo cross-validation. PLS-DA=partial least squares discriminant analysis. Figure 4 Applicability of our model to predict adherence to diverse diets in the INTERMAP UK cohort (A, B) Kernel density estimates of the predicted scores (Tpred) of diet 1, diet 4, and the INTERMAP UK cohort stratified by DASH scores. Dots and squares represent participants from the study cohort, and crosses represent individuals from the INTERMAP UK validation cohort. (C) Tpred of the INTERMAP UK cohort. Box and whiskers plots indicate median with 25th and 75th percentiles (boxes), interval endpoints (notches of boxes), and 1·5 × IQR above or below the 75th and 25th percentiles (whiskers). Crosses indicate outliers—ie, if the predicted values lie outside 1·5 times of the IQR (25th to 75th percentile), corresponding to points lying outside 2·7σ (roughly 0·993 of a normal distribution) either side of the mean. Post-hoc Wilcoxon's signed rank test for pairwise differences (adjusted for multiple testing with Hommel's method) gave the following p values: 0 to 10th percentile vs 45th to 55th percentile p=2·32 × 10−2; 45th to 55th percentile vs 90th to 100th percentile p=4·31 × 10−3; 0 to 10th percentile vs 90th to 100th percentile p=3·53 × 10−6. DASH=Dietary Approaches to Stop Hypertension.

adjusted for multiple testing with Hommel's method) gave the following p values: 0 to 10th percentile vs 45th to 55th percentile p=2·32 × 10−2; 45th to 55th percentile vs 90th to 100th percentile p=4·31 × 10−3; 0 to 10th percentile vs 90th to 100th percentile p=3·53 × 10−6. DASH=Dietary Approaches to Stop Hypertension. Table 1 Macronutrient content and characteristics of the dietary interventions Diet 1 Diet 2 Diet 3 Diet 4 Energy (kcal) 2260 2259 2427 2490 Energy density (kcal/g) 1·2 1·5 1·6 1·9 Proportion of protein 24% 22% 16% 13% Proportion of carbohydrate 51% 51% 46% 44% Total sugar (g) 14 18 22 25 Proportion of fat 23% 24% 35% 42% Saturated fatty acids (g) 5 7 19 20 Monounsaturated fatty acids (g) 8 6 14 12 Polyunsaturated fatty acids (g) 8 5 4 2 Total trans fatty acids (g) 0·5 0·5 1 1 Fibre (g) 45·9 32·1 31·5 13·6 Sodium (mg) 2367 2261 3812 3066 Fruit and vegetables (g) 600 300 180 100 DASH score 37 30 24 11 Specific diet information (foods consumed at specific times) is shown in the appendix (p 5). DASH=Dietary Approaches to Stop Hypertension. Table 2 Baseline characteristics

Diet 1 Diet 2 Diet 3 Diet 4 Energy (kcal) 2260 2259 2427 2490 Energy density (kcal/g) 1·2 1·5 1·6 1·9 Proportion of protein 24% 22% 16% 13% Proportion of carbohydrate 51% 51% 46% 44% Total sugar (g) 14 18 22 25 Proportion of fat 23% 24% 35% 42% Saturated fatty acids (g) 5 7 19 20 Monounsaturated fatty acids (g) 8 6 14 12 Polyunsaturated fatty acids (g) 8 5 4 2 Total trans fatty acids (g) 0·5 0·5 1 1 Fibre (g) 45·9 32·1 31·5 13·6 Sodium (mg) 2367 2261 3812 3066 Fruit and vegetables (g) 600 300 180 100 DASH score 37 30 24 11 Specific diet information (foods consumed at specific times) is shown in the appendix (p 5). DASH=Dietary Approaches to Stop Hypertension. Table 2 Baseline characteristics Data (n=19) Sex Male 10 (53%) Female 9 (47%) Age (years) 55·8 (12·6; 29–65) Ethnic origin White 18 (95%) Asian 1 (5%) Weight (kg) 74·5 (12·5; 52·8–107·9) BMI (kg/m2) 25·6 (3·2; 21·1–33·3) Energy expenditure (kcal/day)* 2099 (351; 1668–2995) Glucose (mmol/L)† 4·8 (0·4; 4·1–5·4) HbA1c (%)† 5·5% (0·1, 5·1–5·8) HbA1c (mmol/mol)† 36·4 (0·9; 32–40) Triglycerides (mmol/L)‡ 0·9 (0·3; 0·5–1·4) Cholesterol (mmol/L)‡ Total 5·1 (0·7; 3·9–6·1) LDL 3·1 (0·7; 1·7–4·2) HDL 1·6 (0·4; 0·9–2·6) Liver function tests (IU/L)‡ Alanine transaminase 21·2 (7·4; 12·3–40·0) Aspartate transaminase 19·5 (3·2; 15·0–24·3) Data are n (%) or mean (SD; range). IU=international units. * Estimated with a physical activity correction of 1·4 in all participants (appendix p 2). † From plasma samples. ‡ From serum samples.

Introduction Intensive insulin therapy is the standard of care in the management of type 1 diabetes.1 Although modern insulin therapy has led to a reduction in the frequency of severe hypoglycaemic events,2 tight glycaemic control remains a predisposing factor to hypoglycaemia and its effect is amplified by duration of the disease.3 Recurrent exposure to hypoglycaemia might lead to attenuated counter-regulatory response to subsequent hypoglycaemic events and, ultimately, impaired hypoglycaemia awareness.4 Frequent hypoglycaemic episodes might have a profound effect on behaviour and diabetes self-management, adversely affecting quality of life.5 The advent of continuous glucose monitoring (CGM) has led to improved glycaemic control and reduced exposure to hypoglycaemia, including severe hypoglycaemia.6, 7 The benefits of hypoglycaemia reduction are enhanced in hypoglycaemia-prone individuals when CGM is integrated with the threshold suspend feature of insulin pumps, which allows insulin delivery to be suspended automatically for up to 2 h when the pre-set glucose threshold is reached8 or predicted.9 Although these technologies have been shown to reduce the incidence of severe hypoglycaemic events, including those leading to hypoglycaemic seizure or coma,10 they do not address the issue of variability in insulin requirements,11 which remains an unmet need in patients with type 1 diabetes. Research in context Evidence before this study

ensor glucose within days was 5·0% (3·0–7·1) lower (p<0·0001), and coefficient of variation of sensor glucose between days was 7·5% (5·3–9·7) lower (p<0·0001; table 2). Total daily insulin was similar between study periods (table 3). Weekly trends in glucose control and insulin delivery are shown in the appendix (p 3). Outcomes at night time (0000 h to 0600 h) and daytime (0601 h to 2359 h) were in concordance with outcomes from the combined day-and-night period (table 4). Night-time use of closed-loop insulin delivery significantly increased the proportion of time with glucose concentration in target range by 17·2 percentage points (95% CI 12·0–22·4, p<0·0001), reduced mean glucose concentration by 0·4 mmol/L (0·1–0·8, p=0·0211), and decreased the burden of hypoglycaemia by 89% (80–94, p<0·0001) compared with the control period. SD and between-night coefficient of variation of sensor glucose were significantly reduced by closed-loop insulin delivery (table 4). Daytime use of closed-loop insulin delivery increased time spent with glucose concentration in the target range by 8·1 percentage points (95% CI 5·3–11·0, p<0·0001) and reduced the burden of hypoglycaemia by 61% (14–75, p=0·0001). Closed-loop insulin delivery significantly reduced the SD and between-day coefficient of variation of sensor glucose (table 4), in line with measured outcomes of night-time glycaemic variability.

The advent of continuous glucose monitoring (CGM) has led to improved glycaemic control and reduced exposure to hypoglycaemia, including severe hypoglycaemia.6, 7 The benefits of hypoglycaemia reduction are enhanced in hypoglycaemia-prone individuals when CGM is integrated with the threshold suspend feature of insulin pumps, which allows insulin delivery to be suspended automatically for up to 2 h when the pre-set glucose threshold is reached8 or predicted.9 Although these technologies have been shown to reduce the incidence of severe hypoglycaemic events, including those leading to hypoglycaemic seizure or coma,10 they do not address the issue of variability in insulin requirements,11 which remains an unmet need in patients with type 1 diabetes. Research in context Evidence before this study We searched PubMed from database inception until Oct 24, 2016, using the search terms “type 1 diabetes” AND (“artificial pancreas” OR “closed-loop”) AND (“home” OR “outpatient”), for reports of randomised controlled trials published in English only. We identified 14 randomised trials assessing the use of closed-loop insulin delivery outside hospital settings. In two randomised home studies in participants with mean HbA1c above 7·5% (58 mmol/mol), long-term (>4 week) use of closed-loop insulin delivery led to a significant decrease in HbA1c and improvement in mean glucose and proportion of time spent within, below, and above the glucose target range (3·9–10·0 mmol/L). However, no studies have so far assessed closed-loop use in non-pregnant adults with HbA1c below 7·5%. Added value of this study

We searched PubMed from database inception until Oct 24, 2016, using the search terms “type 1 diabetes” AND (“artificial pancreas” OR “closed-loop”) AND (“home” OR “outpatient”), for reports of randomised controlled trials published in English only. We identified 14 randomised trials assessing the use of closed-loop insulin delivery outside hospital settings. In two randomised home studies in participants with mean HbA1c above 7·5% (58 mmol/mol), long-term (>4 week) use of closed-loop insulin delivery led to a significant decrease in HbA1c and improvement in mean glucose and proportion of time spent within, below, and above the glucose target range (3·9–10·0 mmol/L). However, no studies have so far assessed closed-loop use in non-pregnant adults with HbA1c below 7·5%. Added value of this study To our knowledge, this study is the first randomised controlled study in free-living adults with type 1 diabetes whose HbA1c is below 7·5%. We showed that, compared with usual insulin pump therapy, day-and-night closed-loop insulin delivery significantly improved overall glycaemic control while reducing the burden of hypoglycaemia. Beneficial effects on glycaemic outcomes included increased time with glucose concentration in target range, reduced time with glucose concentration below and above target range, and decreased mean glucose concentration and glycaemic variability. The findings of increased time in target range and reduced overall hypoglycaemia risks and sensor glucose variability were similarly observed and consistent during night-time and daytime periods of closed-loop use. These outcomes were achieved without change in total insulin delivered. Closed-loop application thereby provides a novel therapeutic approach to optimise glycaemic control in hypoglycaemia-prone adults with HbA1c below 7·5%. Closed-loop application was well tolerated in this population with advanced self-management skills, and might provide clinically significant benefits to their overall diabetes care.

plication thereby provides a novel therapeutic approach to optimise glycaemic control in hypoglycaemia-prone adults with HbA1c below 7·5%. Closed-loop application was well tolerated in this population with advanced self-management skills, and might provide clinically significant benefits to their overall diabetes care. Implications of all the available evidence The use of day-and-night closed-loop insulin delivery might further improve glycaemic control while reducing the risk and burden of hypoglycaemia in adults with type 1 diabetes whose HbA1c is below 7·5%. Results from our study, together with those from previous studies in different target groups, support the benefits of closed-loop insulin delivery in a broad population with type 1 diabetes.

e glycaemic control while reducing the risk and burden of hypoglycaemia in adults with type 1 diabetes whose HbA1c is below 7·5%. Results from our study, together with those from previous studies in different target groups, support the benefits of closed-loop insulin delivery in a broad population with type 1 diabetes. Closed-loop insulin delivery—also known as the artificial pancreas—is a therapeutic approach that is progressing quickly. Closed-loop delivery differs from conventional pump therapy and threshold suspend technology; it has a control algorithm that autonomously increases and decreases subcutaneous insulin delivery in response to real-time sensor glucose levels.12 Results from randomised trials13, 14, 15, 16 of day-and-night closed-loop use during unsupervised free-living conditions in children, adolescents, and adults have shown improved glycaemic outcomes, reduced risk of non-severe hypoglycaemic events, and positive user experience. However, outside of pregnancy,15 none of the studies have focused specifically on patients with well controlled diabetes (HbA1c <7·5% [58 mmol/mol]) who might have a heightened, but masked, risk of hypoglycaemia and glucose variability. In this study, we aimed to investigate whether day-and-night hybrid closed-loop insulin delivery—in which manual administration of prandial bolus was implemented by the user—under free-living conditions in adults with HbA1c below 7·5% can improve glucose control while alleviating the risk of hypoglycaemia, thus informing whether the use and reimbursement of closed-loop systems is justified in this particular population.

range by 8·1 percentage points (95% CI 5·3–11·0, p<0·0001) and reduced the burden of hypoglycaemia by 61% (14–75, p=0·0001). Closed-loop insulin delivery significantly reduced the SD and between-day coefficient of variation of sensor glucose (table 4), in line with measured outcomes of night-time glycaemic variability. Overall mean absolute relative deviation of sensor glucose, using capillary glucose as the reference, was 15·3% (SD 18·2) and median absolute relative deviation was 10·1% (IQR 4·7–19·3), on the basis of 8447 paired capillary-CGM values. Sensor alarm settings were not altered by participants during closed-loop intervention.

ch manual administration of prandial bolus was implemented by the user—under free-living conditions in adults with HbA1c below 7·5% can improve glucose control while alleviating the risk of hypoglycaemia, thus informing whether the use and reimbursement of closed-loop systems is justified in this particular population. Methods Study design and participants In this open-label, randomised, crossover study, we recruited adults (aged ≥18 years) attending diabetes clinics at Addenbrooke's Hospital (Cambridge, UK) and Medical University of Graz (Graz, Austria). Patients were eligible if they had type 1 diabetes (defined according to WHO criteria), non-hypoglycaemic C-peptide concentration less than 100 pmol/L, and HbA1c less than 7·5%; had been using insulin pump therapy for at least 6 months; had knowledge of insulin self-adjustment; and had been self-monitoring their blood glucose concentration at least six times per day. Exclusion criteria included established nephropathy, neuropathy, or proliferative retinopathy; total daily insulin dose of 2·0 U/kg or more; hypoglycaemia unawareness (determined by Gold score ≥4 on the basis of pre-study clinical records); severe visual or hearing impairment; pregnancy; or breastfeeding (see appendix for the full list of inclusion and exclusion criteria). The study was approved by the local ethics committees and national competent authorities in the UK and Austria, and the protocol (phase 2 of APhome04 study) is shown in the appendix. All participants provided written informed consent before the start of study-related procedures.

Methods Study design and participants In this open-label, randomised, crossover study, we recruited adults (aged ≥18 years) attending diabetes clinics at Addenbrooke's Hospital (Cambridge, UK) and Medical University of Graz (Graz, Austria). Patients were eligible if they had type 1 diabetes (defined according to WHO criteria), non-hypoglycaemic C-peptide concentration less than 100 pmol/L, and HbA1c less than 7·5%; had been using insulin pump therapy for at least 6 months; had knowledge of insulin self-adjustment; and had been self-monitoring their blood glucose concentration at least six times per day. Exclusion criteria included established nephropathy, neuropathy, or proliferative retinopathy; total daily insulin dose of 2·0 U/kg or more; hypoglycaemia unawareness (determined by Gold score ≥4 on the basis of pre-study clinical records); severe visual or hearing impairment; pregnancy; or breastfeeding (see appendix for the full list of inclusion and exclusion criteria). The study was approved by the local ethics committees and national competent authorities in the UK and Austria, and the protocol (phase 2 of APhome04 study) is shown in the appendix. All participants provided written informed consent before the start of study-related procedures. Randomisation and masking Participants were randomly assigned (1:1) to receive either day-and-night closed-loop insulin delivery followed by usual pump therapy with blinded CGM, or vice versa. Following the run-in period, the order of the two study periods was randomly determined with an automated web-based programme with locally stratified, randomly permuted blocks of four. Participants and investigators analysing study data were not masked to treatment allocation.

p therapy with blinded CGM, or vice versa. Following the run-in period, the order of the two study periods was randomly determined with an automated web-based programme with locally stratified, randomly permuted blocks of four. Participants and investigators analysing study data were not masked to treatment allocation. Procedures After screening, all participants underwent a 2–4 week run-in period, during which they were trained to use the study insulin pump and CGM device, and calibrated the real-time CGM device according to manufacturer's instructions. Compliance assessment after the run-in period was based on at least 10 days of CGM use in the past 2 weeks.

all participants underwent a 2–4 week run-in period, during which they were trained to use the study insulin pump and CGM device, and calibrated the real-time CGM device according to manufacturer's instructions. Compliance assessment after the run-in period was based on at least 10 days of CGM use in the past 2 weeks. Participants then received insulin via the day-and-night closed-loop system (closed-loop period) for 4 weeks and via usual pump therapy with blinded CGM (control period) for 4 weeks, in the order assigned at randomisation, with a 2–4 week washout period in between. During the washout period, participants returned to their usual care and did not use the study CGM device. Identical study insulin pumps and CGM devices were used during the two treatment periods. Participants used rapid-acting insulin analogue normally applied in their usual clinical care. The built-in bolus wizard of the study insulin pump was used by participants during both treatment periods to calculate insulin boluses at mealtimes and when administering correction boluses. The study was done under free-living conditions without direct or remote supervision by clinical investigators. Participants were not restricted in their dietary intake or daily activities. Support was available at all times to assist participants in case of clinical or technical issues that arose during the study. Standard local hypoglycaemia and hyperglycaemia treatment guidelines were followed.

all mean absolute relative deviation of sensor glucose, using capillary glucose as the reference, was 15·3% (SD 18·2) and median absolute relative deviation was 10·1% (IQR 4·7–19·3), on the basis of 8447 paired capillary-CGM values. Sensor alarm settings were not altered by participants during closed-loop intervention. No significant difference was seen in sensor glucose availability between study periods (97% [IQR 95–99] in closed-loop period vs 96% [91–97] in control period; p=0·10). Day-and-night closed-loop delivery was used for 90% (95% CI 78–89) of the closed-loop period. The user feedback questionnaire was fully completed by 26 participants, and four of the six questions were answered by all participants (appendix). 27 (93%) of 29 participants were happy to have their glucose levels automatically controlled by the closed-loop system. 20 (69%) participants stated that they spent less time managing their diabetes while using the closed-loop system, but seven (24%) disagreed with this statement. 18 (62%) expressed fewer concerns about their glycaemic control while using the closed-loop system. 14 (48%) participants reported improved sleep during the closed-loop period. 23 (88%) of 26 participants reported feeling safe while using the closed-loop system, and 26 (96%) of 27 would recommend it to others.

sion by clinical investigators. Participants were not restricted in their dietary intake or daily activities. Support was available at all times to assist participants in case of clinical or technical issues that arose during the study. Standard local hypoglycaemia and hyperglycaemia treatment guidelines were followed. The FlorenceD2A closed-loop system (University of Cambridge, Cambridge, UK) comprised a model-predictive control algorithm on a smartphone (Galaxy S4, Samsung, South Korea), which communicated wirelessly with a purpose-made translator unit (Triteq, Hungerford, UK) and the study pump (DANA-R Diabecare, Sooil, Seoul, South Korea) through a Bluetooth communication protocol (appendix). The CGM receiver (FreeStyle Navigator II, Abbott Diabetes Care, Alameda, CA, USA) was inserted into the translator, which translated a serial USB protocol into a Bluetooth communication protocol. The calculations used a compartment model of glucose kinetics17 describing the effect of rapid-acting insulin and the carbohydrate content of meals on glucose levels. We applied a hybrid closed-loop approach in which participants were required to count carbohydrates and use a standard bolus calculator for pre-meal boluses according to usual practice. The bolus calculations provided by the study pump's built-in bolus calculator took into account carbohydrate content of meals, insulin on board, and entered capillary blood glucose readings. The algorithm was initialised by pre-programmed basal insulin rates downloaded from the study pump. Participants' bodyweight and total daily insulin dose were entered at set-up. During closed-loop operation, the algorithm adapted itself to a particular participant. The treat-to-target control algorithm aimed to achieve glucose concentrations of 5·8–7·3 mmol/L, and adjusted the actual concentration depending on fasting versus postprandial status and the accuracy of model-based glucose predictions. Control Algorithm version 0.3.46 was used (University of Cambridge, Cambridge, UK) which, compared with that used in our previous study,13 included enhanced adaptation of insulin needs based on analysis of past performance and identification of the time of day when insulin needs are consistently lower or higher. Participants were trained to perform a calibration check before breakfast and evening meal. If sensor glucose was above capillary glucose by more than 3 mmol/L, participants were advised to recalibrate the CGM device.

performance and identification of the time of day when insulin needs are consistently lower or higher. Participants were trained to perform a calibration check before breakfast and evening meal. If sensor glucose was above capillary glucose by more than 3 mmol/L, participants were advised to recalibrate the CGM device. These instructions resulted from in-silico assessment of hypoglycaemia and hyperglycaemia risks using the validated Cambridge simulator.18 Safety rules limited maximum insulin infusion and suspended insulin delivery at sensor glucose at or less than 4·3 mmol/L, or when sensor glucose was rapidly decreasing. During the control period, the display of the study CGM device was masked. Participants were allowed to use their own glucose monitoring devices (CGM or flash glucose monitoring19) if they were part of their pre-study usual care. The control intervention was chosen according to existing clinical practice and participants' preferences.20 The rationale for the control period was to reflect usual clinical practice and national reimbursement policies, and to compare the incremental benefits gained by closed-loop insulin delivery with the therapeutic modality followed by the participants in a pragmatic study design.

ractice and participants' preferences.20 The rationale for the control period was to reflect usual clinical practice and national reimbursement policies, and to compare the incremental benefits gained by closed-loop insulin delivery with the therapeutic modality followed by the participants in a pragmatic study design. At the start of the closed-loop period, participants were admitted to the local clinical research facility for a training session, which covered starting and stopping of the closed-loop system and troubleshooting of technical issues. If sensor glucose readings became unavailable, or in case of other system failures, participants were alerted by an audible alarm and the system restarted the participant's usual insulin delivery rate within 30–60 min to mitigate the risk of insulin under-delivery and over-delivery.21 Participants were instructed to have the low-glucose alarm audible at all times. The sensor glucose alarm threshold for hypoglycaemia was initially set at 3·5 mmol/L, but these settings could be later modified by the participants. A user feedback questionnaire was completed by participants at the end of the closed-loop period.

ants were instructed to have the low-glucose alarm audible at all times. The sensor glucose alarm threshold for hypoglycaemia was initially set at 3·5 mmol/L, but these settings could be later modified by the participants. A user feedback questionnaire was completed by participants at the end of the closed-loop period. Blood samples for HbA1c and C-peptide measurements were taken after enrolment. Plasma C-peptide was measured with chemiluminescence immunoassays (Diasorin Liaison XL [Deutschland GmbH, Dietzenbach, Germany] used in Cambridge; ADVIA Centaur [Siemens Healthcare Diagnostics, USA] used in Graz). HbA1c was measured with ion exchange high-performance liquid chromatography, compliant with the International Federation of Clinical Chemistry and Laboratory Medicine, at study centres (G8 HPLC Analyzer [Tosoh Bioscience, South San Francisco, CA, USA] in Cambridge; Menarini HA-8160 HbA1c auto-analyser [Menarini Diagnostics, Florence, Italy] in Graz).

ange high-performance liquid chromatography, compliant with the International Federation of Clinical Chemistry and Laboratory Medicine, at study centres (G8 HPLC Analyzer [Tosoh Bioscience, South San Francisco, CA, USA] in Cambridge; Menarini HA-8160 HbA1c auto-analyser [Menarini Diagnostics, Florence, Italy] in Graz). Outcomes The primary outcome was the proportion of time during the whole study when sensor glucose concentration was in the target range of 3·9–10·0 mmol/L. Secondary efficacy outcomes were proportion of time with sensor glucose concentration above and below target range; time with glucose concentration below 3·5 mmol/L, 3·3 mmol/L (post hoc), and 2·8 mmol/L, and above 13·9 mmol/L (post hoc) and 16·7 mmol/L; the number of nights and mean duration when sensor glucose was below 3·5 mmol/L for at least 20 min; hypoglycaemia burden (ie, area under the curve when sensor glucose concentration was less than 3·5 mmol/L; post hoc); mean, SD, and coefficient of variation (post hoc) of sensor glucose; total daily, basal, and bolus insulin dose; and weekly trends in glucose control and insulin delivery. The between-day coefficient of variation of sensor glucose was calculated from daily mean glucose values (midnight to midnight). To limit multiple comparison, daytime (0601 h to 2359 h) and night-time (0000 h to 0600 h) endpoints were calculated with data from the respective periods for a subset of outcomes—proportion of time with glucose concentration in target range, above and below target range, below 3·5 mmol/L, and below 2·8 mmol/L; SD of sensor glucose concentration; between-day and between-night coefficient of variation of sensor glucose concentration; and area under the curve when sensor glucose concentration was less than 3·5 mmol/L.

th glucose concentration in target range, above and below target range, below 3·5 mmol/L, and below 2·8 mmol/L; SD of sensor glucose concentration; between-day and between-night coefficient of variation of sensor glucose concentration; and area under the curve when sensor glucose concentration was less than 3·5 mmol/L. Of note, two prespecified secondary outcomes will not be reported. We will not report time spent in target glucose range (3·9–10·0 mmol/L) based on subcutaneous glucose monitoring adjusted for sensor error during the entire home stay because the general consensus is that outcomes based on unadjusted CGM values should be reported.22 Additionally, low blood glucose index will not be reported since it is not generally well understood by non-specialists and is highly correlated with time with glucose concentration below 3·9 mmol/L.23 Safety outcomes were severe hypoglycaemic events, significant ketonaemia (>3·0 mmol/L), and other adverse and serious adverse events. We also assessed the frequency and duration of use of the closed-loop system at home as a measure of utility outcome.

Of note, two prespecified secondary outcomes will not be reported. We will not report time spent in target glucose range (3·9–10·0 mmol/L) based on subcutaneous glucose monitoring adjusted for sensor error during the entire home stay because the general consensus is that outcomes based on unadjusted CGM values should be reported.22 Additionally, low blood glucose index will not be reported since it is not generally well understood by non-specialists and is highly correlated with time with glucose concentration below 3·9 mmol/L.23 Safety outcomes were severe hypoglycaemic events, significant ketonaemia (>3·0 mmol/L), and other adverse and serious adverse events. We also assessed the frequency and duration of use of the closed-loop system at home as a measure of utility outcome. Statistical analysis The power calculation was based on improvements in time spent in glucose concentration target range. Assuming an SD of 18% and mean improvement of time spent in target range of 10%,13, 24 31 participants were needed at the desired 80% power and α level of 0·05 (two-tailed). If the mean improvement was 12%, the required sample size was reduced to 20. We planned to recruit up to 34 participants, aiming for 24 participants to complete the study to allow for dropouts (anticipated dropout rate of 25% based on the investigators' experience and expectation). Participants who dropped out of the study during the run-in period and within the first 2 weeks of the first treatment period were allowed to be replaced.

nts, aiming for 24 participants to complete the study to allow for dropouts (anticipated dropout rate of 25% based on the investigators' experience and expectation). Participants who dropped out of the study during the run-in period and within the first 2 weeks of the first treatment period were allowed to be replaced. We agreed on the statistical analysis plan following completion of the last patient's last visit but before the final dataset was reviewed and analysed. The analyses were done by intention to treat. Efficacy and safety data from all randomised participants, including those who dropped out, were included in the analysis. We compared the respective measurements obtained during the closed-loop period and the control period using a regression model that accounts for period effect. Log-transformed analyses were used for highly skewed endpoints. Values were presented as mean (SD) or median (IQR) for each study period. A 5% significance level was used to declare statistical significance for all comparisons. Outcomes were calculated with GStat software (University of Cambridge, Cambridge, UK), version 2.2.4, and statistical analyses were done with SPSS (IBM Software, Hampshire, UK), version 23. This study is registered with ClinicalTrials.gov, number NCT02727231.

We agreed on the statistical analysis plan following completion of the last patient's last visit but before the final dataset was reviewed and analysed. The analyses were done by intention to treat. Efficacy and safety data from all randomised participants, including those who dropped out, were included in the analysis. We compared the respective measurements obtained during the closed-loop period and the control period using a regression model that accounts for period effect. Log-transformed analyses were used for highly skewed endpoints. Values were presented as mean (SD) or median (IQR) for each study period. A 5% significance level was used to declare statistical significance for all comparisons. Outcomes were calculated with GStat software (University of Cambridge, Cambridge, UK), version 2.2.4, and statistical analyses were done with SPSS (IBM Software, Hampshire, UK), version 23. This study is registered with ClinicalTrials.gov, number NCT02727231. Role of the funding source The funders of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. Abbott Diabetes Care read the manuscript before submission. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. Abbott Diabetes Care read the manuscript before submission. 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 March 21 and June 24, 2016, we recruited 31 participants (figure 1). Two participants withdrew during the run-in period because of issues associated with use of the study pump. 29 eligible participants (17 from Cambridge and 12 from Graz) were randomly assigned. One participant dropped out during the first closed-loop period because of dissatisfaction with study devices and glycaemic control. Of 29 randomised participants, five (17%) used real-time CGM and six (21%) used flash glucose monitoring as part of their usual care (table 1; appendix). The primary outcome of the study—the proportion of time during the whole study period when sensor glucose concentration was in the target range (3·9–10·0 mmol/L)—was 10·5 percentage points higher (95% CI 7·6–13·4; p<0·0001) during the closed-loop period than during the control period (65·6% [SD 8·1] when participants were using usual pump therapy vs 76·2% [6·4] when they used closed-loop; table 2). 24 h sensor glucose and insulin delivery profiles are shown in figure 2. The proportion of time sensor glucose concentration was in the target range seemed unchanged over the 4 week intervention periods (appendix).

ia25 have both been associated with adverse clinical outcomes. We hypothesise that the significant reductions in glucose variability and hypoglycaemia by closed-loop insulin delivery in our study might have implications for clinical outcomes, although this hypothesis will need confirmation by longer and larger studies. Compared with previous unsupervised home-based studies,13, 16 our study revealed new findings, specifically in relation to the improvements of CGM-derived hypoglycaemia parameters. Compared with Thabit and colleagues' study of closed-loop insulin delivery versus sensor-augmented pump therapy in patients with type 1 diabetes and HbA1c between 7·5% and 10·0% (58–86 mmol/mol),13 our study had more pronounced relative reductions in the proportion of time with glucose concentration below 3·9 mmol/L (19% vs 50%) and hypoglycaemia burden (39% vs 73%). These differences might have been attributable to the use of an enhanced adaptive control algorithm and tight glycaemic control at baseline in the present study (mean screening HbA1c 6·9% vs 8·5% in Thabit and colleagues' study13), as well as differences in the control therapies between the two studies. The proportion of time spent with glucose concentration below the target range (ie, 3·9 mmol/L) during usual pump therapy (5·3%, IQR 3·5–10·0) was notably higher in our study than that in Thabit and colleagues' study (3·0%, 1·8–6·1)13 but still lower than that in a study in adults with a mean baseline HbA1c of 6·7% (mean proportion of time spent with sensor glucose below target range 14% [SD 11]).19 Another study15 assessed the use of closed-loop insulin delivery during night time in pregnant women with type 1 diabetes, using a control algorithm with a lower glucose setpoint to accommodate lower glucose targets (3·5–7·8 mmol/L) for this group of patients. In this study,15 use of the closed-loop system increased the primary endpoint (time in target range overnight) by 15 percentage points compared with sensor-augmented pump therapy. The difference between this study15 and ours is the overnight-only use of closed-loop in the former, and the difference in the comparator between the two studies (sensor-augmented pump therapy vs usual pump therapy). Comparability between both studies is also challenging because this specific group of women with type 1 diabetes is highly motivated, albeit over a short time period (ie, during pregnancy), and received intensive dietetic, education, and clinical support.

o studies (sensor-augmented pump therapy vs usual pump therapy). Comparability between both studies is also challenging because this specific group of women with type 1 diabetes is highly motivated, albeit over a short time period (ie, during pregnancy), and received intensive dietetic, education, and clinical support. The additional benefit of closed-loop insulin delivery in individuals with well controlled type 1 diabetes (ie, HbA1c <7·5%) is the reduction of the residual risk of complications, hypoglycaemia, and glycaemic variability, as well as the burden of self-management. Reduction of hypoglycaemic burden has benefits such as improved quality of life and reduced societal cost.5, 29 Existing reimbursement criteria in several countries for CGM are still predominantly focused on HbA1c and might exclude patients with HbA1c below 7·5% because of the dearth of evidence showing efficacy. However, results from our study show efficacy in terms of both improved glycaemic control and reduced risk of hypoglycaemia. Thus, reimbursement of closed-loop technology in patients with HbA1c below 7·5% could be considered justifiable.

patients with HbA1c below 7·5% because of the dearth of evidence showing efficacy. However, results from our study show efficacy in terms of both improved glycaemic control and reduced risk of hypoglycaemia. Thus, reimbursement of closed-loop technology in patients with HbA1c below 7·5% could be considered justifiable. The reduction in hypoglycaemic burden in our study was almost identical to that in Russell and colleagues' study,30 which compared 5 day use of a bihormonal (insulin and glucagon) closed-loop system with usual pump therapy in adults with type 1 diabetes. Participants in Russell and colleagues' study had a baseline HbA1c of 7·1% (SD 0·8), and lower mean glucose concentration and risk of hypoglycaemia during the closed-loop period than during usual pump therapy. The study was done under direct supervision during the closed-loop period but not during the control period. The incremental reduction in hypoglycaemia achieved by bihormonal versus single-hormone closed-loop systems needs to be further assessed in longer, unsupervised, head-to-head, randomised studies to justify the increased complexity and cost. A single-hormone hybrid closed-loop system was approved in September, 2016, by the US Food and Drug Administration on the basis of results from a 3 month study31 in adolescents and adults with mean HbA1c of 7·4% (SD 0·9) at screening. However, this study was non-randomised and did not have a control group, and the length of study periods was not matched (2 week run-in phase, which was used as baseline comparator for a 3 month intervention).

sis of results from a 3 month study31 in adolescents and adults with mean HbA1c of 7·4% (SD 0·9) at screening. However, this study was non-randomised and did not have a control group, and the length of study periods was not matched (2 week run-in phase, which was used as baseline comparator for a 3 month intervention). Results from participants' feedback indicate a high level of trust in the closed-loop system autonomously modulating their glucose concentrations, with the majority agreeing that the time spent on management of diabetes was reduced during the closed-loop period. However, seven (24%) of 29 participants disagreed with this statement, reflecting that user input is still needed for a hybrid closed-loop system. CGM alarms and connectivity issues, especially at night, might have negatively affected participants' experience of the closed-loop system. One participant withdrew from the study because of recurrent technical issues related to the closed-loop system. However, an overall positive endorsement of the closed-loop system was observed, since most participants were willing to recommend the closed-loop system to others.

nts' experience of the closed-loop system. One participant withdrew from the study because of recurrent technical issues related to the closed-loop system. However, an overall positive endorsement of the closed-loop system was observed, since most participants were willing to recommend the closed-loop system to others. The strengths of this study include the randomised, two-centre, two-country, crossover design. So far, none of the home-based studies of closed-loop insulin delivery have focused specifically on patients with HbA1c below 7·5% who might be early adopters of closed-loop technologies striving to further improve control of their diabetes. We did not use remote monitoring or direct supervision, so as to adhere to real-world conditions. Additionally, we did not restrict the participants' dietary habits, activity level, and geographical movements. However, we acknowledge that our study has several limitations. The relatively short study duration might have been insufficient to assess long-term compliance. We excluded participants with hypoglycaemia unawareness, therefore restricting assessment of the closed-loop system in those who might benefit greatly. The prototype nature of the closed-loop system and the number of devices might have increased the participants' device burden and negatively affected some aspects of user feedback. The heterogeneity of sensor use in the control period might have confounded the reported glycaemic outcomes. However, the individualised therapy approaches used in the control period reflect present clinical strategies adopted by this population to achieve their baseline HbA1c, and do not diminish the incremental effects of closed-loop use.20

sor use in the control period might have confounded the reported glycaemic outcomes. However, the individualised therapy approaches used in the control period reflect present clinical strategies adopted by this population to achieve their baseline HbA1c, and do not diminish the incremental effects of closed-loop use.20 To conclude, day-and-night closed-loop insulin delivery in adults with type 1 diabetes and HbA1c below 7·5% significantly improved glycaemic control while reducing the risk of hypoglycaemia. Thus, in adults who are actively engaged with self-management, closed-loop insulin delivery might provide additional benefits, justifying its use in this particular population. The overall positive feedback from participants reflected the acceptance of closed-loop technology during daily diabetes management, albeit with some limitations to its use, which might affect user adherence and experience. Larger and longer studies are needed to validate our findings. Supplementary Material Supplementary appendix

To conclude, day-and-night closed-loop insulin delivery in adults with type 1 diabetes and HbA1c below 7·5% significantly improved glycaemic control while reducing the risk of hypoglycaemia. Thus, in adults who are actively engaged with self-management, closed-loop insulin delivery might provide additional benefits, justifying its use in this particular population. The overall positive feedback from participants reflected the acceptance of closed-loop technology during daily diabetes management, albeit with some limitations to its use, which might affect user adherence and experience. Larger and longer studies are needed to validate our findings. Supplementary Material Supplementary appendix Acknowledgments LB received support from the Swiss National Science Foundation (P1BEP3_165297). Additional support for the artificial pancreas work was from JDRF, National Institute for Health Research Cambridge Biomedical Research Centre, and Wellcome Strategic Award (100574/Z/12/Z). Abbott Diabetes Care supplied discounted CGM devices, sensors, and details of communication protocol to facilitate real-time connectivity. We thank study volunteers for their participation and acknowledge support by the staff at the Addenbrooke's Wellcome Trust Clinical Research Facility. Jasdip Mangat (University Hospitals of Leicester, National Health Service Trust, Leicester, UK) supported development and validation of the closed-loop system. Josephine Hayes (University of Cambridge) provided administrative support. Biochemical assays were done by Keith Burling (UK National Institute for Health Research Cambridge Biomedical Research Centre, Core Biochemical Assay Laboratory).

st, Leicester, UK) supported development and validation of the closed-loop system. Josephine Hayes (University of Cambridge) provided administrative support. Biochemical assays were done by Keith Burling (UK National Institute for Health Research Cambridge Biomedical Research Centre, Core Biochemical Assay Laboratory). Contributors RH coordinated the study. RH, MLE, TRP, HT, LB, and MEW designed the study. LB, HT, SH, HK, JQ-H, and JKM screened and enrolled participants and arranged informed consent from the participants. LB, HT, SH, HK, JQ-H, JKM, MT, and JMA provided patient care and took samples. MEW managed randomisation. LB, HT, MEW, and RH did or supported data analyses, including the statistical analyses. RH designed and implemented the glucose controller. LB, HT, MT, MLE, TRP, and RH interpreted the results. LB, HT, and RH wrote the manuscript. All authors critically reviewed the report. No writing assistance was provided. LB, HT, and RH had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analyses.

, MT, MLE, TRP, and RH interpreted the results. LB, HT, and RH wrote the manuscript. All authors critically reviewed the report. No writing assistance was provided. LB, HT, and RH had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analyses. Declaration of interests SH is a consultant for Novo-Nordisk and the ONSET group, and received speaker and training honoraria from Medtronic. MLE received speaker honoraria from Abbott Diabetes Care, Novo Nordisk, and Animas; and is on advisory panels for Novo Nordisk, Abbott Diabetes Care, Medtronic, Roche, and Cellnovo. RH received speaker honoraria from Eli Lilly and Novo Nordisk and license fees from B Braun and Medtronic; is on advisory panels for Eli Lilly and Novo Nordisk; has served as a consultant to B Braun; and reports patents and patent applications. MEW received license fees from Becton Dickinson, has served as a consultant to Beckton Dickinson, and reports patents and patent applications. MT received speaker honoraria from Novo Nordisk. JKM is a member in the advisory board of Sanofi, Eli Lilly, and Boehringer Ingelheim, and received speaker honoraria from Abbott Diabetes Care, Astra Zeneca, Eli Lilly, Nintamed, Novo Nordisk, Roche Diabetes Care, Servier, and Takeda. TRP is an advisory board member of Novo Nordisk and a consultant for Roche Diabetes Care, Novo Nordisk, Eli Lilly, Infineon, and Carnegie Bank, and is on the speaker's bureau of Novo Nordisk and Astra Zeneca. LB, HT, HK JQ-H, and JMA declare no competing interests. Figure 1 Trial profile

Declaration of interests SH is a consultant for Novo-Nordisk and the ONSET group, and received speaker and training honoraria from Medtronic. MLE received speaker honoraria from Abbott Diabetes Care, Novo Nordisk, and Animas; and is on advisory panels for Novo Nordisk, Abbott Diabetes Care, Medtronic, Roche, and Cellnovo. RH received speaker honoraria from Eli Lilly and Novo Nordisk and license fees from B Braun and Medtronic; is on advisory panels for Eli Lilly and Novo Nordisk; has served as a consultant to B Braun; and reports patents and patent applications. MEW received license fees from Becton Dickinson, has served as a consultant to Beckton Dickinson, and reports patents and patent applications. MT received speaker honoraria from Novo Nordisk. JKM is a member in the advisory board of Sanofi, Eli Lilly, and Boehringer Ingelheim, and received speaker honoraria from Abbott Diabetes Care, Astra Zeneca, Eli Lilly, Nintamed, Novo Nordisk, Roche Diabetes Care, Servier, and Takeda. TRP is an advisory board member of Novo Nordisk and a consultant for Roche Diabetes Care, Novo Nordisk, Eli Lilly, Infineon, and Carnegie Bank, and is on the speaker's bureau of Novo Nordisk and Astra Zeneca. LB, HT, HK JQ-H, and JMA declare no competing interests. Figure 1 Trial profile Figure 2 Median sensor glucose and insulin delivery for the 24 h duration over the study period

Declaration of interests SH is a consultant for Novo-Nordisk and the ONSET group, and received speaker and training honoraria from Medtronic. MLE received speaker honoraria from Abbott Diabetes Care, Novo Nordisk, and Animas; and is on advisory panels for Novo Nordisk, Abbott Diabetes Care, Medtronic, Roche, and Cellnovo. RH received speaker honoraria from Eli Lilly and Novo Nordisk and license fees from B Braun and Medtronic; is on advisory panels for Eli Lilly and Novo Nordisk; has served as a consultant to B Braun; and reports patents and patent applications. MEW received license fees from Becton Dickinson, has served as a consultant to Beckton Dickinson, and reports patents and patent applications. MT received speaker honoraria from Novo Nordisk. JKM is a member in the advisory board of Sanofi, Eli Lilly, and Boehringer Ingelheim, and received speaker honoraria from Abbott Diabetes Care, Astra Zeneca, Eli Lilly, Nintamed, Novo Nordisk, Roche Diabetes Care, Servier, and Takeda. TRP is an advisory board member of Novo Nordisk and a consultant for Roche Diabetes Care, Novo Nordisk, Eli Lilly, Infineon, and Carnegie Bank, and is on the speaker's bureau of Novo Nordisk and Astra Zeneca. LB, HT, HK JQ-H, and JMA declare no competing interests. Figure 1 Trial profile Figure 2 Median sensor glucose and insulin delivery for the 24 h duration over the study period Median (IQR) sensor glucose concentration (A) and insulin delivery (B) during closed-loop period (solid red line and red shaded area) and control period (dashed black line and grey shaded area) for the 24 h duration. The horizontal dashed lines show the lower and upper limits of the glucose target range (3·9–10·0 mmol/L).

Median (IQR) sensor glucose concentration (A) and insulin delivery (B) during closed-loop period (solid red line and red shaded area) and control period (dashed black line and grey shaded area) for the 24 h duration. The horizontal dashed lines show the lower and upper limits of the glucose target range (3·9–10·0 mmol/L). Figure 3 Individual values of mean sensor glucose and proportion of time spent with glucose concentration below the target range for the whole study The size of the circles denotes the proportion of time spent with low glucose (<3·5 mmol/L). Table 1 Baseline characteristics Data (n=29) Sex Female 15 (52%) Male 14 (48%) Age (years) 41 (13) Bodyweight (kg) 72·9 (13·0) BMI (kg/m2) 25·1 (3·0) HbA1c (%) 6·9 (0·5) HbA1c (mmol/mol) 51·7 (4·8) Duration of diabetes (years) 24 (12) Duration using pump (years) 6 (4) Total daily insulin (U/kg/day) 0·5 (0·1) Glucose sensor use No previous glucose sensor use 18 (62%) Real-time continuous glucose monitoring 5 (17%) Flash glucose monitoring 6 (21%) Data are mean (SD) or n (%). Table 2 Overall day-and-night glucose control during closed-loop and control periods based on sensor glucose measurements

Data (n=29) Sex Female 15 (52%) Male 14 (48%) Age (years) 41 (13) Bodyweight (kg) 72·9 (13·0) BMI (kg/m2) 25·1 (3·0) HbA1c (%) 6·9 (0·5) HbA1c (mmol/mol) 51·7 (4·8) Duration of diabetes (years) 24 (12) Duration using pump (years) 6 (4) Total daily insulin (U/kg/day) 0·5 (0·1) Glucose sensor use No previous glucose sensor use 18 (62%) Real-time continuous glucose monitoring 5 (17%) Flash glucose monitoring 6 (21%) Data are mean (SD) or n (%). Table 2 Overall day-and-night glucose control during closed-loop and control periods based on sensor glucose measurements Closed-loop period (n=29) Control period (n=28) Paired difference*or paired ratio†(95% CI) p value Proportion of time with glucose concentration in range (%) 3·9–10·0 mmol/L‡ 76·2% (6·4) 65·6% (8·1) 10·5 (7·6 to 13·4) <0·0001 >10·0 mmol/L 20·4% (6·3) 27·4% (9·6) −6·9 (−10·2 to −3·5) 0·0003 >13·9 mmol/L 3·8% (2·6) 6·9% (3·9) −3·0 (−4·4 to −1·6) 0·0002 >16·7 mmol/L 0·9% (0·8) 2·1% (1·8) −1·2 (−1·9 to −0·6) 0·0009 <3·9 mmol/L 2·9% (2·3 to 4·0) 5·3% (3·5 to 10·0) 0·50 (0·41 to 0·63)† <0·0001 <3·5 mmol/L 1·3% (0·8 to 2·3) 3·4% (1·9 to 7·2) 0·35 (0·26 to 0·47)† <0·0001 <3·3 mmol/L 0·9% (0·5 to 1·7) 2·6% (1·3 to 5·5) 0·30 (0·22 to 0·43)† <0·0001 <2·8 mmol/L 0·3% (0·1 to 0·5) 1·0% (0·5 to 2·6) 0·24 (0·14 to 0·41)† <0·0001 AUCday <3·5 mmol/L (min × mmol/L) 9·1 (3·7 to 18·2) 26·7 (13·1 to 65·5) 0·27 (0·18 to 0·41)† <0·0001 Mean glucose concentration (mmol/L) 7·9 (0·5) 8·3 (0·9) −0·4 (−0·7 to −0·1) 0·0226 SD of glucose concentration (mmol/L) 2·8 (0·4) 3·3 (0·5) −0·5 (−0·7 to −0·3) <0·0001 Coefficient of variation of glucose concentration Within days (%) 35·3 (3·0) 40·3 (5·1) −5·0 (−7·1 to −3·0) <0·0001 Between days (%) 12·8 (3·3) 20·2 (4·6) −7·5 (−9·7 to −5·3) <0·0001 Sensor glucose concentration <3·5 mmol/L for at least 20 min Number of nights 2·1 (1·0) 5·6 (3·5) −3·6 (−4·9 to −2·2) <0·0001 Mean duration of each episode (min) 45 (13) 75 (25) −29 (−38 to −20) <0·0001 Data are mean (SD) or median (IQR), unless otherwise stated. No significant period effect was observed. AUCday=sensor glucose area under the curve per day.

/L for at least 20 min Number of nights 2·1 (1·0) 5·6 (3·5) −3·6 (−4·9 to −2·2) <0·0001 Mean duration of each episode (min) 45 (13) 75 (25) −29 (−38 to −20) <0·0001 Data are mean (SD) or median (IQR), unless otherwise stated. No significant period effect was observed. AUCday=sensor glucose area under the curve per day. * Unless specified otherwise, data are normally distributed and presented as mean difference of closed-loop period minus control period, with 95% CI for mean; a positive value indicates that the measurement was higher during the closed-loop period than in the control period. † Non-normally distributed data are presented as ratio of closed-loop data over control data, with 95% CI for ratio; a value greater than unity indicates that the measurement was higher in the closed-loop period than in the control period. ‡ Primary outcome. Table 3 Insulin delivery over 24 h period Closed-loop period (n=29) Control period (n=28) Paired difference*(95% CI) p value Total daily insulin (U/day) 37·5 (13·8) 37·4 (12·6) 0·8 (−1·0 to 2·6) 0·36 Total bolus insulin (U/day) 18·6 (7·9) 20·2 (8·4) −1·1 (−2·4 to 0·2) 0·11 Total basal insulin (U/day) 18·9 (7·8) 17·2 (5·7) 1·9 (0·7 to 3·1) 0·0038 Data are mean (SD), unless otherwise stated. * Normally distributed data are presented as mean difference of closed-loop period minus control period, with 95% CI for mean; a positive value indicates that the measurement was higher in the closed-loop period than in the control period. Table 4 Night-time and daytime glucose control during closed-loop and control periods based on sensor glucose measurements

* Normally distributed data are presented as mean difference of closed-loop period minus control period, with 95% CI for mean; a positive value indicates that the measurement was higher in the closed-loop period than in the control period. Table 4 Night-time and daytime glucose control during closed-loop and control periods based on sensor glucose measurements Closed-loop period (n=29) Control period (n=28) Paired difference*or paired ratio†(95% CI) p value Night time (0000 h to 0600 h) Proportion of time with glucose concentration in range (%) 3·9–10·0 mmol/L 82·0% (9·7) 64·5% (11·5) 17·2 (12·0 to 22·4) <0·0001 >10·0 mmol/L 14·9% (8·5) 25·4% (11·8) −10·2 (−15·4 to −5·1) 0·0004 <3·9 mmol/L 3·2% (1·6 to 4·0) 9·0% (4·6 to 15·7) 0·33 (0·24 to 0·45)† <0·0001 <3·5 mmol/L 1·1% (0·4 to 2·2) 5·4% (2·9 to 11·4) 0·19 (0·12 to 0·28)† <0·0001 <2·8 mmol/L 0·1% (0·0 to 0·6) 1·7% (0·9 to 5·7) 0·14 (0·08 to 0·26)† <0·0001 AUCday <3·5 mmol/L (mmol/L × min) 5·1 (1·4 to 15·5) 41·5 (21·5 to 122·8) 0·11 (0·06 to 0·20)† <0·0001 Mean glucose concentration (mmol/L) 7·5 (0·6) 8·0 (1·0) −0·4 (−0·8 to −0·1) 0·0211 SD of glucose concentration (mmol/L) 2·5 (0·6) 3·2 (0·6) −0·7 (−1·0 to −0·4) <0·0001 Coefficient of variation of glucose concentration between nights (%) 24·9 (7·1) 34·4 (6·4) −9·6 (−13·2 to −5·9) <0·0001 Daytime (0601 h to 2359 h) Proportion of time with glucose concentration in range (%) 3·9–10·0 mmol/L 74·3% (6·9) 66·1% (8·8) 8·1 (5·3 to 11·0) <0·0001 >10·0 mmol/L 22·2% (7·2) 27·9% (10·4) −5·6 (−8·9 to −2·3) 0·0023 <3·9 mmol/L 2·7% (1·9 to 4·5) 4·4% (2·8 to 8·6) 0·61 (0·49 to 0·76)† 0·0001 <3·5 mmol/L 1·2% (0·7 to 2·2) 2·3% (1·4 to 6·0) 0·47 (0·35 to 0·64)† <0·0001 <2·8 mmol/L 0·2% (0·1 to 0·5) 0·5% (0·2 to 1·3) 0·48 (0·29 to 0·80)† 0·0076 AUCday <3·5 mmol/L (mmol/L × min) 8·3 (3·4 to 14·3) 15·9 (7·5 to 45·7) 0·39 (0·25 to 0·86)† 0·0001 Mean glucose concentration (mmol/L) 8·0 (0·6) 8·4 (1·0) −0·3 (−0·6 to 0·0) 0·0498 SD of glucose concentration (mmol/L) 2·9 (0·4) 3·3 (0·6) −0·5 (−0·7 to −0·2) 0·0002 Coefficient of variation of glucose concentration between days (%) 13·8 (3·0) 20·6 (5·1) −6·8 (−9·1 to −4·6) <0·0001 Data are presented as mean (SD) or median (IQR), unless otherwise stated. No significant period effect was observed. AUCday=sensor glucose area under the curve per day.

3 (0·6) −0·5 (−0·7 to −0·2) 0·0002 Coefficient of variation of glucose concentration between days (%) 13·8 (3·0) 20·6 (5·1) −6·8 (−9·1 to −4·6) <0·0001 Data are presented as mean (SD) or median (IQR), unless otherwise stated. No significant period effect was observed. AUCday=sensor glucose area under the curve per day. * Unless specified otherwise, data are normally distributed and presented as mean difference of closed-loop period minus control period, with 95% CI for mean; a positive value indicates that the measurement was higher in the closed-loop period than in the control period. † Non-normally distributed data are presented as ratio of closed-loop data over control data, with 95% CI for ratio; a value greater than unity indicates that the measurement was higher in the closed-loop period than in the control period.

Introduction Diabetes and high body-mass index (BMI), defined as a BMI greater than or equal to 25 kg/m2, are leading causes of mortality and morbidity globally1 and their prevalence has increased substantially over the past four decades in most countries.2, 3 The global age-standardised adult prevalence of diabetes was reported to be 9·0% in men and 7·9% in women in 2014, affecting about 422 million adults.3 In 2016, the age-standardised adult prevalence of overweight and obesity (those with BMI ≥25 kg/m2) was estimated to be 38·5% in men and 39·2% in women, affecting approximately 2·01 billion adults globally.2

lence of diabetes was reported to be 9·0% in men and 7·9% in women in 2014, affecting about 422 million adults.3 In 2016, the age-standardised adult prevalence of overweight and obesity (those with BMI ≥25 kg/m2) was estimated to be 38·5% in men and 39·2% in women, affecting approximately 2·01 billion adults globally.2 The International Agency for Research on Cancer (IARC) and the World Cancer Research Fund (WCRF) have concluded that there is a causal association between high BMI and colorectal,4 gallbladder,5 pancreas,6 kidney,7 liver,8 endometrial,9 postmenopausal breast,10 ovarian,11 gastric cardia,12 and thyroid cancer,13 as well as oesophageal adenocarcinoma14 and multiple myeloma.13 A study in 2015 estimated that about 3·6% of all cancer cases in 2012 were attributable to high BMI.15 Since then, high BMI has been thought to have a causal relationship with additional site-specific cancers8, 13, 14, 16 and more recent and more detailed global BMI prevalence estimates, based on substantially more data, have become available.2 Diabetes is increasingly recognised as a risk factor for colorectal, pancreatic, liver, gallbladder, breast, and endometrial cancer,17 but the global cancer burden attributable to diabetes has not been quantified. Furthermore, since high BMI is an important risk factor for diabetes, priority setting for public health and clinical interventions requires information on the cancer burden attributable to both high BMI and diabetes. We aimed to estimate the proportion of global cancer incidence in 2012 that was attributable to diabetes and high BMI individually and combined, under varying assumptions about the independence of their effects.

clinical interventions requires information on the cancer burden attributable to both high BMI and diabetes. We aimed to estimate the proportion of global cancer incidence in 2012 that was attributable to diabetes and high BMI individually and combined, under varying assumptions about the independence of their effects. Research in context Evidence before this study We searched MEDLINE via PubMed for articles published up to June 30, 2017, with no language restrictions using the search terms (“Diabetes” OR “Body-mass index” OR “Overweight”, OR “Obesity”), AND (“Cancer risk”, OR “Cancer incidence”), AND “Attributable fraction”. We found one study estimating the burden of cancer associated with type 2 diabetes in 2010 and 2030 in Japan and we found several studies estimating the burden of cancer attributable to high BMI or obesity alone, either in one country or in one country and one cancer site. One previous study quantified the global burden of cancer attributable to high BMI. New, more comprehensive estimates of BMI prevalence have since been published. No previous study has estimated the global burden of cancer attributable to diabetes alone or diabetes and high BMI combined. Added value of this study To our knowledge, this study provides the first estimate of global cancer burden attributable to diabetes alone and to diabetes and high BMI combined, and uses the most comprehensive available estimates of diabetes and high BMI prevalence. We also quantified the global burden of cancer attributable to rises in the prevalence of diabetes and high BMI over time.

first estimate of global cancer burden attributable to diabetes alone and to diabetes and high BMI combined, and uses the most comprehensive available estimates of diabetes and high BMI prevalence. We also quantified the global burden of cancer attributable to rises in the prevalence of diabetes and high BMI over time. Implications of all the available evidence In 2012, about 6% of all incident cancers were attributable to the combined effects of diabetes and high BMI, corresponding to 792 600 cases. As the prevalence of these cancer risk factors increases, clinical and public health efforts should focus on identifying optimal preventive and screening measures for whole populations and individual patients.

nt cancers were attributable to the combined effects of diabetes and high BMI, corresponding to 792 600 cases. As the prevalence of these cancer risk factors increases, clinical and public health efforts should focus on identifying optimal preventive and screening measures for whole populations and individual patients. Methods Study design We reviewed the WCRF continuous update projects, IARC publications, and other published literature that summarised associations of diabetes17 and high BMI with site-specific cancers.4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 We searched MEDLINE via PubMed for articles published up to June 30, 2017, with no language restrictions using the search terms (“Diabetes” OR “Body-mass index” OR “Overweight”, OR “Obesity”), AND (“Cancer risk”, OR “Cancer incidence”), AND “Attributable fraction”. We selected cancers that the WCRF and IARC have judged to have a causal association with high BMI: colorectal, gallbladder, pancreatic, liver, postmenopausal breast, endometrial, kidney, ovarian, stomach cardia, and thyroid cancer, oesophageal adenocarcinoma, and multiple myeloma. For diabetes, we identified published meta-analyses17 of the relative risks (RR) for the association of diabetes with site-specific cancer. The studies included in the meta-analyses had applied rigorous adjustment to control for potential confounding factors, including BMI. The RRs for each site-specific cancer applied in our analysis and their sources are detailed in the appendix (pp 1, 2). For the diabetes analysis we included colorectal, gallbladder, pancreatic, liver, breast, and endometrial cancer.

plied rigorous adjustment to control for potential confounding factors, including BMI. The RRs for each site-specific cancer applied in our analysis and their sources are detailed in the appendix (pp 1, 2). For the diabetes analysis we included colorectal, gallbladder, pancreatic, liver, breast, and endometrial cancer. High BMI has also been proposed to be causally associated with meningioma.13 However, most meningiomas are benign and the incidence of meningioma is not reported in GLOBOCAN. The association between high BMI and oesophageal and stomach cancer is limited to oesophageal adenocarcinoma14 and stomach cardia12 cancer; therefore, we only included these two subtypes in our analysis. Using prevalence of diabetes3 and of categories of BMI2 and RRs for their associations with the cancers identified from published meta-analyses, we estimated the population attributable fraction (PAF) of incident cancers attributable to diabetes and high BMI. For 175 countries in 2012 (appendix p 10), we estimated individual PAFs for each risk factor, as well as two scenarios of diabetes and high BMI combined, one treating their effects as independent and another as overlapping. All analyses were stratified by sex and age group and restricted to people aged 18 years or older. We then estimated the number of cancer cases attributable to diabetes, high BMI, and their combined effect globally by multiplying the PAFs with the number of incident cancers for each age, sex, and country stratum using data from GLOBOCAN.18

atified by sex and age group and restricted to people aged 18 years or older. We then estimated the number of cancer cases attributable to diabetes, high BMI, and their combined effect globally by multiplying the PAFs with the number of incident cancers for each age, sex, and country stratum using data from GLOBOCAN.18 Given the cumulative nature of carcinogenesis, and the importance of risk factor exposure over time, a time lag of several years from exposure to the risk factor and development of the disease is expected. For the association between high BMI and cancer, this lag is commonly assumed to be about 10 years.19 Thus, in our main analysis we calculated cancer incidence in 2012 that we attributed to diabetes and high BMI in 2002. We also estimated cancer incidence due to the change in the prevalence of these two risk factors from 1980 to 2002.

ion between high BMI and cancer, this lag is commonly assumed to be about 10 years.19 Thus, in our main analysis we calculated cancer incidence in 2012 that we attributed to diabetes and high BMI in 2002. We also estimated cancer incidence due to the change in the prevalence of these two risk factors from 1980 to 2002. Data sources We obtained data on the prevalence of diabetes and categories of BMI for 1980 and 2002, stratified by age group (18–19, 20–24, 25–29, 30–34, 35–39, 40–44, 45–49, 50–54, 55–59, 60–64, 65–69, 70–74, 75–79, 80–84, and ≥85 years), sex, and country from estimates2, 3 by the NCD Risk Factor Collaboration (NCD-RisC). BMI data were summarised as prevalence of BMI categories (<18·5, 18·5 to <20, 20 to <25, 25 to <30, 30 to <35, 35 to <40, and ≥40 kg/m2) to characterise the varying shape of the distribution across populations.2 Diabetes was defined as fasting plasma glucose greater than or equal to 7·0 mmol/L, a history of diagnosis of diabetes (we did not differentiate between type 1 and type 2 diabetes), or use of insulin or oral hypoglycaemic drugs. The data sources used by NCD-RisC to estimate BMI and diabetes were checked against a defined set of inclusion criteria, which have been described in detail previously,1, 2 and data were reanalysed according to a common protocol. To avoid potential bias from self-reported data, NCD-RisC only uses data from studies that had measured height and weight or a diabetes biomarker (fasting plasma glucose, 2 h oral glucose tolerance test, or HbA1c). The same criteria and protocol were applied to studies throughout time and across countries. After pooling the data, NCD-RisC fitted a bespoke Bayesian hierarchical model to the data with the Markov chain Monte Carlo algorithm and generated 1000 draws from the posterior distribution for each country-year-sex-age stratum. Details have been reported previously in studies investigating BMI and diabetes.2, 3

ries. After pooling the data, NCD-RisC fitted a bespoke Bayesian hierarchical model to the data with the Markov chain Monte Carlo algorithm and generated 1000 draws from the posterior distribution for each country-year-sex-age stratum. Details have been reported previously in studies investigating BMI and diabetes.2, 3 GLOBOCAN 2012 cancer incidence data18 for the selected cancer sites were available in age groups (15–39, 40–44, 45–49, 50–54, 55–59, 60–64, 65–69, 70–74, and ≥75 years). We used population weighting to ensure that the age groups for diabetes and BMI prevalence were the same as those for cancer incidence. The GLOBOCAN cancer incidence data covered 183 countries and territories, for which both diabetes and BMI estimates were available in 175 of them. We subsequently grouped these 175 countries into nine regions by geographical and national income criteria (appendix p 10). Statistical analysis Most risk factors act proportionally to increase disease risk, therefore we first calculated the proportional reduction of cancer that would occur if exposure to the risk factor was reduced to an alternative scenario, as measured by the PAF.20 The PAF attributable to diabetes and high BMI separately was calculated using the formula21 PAF=∑PiRRi-∑P i′RRi∑PiRRi

Statistical analysis Most risk factors act proportionally to increase disease risk, therefore we first calculated the proportional reduction of cancer that would occur if exposure to the risk factor was reduced to an alternative scenario, as measured by the PAF.20 The PAF attributable to diabetes and high BMI separately was calculated using the formula21 PAF=∑PiRRi-∑P i′RRi∑PiRRi where Pi is the actual prevalence of diabetes or BMI category i, P′i is the prevalence in an alternative scenario, and RRi the adjusted relative risk of site-specific cancer associated with diabetes or the corresponding level of BMI. In our main analysis we estimated the total cancer burden of diabetes and high BMI, and used an optimal prevalence as our alternative scenario—namely zero diabetes prevalence and BMI of 20–25 kg/m2 (used as 22·5 kg/m2 in the calculation), where the cancer risk is assumed to be lowest at the population level. A diabetes prevalence of less than 1% has not been observed,22 so we did a further analysis in which the optimal prevalence of diabetes was 1% rather than zero. We calculated PAFs for 2035 with prevalence in 2025 (projected on the assumption that recent trends continue, as described previously) instead of 2002 prevalence.2, 3, 23 Diabetes and high BMI have increased in prevalence substantially worldwide since 1980.2, 3 We therefore used a second alternative scenario to estimate the cancer burden attributable to these increases. To do this, we replaced the optimal prevalence with the prevalence of diabetes and high BMI in 1980 as the alternative scenario.

nd high BMI have increased in prevalence substantially worldwide since 1980.2, 3 We therefore used a second alternative scenario to estimate the cancer burden attributable to these increases. To do this, we replaced the optimal prevalence with the prevalence of diabetes and high BMI in 1980 as the alternative scenario. We then calculated the PAFs for the combined effects of diabetes and high BMI in two scenarios: diabetes and high BMI as independent risk factors, and a conservative estimate. To calculate combined PAF with high BMI and diabetes as independent risk factors, we used the formula24 PAF =        1 – [(1 – PAFDiabetes) × (1 – PAFHigh BMI)]. For the conservative estimate, we selected the larger of PAFDiabetes and PAFHigh BMI in each age, sex, and country stratum to generate a conservative PAF. This approach assumes complete overlap of pathophysiology of diabetes and high BMI with cancer. We calculated the number of incident cancer cases in 2012 attributable to each risk factor individually and combined as the product of the corresponding PAF and the incident site-specific cancer cases. All analyses were done by sex, age group, and country stratum. To produce aggregated results across age groups, we weighted the age group-specific PAFs by age group-specific cancer incidence by sex and country.

ndividually and combined as the product of the corresponding PAF and the incident site-specific cancer cases. All analyses were done by sex, age group, and country stratum. To produce aggregated results across age groups, we weighted the age group-specific PAFs by age group-specific cancer incidence by sex and country. We propagated the uncertainties of diabetes and BMI prevalence estimates and those of the RRs to the final estimates using a simulation approach. Specifically, we generated 1000 draws for each RR from a log-normal distribution, with mean equal to the reported estimate and SD calculated with the reported confidence interval and 1000 draws from the posterior distributions of diabetes3 and high BMI prevalence.2 We repeated the PAF calculation for each of these draws, resulting in 1000 PAFs which characterised the uncertainty distribution of the output. We report 95% uncertainty intervals (95% UI) for our estimates as the 2·5th to 97·5th percentile of the resultant distributions. All analyses were done with R version 3.2.5.25 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. JP-S, BZ, VK, and JB, had full access to all the data in the study and the corresponding author had final responsibility for the decision to submit for publication.

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. JP-S, BZ, VK, and JB, had full access to all the data in the study and the corresponding author had final responsibility for the decision to submit for publication. Results In 2012, diabetes and high BMI combined were responsible for an estimated 792 600 new cases of cancer worldwide (5·6% of all 14 067 894 cancer cases reported by GLOBOCAN18) in the independent scenario. 280 100 (2·0%) cancer cases were attributable to diabetes and 544 300 (3·9%) to high BMI alone (Figure 1, Figure 2). In the conservative scenario, the two risk factors combined were responsible for 626 900 new cancer cases in 2012. Cancer cases attributable to diabetes and high BMI combined were almost twice as common in women (496 700 cases) as in men (295 900 cases) in the independent scenario.Figure 1 Global cancer cases in 2012 attributable to diabetes and high BMI, individually and combined, in the conservative and independent scenarios, by region BMI=body-mass index. Figure 2 Global site-specific cancer cases in 2012 Cases by (A) diabetes and high BMI, individually and in combination, in the conservative and independent scenarios and (B) region, in the combined independent scenario. BMI=body-mass index.

Results In 2012, diabetes and high BMI combined were responsible for an estimated 792 600 new cases of cancer worldwide (5·6% of all 14 067 894 cancer cases reported by GLOBOCAN18) in the independent scenario. 280 100 (2·0%) cancer cases were attributable to diabetes and 544 300 (3·9%) to high BMI alone (Figure 1, Figure 2). In the conservative scenario, the two risk factors combined were responsible for 626 900 new cancer cases in 2012. Cancer cases attributable to diabetes and high BMI combined were almost twice as common in women (496 700 cases) as in men (295 900 cases) in the independent scenario.Figure 1 Global cancer cases in 2012 attributable to diabetes and high BMI, individually and combined, in the conservative and independent scenarios, by region BMI=body-mass index. Figure 2 Global site-specific cancer cases in 2012 Cases by (A) diabetes and high BMI, individually and in combination, in the conservative and independent scenarios and (B) region, in the combined independent scenario. BMI=body-mass index. In men, 126 700 cases (95% UI 95 900–159 400) were from liver cancer, constituting 42·8% of all cancer cases attributable to diabetes and high BMI combined in the independent scenario; colorectal cancer cases (63 200 cases, 40 600–86 000) were the next largest contributor, constituting 21·4% of the total cases (Figure 1, Figure 2; table 1). In women, there were 147 400 cases (106 700–190 000) of breast cancer, constituting 29·7% of all cancers attributable to diabetes and high BMI; the second largest contributor was endometrial cancer (121 700 cases, 108 600–135 000), which constituted 24·5% of such cases.Table 1 PAF and number of cancer cases attributable to high BMI and diabetes in 2012, individually and in combination, in independent and conservative scenarios

butable to diabetes and high BMI; the second largest contributor was endometrial cancer (121 700 cases, 108 600–135 000), which constituted 24·5% of such cases.Table 1 PAF and number of cancer cases attributable to high BMI and diabetes in 2012, individually and in combination, in independent and conservative scenarios Total number of cases High BMI PAF High BMI cases Diabetes PAF Diabetes cases Independent PAF Independent scenario cases Conservative PAF Conservative scenario cases Men Colorectal 736 000 5·8% (4·2–7·4) 42 200 (30 600–54 800) 2·9% (0·5–5·7) 21 600 (4200–42 100) 8·6% (5·5–11·7) 63 200 (40 600–86 000) 6·3% (4·6–8·0) 46 300 (33 800–58 600) Gallbladder 76 000 9·7% (5·8–13·2%) 7400 (4500–10 100) 7·8% (4·0–11·9) 5900 (3000–9200) 16·7% (11·9–21·8) 12 800 (9100–16 600) 11·7% (8·2–15·4) 9000 (6300–11 800) Liver 543 000 10·1% (5·7–14·7) 54 600 (31 100–79 600) 14·5% (0·8–19·7) 80 200 (54 700–107 800) 23·3% (17·6–29·3) 126 700 (95 900–159 400) 16·5% (12·4–21·2) 89 500 (67 600–115 400) Pancreas 177 000 5·8% (3·9–7·8) 10 300 (6·800–13 700) 12·8% (9·3–16·8) 22 700 (16 200–29 500) 18·0% (14·0–21·6) 31 900 (24 700–38 100) 13·2% (9·7–16·6) 23 300 (17 200–29 300) Kidney 208 000 18·0% (15·5–20·4) 37 400 (32 100–42 300) ·· ·· 18·0% (15·5–20·4) 37 400 (32 100–42 300) 18·0% (15·5–20·4) 37 400 (32 100–42 300) Oesophagus (adenocarcinoma) 31 700 28·7% (22·6–35·0) 9100 (7200–11 100) ·· ·· 28·7% (22·6–35·0) 9100 (7200–11 100) 28·7% (22·6–35·0) 9100 (7200–11 100) Stomach (cardia) 72 700 8·8% (3·0–14·8) 6400 (2200–10 800) ·· ·· 8·8% (3·0–14·8) 6400 (2200–10 800) 8·8% (3·0–14·8) 6400 (2200–10 800) Multiple myeloma 61 900 7·2% (3·3–11·1) 4500 (2100–6900) ·· ·· 7·2% (3·3–11·1) 4500 (2100–6900) 7·2% (3·3–11·1) 4500 (2100–6900) Thyroid 67 000 5·8% (2·8–8·8) 3900 (1900–5900) ·· ·· 5·8% (2·8–8·8) 3900 (1900–5900) 5·8% (2·8–8·8) 3900 (1900–5900) Total 1 973 300 8·9% 175 800 8·5% 130 400 15·0% 295 900 11·6% 229 400 Women Breast 1 656 000 6·9% (4·4–9·4) 114 800 (72 700–156 500) 2·2% (1·3–3·2) 36 200 (21 400–51 600) 8·9% (6·4–11·5) 147 400 (106 700–190 000) 7·2% (4·9–9·8) 120 000 (82 500–161 500) Endometrial 317 000 31·0% (27·1–35·2) 98 400 (86 000–111 500) 10·8% (7·8–13·8) 33 700 (25 100–43 900) 38·4% (34·3–42·6) 121 700 (108 600–135 000) 31·3% (27·4–35·4) 99 100 (87 000–112 200) Colorectal 607 000 7·0% (5·0–9·1) 42 300 (30 200–55 000) 2·8% (0·5–5·3) 16 900 (3200–32 600) 9·7% (6·3–12·7) 58 600 (38 400–77 300) 7·3% (5·2–9·1) 44 200 (31 800–55 400) Gallbladder 101 000 12·9% (7·8–17·6) 13 000 (7900–17 7

0–43 900) 38·4% (34·3–42·6) 121 700 (108 600–135 000) 31·3% (27·4–35·4) 99 100 (87 000–112 200) Colorectal 607 000 7·0% (5·0–9·1) 42 300 (30 200–55 000) 2·8% (0·5–5·3) 16 900 (3200–32 600) 9·7% (6·3–12·7) 58 600 (38 400–77 300) 7·3% (5·2–9·1) 44 200 (31 800–55 400) Gallbladder 101 000 12·9% (7·8–17·6) 13 000 (7900–17 7 00) 7·4% (4·0–11·5) 7600 (3800–11 500) 19·3% (13·6–25·1) 19 400 (13 700–25 200) 13·8% (9·4–18·1) 13 900 (9500–18 300) Liver 223 000 13·5% (7·8–219·4) 30 200 (17 400–43 200) 15·8% (10·9–21·4) 35 300 (24 400–47 200) 27·3% (20·9–33·9) 60 900 (46 500–75 600) 18·8% (14·4–23·8) 42 000 (32 100–53 000) Pancreas 159 000 7·1% (4·6–9·4) 11 200 (7300–15 000) 12·6% (9·2–16·6) 20 000 (14 500–26 200) 19·0% (14·6–22·7) 30 100 (23 200–36 100) 13·1% (9·8–16·5) 20 700 (15 600–26 300) Kidney 118 000 21·3% (18·3–24·1) 25 200 (21 600–28 500) ·· ·· 21·3% (18·3–24·1) 25 200 (21 600–28 500) 21·3% (18·3–24·1) 25 200 (21 600–28 500) Ovarian 235 000 3·9% (0·9–6·7) 9100 (2000–15 800) ·· ·· 3·9% (0·9–6·7) 9100 (2000–15 800) 3·9% (0·9–6·7) 9100 (2000–15 800) Oesophagus (adenocarcinoma) 7300 29·5% (23·1–36·1) 2200 (1700–2600) ·· ·· 29·5% (23·1–36·1) 2200 (1700–2600) 29·5% (23·1–36·1) 2200 (1700–2600) Stomach (cardia) 26 400 11·2% (3·8–18·8) 2900 (1000–5000) ·· ·· 11·2% (3·8–18·8) 2900 (1000–5000) 11·2% (3·8–18·8) 2900 (1000–5000) Multiple myeloma 51 400 8·9% (4·0–13·3) 4400 (2000–6800) 4·9% ·· 8·9% (4·0–13·3) 4400 (2000–6800) 8·9% (4·0–13·3) 4400 (2000–6800) Thyroid 226 400 6·5% (3·2–9·8) 14 800 (7300–22 100) ·· ·· 6·5% (3·2–9·8) 14 800 (7300–22 100) 6·5% (3·2–9·8) 14 800 (7300–22 100) Total 3 727 500 9·9% 368 500 4·9% 149 700 13·3% 496 700 10·7% 398 500 Numbers in parentheses show 95% UI.

2000–6800) 4·9% ·· 8·9% (4·0–13·3) 4400 (2000–6800) 8·9% (4·0–13·3) 4400 (2000–6800) Thyroid 226 400 6·5% (3·2–9·8) 14 800 (7300–22 100) ·· ·· 6·5% (3·2–9·8) 14 800 (7300–22 100) 6·5% (3·2–9·8) 14 800 (7300–22 100) Total 3 727 500 9·9% 368 500 4·9% 149 700 13·3% 496 700 10·7% 398 500 Numbers in parentheses show 95% UI. PAF=population attributable fraction. BMI=body-mass index. Of the six cancers associated with diabetes and 12 associated with high BMI, 15·0% in men and 13·3% in women were attributable to the combined effect of these risk factors in the independent scenario (11·6% in men and 10·7% in women in the conservative scenario; table 1). The PAF varied substantially by cancer site in both sexes. Of all liver cancers, 23·3% (17·6–29·3) in men and 27·3% (20·9–33·9) in women were attributable to diabetes and high BMI combined, compared with just 8·6% (5·5–11·7) of cases of colorectal cancer in men and 9·7% (6·3–12·7) in women. 38·4% (34·3–42·6) of all endometrial cancer cases in 2012 were attributable to these risk factors compared with 3·9% (0·9–6·7) of ovarian cancer cases (table 1).

men were attributable to diabetes and high BMI combined, compared with just 8·6% (5·5–11·7) of cases of colorectal cancer in men and 9·7% (6·3–12·7) in women. 38·4% (34·3–42·6) of all endometrial cancer cases in 2012 were attributable to these risk factors compared with 3·9% (0·9–6·7) of ovarian cancer cases (table 1). There were notable differences in the proportion of cancer cases attributable to diabetes versus high BMI individually. For example, high BMI was responsible for about three times the proportion of breast (6·9%) and endometrial (31·0%) cancers as compared with diabetes (2·2% for breast and 10·8% for endometrial; table 1). By contrast, the proportion of liver (14·5%) and pancreatic (12·8%) cancer in men attributable to diabetes was substantially larger than that attributable to high BMI (10·1% for liver and 5·8% for pancreatic). When using 1% as the optimal diabetes prevalence rather than zero, this resulted in a reduction in cancer cases attributable to diabetes by 6·8% (261 000 vs 280 100). 303 000 (38·2%) of 792 600 cases of cancer attributable to the combined risk of diabetes and high BMI in the independent scenario in 2012 occurred in high-income western countries (Figure 1, Figure 2). East and southeast Asia had the second largest proportion (190 900 [24·1%]) of cases attributable to the combined risk of diabetes and high BMI, and the largest number of cancer cases attributable to diabetes individually (105 500 attributable cases) (figure 2).

ed in high-income western countries (Figure 1, Figure 2). East and southeast Asia had the second largest proportion (190 900 [24·1%]) of cases attributable to the combined risk of diabetes and high BMI, and the largest number of cancer cases attributable to diabetes individually (105 500 attributable cases) (figure 2). The contribution of each cancer site to the regional cancer burden also varied substantially. Of the total cancer burden due to the combination of diabetes and high BMI, liver cancer contributed more than 30·7% in the high-income Asia Pacific region and 53·8% in east and southeast Asia, compared with just 7·0% in central and eastern Europe (figure 2B). By contrast, breast and endometrial cancer contributed about 18·5% of the combined cancer burden in east and southeast Asia and 15·6% in the high-income Asia Pacific region, compared with roughly 40·9% in high-income western countries, central and eastern Europe, and sub-Saharan Africa. There were substantial differences in the PAF of cancer attributable to diabetes and those attributable to high BMI in some regions, for example in women in central Asia, the Middle East, and north Africa (5·6% for diabetes vs 14·3% for high BMI; table 2), and in men in east and southeast Asia (10·0% for diabetes vs 5·6% for high BMI)—where diabetes3 has increased faster than expected by the rise in BMI.2Table 2 Regional cancer cases in 2012 attributable to 2002 prevalence and cancer cases that would have been expected in 2012 had prevalence remained at 1980 levels

e 2), and in men in east and southeast Asia (10·0% for diabetes vs 5·6% for high BMI)—where diabetes3 has increased faster than expected by the rise in BMI.2Table 2 Regional cancer cases in 2012 attributable to 2002 prevalence and cancer cases that would have been expected in 2012 had prevalence remained at 1980 levels Number of cases Cases attributable to 2002 prevalence Proportion of cases attributable to 2002 prevalence Cases attributable to 1980 prevalence Proportion of cases attributable to 1980 prevalence Diabetes Men Central and eastern Europe 114 000 6600 (4200–9600) 5·8% 5400 (2800 −10 000) 4·7% Central Asia and north Africa and the Middle East 56 000 6200 (4600–8000) 11·1% 3900 (1900–6700) 7·0% East and southeast Asia 616 000 61 800 (42 300–82 600) 10·0% 36 400 (14 200–72 700) 5·9% High-income Asia Pacific region 157 000 14 000 (10 300–18 300) 8·9% 10 900 (6300–16 500) 6·9% High-income western countries 385 000 26 000 (18 000–34 500) 6·8% 20 200 (11 900–32 900) 5·2% Latin America and the Caribbean 76 000 6200 (4400–8200) 8·2% 4700 (2700–7700) 6·2% Oceania 800 90 (60–120) 11·3% 50 (20–100) 6·3% South Asia 83 000 6600 (4600–9000) 8·0% 3500 (1500–6700) 4·2% Sub-Saharan Africa 44 000 2900 (2000–4100) 6·6% 1600 (600–3500) 3·6% Women Central and eastern Europe 297 000 16 000 (11 500–21 600) 5·4% 15 000 (8400–24 000) 5·1% Central Asia and north Africa and the Middle East 149 000 8400 (6700–10 200) 5·6% 5300 (2800–9200) 3·6% East and southeast Asia 720 000 43 700 (32 600–56 100) 6·1% 33 800 (15 300–63 500) 4·7% High-income Asia Pacific region 201 000 11 400 (8800–14 700) 5·7% 10 300 (6400–15 200) 5·1% High-income western countries 1 019 000 41 300 (32 200–52 000) 4·1% 36 200 (23 700–54 800) 3·6% Latin America and the Caribbean 254 000 13 400 (10 500–17 100) 5·3% 10 300 (5900–16 500) 4·1% Oceania 2000 130 (90–180) 6·5% 70 (30–150) 3·5% South Asia 283 000 10 800 (7800–14 600) 3·8% 6400 (2800–13 500) 2·3% Sub-Saharan Africa 138 000 4400 (3200–5900) 3·2% 2800 (1300–5700) 2·0% High BMI Men Central and eastern Europe 146 000 18 800 (15 100–22 700) 12·9% 13 400 (10 400–16 900) 9·2% Central Asia and north Africa and the Middle East 67 000 9800 (7200–12 600) 14·6% 6100 (4200–8400) 9·1% East and southeast Asia 711 000 40 000 (25 800–56 100) 5·6% 16 500 (9500–16 500) 2·3% High-income Asia Pacific region 182 000 8600 (6300–11 100) 4·7% 4900 (3500–6800) 2·7% High-income Western countries 502 000 82 200 (65 200–99 000) 16·4% 57 900 (44 900–70 900) 11·5% Latin America and the Caribbean 94

(4200–8400) 9·1% East and southeast Asia 711 000 40 000 (25 800–56 100) 5·6% 16 500 (9500–16 500) 2·3% High-income Asia Pacific region 182 000 8600 (6300–11 100) 4·7% 4900 (3500–6800) 2·7% High-income Western countries 502 000 82 200 (65 200–99 000) 16·4% 57 900 (44 900–70 900) 11·5% Latin America and the Caribbean 94  000 12 300 (9600–15 000) 13·1% 7300 (5500–9600) 7·8% Oceania 800 100 (60–130) 12·5% 60 (40–90) 7·5% South Asia 96 000 2600 (1900–3500) 2·7% 1100 (600–1700) 1·1% Sub-Saharan Africa 46 000 2000 (1300–2800) 4·3% 900 (600–1500) 2·0% Women Central and eastern Europe 348 000 58 700 (49 100–68 500) 16·9% 51 700 (42 600–61 400) 14·9% Central Asia and north Africa and the Middle East 167 000 23 800 (19 100–28 400) 14·3% 16 800 (12 900–21 000) 10·1% East and southeast Asia 815 000 48 000 (38 400–57 700) 5·9% 25 100 (18 500–33 300) 3·1% High-income Asia Pacific region 224 000 10 900 (8600–13 400) 4·9% 8600 (6600–10 800) 3·8% High-income western countries 1 136 000 170 200 (138 000–202 300) 15·0% 124 200 (100 000–149 600) 10·9% Latin America and the Caribbean 281 000 37 700 (30 500–45 000) 13·4% 26 600 (21 000–32 900) 9·5% Oceania 2000 300 (230–370) 15·0% 200 (140–270) 10·0% South Asia 323 000 9800 (7400–12 300) 3·0% 4700 (3000–6700) 1·5% Sub-Saharan Africa 153 000 9700 (7700–11 800) 6·3% 5400 (4100–7000) 3·5% Data are stratified by sex. Numbers in parentheses are 95% UI. BMI=body-mass index.

000 12 300 (9600–15 000) 13·1% 7300 (5500–9600) 7·8% Oceania 800 100 (60–130) 12·5% 60 (40–90) 7·5% South Asia 96 000 2600 (1900–3500) 2·7% 1100 (600–1700) 1·1% Sub-Saharan Africa 46 000 2000 (1300–2800) 4·3% 900 (600–1500) 2·0% Women Central and eastern Europe 348 000 58 700 (49 100–68 500) 16·9% 51 700 (42 600–61 400) 14·9% Central Asia and north Africa and the Middle East 167 000 23 800 (19 100–28 400) 14·3% 16 800 (12 900–21 000) 10·1% East and southeast Asia 815 000 48 000 (38 400–57 700) 5·9% 25 100 (18 500–33 300) 3·1% High-income Asia Pacific region 224 000 10 900 (8600–13 400) 4·9% 8600 (6600–10 800) 3·8% High-income western countries 1 136 000 170 200 (138 000–202 300) 15·0% 124 200 (100 000–149 600) 10·9% Latin America and the Caribbean 281 000 37 700 (30 500–45 000) 13·4% 26 600 (21 000–32 900) 9·5% Oceania 2000 300 (230–370) 15·0% 200 (140–270) 10·0% South Asia 323 000 9800 (7400–12 300) 3·0% 4700 (3000–6700) 1·5% Sub-Saharan Africa 153 000 9700 (7700–11 800) 6·3% 5400 (4100–7000) 3·5% Data are stratified by sex. Numbers in parentheses are 95% UI. BMI=body-mass index. There was substantial heterogeneity in the proportion of cancer cases attributable to diabetes, high BMI, and their combination in the independent scenario at country level. For example, less than 1% of all new cancer cases in Malawi (0·6%) and Tanzania (0·9%) in 2012 were attributable to diabetes and high BMI combined, compared with more than 10% in Egypt (12·0%) and Mongolia (13·9%)—the countries with the largest PAF—reflecting large variations in risk factor prevalence, and in the way that some cancers are more affected by these factors than others (figure 3).Figure 3 Population attributable fraction of all cancer incidence in 2012

compared with more than 10% in Egypt (12·0%) and Mongolia (13·9%)—the countries with the largest PAF—reflecting large variations in risk factor prevalence, and in the way that some cancers are more affected by these factors than others (figure 3).Figure 3 Population attributable fraction of all cancer incidence in 2012 Population attributable fractions shown are those of (A) diabetes, (B) high BMI, and (C) diabetes and high BMI combined as independent risks. Countries shown in grey did not have cancer incidence data. BMI=body-mass index. We calculated that 26·1% of all cancer cases in 2012 attributable to diabetes were due to the increase in diabetes prevalence from 1980 to 2002 (table 2), equating to 77 000 new cases worldwide. 31·9% of cancer cases attributable to high BMI were due to increased prevalence of this risk factor over the same period, accounting for approximately 174 040 cancer cases. The largest proportion of cancer cases attributable to the increase in prevalence of diabetes and high BMI during this period was in low-income and middle-income countries (LMICs) in Asia and sub-Saharan Africa. At the two extremes, just 7% of cancer cases attributable to diabetes were due to increased diabetes prevalence in women in central and eastern Europe, compared with 48% in men in south Asia.

f diabetes and high BMI during this period was in low-income and middle-income countries (LMICs) in Asia and sub-Saharan Africa. At the two extremes, just 7% of cancer cases attributable to diabetes were due to increased diabetes prevalence in women in central and eastern Europe, compared with 48% in men in south Asia. The PAF of cancer attributable to diabetes and high BMI is expected to increase substantially in coming decades (appendix p 5). For example, PAFs for most site-specific cancers would increase by more than 30% in women and 20% in men when using projected 2025 prevalence compared with 2002 prevalence. In men, the PAF for liver cancer would increase by 47% (from 23·3% to 34·3%) and gallbladder cancer would increase by 53% (from 16·7% to 25·5%), while in women, the PAF for ovarian cancer would increase by 38% (from 3·9% to 5·4%). Discussion We estimated that approximately 6% of cancer cases worldwide in 2012 were attributable to diabetes and high BMI, with high BMI being responsible for almost twice as many cases as diabetes. About a third of cancer cases attributable to diabetes and a quarter of cases attributable to high BMI were due to increases in the prevalence of these risk factors from 1980 to 2002. Given the continued rise in the prevalence of these risk factors since 2002,2, 3 the attributable cancer burden is likely to continue to increase in coming decades. Approximately one in four liver and oesophageal adenocarcinomas and 38·4% of endometrial cancers worldwide in 2012 were estimated to be attributable to diabetes and high BMI.

rise in the prevalence of these risk factors since 2002,2, 3 the attributable cancer burden is likely to continue to increase in coming decades. Approximately one in four liver and oesophageal adenocarcinomas and 38·4% of endometrial cancers worldwide in 2012 were estimated to be attributable to diabetes and high BMI. LMICs have had substantial increases in the prevalence of diabetes and high BMI during the past three decades, whereas parts of Europe and the high-income Asia Pacific region have seen more stable age-standardised prevalences (appendix p 7).2, 3 In our analysis LMICs had the largest increases in numbers of cancer cases attributable both to diabetes, and diabetes and high BMI combined, which is particularly important to note because these countries are generally less well equipped to manage the burden of complex non-communicable diseases (NCDs) than high-income countries. Previous studies have quantified the global cancer burden attributable to nine potentially modifiable diet and lifestyle risk factors (PAF 35% in 2001),26 smoking (PAF 21% in 2000),27 high BMI (PAF 3·6% in 2012),15 and common infections (PAF 15·4% in 2012).28 Our findings suggest that 3·9% of global cancer cases in 2012 were attributable to high BMI, taking into account the four additional cancer sites and more comprehensive and up-to-date BMI data compared with previous work.15

(PAF 21% in 2000),27 high BMI (PAF 3·6% in 2012),15 and common infections (PAF 15·4% in 2012).28 Our findings suggest that 3·9% of global cancer cases in 2012 were attributable to high BMI, taking into account the four additional cancer sites and more comprehensive and up-to-date BMI data compared with previous work.15 Proposed biological mechanisms underlying the link between diabetes, high BMI, and cancer include hyperinsulinaemia, hyperglycaemia, chronic inflammation,29 and dysregulation of sex hormone activity. Insulin itself could be oncogenic,30 and results from several analyses showed that people with hyperinsulinaemia were at increased risk of breast and colorectal cancer irrespective of their BMI.31, 32, 33 Prospective studies and large-scale consortia with more accurate assessments of adiposity, diabetes, and metabolic health, which incorporate molecular tools, will be needed to draw conclusions about the underlying mechanisms that link diabetes, high BMI, and cancer, and inform clinical interventions.

eir BMI.31, 32, 33 Prospective studies and large-scale consortia with more accurate assessments of adiposity, diabetes, and metabolic health, which incorporate molecular tools, will be needed to draw conclusions about the underlying mechanisms that link diabetes, high BMI, and cancer, and inform clinical interventions. To our knowledge, this is the only study to have quantified the global burden of cancer attributable to diabetes and to diabetes combined with high BMI, by use of robust evidence from WCRF4, 5, 6, 7, 8, 9, 10, 11, 12, 14 for BMI and high quality meta-analyses for diabetes.17 Our findings are important to policy makers developing coordinated approaches to tackle the rising prevalence of diabetes, high BMI, and all of their sequelae. The cancers judged to have a convincing association with diabetes by the umbrella meta-analysis were restricted to those for which the effect of study bias was expected to be lowest.

tant to policy makers developing coordinated approaches to tackle the rising prevalence of diabetes, high BMI, and all of their sequelae. The cancers judged to have a convincing association with diabetes by the umbrella meta-analysis were restricted to those for which the effect of study bias was expected to be lowest. Our study has some limitations. The precision of the risk estimates used to adjust for common confounders, including diabetes and BMI, might be affected by potential biases such as reverse causality and ascertainment bias, which are believed to affect some estimates of the association between diabetes and cancer.34 We used the same relative risk for age group, sex, and region; more granular risk estimates by age, sex, and stage of diagnosis would allow for greater accuracy at the subgroup level. We quantified the cancer burden attributable to all BMI levels greater than 25 kg/m2. Some researchers have argued that Asian populations might need BMI cutoffs that are different from other populations,35 although meta-analyses of Asian and western cohorts have shown that disease risk increases by similar proportions in Asian and western populations36, 37, 38, 39 and indeed the latest WHO consensus statement on BMI cutoffs, having considered the arguments for region-specific cutoffs, recommended use of similar cutoffs throughout the world.35 The mediated and direct effects of diabetes and high BMI on cancer—which would allow for more accurate estimation of their combined contributions to the cancer burden—have not yet been quantified in the way that has been done for cardiovascular diseases.40 Additionally, the 10-year lag from diabetes and high BMI prevalence to cancer incidence that we used is an imperfect measure of cumulative past risk factor exposure, which is important for cancer burden.41 Our PAF analysis quantified the proportion and number of cancer cases that would be averted if diabetes and high BMI prevalence were reduced to optimal levels.

nd high BMI prevalence to cancer incidence that we used is an imperfect measure of cumulative past risk factor exposure, which is important for cancer burden.41 Our PAF analysis quantified the proportion and number of cancer cases that would be averted if diabetes and high BMI prevalence were reduced to optimal levels. However, if the cancer burden of diabetes and high BMI is removed, these risks could lead to populations developing other disorders such as cardiovascular disease and chronic kidney disease as quantified elsewhere.42 Finally, we assumed an optimal diabetes prevalence of zero, and achieving a prevalence of less than 1% might not be feasible.22 Nonetheless, when we substituted zero for 1% as the optimal diabetes prevalence, the cancer burden attributable to diabetes changed by less than 7% and was still responsible for 261 000 cases.

Finally, we assumed an optimal diabetes prevalence of zero, and achieving a prevalence of less than 1% might not be feasible.22 Nonetheless, when we substituted zero for 1% as the optimal diabetes prevalence, the cancer burden attributable to diabetes changed by less than 7% and was still responsible for 261 000 cases. Trends in diabetes and those in BMI were only partly correlated across regions. For example, in south Asia and possibly east Asia diabetes prevalence has risen faster than would be expected by changes in BMI levels, whereas in northern Europe diabetes prevalence is increasing at a slower rate than might be expected by the changes in BMI. Several factors might be causing these diverse trends. First, regional differences in the prevalence of diabetes might be due to differences in genetic susceptibility or phenotypic variations arising from inadequate fetal and childhood nutrition and growth; earlier onset of β-cell dysfunction could be a differentiating characteristic of Asian populations compared with other groups.43, 44, 45, 46, 47 Second, people who are at high risk of developing diabetes might be identified at an earlier stage in health systems in high-income countries, allowing for earlier intervention with lifestyle and dietary modification or drugs.48 Finally, total caloric intake, dietary composition, and physical activity might affect diabetes risk and contribute to differences in regional trends to a greater extent than would otherwise be expected on the basis of BMI.49

untries, allowing for earlier intervention with lifestyle and dietary modification or drugs.48 Finally, total caloric intake, dietary composition, and physical activity might affect diabetes risk and contribute to differences in regional trends to a greater extent than would otherwise be expected on the basis of BMI.49 Our results suggest that the increases in diabetes and BMI worldwide could lead to a substantial increase in the cancer burden in future decades. For example, when we used 2025 projections for diabetes and BMI prevalence we found that a substantially larger share of cancers would be attributable to these risk factors in the future than in 2012. PAFs for all site-specific cancers would be significantly higher if trends in diabetes and BMI continue as projected, with the largest increases in gallbladder, liver, and endometrial cancers. These projections are particularly alarming in view of the high, and growing, economic cost of cancers and metabolic diseases, and highlight the importance of integrated control measures to tackle common modifiable risk factors, alongside clinician awareness of diabetes and high BMI as established risk factors for common cancers.

ctions are particularly alarming in view of the high, and growing, economic cost of cancers and metabolic diseases, and highlight the importance of integrated control measures to tackle common modifiable risk factors, alongside clinician awareness of diabetes and high BMI as established risk factors for common cancers. Population-based strategies to prevent diabetes and high BMI have great potential impact—not least because many NCDs have overlapping risk factors, comorbidities, and shared sequelae—but have so far often failed, largely because of reluctance by governments and policy makers to pursue structural interventions that tackle key risks for NCDs, such as diet and physical inactivity.1 Future efforts should focus on identifying the most effective clinical interventions to prevent development of NCDs in at-risk groups and their sequelae, such as cancer. Primary care interventions, such as glucose-modifying medications, can be effective in preventing diabetes complications such as macrovascular disease,50 but this approach relies on early identification and close monitoring of people with diabetes, which can be challenging in LMICs that have limited resources. As well as coordinated approaches to halt and reverse the rise in NCDs, global efforts and clinical guidance should reflect the importance of cancer as a sequela of both diabetes and high BMI, and NCD control measures should be integrated into clinical guidelines to identify opportunities to reduce morbidity in this group of patients. Supplementary Material Supplementary appendix

Population-based strategies to prevent diabetes and high BMI have great potential impact—not least because many NCDs have overlapping risk factors, comorbidities, and shared sequelae—but have so far often failed, largely because of reluctance by governments and policy makers to pursue structural interventions that tackle key risks for NCDs, such as diet and physical inactivity.1 Future efforts should focus on identifying the most effective clinical interventions to prevent development of NCDs in at-risk groups and their sequelae, such as cancer. Primary care interventions, such as glucose-modifying medications, can be effective in preventing diabetes complications such as macrovascular disease,50 but this approach relies on early identification and close monitoring of people with diabetes, which can be challenging in LMICs that have limited resources. As well as coordinated approaches to halt and reverse the rise in NCDs, global efforts and clinical guidance should reflect the importance of cancer as a sequela of both diabetes and high BMI, and NCD control measures should be integrated into clinical guidelines to identify opportunities to reduce morbidity in this group of patients. Supplementary Material Supplementary appendix Contributors JP-S and ME conceived the idea of the study. JP-S led the analysis with support from BZ, VK, and JB. ME and MJG supervised the analysis and generating of results. JP-S drafted and finalised the paper with input from all authors. All authors contributed to the analysis, intellectual content, critical revisions to the drafts of the paper and approved the final version. ME had full access to all the data in the study and had final responsibility for the decision to submit for publication.

afted and finalised the paper with input from all authors. All authors contributed to the analysis, intellectual content, critical revisions to the drafts of the paper and approved the final version. ME had full access to all the data in the study and had final responsibility for the decision to submit for publication. Declaration of interests ME reports a charitable grant from the Young Health Programme of AstraZeneca, and personal fees from Third Bridge, Scor, and Prudential, outside the submitted work. All other authors declare no competing interests.

well as total macular volume.2 Not all patients with centre-involving diabetic macular oedema have impaired vision. Therefore, progression of diabetic macular oedema to the centre is better measured as an increase in central subfield thickness (zone 1) than change in visual acuity.Figure 1 ETDRS grid within OCT devices The ETDRS grid divides the macula into nine zones in the right and left eye, with zone 1 the central subfield, zones 2–5 parafoveal zones, and zones 6–9 perifoveal zones. ETDRS=Early Treatment Diabetic Retinopathy Study. OCT=optical coherence tomography. Approximately 8% of people with diabetes have centre-involving diabetic macular oedema and a further 8% have non-central diabetic macular oedema. Standard treatments for people with centre-involving diabetic macular oedema are invasive and include repeated intravitreal injections of anti-vascular endothelial growth factor (anti-VEGF) agents, use of steroids, or laser treatment.3, 4 Less invasive treatment will save patients from these procedures and have a major positive effect on health-care spending. Furthermore, there is an unmet need for preventive and treatment options for non-central diabetic macular oedema to prevent potential visual morbidity due to disease progression.5 Research in context Evidence before this study

Introduction Diabetic macular oedema is the most common cause of moderate visual loss in people with diabetes mellitus.1 The location and quantity of increased retinal thickness due to diabetic macular oedema can be recorded objectively using optical coherence tomography (OCT). All OCT devices have an in-built Early Treatment Diabetic Retinopathy Study (ETDRS) grid with nine zones (subfields) centred on the fovea (figure 1). Centre-involving diabetic macular oedema is defined as presence of oedema in zone 1. Non-centre-involving diabetic macular oedema is oedema restricted to one or more of zones 2–9. The zone of maximum retinal thickness can be monitored over time to assess the course of disease. However, diabetic macular oedema can appear and disappear at any zone and might represent worsening of disease despite resolution of oedema at the zone of maximum retinal thickness. Therefore, response to treatment must be assessed objectively by looking at several variables, including change in retinal thickness in all nine zones as well as total macular volume.2 Not all patients with centre-involving diabetic macular oedema have impaired vision. Therefore, progression of diabetic macular oedema to the centre is better measured as an increase in central subfield thickness (zone 1) than change in visual acuity.Figure 1 ETDRS grid within OCT devices

Approximately 8% of people with diabetes have centre-involving diabetic macular oedema and a further 8% have non-central diabetic macular oedema. Standard treatments for people with centre-involving diabetic macular oedema are invasive and include repeated intravitreal injections of anti-vascular endothelial growth factor (anti-VEGF) agents, use of steroids, or laser treatment.3, 4 Less invasive treatment will save patients from these procedures and have a major positive effect on health-care spending. Furthermore, there is an unmet need for preventive and treatment options for non-central diabetic macular oedema to prevent potential visual morbidity due to disease progression.5 Research in context Evidence before this study We searched PubMed from database inception until Sept 1, 2017, and abstracts from annual meetings of the Association for Research in Vision and Ophthalmology until 2017, with the terms “lightmasks” AND “diabetic retinopathy” OR “diabetic macular oedema” for reports of randomised controlled trials published in English only. We identified no randomised trials. Non-centre-involving diabetic macular oedema can progress to the centre of the macula and cause visual impairment. Standard treatment for centre-involving diabetic macular oedema with visual impairment is repeated intravitreal injections with anti-vascular endothelial growth factor agents. Other treatment options include macular laser and intravitreal steroids. Many investigators have evaluated less invasive treatment options for non-centre-involving macular oedema to delay or prevent disease progression but have been unsuccessful. Rod photoreceptors in the retina consume a maximum amount of oxygen during dark adaptation. In diabetes, the resultant hypoxia can contribute to development and progression of diabetic macular oedema and diabetic retinopathy. Two short-term clinical trials showed that wearing light masks emitting 500–505 nm light through the eyelids to decrease dark adaptation reduced the rate of progression of diabetic retinopathy and early diabetic macular oedema, respectively. In an abstract, 45 healthy volunteers wore an organic light-emitting sleep mask to prevent dark adaptation and showed no safety concerns at 4 months. 24% withdrew from the intervention before 1 month because of light intolerance and sleep disturbance. In another abstract, a light-emitting sleep mask to prevent dark adaptation in six patients with refractory clinically significant diabetic macular oedema showed good acceptability and tolerance for five patients at 6 months. As far as we know, no randomised controlled trials have been done to assess the role of light masks during sleep as a novel treatment for patients with non-central diabetic macular oedema.

th refractory clinically significant diabetic macular oedema showed good acceptability and tolerance for five patients at 6 months. As far as we know, no randomised controlled trials have been done to assess the role of light masks during sleep as a novel treatment for patients with non-central diabetic macular oedema. A study assessing the safety and acceptability of an organic light-emitting sleep mask (Noctura 400 Sleep Mask, PolyPhotonix Medical, Sedgefield, UK) in healthy volunteers (n=45) and patients with diabetic macular oedema (n=15) reported no clinically relevant safety issues at 4 months. 16 withdrew from the study, eight before month 1. The mean change in maximum retinal thickness in eyes with diabetic macular oedema at 4 months was −12·00 μm (range −28·80 to 4·80). A recent publication also showed good compliance with the light masks in diabetic macular oedema at 6 months. Added value of this study

A study assessing the safety and acceptability of an organic light-emitting sleep mask (Noctura 400 Sleep Mask, PolyPhotonix Medical, Sedgefield, UK) in healthy volunteers (n=45) and patients with diabetic macular oedema (n=15) reported no clinically relevant safety issues at 4 months. 16 withdrew from the study, eight before month 1. The mean change in maximum retinal thickness in eyes with diabetic macular oedema at 4 months was −12·00 μm (range −28·80 to 4·80). A recent publication also showed good compliance with the light masks in diabetic macular oedema at 6 months. Added value of this study The CLEOPATRA trial is, to our knowledge, the first randomised controlled trial to evaluate the effect of an organic light-emitting sleep mask as a treatment option for non-centre-involving diabetic macular oedema. The 24-month follow-up period provides data for efficacy, safety, and compliance of wearing these light masks for this condition. The results show that the light masks as used in this study did not provide any discernible clinical benefit. No differences were recorded between wearing and not wearing these light masks in the change in thickness in the zone of maximum retinal thickness, total macular volume, progression of the oedema to the centre, proportion of patients requiring standard treatment for diabetic macular oedema, and progression of diabetic retinopathy. The analysis of compliance highlighted that wearing these light masks over 24 months might also not be a sustainable option, as compliance decreased over time. The results of the study were not accounted for by non-compliance of wearing the light masks. No light mask-related serious adverse events were recorded.

pathy. The analysis of compliance highlighted that wearing these light masks over 24 months might also not be a sustainable option, as compliance decreased over time. The results of the study were not accounted for by non-compliance of wearing the light masks. No light mask-related serious adverse events were recorded. Implications of all the available evidence The CLEOPATRA study provides evidence that the light mask offered to prevent dark adaptation is not recommended as a treatment option for non-centre-involving diabetic macular oedema. Although earlier studies showed short-term improvement in diabetic oedema and diabetic retinopathy using 505 nm light masks, our study shows that compliance wearing the light masks during sleep is challenging and is therefore not a sustainable option. Since laboratory-based evidence of the role of photoreceptors in diabetic retinopathy is increasing, there remains an unmet need to translate this idea into interventions in patients.

ht masks, our study shows that compliance wearing the light masks during sleep is challenging and is therefore not a sustainable option. Since laboratory-based evidence of the role of photoreceptors in diabetic retinopathy is increasing, there remains an unmet need to translate this idea into interventions in patients. During dark adaptation, normal rod photoreceptors in the retina consume nearly all the oxygen available to the eye.6 In patients with diabetes the retinal oxygen supply is compromised and the hypoxic status during dark periods might exacerbate microvascular changes.7 This idea has been substantiated by the fact that oxygen supplementation alleviates diabetic macular oedema in the short term.8 Sivaprasad and Arden7 postulated that if dark adaptation could be prevented, diabetic macular oedema and diabetic retinopathy might be alleviated by decreasing the oxygen demand. Since dark adaptation in man only happens at night during sleep, sleeping in an environment illuminated with 500–505 nm light should suppress rods and prevent or reverse diabetic macular oedema. A proof-of-concept study in 12 patients who slept at night using a mask containing a chemiluminescent source that exposed one eye only to light for 3 months showed that the treatment had no safety issues, was acceptable to patients, and both colour vision and microaneurysm count improved.9 A second study used light-emitting diodes (LEDs) to illuminate one eye with 505 nm light during sleep in 40 patients with bilateral diabetic macular oedema.10 34 patients completed the study and an improvement in retinal function and a decrease in retinal thickness at 6 months was noted. Based on these observations, the Noctura 400 Sleep Mask (PolyPhotonix Medical, Sedgefield, UK) was CE-approved for the treatment of diabetic retinopathy. The long-term effectiveness, compliance, and safety of light masks are unknown. We did a phase 3 clinical trial (CLEOPATRA) to investigate whether offering the light mask to wear over closed eyelids during sleep at night for 24 months could treat and prevent the progression of non-centre-involving diabetic macular oedema.

long-term effectiveness, compliance, and safety of light masks are unknown. We did a phase 3 clinical trial (CLEOPATRA) to investigate whether offering the light mask to wear over closed eyelids during sleep at night for 24 months could treat and prevent the progression of non-centre-involving diabetic macular oedema. Methods Study design and patients The CLEOPATRA study is a phase 3, multicentre, single-blind, parallel-group, randomised controlled trial. Patients were recruited from 15 ophthalmic centres at UK National Health Service (NHS) hospitals. We included adults (aged ≥18 years) with type 1 or 2 diabetes mellitus and clinical and OCT evidence of retinal thickening in at least one non-central ETDRS zone due to diabetic macular oedema with best-corrected visual acuity of more than 55 ETDRS letters, equivalent to 6/18 Snellen. We permitted previous macular laser therapy, intravitreal steroids, or anti-VEGF agents provided the last treatment was given at least 4 months before randomisation.11 Exclusion criteria for eyes were centre-involving diabetic macular oedema, other causes of macular oedema, or coexistent ocular disease that affected or might affect visual acuity or prevent treatment delivery. We also excluded eyes with active proliferative diabetic retinopathy or that were treated previously with panretinal photocoagulation. Systemic exclusion criteria included history of insomnia or any other sleep disturbances.

stent ocular disease that affected or might affect visual acuity or prevent treatment delivery. We also excluded eyes with active proliferative diabetic retinopathy or that were treated previously with panretinal photocoagulation. Systemic exclusion criteria included history of insomnia or any other sleep disturbances. The study was granted approval by the National Research Ethics Committee Service London—Dulwich (13/LO/0145). Trial Steering and Data Monitoring Committees provided independent oversight. A representative of the manufacturer was a non-voting member of the Trial Steering Committee. All eligible patients gave written informed consent before study participation.

ional Research Ethics Committee Service London—Dulwich (13/LO/0145). Trial Steering and Data Monitoring Committees provided independent oversight. A representative of the manufacturer was a non-voting member of the Trial Steering Committee. All eligible patients gave written informed consent before study participation. Randomisation and masking We randomly allocated eligible patients (1:1) to wear during sleep either a light mask or a sham (non-light) mask, using the method of minimisation, concealed before allocation, stratified by HbA1c (<8% [63·89 mmol/mol] or ≥8% [63·90 mmol/mol]), perifoveal (ETDRS zones 6–9) versus parafoveal (ETDRS zones 2–5) baseline thickness in excess of 320 μm in the perifoveal or parafoveal zones, and study site. For patients with the same baseline thickness in excess of 320 μm in the perifoveal or parafoveal zones, the parafoveal zone was chosen. Randomisation was done by collaborating site investigators via the King's Clinical Trials Unit web-based randomisation service. Patients and examining clinicians were aware of the study allocation because of the nature of the intervention. Patients assigned the sham mask had the option of not using it because it became apparent early in the study that many patients were not using it. Outcome assessors including OCT technicians, optometrists, and graders at the independent reading centre based at the Gloucestershire Eye Unit (Gloucester, UK) were unaware of treatment allocation.

am mask had the option of not using it because it became apparent early in the study that many patients were not using it. Outcome assessors including OCT technicians, optometrists, and graders at the independent reading centre based at the Gloucestershire Eye Unit (Gloucester, UK) were unaware of treatment allocation. Procedures The light mask used in the intervention arm was the Noctura 400 Sleep Mask (PolyPhotonix Medical). This CE-certified class 2a device is designed to deliver blue-green light through closed eyelids. The light mask consists of two battery-operated organic LEDs inserted within a fabric mask and placed over the patients' eyes using an adjustable velcro strap. It is operational for a maximum of 8 h therapy per night. The lifetime of the light mask is 84 days, after which time a replacement mask is required. Based on calculations done by the manufacturer, the light mask provides a luminance of 75 photopic cd/m2 (± 10%), equating to 186 scotopic cd/m2. After considering light attenuation through closed eyelids and pupillary diameter, these light masks are expected to cause 40% reduction in rod-circulating current. The decay of mask output over its lifetime is also maintained within 10% of the desired output. The light intensity we used is approximately six orders of magnitude less than for threshold toxicity and two orders below that which causes a 1% change in the melatonin cycle that drives circadian rhythms.

culating current. The decay of mask output over its lifetime is also maintained within 10% of the desired output. The light intensity we used is approximately six orders of magnitude less than for threshold toxicity and two orders below that which causes a 1% change in the melatonin cycle that drives circadian rhythms. The light mask records automatically when it is being worn, providing an accurate measure of compliance. These data were downloaded by study sites when masks were returned. The manufacturer was also sent anonymised data from every returned light mask to measure compliance. We took the pragmatic decision that 6 h/day (4380 h over 2 years) was sufficient to represent 100% compliance and, therefore, represented the level at and above which maximum benefit would be derived. We defined compliance as patients who wore the masks 70% of the time (3066 h, counting time truncated to 6 h/day). If compliance data were missing for a day then we assumed no compliance (ie, the mask had not been worn that night).

d, therefore, represented the level at and above which maximum benefit would be derived. We defined compliance as patients who wore the masks 70% of the time (3066 h, counting time truncated to 6 h/day). If compliance data were missing for a day then we assumed no compliance (ie, the mask had not been worn that night). The trial manager contacted study sites to request they take steps to maximise the rate of mask return, to ensure availability of compliance data. The trial manager followed the trial monitoring plan by undertaking off-site monitoring on a monthly basis. This process included contacting sites at which patients' compliance was less than 40%. The manufacturer also alerted the trial manager when patients had poor compliance. Moreover, during every on-site monitoring visit, sites were asked to reinforce with patients the importance of wearing the masks every night during sleep, and specific patients with issues of compliance were discussed. The Data Monitoring Committee, in closed meetings, reviewed the accumulating compliance data and concluded that a dose-effect of the light masks should also be evaluated by comparing the effect of the light masks at three levels of compliance (50%, 60%, and 70%). The protocol was amended to this effect and approved by the Trial Steering Committee, the sponsor, and the Research Ethics Committee.

compliance data and concluded that a dose-effect of the light masks should also be evaluated by comparing the effect of the light masks at three levels of compliance (50%, 60%, and 70%). The protocol was amended to this effect and approved by the Trial Steering Committee, the sponsor, and the Research Ethics Committee. The clinical assessments schedule is detailed in the appendix (p 1) and in the published protocol.11 We recorded HbA1c at baseline, 12 months, and 24 months. Patients had OCT assessments every 4 months, and these assessments were done twice at 12 months and 24 months to ensure that the treatment effect was distinguished from the test-retest variability. We recorded concomitant diabetic medications, anti-VEGF agents, steroids, and laser treatment throughout the study. We measured refracted best-corrected visual acuity at baseline, 12 months, and 24 months using validated ETDRS visual acuity charts, and we repeated these measurements at baseline to assess test-retest variability. We did three-field colour fundus photography at baseline, 12 months, and 24 months to grade the severity of diabetic retinopathy. Both the examining clinician and graders at the independent reading centre graded anatomical characteristics of the diabetic macular oedema and severity of diabetic retinopathy. We defined an improvement in severity score for diabetic retinopathy as the proportion of patients with an ETDRS severity level of 2 or higher, at 12 months and 24 months (appendix pp 2, 3). We measured sleep disturbances at 12 months and 24 months. We used the Epworth Sleepiness Scale (ESS) to assess changes in daytime sleepiness, with scores ranging from 0 (low level of daytime sleepiness) to 24 (high level of daytime sleepiness),12 and the Pittsburgh Insomnia Rating Scale—20 item version (PIRS-20) to assess changes in insomnia, with scores ranging from 0 (no insomnia) to 60 (worse insomnia).13

ale (ESS) to assess changes in daytime sleepiness, with scores ranging from 0 (low level of daytime sleepiness) to 24 (high level of daytime sleepiness),12 and the Pittsburgh Insomnia Rating Scale—20 item version (PIRS-20) to assess changes in insomnia, with scores ranging from 0 (no insomnia) to 60 (worse insomnia).13 We recorded adverse events at every visit. We analysed differences from baseline to 24 months in ocular and systemic safety profiles with the light mask relative to the sham mask. Two clinicians who were unaware of treatment allocation coded ocular and systemic adverse events. A subset of patients (n=30) also underwent oximetry, multifocal electroretinography, and microperimetry before and after 100% oxygen at baseline and 12 months. This mechanistic component of the study will be reported later. Outcomes The primary outcome was the change from baseline to 24 months in maximum retinal thickness in the study eye with the light mask relative to the sham mask, measured by OCT. For participants with the same maximum baseline retinal thickness in two zones, the zone located in the parafoveal zone was chosen. When these two zones were in the parafoveal zone, the average retinal thickness was taken in subsequent follow-up measurements. For 12-month and 24-month measurements, OCT was done twice and the average of the measurements was taken.

ne retinal thickness in two zones, the zone located in the parafoveal zone was chosen. When these two zones were in the parafoveal zone, the average retinal thickness was taken in subsequent follow-up measurements. For 12-month and 24-month measurements, OCT was done twice and the average of the measurements was taken. A per-protocol secondary analysis excluded data from the point at which any patient was treated for worsening diabetic macular oedema. Additional secondary outcomes assessed at 12 months and 24 months included changes in thickness in the central subfield zone, zones 1–5, and zones 1–9, total macular volume, and morphological characteristics of diabetic macular oedema. We also assessed the change in refracted best-corrected visual acuity from baseline at 12 months and 24 months. Disease progression outcomes included time to occurrence of centre-involving diabetic macular oedema (defined as >300 μm), the proportion of patients progressing to centre-involving diabetic macular oedema of 400 μm or greater (ie, they met eligibility criteria for treatment with anti-VEGF agents in England and Wales), and the number of patients who received standard treatment for diabetic macular oedema at 12 months and 24 months (including anti-VEGF agents, steroids, and macular laser therapy).

diabetic macular oedema of 400 μm or greater (ie, they met eligibility criteria for treatment with anti-VEGF agents in England and Wales), and the number of patients who received standard treatment for diabetic macular oedema at 12 months and 24 months (including anti-VEGF agents, steroids, and macular laser therapy). Statistical analysis The pilot for this intervention8 provided an SD for the change from baseline in retinal thickness of 35·68 μm and informed 20% attrition. The detectable effect size of 15 μm was plausible relative to the 95% CI, and was minimally distinguishable from the 10·2 μm test–retest variation, for which the test–retest mean change over time of 0·9 μm was adequately small. A sample size of 300 patients (150 per treatment group)—with 240 patients analysed—provided 90% power based on a two-sided, unpaired t test at the 5% level of significance. Standardised effect sizes of 0·42 between treatment groups were detectable for secondary outcomes—eg, change in visual acuity. The planned statistical analysis incorporated serial measures and baseline adjustment, ensuring an improvement in power and in the precision of estimated treatment effects on each outcome.

nce. Standardised effect sizes of 0·42 between treatment groups were detectable for secondary outcomes—eg, change in visual acuity. The planned statistical analysis incorporated serial measures and baseline adjustment, ensuring an improvement in power and in the precision of estimated treatment effects on each outcome. We finalised the statistical analysis plan before data lock and agreed it with the oversight committees (Trial Steering and Data Monitoring Committees). We analysed the primary outcome with a linear mixed-effects model, incorporating six 4-monthly post-baseline observations of the outcome over time to 24 months and accommodating the within-participant correlation over time with an unstructured covariance matrix.14 The model included fixed factors for treatment group, HbA1c, and study site, and the continuous baseline of the outcome, each interacting with time. We did a sensitivity assessment of the missing-at-random assumption made in the primary outcome analysis in all patients, with three recommended scenarios affecting either one or both treatment groups, making this an intention-to-treat strategy,15, 16 with the intention-to-treat population comprising all randomised patients with OCT data. We did a per-protocol secondary analysis in which we excluded data from randomised patients at the point at which they were treated with steroids, anti-VEGF agents, or laser therapy, because these treatments could substantially improve retinal thickness after deterioration. We analysed secondary continuous outcomes with the same model specification as for the primary outcome, and with a missing baseline indicator if needed,17 and we reported data as adjusted differences in means. All tests were two-sided at the 5% significance level and effect sizes were interpreted cautiously with 95% CIs. We used the t test to compare means, the χ2 test or Fisher's exact test for single proportions, McNemar's SE for changes in proportions, and the Kaplan-Meier test for cumulative proportions.18

nces in means. All tests were two-sided at the 5% significance level and effect sizes were interpreted cautiously with 95% CIs. We used the t test to compare means, the χ2 test or Fisher's exact test for single proportions, McNemar's SE for changes in proportions, and the Kaplan-Meier test for cumulative proportions.18 We used complier average causal effect (CACE) analysis to estimate efficacy in patients who complied with treatment. We defined compliance, in turn, as wearing the assigned mask 70%, 60%, and 50% of the time, assuming the missing-at-random assumption in the primary outcome model and no effect of randomisation on outcome in non-compliers.19 We did sensitivity analyses of patients who met the requirement for treatment of centre-involving diabetic macular oedema because the central subfield thickness reached 400 μm before the 24-month endpoint, on the time to reaching 400 μm, and on the potential differential variability in retinal thickness between treatment groups (with the Mann-Whitney test). Because Spectralis (Heidelberg Engineering, Heidelberg, Germany) is the only OCT device with automatic real-time tracking, we did a sensitivity analysis including only patients who had OCT outcomes captured with this device at baseline. We also did a sensitivity analysis to exclude outliers defined as 4 SD from expected, which we do not present here because no changes were recorded in primary or secondary outcome analysis conclusions. We used IBM SPSS Statistics version 23 for statistical analyses. This trial is registered with Controlled-Trials.com, number ISRCTN85596558.

We did sensitivity analyses of patients who met the requirement for treatment of centre-involving diabetic macular oedema because the central subfield thickness reached 400 μm before the 24-month endpoint, on the time to reaching 400 μm, and on the potential differential variability in retinal thickness between treatment groups (with the Mann-Whitney test). Because Spectralis (Heidelberg Engineering, Heidelberg, Germany) is the only OCT device with automatic real-time tracking, we did a sensitivity analysis including only patients who had OCT outcomes captured with this device at baseline. We also did a sensitivity analysis to exclude outliers defined as 4 SD from expected, which we do not present here because no changes were recorded in primary or secondary outcome analysis conclusions. We used IBM SPSS Statistics version 23 for statistical analyses. This trial is registered with Controlled-Trials.com, number ISRCTN85596558. Role of the funding source The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The manufacturer of the light mask provided input into protocol development and trained site staff to offer the light mask as per protocol and their instructions for use manual. The manufacturers were sent anonymised data from every returned light mask for measurement of compliance. They provided feedback to the study team on masks that showed low compliance so clinical site staff could be informed and asked to reinforce use. The statisticians (ATP and JCV) had full access to all data in the study and the chief investigator (SS) had final responsibility for the decision to submit the results for publication.

ed feedback to the study team on masks that showed low compliance so clinical site staff could be informed and asked to reinforce use. The statisticians (ATP and JCV) had full access to all data in the study and the chief investigator (SS) had final responsibility for the decision to submit the results for publication. Results Between April 10, 2014, and June 15, 2015, 349 patients were assessed for eligibility. 41 patients did not meet eligibility criteria and were excluded; thus, 308 participants were randomly assigned to receive either the light mask (n=155) or the sham mask (n=153; figure 2).Figure 2 Trial profile OCT=optical coherence tomography. *Includes four patients lost to follow-up who had clinical OCT data. †Includes five patients lost to follow-up who had clinical OCT data.

In conclusion, the light mask as offered in this study is not an effective intervention to prevent or treat patients with non-centre-involving diabetic macular oedema. Future trials should aim to identify better ways of rod suppression to assess the role of rods in diabetic macular oedema and diabetic retinopathy. Supplementary Material Supplementary appendix Acknowledgments The CLEOPATRA Study Group thanks all patients who participated in the study and the many individuals not specifically mentioned who have supported the study. We thank Amy Riddell, Blair McLennan, Gill Lambert, Oliver Pressey, Janice Jimenez, Beverley White-Alao, and the team at King's College Clinical Trials Unit; Polyphotonix Medical from where we purchased the CE-marked compliance measure-incorporated light masks at a reduced price; and the lay panel in the North East Diabetes Research Network, the Comprehensive Local Research Networks, and the UK Clinical Research Network Ophthalmology Subspecialty group. This investigator-initiated study was funded by The Efficacy and Mechanism Evaluation Programme (11/30/02) and managed by the National Institute for Health Research (NIHR) on behalf of the Medical Research Council–NIHR partnership. The research was supported by the NIHR Biomedical Research Centre at Moorfields Eye Hospital NHS Foundation Trust and University College London Institute of Ophthalmology, the NIHR Moorfields Clinical Research Facility, and the UK Clinical Research Collaboration-registered King's Clinical Trials Unit at King's Health Partners, which is part-funded by the NIHR Biomedical Research Centre for Mental Health at South London and Maudsley, National Health Service (NHS) Foundation Trust, and King's College London. We thank team members of the NIHR Evaluation Trials and Studies Coordinating Centre. This Article presents independent research supported by the NIHR. The views expressed in this publication are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health.

Results Between April 10, 2014, and June 15, 2015, 349 patients were assessed for eligibility. 41 patients did not meet eligibility criteria and were excluded; thus, 308 participants were randomly assigned to receive either the light mask (n=155) or the sham mask (n=153; figure 2).Figure 2 Trial profile OCT=optical coherence tomography. *Includes four patients lost to follow-up who had clinical OCT data. †Includes five patients lost to follow-up who had clinical OCT data. Baseline characteristics were well balanced between treatment groups (table 1). The mean age of patients was 57 years (SD 11). 194 (63%) of 308 participants were men. 286 (93%) had baseline maximum retinal thickness in parafoveal zones 2–5 whereas 22 (7%) had maximum retinal thickness in perifoveal zones 6–9. 154 (50%) patients had HbA1c less than 8% (63·89 mmol/mol) at baseline. The average of the two refracted visual acuity measurements at baseline was mean 84·3 (SD 7·3) ETDRS letters, which was equivalent to 6/6 Snellen. 183 (59%) of 308 patients had OCT measurements taken with Spectralis (Heidelberg Engineering), 54 (18%) with Cirrus (Carl Zeiss Meditec, Cambridge, UK), 61 (20%) with Topcon 2000 (Topcon, Tokyo, Japan), and ten (3%) with RS3000 (Nidek, Aichi, Japan). 62 patients did not have primary outcome data (figure 2); no differences in baseline characteristics were noted between patients who dropped out and those who did not, except for study site, which was already adjusted for in the analysis (appendix p 4). This finding was attributable largely to one study site having a high dropout rate.Table 1 Baseline characteristics

gure 2); no differences in baseline characteristics were noted between patients who dropped out and those who did not, except for study site, which was already adjusted for in the analysis (appendix p 4). This finding was attributable largely to one study site having a high dropout rate.Table 1 Baseline characteristics Sham mask (n=153) Light mask (n=155) Age (years) 59·0 (51·0–67·0) 57·0 (51·0–65·0) Sex Men 92 (60%) 102 (66%) Women 61 (40%) 53 (34%) Ethnic origin White 94 (61%) 100 (65%) Black 29 (19%) 27 (17%) Asian 28 (18%) 24 (15%) Other 2 (1%) 4 (3%) Smoker 10 (7%) 13 (8%) Study site Bristol Eye Hospital 6 (4%) 6 (4%) Birmingham Heartlands Hospital 4 (3%) 7 (5%) Sandwell & West Birmingham Hospitals NHS Trust 3 (2%) 3 (2%) Frimley Park Hospital 16 (10%) 16 (10%) Hillingdon Hospital 27 (18%) 27 (17%) King's College Hospital 17 (11%) 18 (12%) Moorfields Eye Hospital 25 (16%) 26 (17%) Central Middlesex Hospital 3 (2%) 1 (1%) Maidstone & Tunbridge Wells Hospital 6 (4%) 6 (4%) Princess Alexandra Hospital, Harlow 6 (4%) 5 (3%) The Royal Wolverhampton NHS Trust 11 (7%) 10 (6%) Brighton & Sussex University Hospitals NHS Trust 1 (1%) 2 (1%) Sunderland Eye Infirmary 13 (8%) 12 (8%) Stoke Mandeville Hospital 5 (3%) 5 (3%) William Harvey Hospital Kent 10 (7%) 11 (7%) Blood pressure (mm Hg) Systolic 140·3 (18·9)* 137·2 (16·5)* Diastolic 81·0 (10·2)* 80·2 (9·4)* Diabetes mellitus Type 1 20 (13%) 31 (20%) Type 2 133 (87%) 124 (80%) Medication Insulin only 22 (14%) 43 (28%) Oral hypoglycaemic agents only 75 (49%) 72 (46%) Insulin and oral hypoglycaemic agents 56 (37%) 39 (25%) Diet-controlled 0 1 (1%) Best-corrected visual acuity (ETDRS letters) 86·0 (81·3–89·0) 85·5 (81·5–89·0) Maximum retinal thickness (μm) 348·8 (24·3) 345·9 (21·6) Total volume (mm3) 8·7 (8·3–9·3) 8·7 (8·3–9·2) HbA1c <8% (<63·89 mmol/mol) 77 (50%) 77 (50%) ≥8% (≥63·90 mmol/mol) 76 (50%) 78 (50%) Severity level (study eye)† 10 2 (1%) 2 (1%) 20 25 (17%) 35 (23%) 35 101 (69%) 93 (60%) 43–47 11 (7%) 15 (10%) 53 0 1 (1%) 61‡ 0 2 (1%) 65‡ 1 (1%) 1 (1%) 71–75‡ 0 1 (1%) 81–85‡ 0 0 90‡ 7 (5%) 4 (3%) Intraocular pressure (mm Hg) 16·0 (14·0–18·0)* 16·0 (14·0–18·0)§ Data are median (IQR), mean (SD), or number of participants (%). ETDRS=Early Treatment Diabetic Retinopathy Study.

5 (23%) 35 101 (69%) 93 (60%) 43–47 11 (7%) 15 (10%) 53 0 1 (1%) 61‡ 0 2 (1%) 65‡ 1 (1%) 1 (1%) 71–75‡ 0 1 (1%) 81–85‡ 0 0 90‡ 7 (5%) 4 (3%) Intraocular pressure (mm Hg) 16·0 (14·0–18·0)* 16·0 (14·0–18·0)§ Data are median (IQR), mean (SD), or number of participants (%). ETDRS=Early Treatment Diabetic Retinopathy Study. * Data missing for one participant. † Data from the independent reading centre; data missing for five participants assigned the sham mask and one allocated the light mask. ‡ Participants with these severity levels should have been excluded. § Data missing for two participants.

5 (23%) 35 101 (69%) 93 (60%) 43–47 11 (7%) 15 (10%) 53 0 1 (1%) 61‡ 0 2 (1%) 65‡ 1 (1%) 1 (1%) 71–75‡ 0 1 (1%) 81–85‡ 0 0 90‡ 7 (5%) 4 (3%) Intraocular pressure (mm Hg) 16·0 (14·0–18·0)* 16·0 (14·0–18·0)§ Data are median (IQR), mean (SD), or number of participants (%). ETDRS=Early Treatment Diabetic Retinopathy Study. * Data missing for one participant. † Data from the independent reading centre; data missing for five participants assigned the sham mask and one allocated the light mask. ‡ Participants with these severity levels should have been excluded. § Data missing for two participants. For the prespecified primary outcome analysis, OCT data were available for 246 (80%) of 308 patients at 24 months, of whom 127 were assigned the light mask and 119 were allocated the sham mask. This number includes five patients assigned the light mask and four patients allocated the sham mask, for whom OCT data were obtained from routine clinical care (ie, the patient attended their clinic appointment but did not attend an intervening research visit). An additional 17 patients assigned the light mask and 14 allocated the sham mask had OCT data from previous timepoints (appendix p 5). Therefore, 277 (90%) of 308 patients were included in the intention-to-treat linear mixed-effects model, of whom 144 had been assigned the light mask and 133 had been allocated the sham mask. At 24 months, in the 246 patients who had data available, no difference was recorded in mean change in maximum retinal thickness between the light mask and sham mask (adjusted difference −0·65 μm, 95% CI −6·90 to 5·59; p=0·84; table 2). Furthermore, no difference was noted in mean change in maximum retinal thickness between treatment groups at any timepoint (appendix p 6).Table 2 Maximum retinal thickness measured by OCT at baseline, 12 months, and 24 months

ight mask and sham mask (adjusted difference −0·65 μm, 95% CI −6·90 to 5·59; p=0·84; table 2). Furthermore, no difference was noted in mean change in maximum retinal thickness between treatment groups at any timepoint (appendix p 6).Table 2 Maximum retinal thickness measured by OCT at baseline, 12 months, and 24 months Maximum retinal thickness (μm) Mean (SE) change from baseline (μm) Adjusted difference (95% CI)*(μm) p value Sham mask Light mask Sham mask Light mask Baseline† 348·8 (24·3), n=153 345·9 (21·6), n=155 .. .. .. .. 12 months 339·1 (35·9), n=121 341·3 (29·7), n=132 −9·5 (3·1) −4·6 (2·5) 1·73 (−5·31 to 8·77) 0·63 24 months 336·3 (29·7), n=119 336·0 (25·5), n=127 −12·9 (2·9) −9·2 (2·5) −0·65 (−6·90 to 5·59) 0·84 Data are mean (SD), number of participants, unless otherwise indicated. 277 patients were included in the linear-mixed effects model. OCT=optical coherence tomography. * Adjusted for HbA1c, study site, and baseline maximum retinal thickness. † Mean maximum baseline retinal thickness for 133 patients assigned the sham mask and included in the linear mixed-effects model was 348·6 μm (SD 24·2) and for 144 patients allocated the light mask it was 345·4 μm (21·2). Mean maximum baseline retinal thickness for 20 patients assigned the sham mask and not included in the model was 350·6 μm (SD 25·6) and for 11 patients allocated the light mask it was 352·9 μm (26·6).

linear mixed-effects model was 348·6 μm (SD 24·2) and for 144 patients allocated the light mask it was 345·4 μm (21·2). Mean maximum baseline retinal thickness for 20 patients assigned the sham mask and not included in the model was 350·6 μm (SD 25·6) and for 11 patients allocated the light mask it was 352·9 μm (26·6). For the per-protocol secondary analysis of change in maximum retinal thickness, in which patients were excluded at the point they began treatment for diabetic macular oedema (laser therapy, steroids, or anti-VEGF agents; appendix p 7), 266 patients were included in the linear mixed-effects model at 12 months and 24 months (the linear mixed-effects model takes into account data at all timepoints). 56 patients needed treatment for diabetic macular oedema, of whom 23 had been assigned the light mask and 33 had been allocated the sham mask. The difference in the cumulative proportion of patients requiring treatment between treatment groups was 8% (95% CI 0–16) at 12 months and 9% (1–18) at 24 months. The change in maximum retinal thickness did not differ between the light mask and sham mask at 12 months and 24 months (adjusted difference at 24 months 3·23 μm, 95% CI −2·11 to 8·58; p=0·23; appendix p 8).

tients requiring treatment between treatment groups was 8% (95% CI 0–16) at 12 months and 9% (1–18) at 24 months. The change in maximum retinal thickness did not differ between the light mask and sham mask at 12 months and 24 months (adjusted difference at 24 months 3·23 μm, 95% CI −2·11 to 8·58; p=0·23; appendix p 8). Median compliance with wearing the light mask was 39·5% (IQR 9·8–78·2) at 4 months, falling to 19·5% (1·9–51·6) at 24 months. When considering the three definitions of compliance (ie, 70%, 60%, and 50%, with 70% compliance meaning the light mask was used 70% of the available time in the study, up to 6 h/day counted daily), the proportions of patients achieving each of these three levels of compliance decreased over time (appendix p 9).

onths. When considering the three definitions of compliance (ie, 70%, 60%, and 50%, with 70% compliance meaning the light mask was used 70% of the available time in the study, up to 6 h/day counted daily), the proportions of patients achieving each of these three levels of compliance decreased over time (appendix p 9). Sensitivity analyses for missing data were done to represent three possible scenarios, to reflect whether departures from the missing-at-random assumption applied within patients assigned the light mask only, within those allocated the sham mask only, and within both treatment groups equally and in the same direction (appendix p 10). The change in maximum retinal thickness did not differ between use of the light mask and the sham mask for all three scenarios. Assuming patients with unobserved outcome data in one or both treatment groups would take values as much as a prespecified 20 μm either side of the adjusted observed effect, all 95% CIs included 0 and excluded −15, thus confirming that the absence of a clinically important light mask effect is robust to missing data (appendix p 11). In the sensitivity analysis for non-compliance, the CACE estimate for compliers defined by 70% compliance was −4·2 (95% CI −44·6 to 36·1), 60% compliance was −3·1 (−32·4 to 26·3), and 50% compliance was −2·5 (−26·7 to 21·7). Across these three definitions of compliers, the results were consistent in estimating a small non-significant intervention effect, which was not close to the detectable effect of 15 μm retinal thickness.

−4·2 (95% CI −44·6 to 36·1), 60% compliance was −3·1 (−32·4 to 26·3), and 50% compliance was −2·5 (−26·7 to 21·7). Across these three definitions of compliers, the results were consistent in estimating a small non-significant intervention effect, which was not close to the detectable effect of 15 μm retinal thickness. In the sensitivity analysis of patients who met the requirement for treatment of centre-involving diabetic macular oedema (ie, retinal thickness reached 400 μm before the 24-month endpoint), the retinal thickness measurement taken just after the participant first reached 400 μm was carried forward to be their final measurement. Nine patients achieved 400 μm in central macula, of whom three had been assigned the sham mask (all achieved this point at 24 months) and six had been allocated the light mask (two achieved this point at 24 months, two at 20 months, and two at 12 months). 161 patients were included in this linear mixed-effects model and the adjusted difference between treatment groups was −0·22 μm (95% CI −8·36 to 7·92; p=0·96). 259 patients were included in the Cox proportional hazards regression time-to-event analysis (ie, time to reaching 400 μm), which was stratified by HbA1c (hazard ratio 2·0, 95% CI 0·5–8·0; p=0·33). For the sensitivity analysis of potential differential variability between the light mask and sham mask over time in the zone of maximum baseline retinal thickness, which was done in 246 patients, the difference between treatment groups was not significant from baseline to 24 months (p=0·38). A sensitivity analysis of the primary outcome in patients who had OCT measurements taken with Spectralis (Heidelberg Engineering) at baseline showed no difference between treatment groups (appendix p 12).

e in 246 patients, the difference between treatment groups was not significant from baseline to 24 months (p=0·38). A sensitivity analysis of the primary outcome in patients who had OCT measurements taken with Spectralis (Heidelberg Engineering) at baseline showed no difference between treatment groups (appendix p 12). In the secondary analyses of retinal thickness and volume, no significant differences were noted between the light mask and sham mask in change from baseline in central subfield thickness, total thickness of central and parafoveal zones, total retinal thickness measured over all nine zones, and total macular volume at 12 months and 24 months (appendix pp 13, 14). Analysis of secondary morphological outcomes showed that significantly more patients assigned the light mask had resolution of diffuse diabetic macular oedema at 12 months (difference between groups in change from baseline, −13%, 95% CI −23 to −2; p=0·0246), but this effect was lost at 24 months (2%, −10 to 14; p=0·75). Foveal cysts were somewhat reduced at 12 months (−12%, 95% CI −25 to 0·1; p=0·052) and 24 months (−14%, −27 to 0·3; p=0·054) in patients assigned the light mask compared with those allocated the sham mask. Changes in visible cysts in the inner ETDRS zones did not differ between treatment groups but the proportion of patients with visible cysts in the outer ETDRS zones was reduced significantly more in patients assigned the light mask compared with those allocated the sham mask (appendix pp 15, 16).

located the sham mask. Changes in visible cysts in the inner ETDRS zones did not differ between treatment groups but the proportion of patients with visible cysts in the outer ETDRS zones was reduced significantly more in patients assigned the light mask compared with those allocated the sham mask (appendix pp 15, 16). The adjusted difference in best-corrected visual acuity between the light mask and sham mask was also not significant (at 12 months, −0·07 ETDRS letters, 95% CI −1·38 to 1·23; p=0·91; at 24 months, 0·13 ETDRS letters, −1·45 to 1·71; p=0·87; appendix pp 13, 14). The proportion of patients showing progression of retinopathy was low, and no difference was recorded between treatment groups at 12 months and 24 months (appendix p 17). With respect to sleep disturbances, ESS scores and PIRS-20 scores did not differ between treatment groups (appendix pp 13, 14). The success of concealing the treatment allocation from primary assessors (OCT technicians and optometrists) was assessed with a guess form. In line with chance, OCT technicians guessed the allocation correctly for 137 (55%) of 248 patients and optometrists for 129 (52%) of 246 patients. The response was based on random choice for 180 (73%) OCT technicians and 231 (94%) optometrists, and 68 (27%) and 15 (4%), respectively, made an educated guess based on a clinical response or adverse event.

s guessed the allocation correctly for 137 (55%) of 248 patients and optometrists for 129 (52%) of 246 patients. The response was based on random choice for 180 (73%) OCT technicians and 231 (94%) optometrists, and 68 (27%) and 15 (4%), respectively, made an educated guess based on a clinical response or adverse event. 58 serious adverse events were recorded, of which 32 were reported in patients assigned the light mask and 26 were noted in those allocated the sham mask; none were related to the active intervention (appendix p 18). 340 adverse events not related to the intervention were reported, of which 172 were noted in patients assigned the light mask and 168 were in those assigned the sham mask (appendix p 19). 72 adverse events were reported as related to the assigned treatment, which included 50 in patients allocated the light mask and 22 in those assigned the sham mask (table 3). The most frequent adverse events related to the assigned treatment were discomfort on the eyes (14 with the light mask vs seven with the sham mask), painful, sticky, or watery eyes (14 vs six), and sleep disturbance (seven vs one).Table 3 Adverse events related to intervention

in those assigned the sham mask (table 3). The most frequent adverse events related to the assigned treatment were discomfort on the eyes (14 with the light mask vs seven with the sham mask), painful, sticky, or watery eyes (14 vs six), and sleep disturbance (seven vs one).Table 3 Adverse events related to intervention Sham mask (n=153) Light mask (n=155) Eyes Corneal abrasion, corneal ulcer 3 0 Mask causing pressure on eyes, pain on eyes, uncomfortable masks 7 14 Sore eyebrows, sore eyelids 0 2 Subconjunctival haemorrhage 0 1 Vision deterioration, disturbance 1 2 Watery eyes, sore eyes, sticky eyes, painful eyes, conjunctivitis 6 14 Neurological Headache, severe persistent headache 0 2 Psychiatric Insomnia 0 1 Musculoskeletal Left-sided neck and skull pain 0 1 Dermatological Scratched face on two occasions getting mask off during sleep 1 0 Sore skin, small lump on side of right eye 0 1 Pod moving around in mask when turns in bed 0 1 Wart 0 1 Other Mask slipping off head 3 3 Sleep disturbance, bad dreams 1 7

tric Insomnia 0 1 Musculoskeletal Left-sided neck and skull pain 0 1 Dermatological Scratched face on two occasions getting mask off during sleep 1 0 Sore skin, small lump on side of right eye 0 1 Pod moving around in mask when turns in bed 0 1 Wart 0 1 Other Mask slipping off head 3 3 Sleep disturbance, bad dreams 1 7 Discussion The CLEOPATRA trial is the first phase 3 randomised controlled trial to evaluate a light mask as an intervention to treat and prevent non-central diabetic macular oedema in a multicentre setting. Our results show that the light mask as offered in this study is not an effective option in the treatment or prevention of progression of non-central diabetic macular oedema. Although objective assessment of the reduction of maximum retinal thickness was our primary outcome, we have made our conclusion based on the primary outcome, per-protocol secondary analysis, and five prespecified sensitivity analyses of the primary outcome, and none of these analyses showed any therapeutic benefit of wearing these light masks. Moreover, because of the dynamic nature of diabetic macular oedema, we considered several secondary outcomes, including reduction in total retinal thickness, macular volume, progression of central subfield thickness to 300 μm or more, and the proportions of patients requiring treatment for new onset centre-involving diabetic macular oedema and of those treated with standard therapy during the trial due to worsening of diabetic macular oedema. None of the changes in these variables was significant between treatment groups, substantiating the results of the primary outcome. Furthermore, no treatment effect was noted in severity of diabetic retinopathy with these light masks. However, the light masks did significantly reduce diffuse diabetic macular oedema and visible cysts in outer ETDRS zones at 12 months, but this effect did not translate to a significant change in retinal thickness and the effect was not sustained at 24 months, suggesting that any positive morphological effects of these light masks on diabetic macular oedema is transient and minimal.

c macular oedema and visible cysts in outer ETDRS zones at 12 months, but this effect did not translate to a significant change in retinal thickness and the effect was not sustained at 24 months, suggesting that any positive morphological effects of these light masks on diabetic macular oedema is transient and minimal. We expected compliance with light masks to be an issue based on findings of the phase 2 study10 and because non-centre-involving diabetic macular oedema is asymptomatic. Therefore, we made several efforts to tackle compliance-related issues in this study. First, we calculated the sample size with a 20% attrition rate, which is higher than most ophthalmic trials. Furthermore, we allowed for OCT measurements from clinic appointments to be used when patients attended the clinic visit and not a clinical trial visit appointment. We had also carefully considered the effect of non-compliance on the potential therapeutic effect of the light masks by incorporating a predefined CACE analysis for non-compliance at three levels—70%, 60%, and 50%. Non-compliance was noted as early as 4 months into the trial and across all three definitions of compliance.

had also carefully considered the effect of non-compliance on the potential therapeutic effect of the light masks by incorporating a predefined CACE analysis for non-compliance at three levels—70%, 60%, and 50%. Non-compliance was noted as early as 4 months into the trial and across all three definitions of compliance. Compliance with use of light masks has varied between studies and can be partly explained by the differences in definitions used in determining compliance levels.10, 20, 21, 22, 23 However, this study is the first randomised trial evaluating the use of a light mask during sleep at night over 24 months and shows that compliance reduces over time in keeping with the adherence patterns of self-management strategies in asymptomatic diabetes.24 Decline in adherence is rapid after the first 6 months of therapy in chronic diseases.25 The compliance levels observed in our study are in keeping with the WHO report25 that shows that the mean adherence to long-term therapy in patients with chronic diseases is approximately 50%. Therefore, further studies should include additional interventions to increase patient engagement in wearing the light masks to evaluate whether this intervention is sustainable over the lifetime of their diabetic eye disease. Both treatment groups showed a gradual mean reduction in the zone of maximum retinal thickness over 24 months. The reduction is within the SD that we used for the sample size calculation. The event rate of progression to centre-involving diabetic macular oedema was also similar to findings of previous reports.2, 5

Compliance with use of light masks has varied between studies and can be partly explained by the differences in definitions used in determining compliance levels.10, 20, 21, 22, 23 However, this study is the first randomised trial evaluating the use of a light mask during sleep at night over 24 months and shows that compliance reduces over time in keeping with the adherence patterns of self-management strategies in asymptomatic diabetes.24 Decline in adherence is rapid after the first 6 months of therapy in chronic diseases.25 The compliance levels observed in our study are in keeping with the WHO report25 that shows that the mean adherence to long-term therapy in patients with chronic diseases is approximately 50%. Therefore, further studies should include additional interventions to increase patient engagement in wearing the light masks to evaluate whether this intervention is sustainable over the lifetime of their diabetic eye disease. Both treatment groups showed a gradual mean reduction in the zone of maximum retinal thickness over 24 months. The reduction is within the SD that we used for the sample size calculation. The event rate of progression to centre-involving diabetic macular oedema was also similar to findings of previous reports.2, 5 The main strength of our study was that we ensured that the primary outcome was corroborated by a predefined sensitivity analysis and secondary analysis to reduce potential systemic biases in this dynamic condition. Other strengths included clear definition of objective endpoints, publication of the protocol, substantial public and patient involvement throughout the study, and strict assessment-assured high-quality data and preplanned analysis for expected non-compliance. The baseline characteristics of the trial population were typical for the intended patient population. The representative multiethnic patient population, together with the multicentre trial design, permit wide generalisability of our results.

ssured high-quality data and preplanned analysis for expected non-compliance. The baseline characteristics of the trial population were typical for the intended patient population. The representative multiethnic patient population, together with the multicentre trial design, permit wide generalisability of our results. Nevertheless, some limitations should be considered when interpreting the results. First, the study only shows that offering the light mask as per this study protocol to suppress rod function is not an effective option to treat non-central diabetic macular oedema. It is possible that the retinal illumination achieved with these devices did not reduce the dark current sufficiently to alter the hypoxic state. Therefore, it is worth evaluating other techniques of rod suppression in diabetic retinopathy and diabetic macular oedema since there is a growing body of scientific evidence that supports the role of photoreceptors in retinal vascular permeability and angiogenesis.26, 27 Other clinical trials of light masks to prevent dark adaptation in diabetic retinopathy are ongoing. The Lahey Light II trial20 (LCID Study Number 2015-020) is evaluating a modified 520 μm LED light mask to prevent dark adaptation in refractory diabetic macular oedema. Two other clinical trials (ISRCTN82148651 and NCT02207712) are ongoing for age-related macular degeneration. However, it is important that studies with long-term follow-up are conducted in these chronic conditions to provide further insight into this intervention. Second, it could be argued that hypoxia might not be a contributing factor in early diabetic macular oedema and that patients with early signs of this disorder might not be the ideal target population. However, oxygen therapy has been shown to ameliorate early diabetic macular oedema, reinforcing the role of hypoxia in this condition.9, 28 A third limitation is that we defined non-centre-involving diabetic macular oedema as a zone of retinal thickness above 320 μm. Although the normative data of some zones on Spectralis OCT could in fact be above 320 μm, we only included eyes with clinical evidence of diabetic macular oedema causing the retinal thickness to be greater than 320 μm. Therefore, we believe that our patient population is representative of early non-central diabetic macular oedema.

normative data of some zones on Spectralis OCT could in fact be above 320 μm, we only included eyes with clinical evidence of diabetic macular oedema causing the retinal thickness to be greater than 320 μm. Therefore, we believe that our patient population is representative of early non-central diabetic macular oedema. In our study, there was no discernible treatment effect in favour of the light mask at 4 months and 8 months when compliance was highest; in fact, any effect was in the opposite direction, which was maintained at 12 months, suggesting that low compliance did not contribute significantly to the overall study result. In conclusion, the light mask as offered in this study is not an effective intervention to prevent or treat patients with non-centre-involving diabetic macular oedema. Future trials should aim to identify better ways of rod suppression to assess the role of rods in diabetic macular oedema and diabetic retinopathy. Supplementary Material Supplementary appendix

and King's College London. We thank team members of the NIHR Evaluation Trials and Studies Coordinating Centre. This Article presents independent research supported by the NIHR. The views expressed in this publication are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health. Contributors SS was the grant holder and chief investigator, contributed to study design and implementation, and wrote the protocol and report. PH was an investigator and contributed to study design and implementation. CM and JK contributed to study design, design of the case report form, and implementation of the study. HH contributed to study implementation and management. ATP and JCV provided statistical input and contributed to study design and the statistical analysis plan. GBA contributed to the hypothesis, study design, and mechanistic evaluation. SG contributed to study implementation. All authors have read and approved the final report.

uted to study implementation and management. ATP and JCV provided statistical input and contributed to study design and the statistical analysis plan. GBA contributed to the hypothesis, study design, and mechanistic evaluation. SG contributed to study implementation. All authors have read and approved the final report. The CLEOPATRA Study Group Trial Coordinators—Frank Ahfat (Department of Ophthalmology, Maidstone Hospital, Maidstone & Tunbridge Wells NHS Trust, Kent, UK); Ajay Bhatnagar, Nirodhini Narendran (Wolverhampton Eye Infirmary, New Cross Hospital, Wolverhampton, UK); Randhir Chavan (Sandwell and West Birmingham NHS Trust, Birmingham, UK); Abosede Cole (Bristol Eye Hospital, Bristol, UK); Roxanne Crosby-Nwaobi, Namritha Patrao, Deepthy Menon, Chris Hogg, Lauren Leitch-Devlin, Catherine Egan, Nisha Shah, Tatiana Mansour, Tunde Peto (National Institute for Health Research Biomedical Research Centre at Moorfields Eye Hospital NHS Foundation Trust, London, UK); Gary Rubin (University College London, London, UK); Haralabos Eleftheriadis (Department of Ophthalmology, King's College Hospital NHS Foundation Trust, London, UK); Jonathan Gibson (Birmingham Heartlands Hospital, Birmingham, UK); Arevik Gulakhszian, Gilli Vafidis (Central Middlesex NHS Trust, London, UK); Edward Hughes (Sussex Eye Hospital, Brighton, UK); Afsar Jafree (Ophthalmology Department, East Kent University Hospital, Kent, UK); Geeta Menon (Ophthalmology Department, Frimley Park Hospital NHS Foundation Trust, Surrey, UK); Priya Prakash (Princess Alexandra Hospital, Harlow, UK); Maria Sandinha (Sunderland Eye Infirmary, Sunderland, UK); and Richard Smith (Buckinghamshire Healthcare NHS Trust, Aylesbury, UK). Independent Reading Centre—Peter Scanlon, Steve Chave, Steve Aldington, Angela Dale (Gloucestershire Eye Unit, Gloucester, UK). Trial Steering Committee—Gillian Hood (Queen Mary, University of London, London, UK); Graham A Hitman (Barts and The London School of Medicine and Dentistry, London, UK); David Crabb (City University, London, UK); Alaistair Denniston (Queen Elizabeth Hospital, Birmingham, UK); Douglas Lewin (lay member; Archbishop, Old Roman Catholic Church, UK); and Ian Grierson (non-voting member; University of Liverpool, Liverpool, UK; represented PolyPhotonix Medical, Sedgefield, UK).

, London, UK); David Crabb (City University, London, UK); Alaistair Denniston (Queen Elizabeth Hospital, Birmingham, UK); Douglas Lewin (lay member; Archbishop, Old Roman Catholic Church, UK); and Ian Grierson (non-voting member; University of Liverpool, Liverpool, UK; represented PolyPhotonix Medical, Sedgefield, UK). Data Monitoring Committee—Sarah Walker (chairman; Medical Research Council, Clinical Trials Unit, University College London, London, UK); Jackie Sturt (King's College London, London, UK); and Debendra Sahu (Southampton NHS Trust, Southampton, UK). Declaration of interests SS has received research grants, travel grants, and speaker fees from, and was an advisory board member for, Novartis, Bayer, Allergan, Roche, Boehringer Ingelheim, and Heidelberg Engineering, outside the submitted work. PH has received research grants, travel grants, and speaker fees from, and was an advisory board member for, Novartis, Bayer, and Allergan, outside the submitted work. All other authors declare no competing interests.

INTRODUCTION In 2014, an estimated 125 million women (5.0%) and 50 million men (2.3%) globally had a BMI≥35kg/m2,(1) making them potentially eligible for bariatric surgery. Bariatric surgery reduces the risk of premature death,(2–4) cardiovascular events,(5, 6) and micro-/macro-vascular diabetes complications.(7, 8) However, there is growing concern about adverse effects on mental health, with increased alcohol and substance abuse after some procedures, as well as signals of an increased suicide risk compared with morbidly obese individuals.(9, 10) Compared with the general population, bariatric surgery patients have been reported to have higher risk of both suicide(11, 12) and nonfatal self-harm.(13) Nonfatal self-harm events are also more common after than before surgery.(12–15)

gnals of an increased suicide risk compared with morbidly obese individuals.(9, 10) Compared with the general population, bariatric surgery patients have been reported to have higher risk of both suicide(11, 12) and nonfatal self-harm.(13) Nonfatal self-harm events are also more common after than before surgery.(12–15) As suicide is rare, it is unlikely that there will ever be a randomised trial of sufficient size and duration to assess suicide risk after bariatric surgery. Further, there are no observational studies on suicide comparing bariatric surgery patients with nonsurgically treated obese controls. The Utah Mortality Study(2) reported an increased risk of “deaths not caused by disease” in patients treated with bariatric surgery compared to age-sex-BMI-matched controls applying for a driver’s licence. The risk of suicide was not statistically significant, but the point estimate was more than twice as high in surgery patients compared to matched controls. In the Utah Obesity Study,(16) no difference in suicide risk over up to 6 years could be detected in the bariatric surgery group (4 suicides) versus a morbidly obese control group seeking but not receiving bariatric surgery (0 suicides). Neither study accounted for baseline psychiatric status (which is likely to be associated with both bariatric surgery exposure and the outcome suicide) between the surgery and control group, nor had they a nonsurgically treated obese control group. A recent Danish cohort study excluded patients with history of psychiatric contacts and reported no difference in suicide rates between bariatric surgery patients versus hospitalised patients with a diagnosis of obesity but without bariatric surgery.(15)

group, nor had they a nonsurgically treated obese control group. A recent Danish cohort study excluded patients with history of psychiatric contacts and reported no difference in suicide rates between bariatric surgery patients versus hospitalised patients with a diagnosis of obesity but without bariatric surgery.(15) We aimed to compare the risk of suicide and nonfatal self-harm in patients with obesity attempting to lose weight with versus without bariatric surgery, accounting for baseline psychiatric status in two Swedish matched cohort studies linked to outcome data from nationwide health registers. METHODS Study Design Matched cohort designs were used to analyse the association between bariatric surgery and the outcomes suicide and nonfatal self-harm. The cohorts used for the current analysis were the Swedish Obese Subjects (SOS) study(3) and a nationwide register linkage combining the Scandinavian Obesity Surgery Registry (SOReg)(17) with the Itrim Health Database, a register including individuals treated with intensive lifestyle modification.(6) The rationale for using two studies was that SOS and SOReg/Itrim have complementary strengths. SOS provides longer follow-up than any other existing controlled study, but used older surgical techniques. SOReg/Itrim included current surgical techniques and an intensively treated control group, but had shorter follow-up.

The rationale for using two studies was that SOS and SOReg/Itrim have complementary strengths. SOS provides longer follow-up than any other existing controlled study, but used older surgical techniques. SOReg/Itrim included current surgical techniques and an intensively treated control group, but had shorter follow-up. SOS and SOReg/Itrim participants were linked to nationwide health registers using the Swedish personal identity number which is unique for each resident. The linkage was performed by officials at the National Board of Health and Welfare and at Statistics Sweden in 2015 and 2016. Setting The Swedish health care system is tax funded and offers universal access, including physicians, psychologists, dietitians and other healthcare specialists. The adult prevalence of BMI≥35kg/m2 in Sweden in 2014 has been estimated to 5–6%.(1) In a global perspective, Sweden had one of the highest percentages of bariatric procedures for the total population in 2013 (0.08% as compared with 0.04% in the US and Canada).(18) In individuals undergoing bariatric surgery, the prevalence of depression, self-harm, and substance abuse at baseline is about twice as high as in the general population in Sweden.(13) The suicide rate in Sweden is similar to the OECD average and that in the United States (12.3, 12.0 and 12.5 per 100,000.(19)

ada).(18) In individuals undergoing bariatric surgery, the prevalence of depression, self-harm, and substance abuse at baseline is about twice as high as in the general population in Sweden.(13) The suicide rate in Sweden is similar to the OECD average and that in the United States (12.3, 12.0 and 12.5 per 100,000.(19) The SOS Study This prospective, nonrandomised, controlled intervention study recruited patients from September 1, 1987, to January 31, 2001(3) via recruitment campaigns in the mass media and at 480 primary healthcare centers. Patients choosing surgery constituted the surgery group. From individuals not choosing surgery, a contemporaneously matched control group was created using 18 matching variables: sex, age, weight, height, waist circumference, hip circumference, systolic blood pressure, serum cholesterol and triglyceride levels, smoking status, diabetes, menopausal status, 4 psychosocial variables with documented associations with death, and 2 personality traits related to treatment preference (data on psychosocial variables and personality traits are provided in eTable1). Matching was not performed at an individual level but an algorithm selected controls so that the current mean values of the matching variables in the control group became as similar as possible to those in the surgery group using the method of sequential treatment assignment.

ersonality traits are provided in eTable1). Matching was not performed at an individual level but an algorithm selected controls so that the current mean values of the matching variables in the control group became as similar as possible to those in the surgery group using the method of sequential treatment assignment. Inclusion/exclusion criteria Study groups had identical inclusion (age 37–60y and BMI≥34kg/m2 in men and ≥38 kg/m2 in women) and exclusion criteria (earlier surgery for gastric or duodenal ulcer, earlier bariatric surgery, gastric ulcer or myocardial infarction during the past 6 months, ongoing or active malignancy during the past 5 years, bulimic eating pattern, drug or alcohol abuse, psychiatric or cooperative problems contraindicating bariatric surgery, and other contraindicating conditions such as chronic glucocorticoid or anti-inflammatory treatment). Interventions The choice of procedure was made by the operating surgeon (265 [13%] gastric bypass, 376 [19%] gastric banding, 1369 [68%] vertical-banded gastroplasty). Open surgery was used in 89% of the patients. Laparoscopic surgery was gradually introduced from 1993 and during the last 2 recruitment years the majority of procedures were performed using this technique. Control patients received the customary nonsurgical obesity treatment at their registration center. No attempt was made to standardise the nonsurgical treatment, which ranged from sophisticated lifestyle intervention to no treatment.

st 2 recruitment years the majority of procedures were performed using this technique. Control patients received the customary nonsurgical obesity treatment at their registration center. No attempt was made to standardise the nonsurgical treatment, which ranged from sophisticated lifestyle intervention to no treatment. The SOReg/Itrim Study SOReg is a nationwide, prospective register for bariatric surgery started in 2007. It has been estimated to cover 98.5% of all bariatric procedures in Sweden.(17) Data are stored electronically and recorded as part of clinical practice. For this study, data were used from intervention years 2007 to 2012. The Itrim Health Database prospectively collects data on individuals who enroll in the commercial weight loss program at 38 Itrim centers across Sweden. Itrim centers use a common IT platform for quarterly follow-up of, for example measured weight, waist circumference, and blood pressure. For this study, data were available from individuals starting the program from January 1, 2006, to December 31, 2013. Inclusion/exclusion criteria In the current report, individuals ≥18y with BMI 30–49.9kg/m2 and baseline weight recorded were included from SOReg and Itrim. There were no mandatory national eligibility criteria for bariatric surgery during the study period, but most county councils recommended BMI≥35 with or BMI≥40kg/m2 without obesity-related comorbidity. In the sample used for this study, 888 surgery patients (4.0%) had a BMI<35kg/m2 (median BMI: 34.1kg/m2).

g and Itrim. There were no mandatory national eligibility criteria for bariatric surgery during the study period, but most county councils recommended BMI≥35 with or BMI≥40kg/m2 without obesity-related comorbidity. In the sample used for this study, 888 surgery patients (4.0%) had a BMI<35kg/m2 (median BMI: 34.1kg/m2). Interventions Surgery participants underwent primary gastric bypass (96.0% of procedures conducted laparoscopically; open surgery was primarily used when a patient had had a previous open abdominal surgery or when complications arose during an initially laparoscopic procedure). Intensive lifestyle participants received the Itrim program including a 3-month weight loss phase with either low or very low calorie diets (eMethods) based on baseline BMI, personal preference, and contraindication status. After the weight loss phase, patients entered a 9-month weight maintenance program including exercise (circuit training at the center 2–3 times/week for 30–45 minutes, and pedometer use to encourage walking), and dietary advice. Behavioral changes were facilitated by a structured program, including twenty 1h group sessions. There were also face-to-face counseling sessions throughout the program.

gram including exercise (circuit training at the center 2–3 times/week for 30–45 minutes, and pedometer use to encourage walking), and dietary advice. Behavioral changes were facilitated by a structured program, including twenty 1h group sessions. There were also face-to-face counseling sessions throughout the program. Covariates in SOS and SOReg/Itrim Demographic data were available on age, sex, and educational level. For SOReg/Itrim, data were retrieved from Statistics Sweden on marital status, disposable income, disability pension (also available for SOS), and unemployment. Measured BMI was available from baseline examinations. Data on healthcare visits for self-harm, substance abuse, and other psychiatric causes, as well as for cardiovascular disease, were retrieved from the National Patient Register (inpatient data from 1969; hospital-based outpatient data from January 1, 2001). Data on psychiatric and anti-diabetic drug use before inclusion were retrieved via self-report in SOS and from the Prescribed Drug Register in SOReg/Itrim (register start date: July 1, 2005). Self-reported drug use in SOS has previously been shown to be reasonably consistent with data from the Prescribed Drug Register.(20) The International Classification of Diseases (ICD) and Anatomical Therapeutical Chemical classification system codes used are provided in eTable2. As missing data on BMI (0.02% [12/61,495]) and education (0.4% [254/61,495]) were rare in SOReg/Itrim, and data were complete for the other variables, patients with missing data were excluded.

lassification of Diseases (ICD) and Anatomical Therapeutical Chemical classification system codes used are provided in eTable2. As missing data on BMI (0.02% [12/61,495]) and education (0.4% [254/61,495]) were rare in SOReg/Itrim, and data were complete for the other variables, patients with missing data were excluded. Outcome and Follow-Up in SOS and SOReg/Itrim The primary outcome in SOS was all cause mortality, for which the study was powered.(3) The outcomes of the current analysis were death by suicide, and death by suicide or nonfatal self-harm, retrieved from the Causes of Death Register and the National Patient Register until December 31, 2013, for SOS and December 31, 2014, for SOReg/Itrim. In the main analysis, we used ICD codes to identify suicide and nonfatal self-harm (ICD9: E950-959, E980-989; ICD10 X60-84, Y10-34, Y870), including both confirmed suicides and deaths from undetermined intent. Participants were followed from the treatment start date until first event, death, emigration, or end of register-based follow-up, whichever came first. SOS controls and Itrim participants who crossed over to bariatric surgery were censored at the cross-over date (SOS n=289; Itrim=335), as were SOS surgery patients who had their procedure reversed to normal anatomy (n=100).

until first event, death, emigration, or end of register-based follow-up, whichever came first. SOS controls and Itrim participants who crossed over to bariatric surgery were censored at the cross-over date (SOS n=289; Itrim=335), as were SOS surgery patients who had their procedure reversed to normal anatomy (n=100). During follow-up, two SOS surgery patients requested to be deleted from the database, and one obtained an unlisted identity number making linkage impossible. In SOS, both groups had identical follow-up with physical examinations and questionnaires at baseline and 0.5, 1, 2, 3, 4, 6, 8, 10, 15 and 20 years. In addition to the follow-up for the research study, SOS patients also had routine follow-up in the public healthcare system (eMethods). Statistical Analysis Outcomes were analysed using survival analysis. Hazard ratios were estimated using Cox regression. The proportional hazard assumption was evaluated by interacting time and treatment. This term was statistically significant for suicide in SOReg/Itrim (P=0.0497). Due to the small number of events, the model was not stratified by follow-up time.

sing survival analysis. Hazard ratios were estimated using Cox regression. The proportional hazard assumption was evaluated by interacting time and treatment. This term was statistically significant for suicide in SOReg/Itrim (P=0.0497). Due to the small number of events, the model was not stratified by follow-up time. In the sequentially matched SOS study, adjustment was made for age, sex, history of self-harm (yes/no), and continuous BMI. In the current SOReg/Itrim analysis, we used coarsened exact matching(21) to match participants by BMI (<35, 35 to <40, 40 to <45, 45 to <50kg/m2), age (18–29, 30–39, 40–49, 50–59, ≥60 years), sex, education level, diabetes, cardiovascular disease, history of self-harm, substance abuse, antidepressant use, anxiolytics use, and history of psychiatric care (yes/no). To minimise loss of information, we allowed matching strata to include different numbers of surgery and intensive lifestyle participants. To compensate for the differential strata sizes, analyses were weighted by the strata size. For example, if there were 2 surgery participants and 4 lifestyle participants in a stratum, then each surgery participant was given the weight 1 and each lifestyle participant the weight 0.5. Additional adjustment was performed for age, BMI, and income as continuous variables, and for marital status (married/unmarried), disability pension (yes/no), and unemployment benefits (yes/no).

articipants in a stratum, then each surgery participant was given the weight 1 and each lifestyle participant the weight 0.5. Additional adjustment was performed for age, BMI, and income as continuous variables, and for marital status (married/unmarried), disability pension (yes/no), and unemployment benefits (yes/no). Subgroup analyses were performed by procedure type (SOS only; analysis by intention-to-treat), psychiatric history, and education level. In SOS, the 10-year weight trajectory was examined in surgery patients with an event versus those without. Statistical analyses were performed using SAS (version 9.4) and Stata (version 14). The SOS study is registered with ClinicalTrials.gov NCT01479452. 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. MN, GB and LMSC had full access to the data in the study. The corresponding author had final responsibility for the decision to submit for publication. RESULTS After recruitment campaigns in the mass media and at primary healthcare centers, 6905 individuals completed an eligibility examination for the SOS study, 5335 were found eligible of which 2010 chose surgical treatment while the contemporaneously matched control group consisted of 2037 individuals not choosing surgery (eFigure1).

uitment campaigns in the mass media and at primary healthcare centers, 6905 individuals completed an eligibility examination for the SOS study, 5335 were found eligible of which 2010 chose surgical treatment while the contemporaneously matched control group consisted of 2037 individuals not choosing surgery (eFigure1). Out of 30,081 SOReg patients who had bariatric surgery during the study period, 26,388 had gastric bypass and were eligible for matching, while 18,365 out of 31,414 intensive lifestyle participants were eligible (eFigure2). After matching, there were 20,256 (77%) gastric bypass and 16,162 (88%) intensive lifestyle participants available for analysis. Baseline characteristics in the two cohorts are presented in Table 1. SOS patients in the surgery group had lower education, more history of hospitalisation for self-harm and cardiovascular disease, and were younger and had a higher BMI compared to controls. Mean body weight changes in the surgery and control group at 2, 10 and 15 years were −23%/0%, −17%/1% and −16%/−1%, respectively. In the SOReg/Itrim cohort, the prevalence of class I, II and III obesity was identical after matching but gastric bypass patients had a higher mean BMI than intensive lifestyle participants. Gastric bypass patients also had lower income, were more often married, on disability pension, unemployed, and using hypnotics or sedatives. The mean 1-year body weight change was −32% in the gastric bypass and −15% in the intensive lifestyle group.

patients had a higher mean BMI than intensive lifestyle participants. Gastric bypass patients also had lower income, were more often married, on disability pension, unemployed, and using hypnotics or sedatives. The mean 1-year body weight change was −32% in the gastric bypass and −15% in the intensive lifestyle group. During 68,528 person-years (median 18; interquartile range 14–21) there were 87 versus 49 suicides or nonfatal self-harm events in the SOS surgery and control group, respectively (adjusted hazard ratio [aHR] 1.78 [95%CI 1.23–2.57]; P=0.0021), of which 9 and 3 were suicides (3.06 [0.79–11.88]; P=0.107; Figure 1). Additional adjustment for baseline diabetes and cardiovascular disease resulted in similar estimates for suicide or nonfatal self-harm (aHR 1.74 [1.20–2.52]; P=0.0033) and for suicide (3.33 [0.86–12.97]; P=0.083). In analyses by primary procedure type, increased risk of suicide or nonfatal self-harm was found for gastric bypass (aHR 3.48 *1.65–7.31+; P=0.0010), gastric banding (2.43 *1.23–4.82+; P=0.011) and vertical-banded gastroplasty (2.25 *1.37–3.71+; P=0.0015) versus controls (Figure 2, Figure 3A). Surgery patients who died by suicide or had a nonfatal self-harm event had similar or lower body weight during follow-up than patients who did not, while there was no difference at baseline (Figure 4).

3–4.82+; P=0.011) and vertical-banded gastroplasty (2.25 *1.37–3.71+; P=0.0015) versus controls (Figure 2, Figure 3A). Surgery patients who died by suicide or had a nonfatal self-harm event had similar or lower body weight during follow-up than patients who did not, while there was no difference at baseline (Figure 4). Poisoning was the most common mode of suicide in SOS (78% [7/9] for surgery versus 100% [3/3] for controls; eTable3) and of nonfatal self-harm (70% [57/81] versus 53% [25/47]; eTable4). Out of 9 suicides in the surgery arm, 5 occurred in gastric bypass patients (2 who had primary gastric bypass, 2 who were converted from vertical-banded gastroplasty, 1 converted from gastric banding; eTable3). Substance abuse was recorded in 48% [39/81] of surgery patients and 28% [13/47] of controls with nonfatal self-harm events (P=0.023; eTable4). During 149,582 person-years (median 3.9; interquartile range 2.8–5.2) there were 341 suicides or nonfatal self-harm events in the SOReg gastric bypass group and 84 in the intensive lifestyle group (aHR 3.16 [2.46–4.06]; P<0.0001), of which 33 and 5 were suicides (5.17 [1.86–14.37]; P=0.0017; Figure 5). As in SOS, poisoning was the most common mode of suicide (79% [26/33] for surgery versus 80% [4/5] for intensive lifestyle; eTable3) and nonfatal self-harm (68% [214/316] versus 59% [47/80]; eTable4). Substance abuse diagnoses were more common after gastric bypass than intensive lifestyle in those with nonfatal self-harm events (51% [162/316] versus 29% [23/80], P=0.0003; eTable4).

6/33] for surgery versus 80% [4/5] for intensive lifestyle; eTable3) and nonfatal self-harm (68% [214/316] versus 59% [47/80]; eTable4). Substance abuse diagnoses were more common after gastric bypass than intensive lifestyle in those with nonfatal self-harm events (51% [162/316] versus 29% [23/80], P=0.0003; eTable4). In subgroup analyses, the risk of suicide or nonfatal self-harm was elevated in both SOS and SOReg/Itrim in surgery patients versus controls in the subgroup free of registered psychiatric disorders and without self-harm history at baseline (Figure 3). The risk was also elevated in both studies in the surgery group versus controls in those with as well as those without university education. DISCUSSION We compared the risk of suicide and nonfatal self-harm after bariatric surgery and nonsurgical obesity treatment in two large matched cohorts, and in both of them, surgery patients were at an increased risk. However, despite certain psychiatric disorders being part of the exclusion criteria in the SOS study, surgery patients had almost twice the prevalence of self-harm history at baseline compared to controls (3.4% *69/2008+ versus 1.9% *38/2037+), and such a history is strongly related to future events.(12) Nevertheless, the increased risk was observed also in the subgroup of patients free of known psychiatric disorders and without self-harm history at baseline, in both SOS and SOReg/Itrim.

at baseline compared to controls (3.4% *69/2008+ versus 1.9% *38/2037+), and such a history is strongly related to future events.(12) Nevertheless, the increased risk was observed also in the subgroup of patients free of known psychiatric disorders and without self-harm history at baseline, in both SOS and SOReg/Itrim. Strengths of this study include access to long-term information on self-harm, substance abuse and other psychiatric disorders in two large matched cohort studies of bariatric surgery and nonsurgically treated obese controls. Nationwide health registers enabled near complete outcome ascertainment for both suicide and nonfatal self-harm resulting in hospital care over up to 8 years in SOReg/Itrim and 27 years in SOS. Furthermore, the two cohorts complemented each other: SOS had very long follow-up, which by necessity means older surgical techniques than in SOReg/Itrim. SOS also had less intensive control treatment than SOReg/Itrim. Trials of bariatric surgery have been criticised for use of comparators of insufficient intensity, and very low calorie diets have been discussed as a component of higher intensity regimens.(22) In a meta-analysis of bariatric surgery trials, weight change during the first 2 years in controls ranged between +1kg to -8kg, while surgery patients lost a mean 20–43kg.(23) At one year in SOReg/Itrim, the intensive lifestyle modification resulted in a weight loss of 15% (18kg) compared to 32% (37kg) for gastric bypass.

a meta-analysis of bariatric surgery trials, weight change during the first 2 years in controls ranged between +1kg to -8kg, while surgery patients lost a mean 20–43kg.(23) At one year in SOReg/Itrim, the intensive lifestyle modification resulted in a weight loss of 15% (18kg) compared to 32% (37kg) for gastric bypass. A limitation of this study is that neither SOS nor SOReg/Itrim were randomised. However, it is unlikely that randomised trials of bariatric surgery will have sufficient power to investigate rare events such as suicide, necessitating observational designs. Both SOS and SOReg/Itrim included obese matched controls attempting to lose weight and accounted for multiple suicide risk factors, but selection bias and residual confounding may still have affected our results. The Itrim participants, in contrast to SOS controls, paid for weight loss treatment, while surgery patients were more likely to be referred and not pay out of pocket, which may be important as suicidal behavior displays a strong socioeconomic gradient. In subgroup analysis by education level, we observed elevated risk of suicide or nonfatal self-harm after surgery in both SOS and SOReg/Itrim in all education level strata. For SOReg/Itrim we also adjusted for income, disability pension, and unemployment to reduce bias from differences in socioeconomic position.

dient. In subgroup analysis by education level, we observed elevated risk of suicide or nonfatal self-harm after surgery in both SOS and SOReg/Itrim in all education level strata. For SOReg/Itrim we also adjusted for income, disability pension, and unemployment to reduce bias from differences in socioeconomic position. No patients in our cohorts had sleeve gastrectomy, a method which is increasingly used. Also, Swedes are predominantly Caucasian. We do not know if our results can be generalised to patients having sleeve gastrectomy or other ethnic groups. Regarding procedure type in SOS, the analyses were conducted according to primary procedure. This may overestimate the risks for vertical-banded gastroplasty and gastric banding, as conversion to gastric bypass was common. Regarding follow-up, due to the long recruitment in SOS and nationwide scope of SOReg/Itrim, it was not possible to provide detailed information on contacts with psychologists and primary care after treatment.

te the risks for vertical-banded gastroplasty and gastric banding, as conversion to gastric bypass was common. Regarding follow-up, due to the long recruitment in SOS and nationwide scope of SOReg/Itrim, it was not possible to provide detailed information on contacts with psychologists and primary care after treatment. Finally, in contrast to SOReg/Itrim there was no between-group difference in suicide or nonfatal self-harm during the first 5 years in SOS. A potential explanation for this is that SOS is a prospective study with at least annual follow-up visits during the first 4 years, in addition to routine follow-up in the healthcare system. SOReg patients had less intensive follow-up. Furthermore, only 13% [265/2008] had gastric bypass in SOS compared with 100% in SOReg. In our analyses by procedure type in SOS, gastric bypass was associated with 3.4 times increased risk, gastric banding 2.4 and vertical-banded gastroplasty 2.3 versus controls. Several mechanisms have been suggested for an increased risk of suicide after bariatric surgery,(11, 24) including disappointment among surgery patients due to insufficient weight loss, subsequent weight re-gain, recurrence of obesity-related comorbidities after initial remission, or that weight loss did not have the expected life-changing effects.(11) However, rather than insufficient weight loss, we found that SOS surgery patients who later died by suicide or had a nonfatal self-harm event had either similar or greater weight loss than those who did not, irrespective of primary procedure type.

or that weight loss did not have the expected life-changing effects.(11) However, rather than insufficient weight loss, we found that SOS surgery patients who later died by suicide or had a nonfatal self-harm event had either similar or greater weight loss than those who did not, irrespective of primary procedure type. Previous studies, including SOS, have shown increased incidence of alcohol and substance abuse after gastric bypass,(9) and this could result in impulsive acts. Also, a certain alcohol intake has been reported to result in higher blood alcohol concentrations after compared to before gastric bypass.(25) The effect on uptake of other substances is largely unknown but it is possible that gastric bypass patients more easily get intoxicated. The higher incidence of alcohol abuse after gastric bypass compared to restrictive procedures(26–28) may partly explain why the hazard ratios for suicide and nonfatal self-harm are higher in SOReg/Itrim (gastric bypass only) than in SOS (13% gastric bypass [265/2008]). In SOS, the risk of alcohol abuse diagnosis was 5 times higher after gastric bypass and twice as high after vertical-banded gastroplasty compared to controls, while no difference was detected after gastric banding.(27)

elf-harm are higher in SOReg/Itrim (gastric bypass only) than in SOS (13% gastric bypass [265/2008]). In SOS, the risk of alcohol abuse diagnosis was 5 times higher after gastric bypass and twice as high after vertical-banded gastroplasty compared to controls, while no difference was detected after gastric banding.(27) Mental health problems are much more prevalent in patients undergoing bariatric surgery than in age-sex-matched general population comparators.(13) The 4-year trajectory of antidepressant use after surgery has been reported to be similar to that in the general population, while steeper for benzodiazepines, hypnotics and sedatives.(13) The association between bariatric surgery, different procedure types, and mental health is not yet well-described based on randomised trials or carefully matched cohort studies with obese control groups attempting to lose weight.(10) In SOS, no difference between surgery and controls in overall psychiatric drug use has been found.(29) For SOReg/Itrim, a higher incidence of hypnotic/sedative use and higher dose increases in prevalent users have recently been reported after gastric bypass versus intensive lifestyle modification.(30)

lose weight.(10) In SOS, no difference between surgery and controls in overall psychiatric drug use has been found.(29) For SOReg/Itrim, a higher incidence of hypnotic/sedative use and higher dose increases in prevalent users have recently been reported after gastric bypass versus intensive lifestyle modification.(30) Other proposed mechanisms behind an association between bariatric surgery and suicide include neuroendocrine alterations, exacerbations of depression and anxiety due to micro-/macro-nutrient deficiencies caused by malabsorption, and psychological mechanisms like maladaptive eating behaviors.(31) In SOS, health-related quality of life has been shown to be improved up to 10 years after surgery compared to baseline, and also higher compared to the control group.(32) However, average improvements may mask deteriorating quality of life in a subset of patients due to, for example, surgical complications or alcohol abuse. For a rare event such as suicide, such a subset does not need to be large to produce statistically significant risk increases.

her compared to the control group.(32) However, average improvements may mask deteriorating quality of life in a subset of patients due to, for example, surgical complications or alcohol abuse. For a rare event such as suicide, such a subset does not need to be large to produce statistically significant risk increases. Our observational findings indicate a need for patient information before surgery regarding self-harm and post-operative psychiatric surveillance, as recently suggested.(33) However, it may be difficult to design such a surveillance system, given the rarity of suicides: we observed 42 suicides in SOS and SOReg over 117,000 person-years after surgery. Hence annual psychiatric surveillance is likely to be inefficient. Restricting surveillance to high risk patients, for example those with baseline psychiatric disorders, would be more efficient but applied to our data this strategy would miss almost 50% dying by suicide. Current international guidelines list active or recent substance abuse as a contraindication to surgery.(34) Psychiatric hospitalisation and self-harm history are considered risk factors for poor outcomes but not a contraindication when appropriate mental health treatment is provided. Further, the European guidelines recommend pre-operative psychological assessment by a psychiatrist or psychologist not just for diagnostic purposes but also to identify areas of vulnerability, and higher-risk patients should be selected for post-operative monitoring.(35)

ppropriate mental health treatment is provided. Further, the European guidelines recommend pre-operative psychological assessment by a psychiatrist or psychologist not just for diagnostic purposes but also to identify areas of vulnerability, and higher-risk patients should be selected for post-operative monitoring.(35) Despite our finding of an increased risk of suicide we do not believe that our findings at this point should discourage use of bariatric surgery, at least not from a survival perspective. Several well-designed observational studies show a survival benefit versus obese controls despite a potential increased suicide risk.(2–4) While the relative risk of suicide is high, the absolute risk is low. For example, in the Utah Mortality Study the incidence of all-cause mortality in surgery and matched control participants was 37.6 and 57.1 per 10,000 person-years, respectively, compared to 2.6 and 0.9 for suicide.(2) Beyond mortality, the many documented and common benefits of bariatric surgery(9, 10) are likely to outweigh our finding of an increased risk of suicide and self-harm, but our observations could help to inform and refine guidelines regarding how surgery candidates are selected and followed over time. In conclusion, we found a positive association between bariatric surgery and suicide or nonfatal self-harm. We also found that history of self-harm was more common in patients choosing surgery than in individuals choosing nonsurgical treatment. Supplementary Material FUNDING

Despite our finding of an increased risk of suicide we do not believe that our findings at this point should discourage use of bariatric surgery, at least not from a survival perspective. Several well-designed observational studies show a survival benefit versus obese controls despite a potential increased suicide risk.(2–4) While the relative risk of suicide is high, the absolute risk is low. For example, in the Utah Mortality Study the incidence of all-cause mortality in surgery and matched control participants was 37.6 and 57.1 per 10,000 person-years, respectively, compared to 2.6 and 0.9 for suicide.(2) Beyond mortality, the many documented and common benefits of bariatric surgery(9, 10) are likely to outweigh our finding of an increased risk of suicide and self-harm, but our observations could help to inform and refine guidelines regarding how surgery candidates are selected and followed over time. In conclusion, we found a positive association between bariatric surgery and suicide or nonfatal self-harm. We also found that history of self-harm was more common in patients choosing surgery than in individuals choosing nonsurgical treatment. Supplementary Material FUNDING Research reported in this publication was supported by the National Institute Of Diabetes And Digestive And Kidney Diseases of the National Institutes of Health under Award Number R01DK105948. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

this publication was supported by the National Institute Of Diabetes And Digestive And Kidney Diseases of the National Institutes of Health under Award Number R01DK105948. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Authors of this study were also supported by the Swedish Research Council (LMSC: K2013-54X-11285-19; MN: K2014-99X-22495-01-3). Further, the SOS study has also been supported by grants from the Swedish Research Council (K2012-55X-22082-01, K2013-54X-11285-19, K2013-99X-22279-01), Sahlgrenska University Hospital ALF research grant, and Diabetesfonden. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. AUTHORS’ CONTRIBUTIONS

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. AUTHORS’ CONTRIBUTIONS LMSC is the principal investigator of the SOS study and MN is the principal investigator of the SOReg/Itrim study. LMSC and MN conceived and coordinated the investigation. MN wrote the first draft of the manuscript. MP and GB were responsible for the preparation of data. GB and MN performed the statistical analyses for the first submission. All the authors undertook revisions and contributed intellectually to the development of this paper. MN and LMSC are the study guarantors. Jonas Söderling performed the statistical analyses for the revisions. CONFLICTS OF INTEREST All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare: CM, MN and JS report receiving consulting fees for participation in the scientific advisory committee of Itrim. LMSC reports receiving lecture fees from Johnson & Johnson, Astra Zeneca and MSD. Further, IN is the previous director of the Scandinavian Obesity Surgery Registry and JO is its current director. ETHICS COMMITTEE APPROVAL

All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare: CM, MN and JS report receiving consulting fees for participation in the scientific advisory committee of Itrim. LMSC reports receiving lecture fees from Johnson & Johnson, Astra Zeneca and MSD. Further, IN is the previous director of the Scandinavian Obesity Surgery Registry and JO is its current director. ETHICS COMMITTEE APPROVAL Seven regional ethical review boards approved the study protocol for the SOS study and informed consent was obtained from all patients. The register linkage of SOReg/Itrim was approved by the regional ethics committee in Stockholm, Sweden, and all analyses were conducted on de-identified data. ABBREVIATIONS BMIbody mass index (kg/m2) LCD/VLCDlow calorie diet/very low calorie diet SORegScandinavian Obesity Surgery Registry SOSstudy Swedish Obese Subjects study Figure 1 Cumulative incidence of suicide and nonfatal self-harm in the Swedish Obese Subjects (SOS) study Hazard ratios adjusted for age, sex, BMI, and history of self-harm Figure 2 Cumulative incidence of suicide and nonfatal self-harm in the Swedish Obese Subjects (SOS) study by primary procedure type Case ascertainment from inpatient care and Causes of Death Register only as the outpatient care component was added in 2001 and gastric bypass was used more in the later part of the SOS recruitment period VBG=vertical-banded gastroplasty

Figure 2 Cumulative incidence of suicide and nonfatal self-harm in the Swedish Obese Subjects (SOS) study by primary procedure type Case ascertainment from inpatient care and Causes of Death Register only as the outpatient care component was added in 2001 and gastric bypass was used more in the later part of the SOS recruitment period VBG=vertical-banded gastroplasty Figure 3A Suicide and nonfatal self-harm in the Swedish Obese Subjects (SOS) cohort overall and by subgroups Adjusted for age, sex, BMI, and history of self-harm. Inpatient care only: Refers to case ascertainment excluding data from the outpatient component from the National Patient Register. Outpatient data were available from 2001 and onwards. Psychiatric history: Baseline characteristics for the subgroup with psychiatric history are provided in eTable5. Figure 3B Suicide and nonfatal self-harm in the SOReg/Itrim cohort overall and by subgroups Matched on age, sex, BMI, education level, cardiovascular disease, diabetes, history of self-harm, substance abuse, visits in psychiatric care, use of antidepressants, and use of anxiolytics. Additional adjustment was made for age, BMI and income as continuous variables, as well as for marital status, disability pension, and unemployment status as binary variables. Incidence rates and hazard ratios are weighted by the strata size to account for the matching.

Figure 3B Suicide and nonfatal self-harm in the SOReg/Itrim cohort overall and by subgroups Matched on age, sex, BMI, education level, cardiovascular disease, diabetes, history of self-harm, substance abuse, visits in psychiatric care, use of antidepressants, and use of anxiolytics. Additional adjustment was made for age, BMI and income as continuous variables, as well as for marital status, disability pension, and unemployment status as binary variables. Incidence rates and hazard ratios are weighted by the strata size to account for the matching. Inpatient care only: Refers to case ascertainment excluding data from the outpatient component from the National Patient Register. Psychiatric history: Baseline characteristics for the subgroup with psychiatric history are provided in eTable5. Figure 4 Weight development over 10 years in surgery patients in the SOS study by suicide and self-harm status (overall and by primary procedure type) Adjustment variables were the same as in the main analysis (age, sex, baseline BMI, and history of self-harm) Figure 5 Cumulative incidence of suicide and nonfatal self-harm in the SOReg/Itrim study comparing gastric bypass with intensive lifestyle modification Matched on age, sex, BMI, education level, cardiovascular disease, diabetes, history of self-harm, substance abuse, visits in psychiatric care, use of antidepressants, and use of anxiolytics. Hazard ratios adjusted for age, BMI, income, marital status, disability pension, and unemployment N for intensive lifestyle group are weighted by the strata size to account for the matching

Figure 5 Cumulative incidence of suicide and nonfatal self-harm in the SOReg/Itrim study comparing gastric bypass with intensive lifestyle modification Matched on age, sex, BMI, education level, cardiovascular disease, diabetes, history of self-harm, substance abuse, visits in psychiatric care, use of antidepressants, and use of anxiolytics. Hazard ratios adjusted for age, BMI, income, marital status, disability pension, and unemployment N for intensive lifestyle group are weighted by the strata size to account for the matching Table 1 Participant characteristics at baseline

Figure 5 Cumulative incidence of suicide and nonfatal self-harm in the SOReg/Itrim study comparing gastric bypass with intensive lifestyle modification Matched on age, sex, BMI, education level, cardiovascular disease, diabetes, history of self-harm, substance abuse, visits in psychiatric care, use of antidepressants, and use of anxiolytics. Hazard ratios adjusted for age, BMI, income, marital status, disability pension, and unemployment N for intensive lifestyle group are weighted by the strata size to account for the matching Table 1 Participant characteristics at baseline Swedish Obese Subjects Recruitment: 1987–2001 SOReg/Itrim Recruitment: 2006–2013 Bariatric Surgerya (n=2008) Controls (n=2037) P Gastric Bypass (n=20,256) Intensive Lifestyle (16,162) P Women, n (%) 1420 (70.7%) 1447 (71.0%) 0.824 16,071 (79.3%) 12,823 (79.3%) 1.0 Age (Years), Mean (SD) 47.2 (5.9) 48.7 (6.3) <0.0001 41.3 (10.5) 41.5 (10.8) 0.125 Body-Mass Index (kg/m2), Mean (SD) 42.4 (4.5) 40.1 (4.7) <0.0001 41.1 (3.9) 40.6 (4.1) <0.0001 University Education, n (%) 256 (12.7%) 431 (21.2%) <0.0001 4660 (23.0%) 3718 (23.0%) 1.0 Married, n (%) – – – 9034 (44.6%) 6837 (42.3%) <0.0001 Income (1000 €), Mean (SD) – – – 23.7 (14.5) 28.2 (19.4) <0.0001 Disability Pension, n (%) 357 (17.8%) 316 (15.5%) 0.053 2366 (11.7%) 997 (6.2%) <0.0001 Unemployment, n (%) – – – 2016 (10.0%) 883 (5.5%) <0.0001 History of Psychiatric Illness, n (%) Self-Harm 69 (3.4%) 38 (1.9%) 0.0019 403 (2.0%) 322 (2.0%) 1.0 Substance Abuse 58 (2.9%) 49 (2.4%) 0.339 294 (1.5%) 235 (1.5%) 1.0 Psychiatric healthcare visitsb 200 (10.0%) 175 (8.6%) 0.133 3083 (15.2%) 2460 (15.2%) 1.0 Use of Antidepressants 133 (6.6%) 114 (5.6%) 0.173 6108 (30.2%) 4873 (30.2%) 1.0 Use of Anxiolytics 98 (4.9%) 88 (4.3%) 0.395 3446 (17.0%) 2750 (17.0%) 1.0 Use of Hypnotics and Sedatives 74 (3.7%) 59 (2.9%) 0.160 4426 (21.9%) 2888 (17.9%) <0.0001 Physical Health Status, n (%) Diabetes 346 (17.2%) 263 (12.9%) 0.0001 1954 (9.6%) 1559 (9.6%) 1.0 Cardiovascular Disease 383 (19.1%) 260 (12.8%) <0.0001 4203 (20.7%) 3354 (20.7%) 1.0 a Primary operations were: 1367 (68.1%) vertical-banded gastroplasty, 376 (18.7%) gastric banding, 265 (13.2%) gastric bypass

9%) <0.0001 Physical Health Status, n (%) Diabetes 346 (17.2%) 263 (12.9%) 0.0001 1954 (9.6%) 1559 (9.6%) 1.0 Cardiovascular Disease 383 (19.1%) 260 (12.8%) <0.0001 4203 (20.7%) 3354 (20.7%) 1.0 a Primary operations were: 1367 (68.1%) vertical-banded gastroplasty, 376 (18.7%) gastric banding, 265 (13.2%) gastric bypass b SOS: Only from inpatient care; SOReg: From both inpatient (6.7% surgery versus 6.5% intensive lifestyle, P=0.439) and hospital-based outpatient care (12.4% surgery versus 12.3% intensive lifestyle, P=0.745) RESEARCH IN CONTEXT Evidence before this study A recent systematic review for the US National Institutes of Health concluded that emerging data indicate an increased risk of suicide, or deaths not caused by disease, after bariatric surgery. The cited observational studies used comparators with obesity who applied for driver’s licenses or were seeking but did not receive bariatric surgery. There are no reports on the risk of suicide after bariatric surgery versus nonsurgical weight loss therapy. Further, previous studies have not accounted for baseline differences in psychiatric status such as history of self-harm, substance abuse and depression. Added value of this study Based on two large long-term matched controlled studies of individuals with obesity intending to lose weight, we found a substantially increased relative risk of suicide or nonfatal self-harm in the surgery group, after accounting for baseline psychiatric status.

RESEARCH IN CONTEXT Evidence before this study A recent systematic review for the US National Institutes of Health concluded that emerging data indicate an increased risk of suicide, or deaths not caused by disease, after bariatric surgery. The cited observational studies used comparators with obesity who applied for driver’s licenses or were seeking but did not receive bariatric surgery. There are no reports on the risk of suicide after bariatric surgery versus nonsurgical weight loss therapy. Further, previous studies have not accounted for baseline differences in psychiatric status such as history of self-harm, substance abuse and depression. Added value of this study Based on two large long-term matched controlled studies of individuals with obesity intending to lose weight, we found a substantially increased relative risk of suicide or nonfatal self-harm in the surgery group, after accounting for baseline psychiatric status. The excess risk after surgery was not explained by insufficient weight loss or weight regain, as individuals dying by suicide or who had hospital treatment for nonfatal self-harm had similar or greater weight loss during follow-up than other patients. Despite our attempts to match and stratify the analyses by baseline history of self-harm, substance abuse, depression and anxiety, we cannot rule out in our nonrandomised studies that the observed increased risk of suicide or nonfatal self-harm after bariatric surgery is simply due to different patient characteristics among individuals who chose surgery instead of nonsurgical weight loss methods.

self-harm, substance abuse, depression and anxiety, we cannot rule out in our nonrandomised studies that the observed increased risk of suicide or nonfatal self-harm after bariatric surgery is simply due to different patient characteristics among individuals who chose surgery instead of nonsurgical weight loss methods. Implications of all the available evidence It is unlikely that there will ever be a randomised trial large and long enough to assess the risk of suicide between bariatric surgery and a nonsurgical intervention. Our matched cohort studies and previous observational studies indicate that bariatric surgery is associated with an increased risk of suicide. The absolute suicide risk is small and the association may be influenced by selection bias and residual confounding. However, the relative risk of suicide and nonfatal self-harm is considerable even when accounting for multiple known suicide risk factors. As bariatric surgery, and especially gastric bypass, is associated with increased risk of alcohol and substance abuse, there is also a plausible mechanism behind an increased suicide risk.

first estimate of global cancer burden attributable to diabetes alone and to diabetes and high BMI combined, and uses the most comprehensive available estimates of diabetes and high BMI prevalence. We also quantified the global burden of cancer attributable to rises in the prevalence of diabetes and high BMI over time. Implications of all the available evidence In 2012, about 6% of all incident cancers were attributable to the combined effects of diabetes and high BMI, corresponding to 804 100 cases. As the prevalence of these cancer risk factors increases, clinical and public health efforts should focus on identifying optimal preventive and screening measures for whole populations and individual patients.

nt cancers were attributable to the combined effects of diabetes and high BMI, corresponding to 804 100 cases. As the prevalence of these cancer risk factors increases, clinical and public health efforts should focus on identifying optimal preventive and screening measures for whole populations and individual patients. Methods Study design We reviewed the WCRF continuous update projects, IARC publications, and other published literature that summarised associations of diabetes17 and high BMI with site-specific cancers.4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 We searched MEDLINE via PubMed for articles published up to June 30, 2017, with no language restrictions using the search terms (“Diabetes” OR “Body-mass index” OR “Overweight”, OR “Obesity”), AND (“Cancer risk”, OR “Cancer incidence”), AND “Attributable fraction”. We selected cancers that the WCRF and IARC have judged to have a causal association with high BMI: colorectal, gallbladder, pancreatic, liver, postmenopausal breast, endometrial, kidney, ovarian, stomach cardia, and thyroid cancer, oesophageal adenocarcinoma, and multiple myeloma. For diabetes, we identified published meta-analyses17 of the relative risks (RR) for the association of diabetes with site-specific cancer. The studies included in the meta-analyses had applied rigorous adjustment to control for potential confounding factors, including BMI. The RRs for each site-specific cancer applied in our analysis and their sources are detailed in the appendix 2 (pp 1, 2). For the diabetes analysis we included colorectal, gallbladder, pancreatic, liver, breast, and endometrial cancer.

ied rigorous adjustment to control for potential confounding factors, including BMI. The RRs for each site-specific cancer applied in our analysis and their sources are detailed in the appendix 2 (pp 1, 2). For the diabetes analysis we included colorectal, gallbladder, pancreatic, liver, breast, and endometrial cancer. High BMI has also been proposed to be causally associated with meningioma.13 However, most meningiomas are benign and the incidence of meningioma is not reported in GLOBOCAN. The association between high BMI and oesophageal and stomach cancer is limited to oesophageal adenocarcinoma14 and stomach cardia12 cancer; therefore, we only included these two subtypes in our analysis. Using prevalence of diabetes3 and of categories of BMI2 and RRs for their associations with the cancers identified from published meta-analyses, we estimated the population attributable fraction (PAF) of incident cancers attributable to diabetes and high BMI. For 175 countries in 2012 (appendix 2 p 10), we estimated individual PAFs for each risk factor, as well as two scenarios of diabetes and high BMI combined, one treating their effects as independent and another as overlapping. All analyses were stratified by sex and age group and restricted to people aged 18 years or older. We then estimated the number of cancer cases attributable to diabetes, high BMI, and their combined effect globally by multiplying the PAFs with the number of incident cancers for each age, sex, and country stratum using data from GLOBOCAN.18

ries. After pooling the data, NCD-RisC fitted a bespoke Bayesian hierarchical model to the data with the Markov chain Monte Carlo algorithm and generated 1000 draws from the posterior distribution for each country-year-sex-age stratum. Details have been reported previously in studies investigating BMI and diabetes.2, 3 GLOBOCAN 2012 cancer incidence data18 for the selected cancer sites were available in age groups (15–39, 40–44, 45–49, 50–54, 55–59, 60–64, 65–69, 70–74, and ≥75 years). We used population weighting to ensure that the age groups for diabetes and BMI prevalence were the same as those for cancer incidence. The GLOBOCAN cancer incidence data covered 183 countries and territories, for which both diabetes and BMI estimates were available in 175 of them. We subsequently grouped these 175 countries into nine regions by geographical and national income criteria (appendix 2 p 10). Statistical analysis Most risk factors act proportionally to increase disease risk, therefore we first calculated the proportional reduction of cancer that would occur if exposure to the risk factor was reduced to an alternative scenario, as measured by the PAF.20 The PAF attributable to diabetes and high BMI separately was calculated using the formula21 PAF=∑PiRRi−∑P'iRRi∑PiRRi

Statistical analysis Most risk factors act proportionally to increase disease risk, therefore we first calculated the proportional reduction of cancer that would occur if exposure to the risk factor was reduced to an alternative scenario, as measured by the PAF.20 The PAF attributable to diabetes and high BMI separately was calculated using the formula21 PAF=∑PiRRi−∑P'iRRi∑PiRRi where Pi is the actual prevalence of diabetes or BMI category i, P′i is the prevalence in an alternative scenario, and RRi the adjusted relative risk of site-specific cancer associated with diabetes or the corresponding level of BMI. In our main analysis we estimated the total cancer burden of diabetes and high BMI, and used an optimal prevalence as our alternative scenario—namely zero diabetes prevalence and BMI of 20–25 kg/m2 (used as 22·5 kg/m2 in the calculation), where the cancer risk is assumed to be lowest at the population level. A diabetes prevalence of less than 1% has not been observed,22 so we did a further analysis in which the optimal prevalence of diabetes was 1% rather than zero. We calculated PAFs for 2035 with prevalence in 2025 (projected on the assumption that recent trends continue, as described previously) instead of 2002 prevalence.2, 3, 23 Diabetes and high BMI have increased in prevalence substantially worldwide since 1980.2, 3 We therefore used a second alternative scenario to estimate the cancer burden attributable to these increases. To do this, we replaced the optimal prevalence with the prevalence of diabetes and high BMI in 1980 as the alternative scenario.

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. JP-S, BZ, VK, and JB, had full access to all the data in the study and the corresponding author had final responsibility for the decision to submit for publication. Results In 2012, diabetes and high BMI combined were responsible for an estimated 804 100 new cases of cancer worldwide (5·7% of all 14 067 894 cancer cases reported by GLOBOCAN18) in the independent scenario. 293 300 (2·1%) cancer cases were attributable to diabetes and 544 300 (3·9%) to high BMI alone (Figure 1, Figure 2). In the conservative scenario, the two risk factors combined were responsible for 629 000 new cancer cases in 2012. Cancer cases attributable to diabetes and high BMI combined were almost twice as common in women (501 600 cases) as in men (302 500 cases) in the independent scenario.Figure 1 Global cancer cases in 2012 attributable to diabetes and high BMI, individually and combined, in the conservative and independent scenarios, by region BMI=body-mass index. Figure 2 Global site-specific cancer cases in 2012 Cases by (A) diabetes and high BMI, individually and in combination, in the conservative and independent scenarios and (B) region, in the combined independent scenario. BMI=body-mass index.

Results In 2012, diabetes and high BMI combined were responsible for an estimated 804 100 new cases of cancer worldwide (5·7% of all 14 067 894 cancer cases reported by GLOBOCAN18) in the independent scenario. 293 300 (2·1%) cancer cases were attributable to diabetes and 544 300 (3·9%) to high BMI alone (Figure 1, Figure 2). In the conservative scenario, the two risk factors combined were responsible for 629 000 new cancer cases in 2012. Cancer cases attributable to diabetes and high BMI combined were almost twice as common in women (501 600 cases) as in men (302 500 cases) in the independent scenario.Figure 1 Global cancer cases in 2012 attributable to diabetes and high BMI, individually and combined, in the conservative and independent scenarios, by region BMI=body-mass index. Figure 2 Global site-specific cancer cases in 2012 Cases by (A) diabetes and high BMI, individually and in combination, in the conservative and independent scenarios and (B) region, in the combined independent scenario. BMI=body-mass index. In men, 126 700 cases (95% UI 95 900–159 400) were from liver cancer, constituting 41·9% of all cancer cases attributable to diabetes and high BMI combined in the independent scenario; colorectal cancer cases (69 800 cases, 56 200–83 700) were the next largest contributor, constituting 23·1% of the total cases (Figure 1, Figure 2; table 1). In women, there were 147 400 cases (106 700–190 000) of breast cancer, constituting 29·4% of all cancers attributable to diabetes and high BMI; the second largest contributor was endometrial cancer (121 700 cases, 108 600–135 000), which constituted 24·3% of such cases.Table 1 PAF and number of cancer cases attributable to high BMI and diabetes in 2012, individually and in combination, in independent and conservative scenarios

butable to diabetes and high BMI; the second largest contributor was endometrial cancer (121 700 cases, 108 600–135 000), which constituted 24·3% of such cases.Table 1 PAF and number of cancer cases attributable to high BMI and diabetes in 2012, individually and in combination, in independent and conservative scenarios Total number of cases High BMI PAF High BMI cases Diabetes PAF Diabetes cases Independent PAF Independent scenario cases Conservative PAF Conservative scenario cases Men Colorectal 736 000 5·8% (4·2–7·4) 42 200 (30 600–54 800) 4·0% (2·9–5·1) 29 000 (21 500–37 600) 9·5% (7·6–11·4) 69 800 (56 200–83 700) 6·5% (5·2–8·0) 48 000 (38 300–59 000) Gallbladder 76 000 9·7% (5·8–13·2) 7400 (4500–10 100) 7·8% (4·0–11·9) 5900 (3000–9200) 16·7% (11·9–21·8) 12 800 (9100–16 600) 11·7% (8·2–15·4) 9000 (6300–11 800) Liver 543 000 10·1% (5·7–14·7) 54 600 (31 100–79 600) 14·5% (0·8–19·7) 80 200 (54 700–107 800) 23·3% (17·6–29·3) 126 700 (95 900–159 400) 16·5% (12·4–21·2) 89 500 (67 600–115 400) Pancreas 177 000 5·8% (3·9–7·8) 10 300 (6800–13 700) 12·8% (9·3–16·8) 22 700 (16 200–29 500) 18·0% (14·0–21·6) 31 900 (24 700–38 100) 13·2% (9·7–16·6) 23 300 (17 200–29 300) Kidney 208 000 18·0% (15·5–20·4) 37 400 (32 100–42 300) .. .. 18·0% (15·5–20·4) 37 400 (32 100–42 300) 18·0% (15·5–20·4) 37 400 (32 100–42 300) Oesophagus (adenocarcinoma) 31 700 28·7% (22·6–35·0) 9100 (7200–11 100) .. .. 28·7% (22·6–35·0) 9100 (7200–11 100) 28·7% (22·6–35·0) 9100 (7200–11 100) Stomach (cardia) 72 700 8·8% (3·0–14·8) 6400 (2200–10 800) .. .. 8·8% (3·0–14·8) 6400 (2200–10 800) 8·8% (3·0–14·8) 6400 (2200–10 800) Multiple myeloma 61 900 7·2% (3·3–11·1) 4500 (2100–6900) .. .. 7·2% (3·3–11·1) 4500 (2100–6900) 7·2% (3·3–11·1) 4500 (2100–6900) Thyroid 67 000 5·8% (2·8–8·8) 3900 (1900–5900) .. ..

00) Stomach (cardia) 72 700 8·8% (3·0–14·8) 6400 (2200–10 800) .. .. 8·8% (3·0–14·8) 6400 (2200–10 800) 8·8% (3·0–14·8) 6400 (2200–10 800) Multiple myeloma 61 900 7·2% (3·3–11·1) 4500 (2100–6900) .. .. 7·2% (3·3–11·1) 4500 (2100–6900) 7·2% (3·3–11·1) 4500 (2100–6900) Thyroid 67 000 5·8% (2·8–8·8) 3900 (1900–5900) .. .. 5·8% (2·8–8·8) 3900 (1900–5900) 5·8% (2·8–8·8) 3900 (1900–5900) Total 1 973 300 8·9% 175 800 9·0% 137 800 15·3% 302 500 11·7% 231 100 Women Breast 1 656 000 6·9% (4·4–9·4) 114 800 (72 700–156 500) 2·2% (1·3–3·2) 36 200 (21 400–51 600) 8·9% (6·4–11·5) 147 400 (106 700–190 000) 7·2% (4·9–9·8) 120 000 (82 500–161 500) Endometrial 317 000 31·0% (27·1–35·2) 98 400 (86 000–111 500) 10·8% (7·8–13·8) 33 700 (25 100–43 900) 38·4% (34·3–42·6) 121 700 (108 600–135 000) 31·3% (27·4–35·4) 99 100 (87 000–112 200) Colorectal 607 000 7·0% (5·0–9·1) 42 300 (30 200–55 000) 3·8% (2·8–4·8) 22 700 (16 600–29 200) 10·5% (8·5–12·6) 63 500 (51 800–76 800) 7·3% (5·8–9·2) 44 600 (34 900–55 900) Gallbladder 101 000 12·9% (7·8–17·6) 13 000 (7900–17 700) 7·4% (4·0–11·5) 7600 (3800–11 500) 19·3% (13·6–25·1) 19 400 (13 700–25 200) 13·8% (9·4–18·1) 13 900 (9500–18 300) Liver 223 000 13·5% (7·8–19·4) 30 200 (17 400–43 200) 15·8% (10·9–21·4) 35 300 (24 400–47 200) 27·3% (20·9–33·9) 60 900 (46 500–75 600) 18·8% (14·4–23·8) 42 000 (32 100–53 000) Pancreas 159 000 7·1% (4·6–9·4) 11 200 (7300–15 000) 12·6% (9·2–16·6) 20 000 (14 500–26 200) 19·0% (14·6–22·7) 30 100 (23 200–36 100) 13·1% (9·8–16·5) 20 700 (15 600–26 300) Kidney 118 000 21·3% (18·3–24·1) 25 200 (21 600–28 500) .. .. 21·3% (18·3–24·1) 25 200 (21 600–28 500) 21·3% (18·3–24·1) 25 200 (21 600–28 500) Ovarian 235 000 3·9% (0·9–6·7) 9100 (2000–15 800) .. .. 3·9% (0·9–6·7) 9100 (2000–15 800) 3·9% (0·9–6·7) 9100 (2000–15 800) Oesophagus (adenocarcinoma) 7300 29·5% (23·1–36·1) 2200 (1700–2600) .. .. 29·5% (23·1–36·1) 2200 (1700–2600) 29·5% (23·1–36·1) 2200 (1700–2600) Stomach (cardia) 26 400 11·2% (3·8–18·8) 2900 (1000–5000) .. .. 11·2% (3·8–18·8) 2900 (1000–5000) 11·2% (3·8–18·8) 2900 (1000–5000) Multiple myeloma 51 400 8·9% (4·0–13·3) 4400 (2000–6800) .. .. 8·9% (4·0–13·3) 4400 (2000–6800) 8·9% (4·0–13·3) 4400 (2000–6800) Thyroid 226 400 6·5% (3·2–9·8) 14 800 (7300–22 100) .. .. 6·5% (3·2–9·8) 14 800 (7300–22 100) 6·5% (3·2–9·8) 14 800 (7300–22 100) Total 3 727 500 9·9% 368 500 5·1% 155 500 13·5% 501 600 10·7% 397 900 Numbers in parentheses show 95% UI. PAF=population attributable fraction. BMI=body-mass index.

·0–13·3) 4400 (2000–6800) Thyroid 226 400 6·5% (3·2–9·8) 14 800 (7300–22 100) .. .. 6·5% (3·2–9·8) 14 800 (7300–22 100) 6·5% (3·2–9·8) 14 800 (7300–22 100) Total 3 727 500 9·9% 368 500 5·1% 155 500 13·5% 501 600 10·7% 397 900 Numbers in parentheses show 95% UI. PAF=population attributable fraction. BMI=body-mass index. Of the six cancers associated with diabetes and 12 associated with high BMI, 15·3% in men and 13·5% in women were attributable to the combined effect of these risk factors in the independent scenario (11·7% in men and 10·7% in women in the conservative scenario; table 1). The PAF varied substantially by cancer site in both sexes. Of all liver cancers, 23·3% (17·6–29·3) in men and 27·3% (20·9–33·9) in women were attributable to diabetes and high BMI combined, compared with just 9·5% (7·6–11·4) of cases of colorectal cancer in men and 10·5% (8·5–12·6) in women. 38·4% (34·3–42·6) of all endometrial cancer cases in 2012 were attributable to these risk factors compared with 3·9% (0·9–6·7) of ovarian cancer cases (table 1).

en were attributable to diabetes and high BMI combined, compared with just 9·5% (7·6–11·4) of cases of colorectal cancer in men and 10·5% (8·5–12·6) in women. 38·4% (34·3–42·6) of all endometrial cancer cases in 2012 were attributable to these risk factors compared with 3·9% (0·9–6·7) of ovarian cancer cases (table 1). There were notable differences in the proportion of cancer cases attributable to diabetes versus high BMI individually. For example, high BMI was responsible for about three times the proportion of breast (6·9%) and endometrial (31·0%) cancers as compared with diabetes (2·2% for breast and 10·8% for endometrial; table 1). By contrast, the proportion of liver (14·5%) and pancreatic (12·8%) cancer in men attributable to diabetes was substantially larger than that attributable to high BMI (10·1% for liver and 5·8% for pancreatic). When using 1% as the optimal diabetes prevalence rather than zero, this resulted in a reduction in cancer cases attributable to diabetes by 6·6% (274 000 vs 293 300). 313 000 (38·9%) of 804 100 cases of cancer attributable to the combined risk of diabetes and high BMI in the independent scenario in 2012 occurred in high-income western countries (Figure 1, Figure 2). East and southeast Asia had the second largest proportion (191 900 [23·8%]) of cases attributable to the combined risk of diabetes and high BMI, and the largest number of cancer cases attributable to diabetes individually (108 700 attributable cases) (figure 2).

ed in high-income western countries (Figure 1, Figure 2). East and southeast Asia had the second largest proportion (191 900 [23·8%]) of cases attributable to the combined risk of diabetes and high BMI, and the largest number of cancer cases attributable to diabetes individually (108 700 attributable cases) (figure 2). The contribution of each cancer site to the regional cancer burden also varied substantially. Of the total cancer burden due to the combination of diabetes and high BMI, liver cancer contributed more than 29·6% in the high-income Asia Pacific region and 53·1% in east and southeast Asia, compared with just 6·9% in central and eastern Europe (figure 2B). By contrast, breast and endometrial cancer contributed about 18·3% of the combined cancer burden in east and southeast Asia and 15·1% in the high-income Asia Pacific region, compared with roughly 40·5% in high-income western countries, central and eastern Europe, and sub-Saharan Africa. There were substantial differences in the PAF of cancer attributable to diabetes and those attributable to high BMI in some regions, for example in women in central Asia, the Middle East, and north Africa (5·8% for diabetes vs 14·3% for high BMI; table 2), and in men in east and southeast Asia (10·3% for diabetes vs 5·6% for high BMI)—where diabetes3 has increased faster than expected by the rise in BMI.2Table 2 Regional cancer cases in 2012 attributable to 2002 prevalence and cancer cases that would have been expected in 2012 had prevalence remained at 1980 levels

e 2), and in men in east and southeast Asia (10·3% for diabetes vs 5·6% for high BMI)—where diabetes3 has increased faster than expected by the rise in BMI.2Table 2 Regional cancer cases in 2012 attributable to 2002 prevalence and cancer cases that would have been expected in 2012 had prevalence remained at 1980 levels Number of cases Cases attributable to 2002 prevalence Proportion of cases attributable to 2002 prevalence Cases attributable to 1980 prevalence Proportion of cases attributable to 1980 prevalence Diabetes Men Central and eastern Europe 114 000 7400 (5800–9300) 6·5% 6000 (3800–9900) 5·3% Central Asia and north Africa and the Middle East 56 000 6500 (4900–8300) 11·6% 4100 (2100–6900) 7·3% East and southeast Asia 616 000 63 600 (44 100–84 400) 10·3% 37 500 (15 300–73 800) 6·1% High-income Asia Pacific region 157 000 15 000 (11 300–19 300) 9·6% 11 800 (7200–17 400) 7·5% High-income western countries 385 000 28 700 (20 700-37 200) 7·5% 22 500 (14 200–35 200) 5·8% Latin America and the Caribbean 76 000 6700 (4900–8600) 8·8% 5100 (3100–8100) 6·7% Oceania 800 90 (60–120) 11·3% 50 (20–100) 6·3% South Asia 83 000 7100 (5400–8800) 8·6% 3700 (1200–6700) 4·5% Sub-Saharan Africa 44 000 3000 (2200–4100) 6·8% 1700 (700–3500) 3·9% Women Central and eastern Europe 297 000 16 800 (13 200–20 800) 5·7% 15 800 (9600–24 000) 5·3% Central Asia and north Africa and the Middle East 149 000 8600 (6900–10 400) 5·8% 5500 (3000–9400) 3·7% East and southeast Asia 720 000 45 100 (34 000–57 300) 6·3% 35 000 (16 500–64 700) 4·9% High-income Asia Pacific region 201 000 12 100 (9500–15 400) 6·0% 10 900 (7000–15 800) 5·4% High-income western countries 1 019 000 43 200 (34 100–53 900) 4·2% 38 000 (25 500–56 600) 3·7% Latin America and the Caribbean 254 000 13 900 (11 000–17 600) 5·5% 10 700 (6300–16 900) 4·2% Oceania 2000 130 (90–180) 6·5% 70 (30–150) 3·5% South Asia 283 000 11 100 (8400–14 400) 3·9% 6600 (3000–13 500) 2·3% Sub-Saharan Africa 138 000 4500 (3400–5800) 3·3% 2900 (1400–5700) 2·1% High BMI Men Central and eastern Europe 146 000 18 800 (15 100–22 700) 12·9% 13 400 (10 400–16 900) 9·2% Central Asia and north Africa and the Middle East 67 000 9800 (7200–12 600) 14·6% 6100 (4200–8400) 9·1% East and southeast Asia 711 000 40 000 (25 800–56 100) 5·6% 16 500 (9500–23 500) 2·3% High-income Asia Pacific region 182 000 8600 (6300–11 100) 4·7% 4900 (3500–6800) 2·7% High-income Western countries 502 000 82 200 (65 200–99 000) 16·4% 57 900 (44 900–70 900) 11·5% Latin America and the Caribbean 94 000

00–8400) 9·1% East and southeast Asia 711 000 40 000 (25 800–56 100) 5·6% 16 500 (9500–23 500) 2·3% High-income Asia Pacific region 182 000 8600 (6300–11 100) 4·7% 4900 (3500–6800) 2·7% High-income Western countries 502 000 82 200 (65 200–99 000) 16·4% 57 900 (44 900–70 900) 11·5% Latin America and the Caribbean 94 000 12 300 (9600–15 000) 13·1% 7300 (5500–9600) 7·8% Oceania 800 100 (60–130) 12·5% 60 (40–90) 7·5% South Asia 96 000 2600 (1900–3500) 2·7% 1100 (600–1700) 1·1% Sub-Saharan Africa 46 000 2000 (1300–2800) 4·3% 900 (600–1500) 2·0% Women Central and eastern Europe 348 000 58 700 (49 100–68 500) 16·9% 51 700 (42 600–61 400) 14·9% Central Asia and north Africa and the Middle East 167 000 23 800 (19 100–28 400) 14·3% 16 800 (12 900–21 000) 10·1% East and southeast Asia 815 000 48 000 (38 400–57 700) 5·9% 25 100 (18 500–33 300) 3·1% High-income Asia Pacific region 224 000 10 900 (8600–13 400) 4·9% 8600 (6600–10 800) 3·8% High-income western countries 1 136 000 170 200 (138 000–202 300) 15·0% 124 200 (100 000–149 600) 10·9% Latin America and the Caribbean 281 000 37 700 (30 500–45 000) 13·4% 26 600 (21 000–32 900) 9·5% Oceania 2000 300 (230–370) 15·0% 200 (140–270) 10·0% South Asia 323 000 9800 (7400–12 300) 3·0% 4700 (3000–6700) 1·5% Sub-Saharan Africa 153 000 9700 (7700–11 800) 6·3% 5400 (4100–7000) 3·5% Data are stratified by sex. Numbers in parentheses are 95% UI. BMI=body-mass index.

12 300 (9600–15 000) 13·1% 7300 (5500–9600) 7·8% Oceania 800 100 (60–130) 12·5% 60 (40–90) 7·5% South Asia 96 000 2600 (1900–3500) 2·7% 1100 (600–1700) 1·1% Sub-Saharan Africa 46 000 2000 (1300–2800) 4·3% 900 (600–1500) 2·0% Women Central and eastern Europe 348 000 58 700 (49 100–68 500) 16·9% 51 700 (42 600–61 400) 14·9% Central Asia and north Africa and the Middle East 167 000 23 800 (19 100–28 400) 14·3% 16 800 (12 900–21 000) 10·1% East and southeast Asia 815 000 48 000 (38 400–57 700) 5·9% 25 100 (18 500–33 300) 3·1% High-income Asia Pacific region 224 000 10 900 (8600–13 400) 4·9% 8600 (6600–10 800) 3·8% High-income western countries 1 136 000 170 200 (138 000–202 300) 15·0% 124 200 (100 000–149 600) 10·9% Latin America and the Caribbean 281 000 37 700 (30 500–45 000) 13·4% 26 600 (21 000–32 900) 9·5% Oceania 2000 300 (230–370) 15·0% 200 (140–270) 10·0% South Asia 323 000 9800 (7400–12 300) 3·0% 4700 (3000–6700) 1·5% Sub-Saharan Africa 153 000 9700 (7700–11 800) 6·3% 5400 (4100–7000) 3·5% Data are stratified by sex. Numbers in parentheses are 95% UI. BMI=body-mass index. There was substantial heterogeneity in the proportion of cancer cases attributable to diabetes, high BMI, and their combination in the independent scenario at country level. For example, less than 1% of all new cancer cases in Malawi (0·6%) and Tanzania (0·9%) in 2012 were attributable to diabetes and high BMI combined, compared with more than 10% in Egypt (12·0%) and Mongolia (13·9%)—the countries with the largest PAF—reflecting large variations in risk factor prevalence, and in the way that some cancers are more affected by these factors than others (figure 3).Figure 3 Population attributable fraction of all cancer incidence in 2012

compared with more than 10% in Egypt (12·0%) and Mongolia (13·9%)—the countries with the largest PAF—reflecting large variations in risk factor prevalence, and in the way that some cancers are more affected by these factors than others (figure 3).Figure 3 Population attributable fraction of all cancer incidence in 2012 Population attributable fractions shown are those of (A) diabetes, (B) high BMI, and (C) diabetes and high BMI combined as independent risks. Countries shown in grey did not have cancer incidence data. BMI=body-mass index. We calculated that 25·8% of all cancer cases in 2012 attributable to diabetes were due to the increase in diabetes prevalence from 1980 to 2002 (table 2), equating to 75 600 new cases worldwide. 31·9% of cancer cases attributable to high BMI were due to increased prevalence of this risk factor over the same period, accounting for approximately 174 040 cancer cases. The largest proportion of cancer cases attributable to the increase in prevalence of diabetes and high BMI during this period was in low-income and middle-income countries (LMICs) in Asia and sub-Saharan Africa. At the two extremes, just 3% of cancer cases attributable to diabetes were due to increased diabetes prevalence in women in central and eastern Europe, compared with 57·2% in men in east and southeast Asia.

high BMI during this period was in low-income and middle-income countries (LMICs) in Asia and sub-Saharan Africa. At the two extremes, just 3% of cancer cases attributable to diabetes were due to increased diabetes prevalence in women in central and eastern Europe, compared with 57·2% in men in east and southeast Asia. The PAF of cancer attributable to diabetes and high BMI is expected to increase substantially in coming decades (appendix 2 p 5). For example, PAFs for most site-specific cancers would increase by more than 30% in women and 20% in men when using projected 2025 prevalence compared with 2002 prevalence. In men, the PAF for liver cancer would increase by 47% (from 23·3% to 34·3%) and gallbladder cancer would increase by 53% (from 16·7% to 25·5%), while in women, the PAF for ovarian cancer would increase by 38% (from 3·9% to 5·4%). Discussion We estimated that approximately 6% of cancer cases worldwide in 2012 were attributable to diabetes and high BMI, with high BMI being responsible for almost twice as many cases as diabetes. About a third of cancer cases attributable to diabetes and a quarter of cases attributable to high BMI were due to increases in the prevalence of these risk factors from 1980 to 2002. Given the continued rise in the prevalence of these risk factors since 2002,2, 3 the attributable cancer burden is likely to continue to increase in coming decades. Approximately one in four liver and oesophageal adenocarcinomas and 38·4% of endometrial cancers worldwide in 2012 were estimated to be attributable to diabetes and high BMI.

rise in the prevalence of these risk factors since 2002,2, 3 the attributable cancer burden is likely to continue to increase in coming decades. Approximately one in four liver and oesophageal adenocarcinomas and 38·4% of endometrial cancers worldwide in 2012 were estimated to be attributable to diabetes and high BMI. LMICs have had substantial increases in the prevalence of diabetes and high BMI during the past three decades, whereas parts of Europe and the high-income Asia Pacific region have seen more stable age-standardised prevalences (appendix 2 p 7).2, 3 In our analysis LMICs had the largest increases in numbers of cancer cases attributable both to diabetes, and diabetes and high BMI combined, which is particularly important to note because these countries are generally less well equipped to manage the burden of complex non-communicable diseases (NCDs) than high-income countries. Previous studies have quantified the global cancer burden attributable to nine potentially modifiable diet and lifestyle risk factors (PAF 35% in 2001),26 smoking (PAF 21% in 2000),27 high BMI (PAF 3·6% in 2012),15 and common infections (PAF 15·4% in 2012).28 Our findings suggest that 3·9% of global cancer cases in 2012 were attributable to high BMI, taking into account the four additional cancer sites and more comprehensive and up-to-date BMI data compared with previous work.15

nd high BMI prevalence to cancer incidence that we used is an imperfect measure of cumulative past risk factor exposure, which is important for cancer burden.41 Our PAF analysis quantified the proportion and number of cancer cases that would be averted if diabetes and high BMI prevalence were reduced to optimal levels. However, if the cancer burden of diabetes and high BMI is removed, these risks could lead to populations developing other disorders such as cardiovascular disease and chronic kidney disease as quantified elsewhere.42 Finally, we assumed an optimal diabetes prevalence of zero, and achieving a prevalence of less than 1% might not be feasible.22 Nonetheless, when we substituted zero for 1% as the optimal diabetes prevalence, the cancer burden attributable to diabetes changed by less than 7% and was still responsible for 274 000 cases.

Finally, we assumed an optimal diabetes prevalence of zero, and achieving a prevalence of less than 1% might not be feasible.22 Nonetheless, when we substituted zero for 1% as the optimal diabetes prevalence, the cancer burden attributable to diabetes changed by less than 7% and was still responsible for 274 000 cases. Trends in diabetes and those in BMI were only partly correlated across regions. For example, in south Asia and possibly east Asia diabetes prevalence has risen faster than would be expected by changes in BMI levels, whereas in northern Europe diabetes prevalence is increasing at a slower rate than might be expected by the changes in BMI. Several factors might be causing these diverse trends. First, regional differences in the prevalence of diabetes might be due to differences in genetic susceptibility or phenotypic variations arising from inadequate fetal and childhood nutrition and growth; earlier onset of β-cell dysfunction could be a differentiating characteristic of Asian populations compared with other groups.43, 44, 45, 46, 47 Second, people who are at high risk of developing diabetes might be identified at an earlier stage in health systems in high-income countries, allowing for earlier intervention with lifestyle and dietary modification or drugs.48 Finally, total caloric intake, dietary composition, and physical activity might affect diabetes risk and contribute to differences in regional trends to a greater extent than would otherwise be expected on the basis of BMI.49

ctions are particularly alarming in view of the high, and growing, economic cost of cancers and metabolic diseases, and highlight the importance of integrated control measures to tackle common modifiable risk factors, alongside clinician awareness of diabetes and high BMI as established risk factors for common cancers. Population-based strategies to prevent diabetes and high BMI have great potential impact—not least because many NCDs have overlapping risk factors, comorbidities, and shared sequelae—but have so far often failed, largely because of reluctance by governments and policy makers to pursue structural interventions that tackle key risks for NCDs, such as diet and physical inactivity.1 Future efforts should focus on identifying the most effective clinical interventions to prevent development of NCDs in at-risk groups and their sequelae, such as cancer. Primary care interventions, such as glucose-modifying medications, can be effective in preventing diabetes complications such as macrovascular disease,50 but this approach relies on early identification and close monitoring of people with diabetes, which can be challenging in LMICs that have limited resources. As well as coordinated approaches to halt and reverse the rise in NCDs, global efforts and clinical guidance should reflect the importance of cancer as a sequela of both diabetes and high BMI, and NCD control measures should be integrated into clinical guidelines to identify opportunities to reduce morbidity in this group of patients.

inated approaches to halt and reverse the rise in NCDs, global efforts and clinical guidance should reflect the importance of cancer as a sequela of both diabetes and high BMI, and NCD control measures should be integrated into clinical guidelines to identify opportunities to reduce morbidity in this group of patients. A previous version of this Article has been retracted, for changes made see appendix 1 Supplementary Material Supplementary appendix 1 Supplementary appendix 2 Contributors JP-S and ME conceived the idea of the study. JP-S led the analysis with support from BZ, VK, and JB. ME and MJG supervised the analysis and generating of results. JP-S drafted and finalised the paper with input from all authors. All authors contributed to the analysis, intellectual content, critical revisions to the drafts of the paper and approved the final version. ME had full access to all the data in the study and had final responsibility for the decision to submit for publication. Declaration of interests ME reports a charitable grant from the Young Health Programme of AstraZeneca, and personal fees from Third Bridge, Scor, and Prudential, outside the submitted work. All other authors declare no competing interests.

Introduction High BMI contributed to an estimated 4 million deaths globally in 2015.1 Several major studies and meta-analyses have found strong associations between BMI and all-cause mortality; most have described a U-shaped association with minimum mortality in the healthy weight (20–25 kg/m2) range;2, 3, 4, 5 a 2013 meta-analysis suggested that overweight might be protective6 but concerns were raised about whether the study had adequately accounted for age, reverse causality, and confounding by smoking.7 A recent large meta-analysis, which explored the effect of different methodological decisions on results, observed a higher nadir of the BMI–mortality curve when studies with short follow-up (and thus increased susceptibility to reverse causality) were included, and when ever-smokers were included.2

unding by smoking.7 A recent large meta-analysis, which explored the effect of different methodological decisions on results, observed a higher nadir of the BMI–mortality curve when studies with short follow-up (and thus increased susceptibility to reverse causality) were included, and when ever-smokers were included.2 There is less evidence about the associations between BMI and cause-specific mortality outcomes. Among studies investigating cardiovascular mortality, increased BMI has generally been associated with increased risk,4, 5, 8, 9 but it is unclear whether risk begins to increase with overweight5 or only with obesity,10 and whether underweight affects risk.4, 9 U-shaped or J-shaped associations have been found between BMI and all-cancer mortality4 but there is evidence of variation by cancer site based on studies looking at a range of site-specific cancers.11 Drawing out patterns from existing evidence is complicated by variation in study settings and populations; in analytical strategy, including inclusion and exclusion of smokers, inclusion and exclusion of early follow-up time, and handling of pre-existing disease; and in the range and granularity of outcomes considered. Research in context Evidence before the study

There is less evidence about the associations between BMI and cause-specific mortality outcomes. Among studies investigating cardiovascular mortality, increased BMI has generally been associated with increased risk,4, 5, 8, 9 but it is unclear whether risk begins to increase with overweight5 or only with obesity,10 and whether underweight affects risk.4, 9 U-shaped or J-shaped associations have been found between BMI and all-cancer mortality4 but there is evidence of variation by cancer site based on studies looking at a range of site-specific cancers.11 Drawing out patterns from existing evidence is complicated by variation in study settings and populations; in analytical strategy, including inclusion and exclusion of smokers, inclusion and exclusion of early follow-up time, and handling of pre-existing disease; and in the range and granularity of outcomes considered. Research in context Evidence before the study Two meta-analyses published in 2016 reviewed studies of associations between BMI and all-cause mortality. The first found a J-shaped association with lowest risk at a BMI of 23–24 kg/m2 among never-smokers. Inclusion of smokers and people with existing but undiagnosed illnesses were identified as important potential sources of bias. The second, by the Global BMI Mortality Collaboration, was an individual patient data meta-analysis and found a similar pattern, largely consistent across four continents. In the absence of any broad systematic reviews examining the association of BMI and cause-specific mortality outcomes, we searched PubMed for articles published in the past 10 years (Jan 1, 2007, to Dec 31, 2017) and retrieved 67 studies that investigated associations between BMI (as a continuous variable or in at least three categories) and one or more cause-specific mortality outcomes. The search string was (“body mass index” OR bmi OR obes*OR overweight) AND (mortality OR death); inclusion and exclusion criteria are listed in the appendix. The 67 included studies are described in the appendix; deaths from cardiovascular disease, cancer, and respiratory disease were most often studied, whereas only a small number of studies investigated deaths from other causes—namely, diabetes, external causes, liver or digestive diseases, kidney diseases, and infectious diseases. Only six studies investigated four or more of these categories of causes. Findings for cause-specific mortality outcomes from studies based in European, North American, Australian, and trans-continental settings (probably the most comparable to our data) are summarised in the appendix. Positive or J-shaped associations were observed for most cardiovascular disease mortality outcomes; associations between BMI and cancer mortality were smaller overall but varied by cancer site; inverse or U-shaped associations were reported for respiratory deaths, and for other outcomes there was limited evidence. Analogous results from 30 studies in Asian settings are also summarised in the appendix.

ality outcomes; associations between BMI and cancer mortality were smaller overall but varied by cancer site; inverse or U-shaped associations were reported for respiratory deaths, and for other outcomes there was limited evidence. Analogous results from 30 studies in Asian settings are also summarised in the appendix. Added value of this study To the best of our knowledge, this is one of the largest single cohort studies of its kind to date that quantifies the associations between BMI and a comprehensive range of mortality outcomes at three levels of granularity, including several outcomes for which few previous data are available. We built on previous evidence by using flexible spline models to examine non-linearity in detail, and we also investigated potential effect modification. We used consistent methodology throughout to minimise confounding and reverse causality, and did extensive sensitivity analyses. Implications of all the available evidence

To the best of our knowledge, this is one of the largest single cohort studies of its kind to date that quantifies the associations between BMI and a comprehensive range of mortality outcomes at three levels of granularity, including several outcomes for which few previous data are available. We built on previous evidence by using flexible spline models to examine non-linearity in detail, and we also investigated potential effect modification. We used consistent methodology throughout to minimise confounding and reverse causality, and did extensive sensitivity analyses. Implications of all the available evidence Important associations exist between BMI and almost every category of mortality outcome. In contrast with some previous evidence suggesting that overweight might be protective, we found that risk began to increase above 21–25 kg/m2 for most outcomes, including all-cause mortality, cardiovascular disease, and cancer. However, for mental and behavioural, neurological, and external causes, only lower BMI was associated with increased risk. We found strong evidence of effect modification by age; further work is needed to explore the drivers of this effect and thus understand whether healthy weight recommendations might need to take age into consideration.

nd behavioural, neurological, and external causes, only lower BMI was associated with increased risk. We found strong evidence of effect modification by age; further work is needed to explore the drivers of this effect and thus understand whether healthy weight recommendations might need to take age into consideration. We aimed to examine in detail the association of BMI with all-cause and cause-specific mortality outcomes using a large, single, contemporary, population-based cohort. We applied a consistent approach to minimise reverse causality and residual confounding. We also investigated effect modification by key individual-level characteristics and estimated absolute effects of BMI on mortality outcomes. Methods Study design and setting We did a cohort study using prospectively collected data from the UK Clinical Practice Research Datalink (CPRD) linked to national death registration data. The CPRD contains primary care records from general practitioners (GPs) covering around 9% of the UK population; linkage to death registration data, including the date and causes of death, was available for 80% of GP clinics in England. CPRD data, including the linked subset, have been shown to be broadly representative of the general population in terms of age, sex, ethnicity, and BMI.12, 13, 14, 15 The study protocol was approved by the London School of Hygiene & Tropical Medicine Ethics Committee (14389) and the Independent Scientific Advisory Committee for MHRA Database Research (protocol number 16_174, approved Aug 24, 2016, and provided in the appendix).

Methods Study design and setting We did a cohort study using prospectively collected data from the UK Clinical Practice Research Datalink (CPRD) linked to national death registration data. The CPRD contains primary care records from general practitioners (GPs) covering around 9% of the UK population; linkage to death registration data, including the date and causes of death, was available for 80% of GP clinics in England. CPRD data, including the linked subset, have been shown to be broadly representative of the general population in terms of age, sex, ethnicity, and BMI.12, 13, 14, 15 The study protocol was approved by the London School of Hygiene & Tropical Medicine Ethics Committee (14389) and the Independent Scientific Advisory Committee for MHRA Database Research (protocol number 16_174, approved Aug 24, 2016, and provided in the appendix). Participants, exposures, and outcomes We included all individuals with BMI data collected at age 16 years and older and with subsequent follow-up time available; BMI data were processed as described elsewhere.12 Exposure was assigned as the earliest BMI recorded during CPRD research-standard follow-up (appendix);14 in the absence of BMI recorded at the start of follow-up, we used the most recent historical BMI record (if available) and updated it at the date of the first BMI during follow-up.

rocessed as described elsewhere.12 Exposure was assigned as the earliest BMI recorded during CPRD research-standard follow-up (appendix);14 in the absence of BMI recorded at the start of follow-up, we used the most recent historical BMI record (if available) and updated it at the date of the first BMI during follow-up. To minimise reverse causality (disease leading to weight change), we excluded the first 5 years of follow-up after the BMI record. Follow-up began at whichever was the latest of: start of CPRD research-standard follow-up, the 5-year anniversary of the first BMI record, or on Jan 1, 1998 (start date for death registration data); follow-up ended at death or on March 8, 2016 (end date for death registration data).

s of follow-up after the BMI record. Follow-up began at whichever was the latest of: start of CPRD research-standard follow-up, the 5-year anniversary of the first BMI record, or on Jan 1, 1998 (start date for death registration data); follow-up ended at death or on March 8, 2016 (end date for death registration data). The outcomes were all-cause mortality and cause-specific mortality based on the International Classification of Diseases, 10th revision (ICD-10) code recorded as the underlying cause of death. We used a three-level hierarchical classification of causes of death as used by the Global Burden of Diseases, Injuries, and Risk Factors Study.16 We studied all Level 1 outcomes (communicable diseases, non-communicable diseases, and injuries and external causes); all Level 2 non-communicable disease outcomes (high-level disease groupings such as cancer and cardiovascular disease), and selected Level 3 outcomes (specific disease and injury types, such as lung cancer and heart failure) that were either common causes of death in the UK,17 or were a priori expected to have important associations with BMI (the full list of outcomes with ICD codes is provided in the appendix).

cardiovascular disease), and selected Level 3 outcomes (specific disease and injury types, such as lung cancer and heart failure) that were either common causes of death in the UK,17 or were a priori expected to have important associations with BMI (the full list of outcomes with ICD codes is provided in the appendix). Statistical analysis Cox regression models with an age timescale were fitted for all-cause mortality and for each cause-specific mortality outcome, censoring deaths from competing causes.18 BMI was initially fitted in WHO categories,19 then in finer BMI categories as used by the Global BMI Mortality Collaboration (to aid comparison with work based on this classification),4 and then as a restricted cubic spline (smooth curve). Fully adjusted models were stratified by sex and adjusted for baseline age, smoking, alcohol use, diabetes, index of multiple deprivation (a measure of socioeconomic status20), and calendar period. We excluded individuals with missing smoking status (n=25 373 [0·7%]) or alcohol status (n=244 848 [6·7%]). Further details about parametrisation of covariates are provided in the appendix. Simpler linear or piecewise-linear models were next fitted for the Level 1 and Level 2 outcomes to quantify associations: where there was evidence of non-linearity, a two-line piecewise linear model with a single change point was estimated by trying all possible values for the change point and choosing the value with highest likelihood.

piecewise-linear models were next fitted for the Level 1 and Level 2 outcomes to quantify associations: where there was evidence of non-linearity, a two-line piecewise linear model with a single change point was estimated by trying all possible values for the change point and choosing the value with highest likelihood. We fitted interactions to investigate effect modification by sex, current age, smoking, index of multiple deprivation quintile, and (among those with available data) ethnicity. To quantify absolute effects, we estimated the expected age of death for men and women aged 40 years by BMI category using a simplified Poisson model including BMI category, age, sex, and interactions between these variables. Age 40 years was chosen as this was the approximate median age at entry and there were few deaths at younger ages; full details of the methods are provided in the appendix. Population attributable fractions were calculated by combining hazard ratios (HRs) from the nine-category BMI model with observed deaths in each BMI category.21 Cumulative incidences of mortality from cardiovascular disease, cancer, neurological causes, and respiratory causes were also calculated by BMI category by use of competing risks methods.18

ons were calculated by combining hazard ratios (HRs) from the nine-category BMI model with observed deaths in each BMI category.21 Cumulative incidences of mortality from cardiovascular disease, cancer, neurological causes, and respiratory causes were also calculated by BMI category by use of competing risks methods.18 Sensitivity analyses First, we varied the amount of initial follow-up time after the BMI record that was excluded between 0 years and 10 years; in the primary analysis the first 5 years were excluded to minimise reverse causality. Second, we excluded individuals with prevalent cancer or cardiovascular disease at the start of follow-up for those respective mortality outcomes, to further explore possible reverse causality; we also excluded individuals with previous chronic obstructive pulmonary disorder from the respiratory mortality analysis; dementia or Alzheimer's disease from the neurological mortality analysis; and depression, bipolar disorder, or schizophrenia from the analyses of deaths from mental and behavioural causes and from self-harm and interpersonal violence. Third, we considered alternative non-linear parametrisations of BMI—namely, very fine categories of 1 kg/m2 width (18·0–18·9 kg/m2, 19·0–19·9 kg/m2, and so on), and a two-term fractional polynomial.22 Fourth, we adjusted for ethnicity where such data were available. Fifth, we dropped adjustment for diabetes, then excluded patients with diabetes entirely, in case diabetes might act as an intermediary. Sixth, we dropped BMIs recorded before the start of research-standard CPRD follow-up. Seventh, we restricted the analysis to patients who had a BMI record less than 12 months after registration (more likely because of administrative reasons rather than clinically motivated). Finally, we explored the effect of missing BMI by restricting analyses to more recent calendar periods in which BMI completeness was higher.

e restricted the analysis to patients who had a BMI record less than 12 months after registration (more likely because of administrative reasons rather than clinically motivated). Finally, we explored the effect of missing BMI by restricting analyses to more recent calendar periods in which BMI completeness was higher. Role of the funding source The sponsor had no role in study 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.

e restricted the analysis to patients who had a BMI record less than 12 months after registration (more likely because of administrative reasons rather than clinically motivated). Finally, we explored the effect of missing BMI by restricting analyses to more recent calendar periods in which BMI completeness was higher. Role of the funding source The sponsor had no role in study 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 data from 8 093 746 individuals from CPRD practices in England, 3 632 674 were included in the study after excluding data from individuals with no linked mortality data, BMI data, or BMI data outside of the plausible range (15–50 kg/m2), and from individuals followed up for less than 5 years after BMI measure (appendix). Of individuals included in the study, 1 969 648 were never-smokers. 367 512 deaths were observed (188 057 among never-smokers; table 1; appendix). A positive association was observed between age and BMI (median age 25 years in underweight individuals, 33 years in healthy-weight individuals, 42 years in overweight individuals, and 43 years in obese individuals), in keeping with expected trends;23 70·3% of underweight individuals were women, whereas 56·1% of obese individuals were men (table 1).Table 1 Characteristics of study population at the time of BMI measure by WHO BMI category,19 restricted to individuals with follow-up available from 5 years after the BMI record

duals), in keeping with expected trends;23 70·3% of underweight individuals were women, whereas 56·1% of obese individuals were men (table 1).Table 1 Characteristics of study population at the time of BMI measure by WHO BMI category,19 restricted to individuals with follow-up available from 5 years after the BMI record Underweight (<18·5 kg/m2); n=112 077 Healthy weight (18·5–24·9 kg/m2); n=1 793 989 Overweight (25·0–29·9 kg/m2); n=1 151 359 Obese (≥30·0 kg/m2); n=575 249* Overall; n=3 632 674 Time from BMI record date to end of follow-up (years) Mean (SD) 12·2 (5·4) 13·1 (5·6) 12·6 (5·4) 11·5 (5·1) 12·7 (5·5) Median (IQR) 11 (7·8–15·6) 12 (8·4–17·2) 11·6 (8·1–16·1) 10·4 (7·4–14·4) 11·6 (8·1–16·3) Total follow-up included (millions of person-years) 0·756 13·614 8·248 3·557 26·176 Age (years) <30 73 653 (65·7%) 769 152 (42·9%) 271 837 (23·6%) 124 965 (21·7%) 1 239 607 (34·1%) 30–39 15 780 (14·1%) 394 672 (22·0%) 253 964 (22·1%) 123 093 (21·4%) 787 509 (21·7%) 40–49 7529 (6·7%) 243 730 (13·6%) 211 796 (18·4%) 116 511 (20·3%) 579 566 (16·0%) 50–59 4531 (4·0%) 163 210 (9·1%) 178 411 (15·5%) 99 176 (17·2%) 445 328 (12·3%) 60–69 4221 (3·8%) 117 085 (6·5%) 137 148 (11·9%) 70 164 (12·2%) 328 618 (9·0%) 70–79 3937 (3·5%) 76 051 (4·2%) 76 680 (6·7%) 33 964 (5·9%) 190 632 (5·2%) ≥80 2426 (2·2%) 30 089 (1·7%) 21 523 (1·9%) 7376 (1·3%) 61 414 (1·7%) Median (IQR) 24·7 (19·2–35·5) 32·7 (24·3–47·0) 42·2 (30·6–56·8) 43·3 (31·6–56·5) 36·9 (26·6–52·4) Sex Women 78 745 (70·3%) 1 068 110 (59·5%) 518 896 (45·1%) 322 647 (56·1%) 1 988 398 (54·7%) Men 33 332 (29·7%) 725 879 (40·5%) 632 463 (54·9%) 252 602 (43·9%) 1 644 276 (45·3%) Smoking status Never-smoker 59 327 (52·9%) 989 704 (55·2%) 617 698 (53·6%) 302 919 (52·7%) 1 969 648 (54·2%) Current smoker 41 272 (36·8%) 559 478 (31·2%) 312 427 (27·1%) 147 707 (25·7%) 1 060 884 (29·2%) Ex-smoker 9828 (8·8%) 231 269 (12·9%) 214 542 (18·6%) 121 130 (21·1%) 576 769 (15·9%) Data missing 1650 (1·5%) 13 538 (0·8%) 6692 (0·6%) 3493 (0·6%) 25 373 (0·7%) Alcohol use Non-drinker 27 058 (24·1%) 277 594 (15·5%) 169 072 (14·7%) 105 930 (18·4%) 579 654 (16·0%) Current drinker, 1–2 units per day 56 496 (50·4%) 1 083 931 (60·4%) 674 329 (58·6%) 317 722 (55·2%) 2 132 478 (58·7%) Current drinker, 3–6 units per day 5337 (4·8%) 170 458 (9·5%) 148 318 (12·9%) 56 156 (9·8%) 380 269 (10·5%) Current drinker, ≥7 units per day 1835 (1·6%) 28 469 (1·6%) 21 958 (1·9%) 11 634 (2·0%) 63 896 (1·8%) Current drinker, unknown level 5250 (4·7%) 76 572 (4·3%) 47 688 (4·1%) 25 650 (4·5%) 15

·2%) 2 132 478 (58·7%) Current drinker, 3–6 units per day 5337 (4·8%) 170 458 (9·5%) 148 318 (12·9%) 56 156 (9·8%) 380 269 (10·5%) Current drinker, ≥7 units per day 1835 (1·6%) 28 469 (1·6%) 21 958 (1·9%) 11 634 (2·0%) 63 896 (1·8%) Current drinker, unknown level 5250 (4·7%) 76 572 (4·3%) 47 688 (4·1%) 25 650 (4·5%) 15 5 160 (4·3%) Ex-drinker 2396 (2·1%) 31 974 (1·8%) 24 398 (2·1%) 17 601 (3·1%) 76 369 (2·1%) Data missing 13 705 (12·2%) 124 991 (7·0%) 65 596 (5·7%) 40 556 (7·1%) 244 848 (6·7%) Any previous diabetes diagnosis 885 (0·8%) 25 896 (1·4%) 38 903 (3·4%) 39 323 (6·8%) 105 007 (2·9%) Index of multiple deprivation Quintile 1 (low) 21 735 (19·4%) 428 458 (23·9%) 260 902 (22·7%) 102 681 (17·8%) 813 776 (22·4%) Quintile 2 21 840 (19·5%) 392 644 (21·9%) 255 758 (22·2%) 115 817 (20·1%) 786 059 (21·6%) Quintile 3 23 046 (20·6%) 374 109 (20·9%) 242 194 (21·0%) 121 034 (21·0%) 760 383 (20·9%) Quintile 4 22 960 (20·5%) 329 436 (18·4%) 214 904 (18·7%) 120 965 (21·0%) 688 265 (18·9%) Quintile 5 (high) 22 324 (19·9%) 267 155 (14·9%) 176 289 (15·3%) 114 098 (19·8%) 579 866 (16·0%) Ethnicity White 36 317 (32·4%) 619 968 (34·6%) 428 065 (37·2%) 235 071 (40·9%) 1 319 421 (36·3%) South Asian 5432 (4·8%) 48 835 (2·7%) 28 318 (2·5%) 10 835 (1·9%) 93 420 (2·6%) Black 1474 (1·3%) 23 675 (1·3%) 20 772 (1·8%) 14 292 (2·5%) 60 213 (1·7%) Other 2244 (2·0%) 21 529 (1·2%) 10 279 (0·9%) 4170 (0·7%) 38 222 (1·1%) Mixed 648 (0·6%) 8357 (0·5%) 4499 (0·4%) 2374 (0·4%) 15 878 (0·4%) Data missing 65 962 (58·9%) 1 071 625 (59·7%) 659 426 (57·3%) 308 507 (53·6%) 2 105 520 (58·0%) Calendar year of first available BMI record <1989 803 (0·7%) 18 480 (1·0%) 11 300 (1·0%) 4414 (0·8%) 34 997 (1·0%) 1990–94 16 549 (14·8%) 379 351 (21·1%) 233 344 (20·3%) 83 708 (14·6%) 712 952 (19·6%) 1995–99 22 241 (19·8%) 400 716 (22·3%) 252 260 (21·9%) 111 507 (19·4%) 786 724 (21·7%) 2000–04 28 235 (25·2%) 416 163 (23·2%) 279 589 (24·3%) 153 742 (26·7%) 877 729 (24·2%) 2005–09 36 201 (32·3%) 479 627 (26·7%) 311 622 (27·1%) 181 477 (31·5%) 1 008 927 (27·8%) ≥2010 8048 (7·2%) 99 652 (5·6%) 63 244 (5·5%) 40 401 (7·0%) 211 345 (5·8%) Data are n (%), mean (SD), or median (IQR).

7 (19·4%) 786 724 (21·7%) 2000–04 28 235 (25·2%) 416 163 (23·2%) 279 589 (24·3%) 153 742 (26·7%) 877 729 (24·2%) 2005–09 36 201 (32·3%) 479 627 (26·7%) 311 622 (27·1%) 181 477 (31·5%) 1 008 927 (27·8%) ≥2010 8048 (7·2%) 99 652 (5·6%) 63 244 (5·5%) 40 401 (7·0%) 211 345 (5·8%) Data are n (%), mean (SD), or median (IQR). * Among 575 249 obese individuals, 405 005 (70·4%), had obesity class 1 (BMI 30–34·9 kg/m2), 121 891 (21·2%) had obesity class 2 (BMI 35–39·9 kg/m2), and 48 353 (8·4%) had obesity class 3 (BMI ≥40 kg/m2). A further breakdown of these characteristics by sex as well as BMI category is given in the appendix. Characteristics are at the time of the first BMI record used in study where applicable; smoking was assigned by use of the record from same date as the BMI record or within 1 year before where available (for 3 000 050 [83%] patients), or by use of the nearest record in year after the BMI record (160 790 [4%]), or by use of the nearest record >1 year before the BMI record (285 817 [8%]), or by use of the nearest record >1 year after the BMI record (160 427 [4%]); a similar algorithm was used for alcohol; for ethnicity, the earliest available record was used.

e of the nearest record in year after the BMI record (160 790 [4%]), or by use of the nearest record >1 year before the BMI record (285 817 [8%]), or by use of the nearest record >1 year after the BMI record (160 427 [4%]); a similar algorithm was used for alcohol; for ethnicity, the earliest available record was used. Associations between BMI and mortality were J-shaped for all-cause, communicable, and non-communicable disease mortality. For injuries and external causes, a marked increase in risk was observed at low BMIs, but minimal elevation in risk at higher BMIs (figure 1). Restriction to never-smokers slightly attenuated the association at low BMIs for all these outcomes (figure 1) and also for cancer mortality (appendix). The nadir for all-cause mortality risk among never-smokers was estimated from piecewise linear models to be at a BMI of 25 kg/m2 (table 2).Figure 1 All-cause mortality and Level 1 cause-specific mortality outcomes in total study population (A) and in never-smokers only (B)

e 1) and also for cancer mortality (appendix). The nadir for all-cause mortality risk among never-smokers was estimated from piecewise linear models to be at a BMI of 25 kg/m2 (table 2).Figure 1 All-cause mortality and Level 1 cause-specific mortality outcomes in total study population (A) and in never-smokers only (B) We used a three-level hierarchical classification of causes of death as used by the Global Burden of Diseases, Injuries, and Risk Factors Study.16 All Level 1 outcomes (communicable diseases, non-communicable diseases, and injuries and external causes) were studied. 5-year exclusion period applied for person-time and events after a BMI record. Dashed vertical lines represent WHO BMI category thresholds of 18·5 kg/m2 (underweight to healthy), 25 kg/m2 (healthy weight to overweight), and 30 kg/m2 (overweight to obese). Estimates adjusted for age at BMI record, deprivation, calendar year, diabetes, alcohol status, and smoking (all as defined at date of BMI measure) and stratified for sex. The p values for overall association and p values for non-linearity were less than 0·0001 for all outcomes, in both full and never-smoker populations. HR=hazard ratio. Table 2 Estimated change points in the association between BMI and mortality among never-smokers, and associations with mortality below and above the change point, from piecewise two-line models for the 5-year post-BMI exclusion period

We used a three-level hierarchical classification of causes of death as used by the Global Burden of Diseases, Injuries, and Risk Factors Study.16 All Level 1 outcomes (communicable diseases, non-communicable diseases, and injuries and external causes) were studied. 5-year exclusion period applied for person-time and events after a BMI record. Dashed vertical lines represent WHO BMI category thresholds of 18·5 kg/m2 (underweight to healthy), 25 kg/m2 (healthy weight to overweight), and 30 kg/m2 (overweight to obese). Estimates adjusted for age at BMI record, deprivation, calendar year, diabetes, alcohol status, and smoking (all as defined at date of BMI measure) and stratified for sex. The p values for overall association and p values for non-linearity were less than 0·0001 for all outcomes, in both full and never-smoker populations. HR=hazard ratio. Table 2 Estimated change points in the association between BMI and mortality among never-smokers, and associations with mortality below and above the change point, from piecewise two-line models for the 5-year post-BMI exclusion period BMI change point, kg/m2 (95% CI) HR per 5 kg/m2 BMI increase below change point*(95% CI) HR per 5 kg/m2 BMI increase above change point (95% CI) All-cause mortality 25 (25–25) 0·81 (0·80–0·82) 1·21 (1·20–1·22) Level 1 outcomes Communicable diseases 26 (26–26) 0·73 (0·71–0·76) 1·28 (1·24–1·31) Non-communicable diseases 25 (25–25) 0·83 (0·81–0·84) 1·22 (1·21–1·23) Injuries and external causes 27 (26–28) 0·75 (0·71–0·80) 1·10 (1·04–1·17) Level 2 outcomes (ICD-10 chapters/codes) Cancers (C) 21 (20–25) 0·88 (0·80–0·97) 1·13 (1·12–1·14) Blood and endocrine (D50–89, E) 22 (22–29) 0·43 (0·35–0·54) 1·42 (1·37–1·48) Mental and behavioural (F) 24 (21–25) 0·31 (0·22–0·44) 1·05 (0·86–1·27) Neurological (G) 26 (25–27) 0·68 (0·66–0·70) 0·98 (0·96–1·01) Cardiovascular (I) 25 (25–25) 0·89 (0·87–0·91) 1·29 (1·27–1·30) Respiratory (J23–99) 25 (24–25) 0·53 (0·50–0·56) 1·25 (1·21–1·29) Liver cirrhosis (K70·3/71·7/74·3–6) 23 (22–27) 0·75 (0·48–1·16) 1·44 (1·33–1·55) Digestive (K, excluding cirrhosis) 24 (22–25) 0·79 (0·72–0·86) 1·32 (1·28–1·36) Musculoskeletal (M) 24 (24–25) 0·45 (0·39–0·53) 1·23 (1·15–1·32) Urogenital (N) 25 (24–25) 0·84 (0·77–0·93) 1·45 (1·39–1·51) Accident, transport-related (V) NA* 1·00 (0·90–1·11) .. Accident, excluding transport (W/X00–59) 27 (26–28) 0·71 (0·66–0·77) 1·17 (1·09–1·26) Self-harm and interpersonal violence (X60–Y09) NA* 0·87 (0·80–0·94) .. HR=hazard ratio. ICD-10=International Classification of Diseases, 10th revision. NA=not available.

39–1·51) Accident, transport-related (V) NA* 1·00 (0·90–1·11) .. Accident, excluding transport (W/X00–59) 27 (26–28) 0·71 (0·66–0·77) 1·17 (1·09–1·26) Self-harm and interpersonal violence (X60–Y09) NA* 0·87 (0·80–0·94) .. HR=hazard ratio. ICD-10=International Classification of Diseases, 10th revision. NA=not available. * For transport-related accidents, and self-harm and interpersonal violence, there was little or no evidence against linearity (figure 2) so a single linear effect without change point was estimated.

39–1·51) Accident, transport-related (V) NA* 1·00 (0·90–1·11) .. Accident, excluding transport (W/X00–59) 27 (26–28) 0·71 (0·66–0·77) 1·17 (1·09–1·26) Self-harm and interpersonal violence (X60–Y09) NA* 0·87 (0·80–0·94) .. HR=hazard ratio. ICD-10=International Classification of Diseases, 10th revision. NA=not available. * For transport-related accidents, and self-harm and interpersonal violence, there was little or no evidence against linearity (figure 2) so a single linear effect without change point was estimated. Estimated associations between BMI and more specific mortality outcomes are shown from non-linear spline models (figure 2) and from linear and piecewise linear models (table 2) in never-smokers (analagous results in the full study population including smokers are provided in the appendix). For 11 of 13 Level 2 mortality outcomes, there was evidence of non-linearity, with two main patterns seen: for eight outcomes (cancer, cardiovascular, respiratory, blood/endocrine, liver cirrhosis, other digestive, musculoskeletal, and urogenital causes), we estimated the mortality risk to reach a nadir at BMIs in the range of 21–25 kg/m2, with inverse associations below, and positive associations above, although the magnitude of associations varied; for three outcomes (mental and behavioural, neurological, and accidental [non-transport-related]), we found inverse associations below a BMI of 24–27 kg/m2, with little evidence of association at higher BMIs. We estimated a linear inverse association between BMI and deaths from self-harm and interpersonal violence (HR 0·87 per 5 kg/m2 increase; 95% CI 0·80–0·94); we found no evidence of association between BMI and deaths from transport-related accidents.Figure 2 Association between BMI and Level 2 and Level 3 cause-specific mortality outcomes among never-smokers (organised by ICD-10 code)

rom self-harm and interpersonal violence (HR 0·87 per 5 kg/m2 increase; 95% CI 0·80–0·94); we found no evidence of association between BMI and deaths from transport-related accidents.Figure 2 Association between BMI and Level 2 and Level 3 cause-specific mortality outcomes among never-smokers (organised by ICD-10 code) We used a three-level hierarchical classification of causes of death as used by the Global Burden of Diseases, Injuries, and Risk Factors Study.16 We studied all Level 2 non-communicable disease outcomes and selected Level 3 outcomes that were either common causes of death in the UK or were a priori expected to have important associations with BMI. 5-year exclusion period applied for person-time and events after a BMI record; estimates adjusted for age, deprivation, calendar year, diabetes, alcohol status (all as defined at date of BMI measure) and stratified for sex. HR=hazard ratio. ICD-10=International Classification of Diseases, 10th revision. The overall pattern of association between BMI and cardiovascular death was similar for most Level 3 cardiovascular outcomes but was more muted for cerebrovascular deaths (figure 2). Broadly positive associations were observed between BMI and 12 of 15 site-specific cancer mortality outcomes, with evidence of non-linearity for oesophageal, stomach, uterus, prostate, and kidney cancers, generally reflecting a weak association at the lowest BMIs, although for oesophageal cancer there was a J-shaped association. BMI was not associated with deaths from lung cancer, brain or CNS cancer, or malignant melanoma.

s, with evidence of non-linearity for oesophageal, stomach, uterus, prostate, and kidney cancers, generally reflecting a weak association at the lowest BMIs, although for oesophageal cancer there was a J-shaped association. BMI was not associated with deaths from lung cancer, brain or CNS cancer, or malignant melanoma. Results from categorical BMI models are in the appendix. High BMI was more strongly associated with overall and cardiovascular mortality in men than in women (figure 3; appendix). Most associations between BMI and mortality were attenuated at older ages (figure 3; appendix). In a post-hoc analysis we estimated the nadir of all-cause mortality risk to be 23 kg/m2 at age younger than 70 years, rising to 25 kg/m2 at age 70 years and older. We found no evidence of effect modification by deprivation (appendix). Investigation of effect modification by ethnicity was limited by low power and we only considered all-cause mortality; despite some difference in the observed patterns by ethnicity, particularly at low BMI, there was insufficient evidence to rule out chance variation (appendix).Figure 3 Association between BMI and all-cause mortality among never-smokers, by sex (A) and age (B) 5-year exclusion period applied for person-time and events after a BMI record; estimates adjusted for age, deprivation, calendar year, diabetes, and alcohol status (all as defined at date of BMI measure) and stratified by sex. HR=hazard ratio.

High BMI was more strongly associated with overall and cardiovascular mortality in men than in women (figure 3; appendix). Most associations between BMI and mortality were attenuated at older ages (figure 3; appendix). In a post-hoc analysis we estimated the nadir of all-cause mortality risk to be 23 kg/m2 at age younger than 70 years, rising to 25 kg/m2 at age 70 years and older. We found no evidence of effect modification by deprivation (appendix). Investigation of effect modification by ethnicity was limited by low power and we only considered all-cause mortality; despite some difference in the observed patterns by ethnicity, particularly at low BMI, there was insufficient evidence to rule out chance variation (appendix).Figure 3 Association between BMI and all-cause mortality among never-smokers, by sex (A) and age (B) 5-year exclusion period applied for person-time and events after a BMI record; estimates adjusted for age, deprivation, calendar year, diabetes, and alcohol status (all as defined at date of BMI measure) and stratified by sex. HR=hazard ratio. The expected age of death for a 40-year-old never-smoker with healthy weight was 82·2 years for men and 84·3 years for women (table 3). Underweight, overweight, and obesity were all associated with reductions in these life expectancies: obesity overall was associated with shortening of life expectancy by 4·2 years in men and by 3·5 years in women; class 3 obesity was associated with shortening of life expectancy by 9·1 years in men and by 7·7 years in women.Table 3 Expected age at death for a never-smoker aged 40 years by WHO BMI category,19 and estimated reduction in life expectancy compared with an individual of healthy weight

in men and by 3·5 years in women; class 3 obesity was associated with shortening of life expectancy by 9·1 years in men and by 7·7 years in women.Table 3 Expected age at death for a never-smoker aged 40 years by WHO BMI category,19 and estimated reduction in life expectancy compared with an individual of healthy weight Men Women Expected age of death at age 40 years (years) Reduction in life expectancy (years) Expected age of death at age 40 years (years) Reduction in life expectancy (years) Underweight (<18·5 kg/m2) 77·9 4·3 79·8 4·5 Healthy weight (18·5–24·9 kg/m2) 82·2 .. 84·3 .. Overweight (25·0–29·9 kg/m2) 81·2 1·0 83·5 0·8 Obese (all, ≥30·0 kg/m2) 78·0 4·2 80·9 3·5 Obese class 1 (30·0–34·9 kg/m2) 78·7 3·4 81·9 2·4 Obese class 2 (35·0–39·9 kg/m2) 76·2 5·9 79·6 4·7 Obese class 3 (≥40·0 kg/m2) 73·1 9·1 76·6 7·7 Expected age of death at age 40 years estimated from a Poisson model for overall survival with six-category BMI variable, 5-year age bands, sex, and interaction terms for BMI with age at BMI, and BMI with sex (see appendix for details); estimates assume mortality observed in the study remains constant. Reduction in life expectancy is calculated as expected age of death minus expected age at death in the healthy weight category. Assuming causality, we estimated that 4·3% of all deaths might be attributable to obesity, and 5·5% to overweight including obesity (appendix). Cumulative incidences of cause-specific mortality outcomes by BMI are shown in the appendix.

Men Women Expected age of death at age 40 years (years) Reduction in life expectancy (years) Expected age of death at age 40 years (years) Reduction in life expectancy (years) Underweight (<18·5 kg/m2) 77·9 4·3 79·8 4·5 Healthy weight (18·5–24·9 kg/m2) 82·2 .. 84·3 .. Overweight (25·0–29·9 kg/m2) 81·2 1·0 83·5 0·8 Obese (all, ≥30·0 kg/m2) 78·0 4·2 80·9 3·5 Obese class 1 (30·0–34·9 kg/m2) 78·7 3·4 81·9 2·4 Obese class 2 (35·0–39·9 kg/m2) 76·2 5·9 79·6 4·7 Obese class 3 (≥40·0 kg/m2) 73·1 9·1 76·6 7·7 Expected age of death at age 40 years estimated from a Poisson model for overall survival with six-category BMI variable, 5-year age bands, sex, and interaction terms for BMI with age at BMI, and BMI with sex (see appendix for details); estimates assume mortality observed in the study remains constant. Reduction in life expectancy is calculated as expected age of death minus expected age at death in the healthy weight category. Assuming causality, we estimated that 4·3% of all deaths might be attributable to obesity, and 5·5% to overweight including obesity (appendix). Cumulative incidences of cause-specific mortality outcomes by BMI are shown in the appendix. Exclusion of more initial follow-up time attenuated inverse associations at low BMI for most outcomes; exclusion of prevalent disease gave similar results to exclusion of early person-time (appendix). Other sensitivity analyses made little difference to estimated associations between BMI and mortality (appendix).

Assuming causality, we estimated that 4·3% of all deaths might be attributable to obesity, and 5·5% to overweight including obesity (appendix). Cumulative incidences of cause-specific mortality outcomes by BMI are shown in the appendix. Exclusion of more initial follow-up time attenuated inverse associations at low BMI for most outcomes; exclusion of prevalent disease gave similar results to exclusion of early person-time (appendix). Other sensitivity analyses made little difference to estimated associations between BMI and mortality (appendix). Discussion We observed a J-shaped association between BMI and all-cause mortality, with lowest mortality at 25 kg/m2. BMI was associated with mortality risk from every main category of cause except for transport-related accidents. There were three broad patterns of association: for cancer, cardiovascular, respiratory, blood and endocrine, digestive, musculoskeletal, and urogenital causes of death, we found J-shaped associations with nadirs at 21–25 kg/m2, and varying magnitudes of association; for mental and behavioural, neurological, and accidental (non-transport-related) causes, BMI was inversely associated with mortality up to 24–27 kg/m2 with little association above this point; and BMI was inversely and linearly associated with deaths from self-harm and interpersonal violence. Associations between high BMI and several mortality outcomes attenuated with age and were stronger in men than in women. Obesity was associated with a 4·2-year reduction in remaining life expectancy for a male 40-year-old never-smoker and a 3·5-year reduction for a female 40-year-old never smoker, when compared with individuals of healthy weight, with longest reductions in life expectancy estimated for the most severely obese (class 3) individuals; underweight was associated with a reduction in life expectancy of more than 4 years.

-old never-smoker and a 3·5-year reduction for a female 40-year-old never smoker, when compared with individuals of healthy weight, with longest reductions in life expectancy estimated for the most severely obese (class 3) individuals; underweight was associated with a reduction in life expectancy of more than 4 years. The J-shaped association we observed between BMI and all-cause mortality was consistent with results from some major studies,2, 4 but others have estimated a reduced risk among overweight individuals compared with those of healthy weight.6 Reverse causality and residual confounding (in particular by smoking) might partly explain the discrepant results between studies; associations between overweight and mortality are attenuated in studies with short follow-up, or in studies that include smokers.2 We also observed clear heterogeneity in the associations between BMI and different causes of death, and strong interactions with age, meaning that the association of all-cause mortality with BMI in any individual study will be affected by the age and cause-of-death distributions in the source population.

lude smokers.2 We also observed clear heterogeneity in the associations between BMI and different causes of death, and strong interactions with age, meaning that the association of all-cause mortality with BMI in any individual study will be affected by the age and cause-of-death distributions in the source population. In the absence of broad evidence summarising cause-specific mortality outcomes, we did a systematic literature review and identified 67 relevant studies from 2007 to 2017 for comparison (appendix). Most studies either excluded an initial follow-up period of 1–10 years (n=20) or excluded people with previous disease (n=11), or excluded both (n=13). Most studies of overall cardiovascular mortality from similar settings found approximately J-shaped associations, consistent with our results (appendix); only a few examined more specific cardiovascular mortality outcomes. J-shaped associations were similarly seen for deaths from coronary heart disease and heart failure, and, in some studies, for cerebrovascular deaths;4, 5 however, more modest associations between BMI and cerebrovascular death were also observed, as in the present analysis.24 Ischaemic and haemorrhagic stroke were rarely distinguished, despite their potentially different associations with BMI.5 Most previous evidence about cancer mortality has focused on any cancer type or on breast, colorectal, lung, and prostate cancers, with less common cancers infrequently studied (appendix). Studies into prostate cancer mortality generally imposed linearity or used few BMI categories; our results suggest important non-linearity in the association, with a levelling off or reduction in risk at the highest BMIs. Various studies agreed with our finding of no association between BMI and lung cancer mortality, whereas others found both strong positive and negative associations, possibly reflecting the complications of adequately accounting for smoking. Few studies looked at mortality outcomes other than those from cardiovascular disease and cancer (appendix).

r finding of no association between BMI and lung cancer mortality, whereas others found both strong positive and negative associations, possibly reflecting the complications of adequately accounting for smoking. Few studies looked at mortality outcomes other than those from cardiovascular disease and cancer (appendix). The raised risks of many outcomes at low BMI, coupled with the fact that mental health conditions showed the strongest inverse associations with low BMI, might indicate pervasive effects of mental health problems on a range of outcomes, through pathways that could include poorer self-care and less access to or use of health-care services, or both. The persistence of inverse associations between BMI and deaths from self-harm and interpersonal violence even in sensitivity analyses in which follow-up was started up to 10 years after BMI recording, or when individuals with previously recorded mental illness were excluded, argues against reverse causality. However, it remains possible that depression and related diseases leading to appetite suppression even over a long time period or without a formal diagnosis could partly explain this finding. Imposing a longer period between BMI recording and study entry tended to attenuate associations between low BMI and outcomes; this might have been observed because a side-effect of this approach is to reduce the amount of person-time included at young ages, and we separately found the strongest associations between low BMI and mortality to be in younger people. The analyses stratified by age also suggested that mortality was minimised in older individuals at higher BMI, perhaps indicating increased importance of nutritional reserves in older age. This finding might suggest that healthy weight recommendations need to account for age, but further work is needed to establish whether increased weight is actually beneficial for older individuals: there is increased risk of reverse causation in older people because of the increased prevalence of most diseases, and BMI might be compromised as a measure of adiposity in the oldest individuals because of complications from loss of muscle mass.25

whether increased weight is actually beneficial for older individuals: there is increased risk of reverse causation in older people because of the increased prevalence of most diseases, and BMI might be compromised as a measure of adiposity in the oldest individuals because of complications from loss of muscle mass.25 In this study, we systematically analysed outcomes at different levels of granularity, used consistent methodology to deal with confounding and reverse causality, and did a wide range of sensitivity analyses. A particular strength was that the size of the study allowed us to retain power while restricting analyses to never-smokers and thus minimising confounding by smoking. BMI data in the CPRD have good validity and representativeness.12 In preliminary analyses, we validated linked mortality data by comparing these with data recorded directly in primary care records; dates of death agreed to within 1 month in 97% of cases.

yses to never-smokers and thus minimising confounding by smoking. BMI data in the CPRD have good validity and representativeness.12 In preliminary analyses, we validated linked mortality data by comparing these with data recorded directly in primary care records; dates of death agreed to within 1 month in 97% of cases. 25% of individuals who were otherwise eligible were excluded because they had no BMI records available. This complete case analysis approach is valid provided that absence of data is conditionally independent of the outcome under study;26 we considered this approach more appropriate than multiple imputation, because underweight and overweight individuals would be more likely to have their BMI recorded in primary care, contradicting the required missing at random assumption.27 Several sensitivity analyses suggested that missing BMI data had little effect on our estimates: there was little change in results when analyses were restricted to more recent calendar years, despite BMI data completeness increasing from 66% to 80% in 2000–15. There might have been inaccuracies in cause of death recording. Different physicians could differently interpret the underlying cause of death, and certification of deaths in hospital could be completed by one of several doctors in a team, increasing the risk of classification errors.28 Some causes of death might be particularly prone to misclassification: we noted a relatively large number of deaths attributed to pneumonia, which might have been secondary to other morbidities. Clear national guidance about certification should have helped reduce errors.28 If any misclassification was unrelated to BMI, the expected impact on our analysis would be loss of power. We might also have missed deaths that were not registered in the UK because of emigration, but these are likely to represent a small proportion of deaths. We had no data about diet, physical activity, or cardiorespiratory fitness, with which we could have further explored causal pathways and confounding,29, 30 and we had no data about measures of adiposity other than BMI.

red in the UK because of emigration, but these are likely to represent a small proportion of deaths. We had no data about diet, physical activity, or cardiorespiratory fitness, with which we could have further explored causal pathways and confounding,29, 30 and we had no data about measures of adiposity other than BMI. In conclusion, BMI had a J-shaped association with overall mortality, and BMI outside the healthy range was associated with up to several years of lost lifespan, with most of the absolute mortality burden driven by obesity (BMI ≥30 kg/m2). However, the overall association between BMI and mortality was driven by varying associations with individual cause-specific mortality outcomes, including predominantly inverse associations for mental and behavioural, neurological, and external causes. Associations between BMI and mortality varied by age; an improved understanding of the reasons for this interaction could help inform age-specific public health recommendations. Supplementary Material Supplementary appendix Acknowledgments KB holds a Sir Henry Dale fellowship jointly funded by the Wellcome Trust and the Royal Society (grant number 107731/Z/15/Z). Contributors KB had the initial idea, analysed the data, and wrote the paper. All authors contributed to study design, commenting on drafts, and revisions.

Supplementary Material Supplementary appendix Acknowledgments KB holds a Sir Henry Dale fellowship jointly funded by the Wellcome Trust and the Royal Society (grant number 107731/Z/15/Z). Contributors KB had the initial idea, analysed the data, and wrote the paper. All authors contributed to study design, commenting on drafts, and revisions. Declaration of interests KB reports grants from the Wellcome Trust and the Royal Society, during the conduct of the study, and grants from the Medical Research Council (MRC) and British Heart Foundation, outside the submitted work. IJD reports grants from GlaxoSmithKline and owns shares in the company; and reports grants from the National Institute of Health Research and Association of the British Pharmaceutical Industry, outside the submitted work. LS reports grants from the Wellcome Trust, MRC, the National Institute for Health Research (NIHR), GlaxoSmithKline, British Heart Foundation, and Diabetes UK, outside the submitted work. LS is a trustee of the British Heart Foundation. Id-S-S and DAL declare no competing interests.

Research in context Evidence before this study We searched PubMed, CINAHL, EMBASE, Web of Science, and Google Scholar for systematic reviews and published original studies on non-pharmacological interventions for the prevention and control of type 2 diabetes published up to July, 2015, with a particular focus on mHealth and community mobilisation interventions in low-income and middle-income countries. We used the search terms “mHealth”, “digital interventions”, “community interventions”, “community groups”, “peer support”, “peer education”, and “community participation” in combination with “diabetes management”, “diabetes prevention”, and “chronic disease”. There were no language restrictions used. Because we wanted to understand both the nature of existing interventions and their effectiveness, we were interested in a range of studies including randomised controlled trials, pilot studies, case-control studies, and qualitative research. Previous studies have shown that mHealth techniques might affect health behaviours, including treatment adherence, weight loss, diet, and exercise, and might reduce the incidence of type 2 diabetes among high-risk individuals. Good evidence exists for group-support and peer-support lifestyle interventions in the prevention or delaying of the onset of type 2 diabetes. Evidence supports the involvement of the community and peer support as a cost-effective means of promoting lifestyle changes in high-income settings, but research in resource-poor settings is lacking. Evidence on the effects of mHealth and community mobilisation on diabetes and related risk factors among the general population, as opposed to high-risk individuals, is yet to emerge from low-income or middle-income countries.

lifestyle changes in high-income settings, but research in resource-poor settings is lacking. Evidence on the effects of mHealth and community mobilisation on diabetes and related risk factors among the general population, as opposed to high-risk individuals, is yet to emerge from low-income or middle-income countries. Added value of this study The Bangladesh DMagic trial provides the first large-scale, population-level evidence concerning the effectiveness and cost-effectiveness of mHealth and community mobilisation interventions for reducing the prevalence of intermediate hyperglycaemia and type 2 diabetes and the incidence of type 2 diabetes. Both interventions were acceptable to participants and achieved large population coverage. The mHealth intervention increased knowledge and awareness of type 2 diabetes and its risk factors but had no detectable impact on disease outcomes. Community mobilisation using a participatory learning and action (PLA) approach not only increased knowledge and awareness of disease, but also significantly reduced population prevalence of diabetes and intermediate hyperglycaemia and the incidence of type 2 diabetes among an intermediate hyperglycaemic cohort. National scale-up of PLA in Bangladesh could prevent about 240 000 cases of type 2 diabetes and intermediate hyperglycaemia each year, representing savings in health-care costs of INT$132 million per year. Implications of all the available evidence

The Bangladesh DMagic trial provides the first large-scale, population-level evidence concerning the effectiveness and cost-effectiveness of mHealth and community mobilisation interventions for reducing the prevalence of intermediate hyperglycaemia and type 2 diabetes and the incidence of type 2 diabetes. Both interventions were acceptable to participants and achieved large population coverage. The mHealth intervention increased knowledge and awareness of type 2 diabetes and its risk factors but had no detectable impact on disease outcomes. Community mobilisation using a participatory learning and action (PLA) approach not only increased knowledge and awareness of disease, but also significantly reduced population prevalence of diabetes and intermediate hyperglycaemia and the incidence of type 2 diabetes among an intermediate hyperglycaemic cohort. National scale-up of PLA in Bangladesh could prevent about 240 000 cases of type 2 diabetes and intermediate hyperglycaemia each year, representing savings in health-care costs of INT$132 million per year. Implications of all the available evidence The effect size of the PLA community mobilisation on blood glucose is compelling and was robust to sensitivity analysis. Based on the philosophy of Paulo Freire and building on earlier evidence of effectiveness on maternal, neonatal, and child health, ours is the first study to show effectiveness of PLA on risk of type 2 diabetes. The observed effect of facilitated discussion, mutual learning, and collective action is an important challenge to the individualised nature of behavioural interventions, which have shown little success in reducing diabetes risk in general populations. However, as the first study of its kind, replication studies in Bangladesh and elsewhere should be a research priority. The absence of major quantifiable changes in behavioural indicators related to diet, physical activity, and care seeking demands further exploration and hints at complex, ecological mechanisms of action. Lack of evidence of an effect of mHealth on disease outcomes in our general population contrasts with findings from mHealth interventions that target high-risk individuals in other settings. Mixed-methods implementation research will be essential to better understand and develop population-level interventions that stimulate contextually specific actions to prevent and control type 2 diabetes.

population contrasts with findings from mHealth interventions that target high-risk individuals in other settings. Mixed-methods implementation research will be essential to better understand and develop population-level interventions that stimulate contextually specific actions to prevent and control type 2 diabetes. Introduction The global prevalence of diabetes was estimated to be 9% among adults in 2016 and about 75% of people living with diabetes were in low-income and middle-income countries (LMICs).1 Roughly 20–30% of adults in rural areas of Bangladesh have abnormal fasting glucose or impaired glucose tolerance (together termed intermediate hyperglycaemia) and about 10% have diabetes,2, 3, 4 with the prevalence of diabetes (mostly type 2 diabetes) expected to reach 24–34% by 2030.5 Despite the large burden of diabetes and intermediate hyperglycaemia in Bangladesh, awareness and knowledge is low6 and effective strategies to prevent and control diabetes are urgently needed.

nd about 10% have diabetes,2, 3, 4 with the prevalence of diabetes (mostly type 2 diabetes) expected to reach 24–34% by 2030.5 Despite the large burden of diabetes and intermediate hyperglycaemia in Bangladesh, awareness and knowledge is low6 and effective strategies to prevent and control diabetes are urgently needed. Lifestyle and non-pharmacological interventions can prevent or delay the onset of type 2 diabetes.7 Individual targeted strategies that use mobile phone technology (mHealth) have been shown to reduce the incidence of type 2 diabetes in high-risk individuals,8 but have not been shown to affect behaviour change and diabetes status among a general, rural population. Community-based and peer support interventions might be a cost-effective means of promoting lifestyle changes in LMICs,9, 10, 11 although a recent trial in India showed no effect on disease outcomes.12 Participatory learning and action (PLA) is a specific approach to community mobilisation that engages communities to identify and address their own local problems. It has been shown to improve maternal and newborn survival in LMICs13 and might also improve child health14, 15 and women's reproductive health.16 We aimed to separately assess the effects of mHealth health messaging and PLA community mobilisation on the prevalence of intermediate hyperglycaemia and diabetes among the general adult population in rural Bangladesh, and to assess the effect of these interventions on the incidence of diabetes among people with intermediate hyperglycaemia within the study population.

th health messaging and PLA community mobilisation on the prevalence of intermediate hyperglycaemia and diabetes among the general adult population in rural Bangladesh, and to assess the effect of these interventions on the incidence of diabetes among people with intermediate hyperglycaemia within the study population. Methods Study design and participants The DMagic (Diabetes Mellitus: Action through community Groups or mHealth Information for better Control of population blood glucose, risk factors, knowledge and care seeking) trial was done in 96 villages (total population about 125 000) across four selected subdistricts (upazillas) in Faridpur district, Bangladesh, from June 27, 2015 to June 28, 2018. 24 villages with population between 750 and 2500 were selected in each upazilla to minimise contamination via contiguous borders between villages, with buffer, non-study villages separating most clusters. Intervention mapping at the beginning of the study revealed no recent or ongoing community-based programmes specifically designed to reduce the burden of non-communicable diseases, including type 2 diabetes.

contamination via contiguous borders between villages, with buffer, non-study villages separating most clusters. Intervention mapping at the beginning of the study revealed no recent or ongoing community-based programmes specifically designed to reduce the burden of non-communicable diseases, including type 2 diabetes. DMagic was a three-arm, stratified, cluster-randomised controlled trial in which villages were the units of randomisation and men and non-pregnant women aged 30 years and older were the units of analysis. We did 10 months of formative research and intervention development, including a baseline survey of intermediate hyperglycaemia, diabetes, and non-communicable disease risk factors, and piloting of the PLA intervention; an 18-month intervention phase; and an 8-month post-intervention phase, including an end-of-study survey and analysis. Process evaluation was done concurrently to describe the intervention implementation and explore mechanisms of effect. The trial received ethical approval from University College London, London, UK (4766/002) and the Diabetic Association of Bangladesh, Dhaka, Bangladesh (BADAS-ERC/EC/t5100246). All survey participants provided informed consent through signature or thumbprint. The study protocol has been previously published.17 Data collection, management, and analytical procedures were monitored by an independent data monitoring committee. Trial management was also reviewed by an independent trial steering committee.

ey participants provided informed consent through signature or thumbprint. The study protocol has been previously published.17 Data collection, management, and analytical procedures were monitored by an independent data monitoring committee. Trial management was also reviewed by an independent trial steering committee. Randomisation and masking At the outset of the study, we held a public orientation meeting in Faridpur where we obtained community consent and, using stratified randomisation, the 96 villages were randomly allocated (1:1:1) to the mHealth intervention, the community mobilisation (PLA) intervention, or control, with each upazilla constituting one stratum.17 The name of each village was written on pieces of paper, colour-coded by upazilla, which when folded were indistinguishable from each other. For each upazilla, the 24 folded pieces of paper were placed in a bottle and then drawn by community leaders and representatives at the public orientation meeting. The first eight villages per upazilla drawn from the bottle were allocated to Arm A, the next eight villages to Arm B, and the final eight villages to Arm C. After all 96 villages had been allocated (32 to each trial arm), each of the three arms were randomly assigned to either the mHealth intervention, community mobilisation (PLA) intervention, or the control group by simultaneous drawing of arm letter and intervention allocation from two separate bottles. Because of the nature of the interventions being tested, the intervention team could not be masked to allocation. The data collection team was masked to allocation at the cluster and individual level during the baseline survey. Primary outcome analysis was done masked to allocation.

llocation from two separate bottles. Because of the nature of the interventions being tested, the intervention team could not be masked to allocation. The data collection team was masked to allocation at the cluster and individual level during the baseline survey. Primary outcome analysis was done masked to allocation. Procedures The villages allocated to the control group received usual care, which in this context is care seeking in government or private facilities (which is often associated with out-of-pocket payment for blood glucose testing, consultations, and treatments), and little or no preventative public health campaigning. The mHealth intervention consisted of twice-weekly health behaviour and awareness-raising voice messages sent to participants' mobile phones over a period of 14 months. Message content included information on signs, symptoms, prevention, and care for type 2 diabetes, and provided examples of strategies to reduce the risk of type 2 diabetes and its complications (appendix). Message content was informed by formative research and behaviour change theories,18, 19 and was reviewed by medical experts. Messages were about 1 min duration and had various formats, including mini-dramas, dialogues, and songs. The intervention was available to anyone with access to a mobile phone in the intervention areas (>95%) who volunteered their mobile phone number to community recruiters at the beginning of the intervention period or at 3 months into the intervention period.

various formats, including mini-dramas, dialogues, and songs. The intervention was available to anyone with access to a mobile phone in the intervention areas (>95%) who volunteered their mobile phone number to community recruiters at the beginning of the intervention period or at 3 months into the intervention period. The PLA intervention entailed monthly group meetings, with an average of 27 members per group, led by a lay facilitator who guided participants through a four-phase PLA cycle focused on type 2 diabetes prevention and control. Groups were open to all community members, and people with type 2 diabetes or who were deemed to be at high risk of non-communicable diseases were particularly encouraged to attend. Through the PLA cycle, community members identified behavioural, social, and environmental threats to their health and barriers to healthy lifestyles, prioritised these, and then planned, implemented, and evaluated strategies to address these threats. Awareness raising, exercising in groups, local coordination of blood sugar testing, income generation, and kitchen gardening to increase access to healthy food were popular strategies. Facilitators were locally recruited men and women who had completed higher secondary certificate level of education and were recruited by the study team following assessment of their communication skills, motivation, and familiarity with the study areas. Facilitators received a total of 14 days' training about PLA and community entry, group facilitation, and the basics of type 2 diabetes symptoms, prevention, and control. Each facilitator was responsible for running six to nine PLA groups each month. In addition, an equal number of men's and women's groups were established within each village, with a total of 122 groups facilitated by 16 facilitators (eight men, eight women) across 32 villages. Joint meetings of men's and women's groups were encouraged after phase 1 of the PLA cycle (ie, after identification and prioritisation of health determinants).

and women's groups were established within each village, with a total of 122 groups facilitated by 16 facilitators (eight men, eight women) across 32 villages. Joint meetings of men's and women's groups were encouraged after phase 1 of the PLA cycle (ie, after identification and prioritisation of health determinants). Training of informal health workers in the prevention and control of type 2 diabetes was done by the Diabetic Association of Bangladesh across all intervention and control villages during the intervention period. Project mapping of services in the study areas identified all informal care providers (eg, village doctors, pharmacy owners), who were then invited to participate in service-strengthening activities on a voluntary basis. This service strengthening included day-long workshops and provision of guidelines to cadres of largely unregulated care providers who had not received formal accredited training but might have had some degree of informal training through apprentices, workshops, or seminars. These informal care providers are typically the first point of care in health seeking by individuals in rural Bangladesh.20

nes to cadres of largely unregulated care providers who had not received formal accredited training but might have had some degree of informal training through apprentices, workshops, or seminars. These informal care providers are typically the first point of care in health seeking by individuals in rural Bangladesh.20 A sampling frame of all permanent residents aged 30 years and older was developed from a household census done between Aug 21, and Oct 28, 2015. 143 households with at least one eligible resident were then selected from each village by use of probability proportional to size sampling. A single eligible adult was selected from each of the 143 households for inclusion in the survey via simple random sampling. A baseline cross-sectional survey among the sampled individuals to obtain sociodemographic characteristics, behaviours, and knowledge of type 2 diabetes was done between Jan 23, and May 30, 2016. The survey included an overnight fasting blood glucose measurement in whole capillary blood obtained by finger prick in the middle or ring finger. All individuals without diagnosed type 2 diabetes then received a 75 g glucose load dissolved in 250 mL water. A 2-h post-prandial repeat capillary blood test was done to determine glucose tolerance status and differentiate between individuals with intermediate hyperglycaemia (defined as impaired fasting glucose or impaired glucose tolerance) and those with type 2 diabetes, based on WHO criteria (appendix).21 These baseline data were used to identify an intermediate hyperglycaemia cohort and to compare sociodemographic characteristics between the three trial groups.

ermediate hyperglycaemia (defined as impaired fasting glucose or impaired glucose tolerance) and those with type 2 diabetes, based on WHO criteria (appendix).21 These baseline data were used to identify an intermediate hyperglycaemia cohort and to compare sociodemographic characteristics between the three trial groups. Following intervention, the sampling frame was updated and a new random sample of adults aged 30 years and older was selected via the same sampling method as used at baseline. By chance, approximately 25% of individuals sampled for the end-of-study survey had also been included in the baseline survey. In addition, all individuals identified with intermediate hyperglycaemia at baseline were followed-up in the end-of-study survey to measure type 2 diabetes incidence in this cohort. An end-of-study survey of sociodemographic data, knowledge and behaviours, and anthropometric measures of weight, height, blood pressure, and fasting and 2 h post-prandial blood glucose measures was completed in the random cross-sectional sample and the baseline intermediate hyperglycaemia cohort between Jan 16, and April 30, 2018, using the same methods as at baseline.

wledge and behaviours, and anthropometric measures of weight, height, blood pressure, and fasting and 2 h post-prandial blood glucose measures was completed in the random cross-sectional sample and the baseline intermediate hyperglycaemia cohort between Jan 16, and April 30, 2018, using the same methods as at baseline. Data were collected by 16 pairs of fieldworkers (one man and one woman in each pair) with at least secondary education, who underwent extensive training in survey methods, including supervised field practice. Most data collection took place in testing centres established by the field team for the purposes of the study, with additional data collection with pretested questionnaires taking place at respondents' homes.17, 22 Data collectors were supervised by four field supervisors responsible for observing and verifying data. Data quality-control measures were implemented within the direct digital data capture system used (eg, range and consistency checks), through repeat measures by supervisors on a random basis, and where outlier data were detected on data inspection. Data collectors, supervisors, and managers were unaware of randomisation assignments at baseline, but might have been able to deduce assignment during data collection at the end of the study. Access to end-of-study data was restricted to the monitoring and evaluation managers until collection was complete, at which point the data were available for masked analysis by the lead author (EF).

umption (table 4). Overall quality-of-life score did not differ between study groups, and the crude significant difference in self-rated health measured on a scale of 0 (worst health) to 100 (best health) between the PLA and control groups was attenuated and non-significant when adjusted for household wealth (table 4). Prespecified sensitivity analyses including adjustment for a possible screening effect among individuals included in the baseline and end-of-study surveys and multilevel multiple imputations for missing blood glucose measurements had negligible effect on the primary analyses (appendix). In view of the findings for the PLA intervention, we did additional post-hoc sensitivity analyses. First, we assessed potential sampling error and imbalances in key sociodemographic characteristics between study groups by running a predictive regression of our primary outcomes on sociodemographic characteristics at baseline and then checked whether the significant variables in this predictive regression differed between treatment groups. Apart from wealth quintiles, which we had prespecified as a covariate in our model, no other sociodemographic measures were significantly different between treatment groups, lending support to our primary findings (data not shown) and suggesting that our random sampling approach was effective and consistent across study groups.

n assignments at baseline, but might have been able to deduce assignment during data collection at the end of the study. Access to end-of-study data was restricted to the monitoring and evaluation managers until collection was complete, at which point the data were available for masked analysis by the lead author (EF). Process evaluation data will be reported in detail elsewhere. We collected data in all four intervention upazillas of Faridpur district, including small group discussions with men and women attendees, with small groups of men and women with type 2 diabetes, and with group non-attenders. Later we met groups of men and women to explore triangulation and seek consensus on community changes. We also met with men's and women's group facilitators. In some of the groups we used participatory photography where they had identified and represented issues of importance to them using mobile phone cameras. Focus group discussions were digitally recorded and one author (KAk) made notes about the findings in English and translated field observation notes to English for analysis. Key themes around individual, household, and community change were compared with the theory of change drafted after the formative phase of research. Outcomes We prespecified two primary outcomes: the prevalence of intermediate hyperglycaemia and type 2 diabetes at the end of the study and the 2-year cumulative incidence of type 2 diabetes among the cohort with baseline intermediate hyperglycaemia.

Process evaluation data will be reported in detail elsewhere. We collected data in all four intervention upazillas of Faridpur district, including small group discussions with men and women attendees, with small groups of men and women with type 2 diabetes, and with group non-attenders. Later we met groups of men and women to explore triangulation and seek consensus on community changes. We also met with men's and women's group facilitators. In some of the groups we used participatory photography where they had identified and represented issues of importance to them using mobile phone cameras. Focus group discussions were digitally recorded and one author (KAk) made notes about the findings in English and translated field observation notes to English for analysis. Key themes around individual, household, and community change were compared with the theory of change drafted after the formative phase of research. Outcomes We prespecified two primary outcomes: the prevalence of intermediate hyperglycaemia and type 2 diabetes at the end of the study and the 2-year cumulative incidence of type 2 diabetes among the cohort with baseline intermediate hyperglycaemia. Secondary outcomes were mean diastolic and systolic blood pressure, prevalence of hypertension, hypertension control (among those with known hypertension), BMI, prevalence of overweight and obesity and abdominal obesity (waist to hip ratio >0·9 for men and >0·85 for women), health related quality of life (using EQ-5D score), physical activity, fruit and vegetable consumption, and knowledge of the causes, symptoms, complications, prevention and control of type 2 diabetes. Additional secondary outcomes among people with type 2 diabetes were self-awareness of diabetes status and, among those with known diabetes, prevalence of diabetes control, psychological distress (with SRQ-20 screening tool), and receipt of professional medical treatment or advice for diabetes. Full specification of secondary outcomes and methods of assessment have been described previously.17

areness of diabetes status and, among those with known diabetes, prevalence of diabetes control, psychological distress (with SRQ-20 screening tool), and receipt of professional medical treatment or advice for diabetes. Full specification of secondary outcomes and methods of assessment have been described previously.17 Additional pre-specified outcomes were intervention costs, incremental cost-effectiveness ratios and costs per disability adjusted life years (DALYs) averted for any effective intervention, and process indicators of intervention coverage and qualitative assessments of behaviour change caused by the interventions. An additional outcome of diabetes-only prevalence (excluding intermediate hyperglycaemia) was included post-hoc. Key findings from the qualitative, process assessments of behaviour change are sumarised below, but full analyses are to be reported in a forthcoming publication. Statistical analysis We estimated that a target sample of 143 adults per village (total 13 728, including 10% oversample for non-response) would provide 80% power at 95% confidence to detect a minimum 21·5% reduction in combined prevalence of type 2 diabetes and intermediate hyperglycaemia and 78% power to detect a 33% reduction in cumulative incidence of type 2 diabetes among the baseline intermediate hyperglycaemia cohort in intervention clusters relative to control clusters.17

t 95% confidence to detect a minimum 21·5% reduction in combined prevalence of type 2 diabetes and intermediate hyperglycaemia and 78% power to detect a 33% reduction in cumulative incidence of type 2 diabetes among the baseline intermediate hyperglycaemia cohort in intervention clusters relative to control clusters.17 Analysis of intervention effect on the combined prevalence of type 2 diabetes and intermediate hyperglycaemia included all individuals who provided blood glucose measurements in the end-of-study sample survey. This included individuals who provided a random blood glucose measure on the basis of self-reported diagnosis of type 2 diabetes by a medical professional. For the cumulative incidence of type 2 diabetes among the intermediate hyperglycaemia cohort, the analysis population was all individuals for whom a baseline blood glucose measurement revealed intermediate hyperglycaemia and for whom an end-of-study blood glucose measurement was taken.

abetes by a medical professional. For the cumulative incidence of type 2 diabetes among the intermediate hyperglycaemia cohort, the analysis population was all individuals for whom a baseline blood glucose measurement revealed intermediate hyperglycaemia and for whom an end-of-study blood glucose measurement was taken. All analyses were done on an intention-to-treat basis. Comparison between interventions relative to control used random-effects logistic regression, allowing for the stratified and clustered nature of the data. An additive model was used to estimate absolute differences for the primary outcomes. Baseline data were examined by the research team for imbalance between anonymised randomised villages and it was agreed with the data monitoring committee that an apparent marginal difference in baseline household wealth quintiles (derived from principal component analysis of household assets) between study groups would be adjusted for in end-of-study multivariate analyses. Prespecified sensitivity analysis of primary outcomes assessed effects of missing data using multiple imputation and screening effects of individuals being included in the baseline and end-of-study surveys. Post-hoc sensitivity analysis assessed sampling error, enumerator bias, and blood glucose measurement bias (ie, by running analysis on continuous blood glucose measurements and by separately applying different arbitrary fasting blood glucose cut-offs of 5·5 mmol/L, 6·3 mmol/L, and 7·8 mmol/L and 2-h blood glucose cut-offs of 6·8 mmol/L and 10·4 mmol/L for classifications of intermediate hyperglycaemia or diabetes). We also restricted primary outcome analysis to individuals with normal blood glucose levels at baseline who also happened to be included in the end-of-study survey. In view of the high prevalence of intermediate hyperglycaemia and type 2 diabetes and following recommendation from reviewers, we calculated a post-hoc estimate of PLA effect size compared to control as relative risk by use of log binomial models estimated by generalised estimating equations with robust SEs to account for clustering and then calculated the number needed to treat (NNT). Finally, in view of the clinical relevance of type 2 diabetes as an outcome in its own right (ie, not combined with intermediate hyperglycaemia), we did a post-hoc analysis in which we assessed intervention effects on a diabetes-only outcome.

account for clustering and then calculated the number needed to treat (NNT). Finally, in view of the clinical relevance of type 2 diabetes as an outcome in its own right (ie, not combined with intermediate hyperglycaemia), we did a post-hoc analysis in which we assessed intervention effects on a diabetes-only outcome. Primary analysis was done by the trial principal investigator (EF), who was masked to treatment allocation and who reported the results to the data monitoring committee and chair of the trial steering committee, after which the identities of the trial groups were revealed, and analysis continued unmasked. Prespecified secondary outcomes analyses were based on complete data only. Comparative analysis used random-effects logistic regression for binary outcomes and mixed-effects linear regression for continuous outcomes, each allowing for clustering and upazilla stratification. Continuous outcome measures with a skewed distribution were log-transformed before regression analysis. All quantitative analyses were done using STATA/SE version 15.1.

c regression for binary outcomes and mixed-effects linear regression for continuous outcomes, each allowing for clustering and upazilla stratification. Continuous outcome measures with a skewed distribution were log-transformed before regression analysis. All quantitative analyses were done using STATA/SE version 15.1. Intervention implementation and coverage was estimated from process evaluation data. Total cost and cost-effectiveness analysis of the DMagic interventions was done initially from a provider (health system) perspective, including costs to the programme provider and public health-care providers (ie, costs associated with increasing service demand and utilisation). Incremental cost-effectiveness ratios were calculated in terms of cost per case of intermediate hyperglycaemia and type 2 diabetes prevented in the general population sample, cost per case of type 2 diabetes prevented among the high-risk cohort with intermediate hyperglycaemia at baseline, and costs per DALY averted. All costs were adjusted for inflation, discounted at 3% per year, and converted to 2018 international dollars (INT$). We calculated DALYs averted using the Global Burden of Disease study approach (appendix);23, 24 a detailed description of the economic evaluation methodology is presented elsewhere.25 Since our interventions and evaluation of them were population based (as opposed to an individual-level study) and, for the most part, we did not follow up the same individuals from baseline to end of study, we were not able to estimate quality-adjusted life years (QALYs) gained.

n methodology is presented elsewhere.25 Since our interventions and evaluation of them were population based (as opposed to an individual-level study) and, for the most part, we did not follow up the same individuals from baseline to end of study, we were not able to estimate quality-adjusted life years (QALYs) gained. This trial is registered with the ISRCTN registry, number ISRCTN41083256, and is completed. The trial was registered on March 30, 2016, which was after the initial phase of the study but before any intervention delivery began. In view of the study design, it was not feasible to register the trial before this date, since cluster selection was purposeful (ie, based on population size and non-contiguous borders with other clusters) and identification of clusters was integral to project design and planning. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing 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.

This trial is registered with the ISRCTN registry, number ISRCTN41083256, and is completed. The trial was registered on March 30, 2016, which was after the initial phase of the study but before any intervention delivery began. In view of the study design, it was not feasible to register the trial before this date, since cluster selection was purposeful (ie, based on population size and non-contiguous borders with other clusters) and identification of clusters was integral to project design and planning. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing 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 All 96 villages agreed to participate and were randomly assigned, 32 to each of the three trial groups. Survey and/or anthropometric baseline data were collected from 12280 (89·5%) of 13 684 individuals between Jan 23, and May 30, 2016. The target sample was slightly smaller than expected because two villages only had 128 and 114 eligible individuals living in separate households. From the baseline survey, 2470 individuals were identified with intermediate hyperglycaemia. Baseline characteristics were similar among all study groups (table 1, appendix), apart from a small difference in household wealth (higher in PLA villages), which was later adjusted for in multivariate analyses.Table 1 Baseline characteristics

survey, 2470 individuals were identified with intermediate hyperglycaemia. Baseline characteristics were similar among all study groups (table 1, appendix), apart from a small difference in household wealth (higher in PLA villages), which was later adjusted for in multivariate analyses.Table 1 Baseline characteristics Control mHealth PLA Cluster level Villages (clusters) 32 32 32 Mean village population aged ≥30 years (SD) 521 (189) 551 (152) 548 (225) Mean number of households (SD) 269 (97) 282 (79) 285 (112) Individual level Total participants who completed survey 4048 4071 4021 Age, years 30–39 1391 (34%) 1329 (33%) 1388 (35%) 40–49 991 (24%) 1068 (26%) 992 (25%) 50–59 767 (19%) 765 (19%) 761 (19%) 60–69 648 (16%) 659 (16%) 610 (15%) 70–100 251 (6%) 250 (6%) 270 (7%) Sex Men 1950 (48%) 1845 (45%) 1889 (47%) Women 2098 (52%) 2226 (55%) 2132 (53%) Education None 2116 (52%) 1950 (48%) 1905 (47%) Primary 921 (23%) 852 (21%) 1004 (25%) Secondary 989 (24%) 1231 (30%) 1086 (27%) Tertiary 22 (1%) 38 (1%) 26 (1%) Literacy Literate 1455 (36%) 1649 (41%) 1561 (39%) Illiterate 2593 (64%) 2422 (59%) 2460 (61%) Marital status Married 3528 (87%) 3574 (88%) 3530 (88%) Not married 520 (13%) 497 (12%) 491 (12%) Religion Muslim 3660 (90%) 3674 (90%) 3666 (91%) Other 388 (10%) 397 (10%) 355 (9%) Occupation Not working 2216 (55%) 2323 (57%) 2164 (54%) Manual labour 1362 (34%) 1296 (32%) 1376 (34%) Non-manual labour 468 (12%) 452 (11%) 481 (12%) Missing data 2 (<1%) 0 0 Wealth quintile Most poor 845 (21%) 910 (22%) 676 (17%) Very poor 781 (19%) 896 (22%) 768 (19%) Poor 822 (20%) 821 (20%) 798 (20%) Less poor 855 (21%) 722 (18%) 826 (21%) Least poor 745 (18%) 722 (18%) 953 (24%) Total participants with baseline data for glycaemic status* 4070 4063 4054 Glycaemic status Normal 2860 (70%) 2816 (69%) 2801 (69%) Impaired fasting glucose† 200 (5%) 186 (5%) 203 (5%) Impaired glucose tolerance† 632 (16%) 655 (16%) 594 (15%) Diabetes‡ 378 (9%) 406 (10%) 456 (11%) PLA=participatory learning and action.

Total participants with baseline data for glycaemic status* 4070 4063 4054 Glycaemic status Normal 2860 (70%) 2816 (69%) 2801 (69%) Impaired fasting glucose† 200 (5%) 186 (5%) 203 (5%) Impaired glucose tolerance† 632 (16%) 655 (16%) 594 (15%) Diabetes‡ 378 (9%) 406 (10%) 456 (11%) PLA=participatory learning and action. * Glycaemic status missing for 38 participants in the control group, 30 participants in the mHealth group, and 25 participants in the PLA group (excluded from totals). † Intermediate hyperglycaemia was defined as impaired fasting glucose or impaired glucose tolerance. ‡ Based on blood glucose measurement or self-reported previous medical diagnosis.21 End-of-study data were collected from 11 454 (83·7%) of 13 687 individuals between Jan 16, and April 30, 2018. The end-of-study cross-sectional sample was similar in terms of sociodemographic characteristics to the baseline sample and between study groups (appendix). Of the 2470 individuals with intermediate hyperglycaemia at baseline, 2100 were followed-up (85%; figure 1). Non-responders were more likely to be men (1712 [23%] of 7520 men vs 721 [9%] of 7854 women; p<0·0001; figure 1) and men non-responders were younger than men responders (mean difference 2·0 years; p<0·0001), whereas women non-responders were slightly older than women responders (mean difference 3·1 years; p<0·0001). The same patterns of non-response were seen across all study groups.Figure 1 Trial profile

men; p<0·0001; figure 1) and men non-responders were younger than men responders (mean difference 2·0 years; p<0·0001), whereas women non-responders were slightly older than women responders (mean difference 3·1 years; p<0·0001). The same patterns of non-response were seen across all study groups.Figure 1 Trial profile Anthropometry includes all physical measures—ie, blood glucose, blood pressure, weight, height, and waist and hip circumferences. PLA=participatory learning and action. *Other reasons were physical or mental disability or severe illness preventing participation, incarceration, or temporary residence only. Exposure to the interventions was fairly high (table 2). In the mHealth intervention group, 2613 (61%) of 4320 villagers reported ever receiving a mobile phone message; in the PLA group, 3136 (74%) of 4247 reported ever participating in a group.Table 2 Intervention coverage process indicators PLA (n=4247) mHealth (n=4320) Control (n=4292) Ever participated in PLA community group 3136 (74%) 1 (<1%) 1 (<1%) Knows someone who attended PLA community group 3046 (72%) 10 (<1%) 0 (<1%) Ever received mHealth message 55 (1%) 2613 (60%) 7 (<1%) Knows someone who received mHealth messages 32 (1%) 2845 (66%) 6 (<1%) Data are n (%). PLA=participatory learning and action.

participated in PLA community group 3136 (74%) 1 (<1%) 1 (<1%) Knows someone who attended PLA community group 3046 (72%) 10 (<1%) 0 (<1%) Ever received mHealth message 55 (1%) 2613 (60%) 7 (<1%) Knows someone who received mHealth messages 32 (1%) 2845 (66%) 6 (<1%) Data are n (%). PLA=participatory learning and action. At the end-of-study survey, the combined prevalence of type 2 diabetes and intermediate hyperglycaemia was significantly lower in the PLA group compared with the control group (table 3, figure 2). The 2-year cumulative incidence of type 2 diabetes among the intermediate hyperglycaemia cohort in the PLA group compared with the control group was also significantly reduced (table 3, figure 2). Our post-hoc analysis of the effect on type 2 diabetes prevalence only (ie, not combined with intermediate hyperglycaemia) showed that the PLA intervention reduced the prevalence of type 2 diabetes by 48% relative to control (305 [8%] of 3757 vs 493 [13%] of 3821; adjusted [stratified, clustered design, and wealth] odds ratio [aOR] 0·52, 0·38–0·71; p<0·0001). There was no evidence of an effect of the mHealth intervention on the combined prevalence of type 2 diabetes and intermediate hyperglycaemia or the incidence of type 2 diabetes in the intermediate hyperglycaemia cohort compared with control villages (table 3, figure 2). There was also no evidence of an effect of the mHealth intervention on the prevalence of type 2 diabetes alone in our post-hoc analysis (563 [15%] of 3797 vs 493 [13%] of 3821; adjusted [stratified, clustered design, and wealth] aOR 1·18, 0·95–1·48; p=0·139).Table 3 Primary outcome measures at end of study

, figure 2). There was also no evidence of an effect of the mHealth intervention on the prevalence of type 2 diabetes alone in our post-hoc analysis (563 [15%] of 3797 vs 493 [13%] of 3821; adjusted [stratified, clustered design, and wealth] aOR 1·18, 0·95–1·48; p=0·139).Table 3 Primary outcome measures at end of study Control mHealth PLA Population prevalence of intermediate hyperglycaemia and diabetes*† Total population 3821 3797 3757 Normoglycaemic 1960 (51·3%) 2003 (52·8%) 2686 (71·5%) Diabetes or intermediate hyperglycaemic 1861 (48·7%) 1794 (47·2%) 1071 (28·5%) Relative difference (odds ratio [95% CI]) Unadjusted (allowing for stratified clustered design) Reference 0·94 (0·75–1·18; p=0·611) 0·37 (0·27–0·49; p<0·0001) Adjusted for household wealth quintile and allowing for stratified clustered design Reference 0·93 (0·74–1·16; p=0·513) 0·36 (0·27–0·48; p<0·0001) Absolute risk difference, % points (95% CI) Unadjusted (allowing for stratified clustered design Reference –1·4 (−6·9 to 4·1; p=0·626) –20·1 (−26·1 to −14·0; p<0·0001) Adjusted for household wealth quintile and allowing for stratified clustered design Reference –1·7 (−7·3 to 3·9; p=0·542) –20·7 (−26·7 to −14·6; p<0·0001) 2 year cumulative incidence among intermediate hyperglycaemic cohort† Total population 712 717 665 Normoglycaemic 249 (35·0%) 280 (39·1%) 407 (61·2%) Intermediate hyperglycaemic 337 (47·3%) 315 (43·9%) 199 (29·9%) Diabetes 126 (17·7%) 122 (17·0%) 59 (8·9%) Relative difference (odds ratio [95% CI]) Unadjusted (allowing for stratified clustered design) Reference 0·99 (0·70–1·39; p=0·941) 0·41 (0·24–0·68; p=0·0005) Adjusted for household wealth quintile and allowing for stratified clustered design Reference 1·02 (0·73–1·43; p=0·912) 0·39 (0·24–0·65; p=0·0003) Absolute risk difference, % points (95% CI) Unadjusted (allowing for stratified clustered design) Reference –0·04 (−5·3 to 5·3; p=0·987) –8·4 (−13·8 to −3·0; p=0·0023) Adjusted for household wealth quintile and allowing for stratified clustered design Reference 0·36 (−4·7 to 5·5; p=0·889) –8·7 (−14·0 to −3·5; p=0·0011) Data are n or n (%), unless otherwise indicated. All p-value comparisons are versus control. PLA=participatory learning and action.

987) –8·4 (−13·8 to −3·0; p=0·0023) Adjusted for household wealth quintile and allowing for stratified clustered design Reference 0·36 (−4·7 to 5·5; p=0·889) –8·7 (−14·0 to −3·5; p=0·0011) Data are n or n (%), unless otherwise indicated. All p-value comparisons are versus control. PLA=participatory learning and action. * Coefficient of variation for diabetes and intermediate hyperglycaemia is 0·346. † Anthropometry participants with missing blood glucose data: eight in the control group, 15 in the mHealth group, and 41 in the PLA group. ‡ Anthropometry participants with missing blood glucose data: one in the control group, two in the mHealth group, and three in the PLA group. Figure 2 Allocation box-plots showing (A) a cluster-level summary of intermediate hyperglycaemia and diabetes at end of study and (B) cluster-level 2 year cumulative incidence of diabetes among the hyperglycaemic cohort Box-plot shows median, minimum and maximum values and interquartile range.

‡ Anthropometry participants with missing blood glucose data: one in the control group, two in the mHealth group, and three in the PLA group. Figure 2 Allocation box-plots showing (A) a cluster-level summary of intermediate hyperglycaemia and diabetes at end of study and (B) cluster-level 2 year cumulative incidence of diabetes among the hyperglycaemic cohort Box-plot shows median, minimum and maximum values and interquartile range. Increases in ability to report one or more valid causes, symptoms, complications, and strategies for prevention and control of diabetes were observed in both intervention groups compared with control, with the effect consistently greatest in the PLA group (table 4). Self-awareness of diabetes status among individuals identified as having type 2 diabetes by blood glucose testing was five-times higher in the PLA group than in the control group (aOR 5·09, 2·95–8·79; p<0·0001), but the smaller increase in the mHealth group compared with control was not significant (aOR 1·44, 0·94–2·19; p=0·092). Improvements in diabetes control among people with known diabetes in the intervention groups compared with control were not significant and there was no intervention effect on receipt of professional care for diabetes among those aware of their disease (table 4).Table 4 Secondary outcome measures at end of study

·092). Improvements in diabetes control among people with known diabetes in the intervention groups compared with control were not significant and there was no intervention effect on receipt of professional care for diabetes among those aware of their disease (table 4).Table 4 Secondary outcome measures at end of study Study group Unadjusted coefficient or odds ratio* Adjusted coefficient or odds ratio† PLA (n=3798) mHealth (n=3812) Control (n=3829) PLA vs control mHealth vs control PLA vs control mHealth vs control Mean diastolic blood pressure (mm Hg) 73·8 (11·2) 72·9 (11·2) 73·9 (11·2) –0·25 (−1·78 to 1·28; p=0·748) –1·12 (−2·59 to 0·35; p=0·135) –0·50 (−2·05 to 1·05; p=0·526) –1·18 (−2·58 to 0·22; p=0·098) Mean systolic blood pressure (mm Hg) 125·3 (19·6) 124·4 (20·0) 125·5 (20·8) –0·19 (−2·54 to 2·17; p=0·877) –1·05 (−3·46 to 1·35; p=0·391) –0·53 (−2·90 to 1·83; p=0·661) –1·10 (−3·44 to 1·25; p=0·360) Hypertension 958 (25%) 854 (22%) 899 (23%) 1·08 (0·86 to 1·35; p=0·516) 0·94 (0·78 to 1·14; p=0·517) 1·02 (0·81 to 1·29; p=0·837) 0·93 (0·77 to 1·14; p=0·483) Hypertension control (among those with known hypertension) 131/332 (39%) 122/313 (39%) 90/258 (35%) 1·21 (0·79 to 1·87; p=0·375) 1·12 (0·79 to 1·59; p=0·523) 1·21 (0·78 to 1·87; p=0·389) 1·13 (0·80 to 1·60; p=0·501) Mean BMI (kg/m2) 22·0 (3·6) 21·9 (3·6) 21·9 (3·6) 0·08 (−0·21 to 0·38; p=0·572) –0·05 (−0·28 to 0·17; p=0·631) –0·09 (−0·39 to 0·20; p=0·543) –0·09 (−0·29 to 0·12; p=0·410) Overweight or obese (waist:hip ratio >0·9 for men and >0·85 for women) 1341 (35%) 1265 (33%) 1326 (35%) 1·03 (0·88 to 1·20; p=0·720) 0·94 (0·83 to 1·06; p=0·286) 0·94 (0·80 to 1·11; p=0·459) 0·92 (0·82 to 1·03; p=0·151) Abdominal obesity 1964 (52%) 1952 (51%) 1922 (50%) 1·05 (0·74 to 1·48; p=0·791) 1·03 (0·73 to 1·46; p=0·851) 0·98 (0·70 to 1·37; p=0·914) 1·02 (0·72 to 1·44; p=0·916) Median EQ-5D‡§ 0·80 (0·73–1·0) 0·85 (0·73–1·0) 0·80 (0·73–1·0) 0·03 (−0·02 to 0·08; p=0·253) –0·01 (−0·06 to 0·04; p=0·687) 0·03 (−0·02 to 0·07; p=0·261) –0·02 (−0·06 to 0·03; p=0·471) Mean self-rated health§ 77·9 (15·5) 76·7 (16·0) 74·5 (16·1) 3·28 (0·37 to 6·19; p=0·027) 2·12 (−0·78 to 5·02; p=0·153) 2·76 (−0·16 to 5·68; p=0·064) 1·96 (−0·97 to 4·88; p=0·189) Median SRQ-20 score among adults aged ≥30 years with self-reported diabetes 8 (5–12); n=137 6 (4–9); n=144 8 (5–13); n=94 –0·004 (−0·20 to 0·19; p=0·971) –0·17 (−0·36 to 0·02; p=0·085) 0·006 (−0·19 to 0·20; p=0·949) –0·16 (−0·35 to 0·03; p=0·105) Ability to report one or more valid cause of diabetes§

to 4·88; p=0·189) Median SRQ-20 score among adults aged ≥30 years with self-reported diabetes 8 (5–12); n=137 6 (4–9); n=144 8 (5–13); n=94 –0·004 (−0·20 to 0·19; p=0·971) –0·17 (−0·36 to 0·02; p=0·085) 0·006 (−0·19 to 0·20; p=0·949) –0·16 (−0·35 to 0·03; p=0·105) Ability to report one or more valid cause of diabetes§ 3646 (96%) 2981 (78%) 2153 (57%) 36·7 (18·2 to 73·7; p<0·0001) 3·72 (2·06 to 6·73; p<0·0001) 35·7 (17·7 to 71·9; p<0·0001) 3·77 (2·05 to 6·91; p<0·0001) Ability to report one or more valid symptom of diabetes§ 3659 (96%) 3205 (84%) 2452 (65%) 25·1 (11·8 to 53·2; p<0·0001) 4·34 (2·07 to 9·10; p=0·0001) 24·0 (11·3 to 50·9; p<0·0001) 4·37 (2·07 to 9·24; p=0·0001) Ability to report one or more valid complication of diabetes§ 3649 (96%) 3084 (81%) 2161 (57%) 36·5 (18·5 to 72·0; p<0·0001) 5·29 (2·58 to 10·9; p<0·0001) 35·4 (17·8 to 70·4; p<0·0001) 5·42 (2·60 to 11·3; p<0·0001) Ability to recognise one or more valid complication of diabetes when prompted§ 3723 (98%) 3367 (89%) 2932 (77%) 19·7 (8·24 to 46·9; p<0·0001) 3·87 (1·49 to 10·1; p=0·0056) 18·3 (7·66 to 43·9; p<0·0001) 3·88 (1·47 to 10·2; p=0·0063) Ability to report one or more valid way to prevent diabetes§ 3608 (95%) 3294 (87%) 2626 (69%) 10·6 (5·74 to 19·4; p<0·0001) 4·28 (2·10 to 8·68; p=0·0001) 10·0 (5·44 to 18·5; p<0·0001) 4·31 (2·10 to 8·85; p=0·0001) Ability to report one or more valid way to control diabetes§ 3619 (95%) 3354 (88%) 2796 (74%) 8·81 (4·64 to 16·7; p<0·0001) 3·94 (1·92 to 8·08; p=0·0002) 8·36 (4·42 to 15·8; p<0·0001) 3·93 (1·90 to 8·12; p=0·0002) Diabetes control among those with known of diabetes 68/123 (55%) 57/134 (43%) 33/86 (38%) 2·17 (0·97 to 4·87; p=0·060) 1·22 (0·51 to 2·94; p=0·657) 2·24 (0·97 to 5·16; p=0·058) 1·21 (0·50 to 2·92; p=0·666) Self-reported awareness of diabetes status among all those identified as having diabetes by objective blood glucose test 143/342 (42%) 156/652 (24%) 100/580 (17%) 5·10 (2·97 to 8·76; p<0·0001) 1·42 (0·91 to 2·20; p=0·119) 5·09 (2·95 to 8·79; p<0·0001) 1·44 (0·94 to 2·19; p=0·092) Receipt of professional treatment or advice for diabetes among those with diabetes and aware of their status 114/143 (80%) 128/156 (82%) 84/100 (84%) 0·85 (0·35 to 2·02; p=0·708) 0·97 (0·43 to 2·19; p=0·947) 0·78 (0·32 to 1·92; p=0·584) 0·93 (0·41 to 2·13; p=0·873) Average ≥150 min physical activity per week§ 2844 (75%) 2891 (76%) 2923 (77%) 0·84 (0·54 to 1·30; p=0·435) 1·00 (0·63 to 1·59; p=0·993) 0·83 (0·53 to 1·30; p=0·418) 0·98 (0·62 to 1·57; p=0·945) Mean number of

(0·35 to 2·02; p=0·708) 0·97 (0·43 to 2·19; p=0·947) 0·78 (0·32 to 1·92; p=0·584) 0·93 (0·41 to 2·13; p=0·873) Average ≥150 min physical activity per week§ 2844 (75%) 2891 (76%) 2923 (77%) 0·84 (0·54 to 1·30; p=0·435) 1·00 (0·63 to 1·59; p=0·993) 0·83 (0·53 to 1·30; p=0·418) 0·98 (0·62 to 1·57; p=0·945) Mean number of portions of fruit and/ or vegetables consumed per day¶ 4·0 (2·3) 3·4 (1·6) 3·6 (1·6) 0·35 (−0·11 to 0·81; p=0·133) –0·18 (−0·54 to 0·18; p=0·335) 0·29 (−0·10 to 0·69; p=0·143) –0·19 (−0·53 to 0·15; p=0·274) Data are mean (SD), median (IQR), n/N (%), or n (%), unless otherwise indicated. Measures of effect are beta coefficients where the measure is continuous (eg, mean or median) and are odds ratios where the measure is %. PLA=participatory learning and action. SRQ-20=Self-Reporting Questionnaire 20-Item. * Accounting for stratified clustered design. † Adjusted for household wealth quintile and accounting for stratified, clustered design. ‡ EuroQol-5D (EQ-5D) using UK tariffs. § Survey totals PLA=3795; mHealth=3802; and control=3786. ¶ Survey totals PLA=3792; mHealth=3802; control=3777. There was no evidence of an effect of either intervention on blood pressure, overweight and obesity, or self-reported physical activity or fruit and vegetable consumption (table 4). Overall quality-of-life score did not differ between study groups, and the crude significant difference in self-rated health measured on a scale of 0 (worst health) to 100 (best health) between the PLA and control groups was attenuated and non-significant when adjusted for household wealth (table 4).

m wealth quintiles, which we had prespecified as a covariate in our model, no other sociodemographic measures were significantly different between treatment groups, lending support to our primary findings (data not shown) and suggesting that our random sampling approach was effective and consistent across study groups. We then explored potential blood glucose measurement biases and errors. We checked the effect of PLA on the continuous fasting and 2-h blood glucose measures, which showed significant reductions in mean population glucose compared with control (appendix). We also ran a series of regression models on the primary outcomes using different, arbitrary cut-off points for categorisation of intermediate hyperglycaemia and type 2 diabetes; again, our finding of a strong association of reduced prevalence and incidence with PLA exposure remained irrespective of cut-off values used (data not shown). Finally, to rule out potential enumerator effects (ie, potential bias in fieldworker measurements), we included enumerator identifiers as fixed effects in our regression of intervention exposure and fasting and 2 h blood glucose, treated as continuous outcomes. By doing so were able to show that our findings were robust to the inclusion of enumerator fixed effects (appendix) and that the enumerator fixed effects were statistically null (fasting blood glucose enumerator F-test=0·002, p=0·9642; 2 h blood glucose enumerator F-test=0·966, p=0·3257).

reated as continuous outcomes. By doing so were able to show that our findings were robust to the inclusion of enumerator fixed effects (appendix) and that the enumerator fixed effects were statistically null (fasting blood glucose enumerator F-test=0·002, p=0·9642; 2 h blood glucose enumerator F-test=0·966, p=0·3257). We also restricted our incidence analysis to 1457 individuals with normoglycaemic blood glucose measurements at baseline who happened to also be included in our end-of-study sample. Our random-effects regression findings showed a 72% reduction in the combined incidence of intermediate hyperglycaemia and type 2 diabetes among normoglycaemic individuals in PLA villages relative to control (OR 0·28, 0·19–0·42; p<0·0001), suggesting that PLA was also effective among a normoglycaemic population. For the combined prevalence outcome, the relative risk for PLA compared with control was 0·61 (95% CI 0·50–0·73), with an NNT of 4·75 (95% CI 3·71–6·61). For the 2 year cumulative incidence outcome, the relative risk was 0·49 (0·31–0·80), with an NNT of 10·68 (7·02–22·32).

We also restricted our incidence analysis to 1457 individuals with normoglycaemic blood glucose measurements at baseline who happened to also be included in our end-of-study sample. Our random-effects regression findings showed a 72% reduction in the combined incidence of intermediate hyperglycaemia and type 2 diabetes among normoglycaemic individuals in PLA villages relative to control (OR 0·28, 0·19–0·42; p<0·0001), suggesting that PLA was also effective among a normoglycaemic population. For the combined prevalence outcome, the relative risk for PLA compared with control was 0·61 (95% CI 0·50–0·73), with an NNT of 4·75 (95% CI 3·71–6·61). For the 2 year cumulative incidence outcome, the relative risk was 0·49 (0·31–0·80), with an NNT of 10·68 (7·02–22·32). Both interventions were delivered as per our published protocol, although the mHealth intervention started about 4 months later than originally planned and thus ran for just 14 months (instead of 18 months) due to technical and bureaucratic delays in establishing a system to deliver messages. The 122 PLA groups ran from June, 2016, to December, 2017, with a population coverage of one group per 342 people. In the PLA group, 3136 (74%) of 4247 people in the end-of-study survey population reported participating in group meetings (table 2). Repeat attendance at groups was very high, with about 90% of attenders participating in multiple meetings (data not shown). Roughly 9000 individuals (22% of the total population of around 41 667) provided their mobile phone numbers to receive the mHealth intervention and messages were delivered to about 7400 (82%) of these (around 51% of eligible adults). A total of 120 different messages were developed and sent between Oct 21, 2016 and Dec 24, 2017. An average of 60% of the first 26 messages were received; thereafter we improved the delivery systems, which increased receipt to 86% for the remaining 94 messages. More than 50% of 100 randomly selected message recipients who participated in a process evaluation survey reported sharing messages with others and two-thirds of the end-of-study population in mHealth villages received or knew someone who received the messages (table 2). Contamination between trial groups was minimal.

More than 50% of 100 randomly selected message recipients who participated in a process evaluation survey reported sharing messages with others and two-thirds of the end-of-study population in mHealth villages received or knew someone who received the messages (table 2). Contamination between trial groups was minimal. Details of our qualitative findings will be reported elsewhere. In summary, these data suggest that changes occurred at the individual, household, and community levels in response to the PLA intervention. As awareness of how to prevent type 2 diabetes increased, and individuals were motivated to change their diets, reporting reduced rice, oil, sugar, and salt consumption and increased consumption of a variety of vegetables. Kitchen gardens were encouraged by groups and increased access to vegetables. Both attenders and non-attenders reported an increase in intentional physical activity in PLA intervention areas, particularly among women, as it became more acceptable for them to walk in groups to prevent or control diabetes. Formative data show that few women felt comfortable to walk outside their households before either intervention. In our end-of-study survey, the median self-reported time spent doing physical activity each week was 60 min greater in PLA clusters than in control clusters (ratio of geometric mean of activity time 1·12 [95% CI 0·88–1·43]), although this difference was not significant when we accounted for the stratified, clustered survey design. When men and women from the same household participated in groups, behaviour change was easier to initiate and sustain. PLA groups and their activities made community members feel in control of their health and able to prevent diabetes. The PLA intervention also destigmatised blood glucose testing and healthy behaviours such as physical activity, reduced or no sugar in tea, and healthy eating while socialising.

to initiate and sustain. PLA groups and their activities made community members feel in control of their health and able to prevent diabetes. The PLA intervention also destigmatised blood glucose testing and healthy behaviours such as physical activity, reduced or no sugar in tea, and healthy eating while socialising. Total and average annual costs of the PLA intervention were INT$601 484 and $240 594, respectively. Total and average annual costs of the mHealth intervention were $312 630 and $125 052, respectively. The average annual costs of the PLA and mHealth per beneficiary (adults ≥30 years) covered were $14 and $7, respectively, with costs per total population (all ages) being $6 and $3, respectively. The incremental cost-effectiveness ratios for PLA were $316 per case of intermediate hyperglycaemia or type 2 diabetes prevented (or $124 per DALY averted) and $6518 per case of type 2 diabetes prevented (or $2551 per DALY averted) among individuals with intermediate hyperglycaemia at baseline.

$6 and $3, respectively. The incremental cost-effectiveness ratios for PLA were $316 per case of intermediate hyperglycaemia or type 2 diabetes prevented (or $124 per DALY averted) and $6518 per case of type 2 diabetes prevented (or $2551 per DALY averted) among individuals with intermediate hyperglycaemia at baseline. Discussion We assessed two community interventions to prevent and control type 2 diabetes and intermediate hyperglycaemia in rural Bangladesh. Facilitated PLA community mobilisation led to large, significant reductions in the combined prevalence of type 2 diabetes and intermediate hyperglycaemia and 2-year incidence of type 2 diabetes among an intermediate hyperglycaemia cohort. The mHealth intervention had no effect on diabetes status. Both interventions were associated with improvements in diabetes knowledge, but had no apparent impact on blood pressure, overweight and obesity, or on recalled fruit and vegetable consumption or physical activity.

s among an intermediate hyperglycaemia cohort. The mHealth intervention had no effect on diabetes status. Both interventions were associated with improvements in diabetes knowledge, but had no apparent impact on blood pressure, overweight and obesity, or on recalled fruit and vegetable consumption or physical activity. The effect size of the PLA community mobilisation on blood glucose is surprising, especially in the absence of major quantifiable changes in behavioural indicators related to diet, physical activity, and care seeking. We therefore did several additional data checks and sensitivity analyses to identify alternative explanations for the observed effects. However, our results are robust to examinations for sampling errors, enumerator effects, measurement biases, and response bias. Furthermore, the observed effectiveness of PLA on mean population blood glucose measures suggests that our findings are not an artefact of blood glucose cut-off values used. The effectiveness of the intervention on normoglycaemic individuals adds to the evidence of PLA effectiveness in the general population and not only high-risk individuals. Overall, therefore, our findings are compelling, but replication is needed in other populations in Bangladesh and elsewhere.

se cut-off values used. The effectiveness of the intervention on normoglycaemic individuals adds to the evidence of PLA effectiveness in the general population and not only high-risk individuals. Overall, therefore, our findings are compelling, but replication is needed in other populations in Bangladesh and elsewhere. Although very different in terms of mode of delivery and, ultimately, impact, there was substantial overlap in terms of the content and focus of our mHealth and PLA interventions. Our findings contribute to existing literature on mHealth interventions in LMICs. Despite potential for relatively low-cost scalability,26 mHealth interventions have often been criticised for a lack of robust assessment in terms of health outcomes.27, 28 Our twice-weekly voice message intervention had a strong basis in behaviour change theory, high population coverage (albeit lower than the PLA intervention), and important impacts on knowledge and awareness of type 2 diabetes, but did not change disease outcomes at the population level. Our findings on mHealth effectiveness differ from those reported by Ramachandran and colleagues,8 who noted significant reductions in the incidence of type 2 diabetes among urban working men with impaired glucose tolerance who received prescribed lifestyle changes followed by tailored text messages in southeast India. Our findings also differ to those reported by Islam and colleagues,29 who observed improved glycaemic control among patients with type 2 diabetes randomly assigned to receive mobile text messages over a 6-month period.29 Although details of the intervention and trial design of Islam and colleagues' study are unclear, differences in population demographics of Ramachandran and colleagues' and Islam and colleagues' trials, the individual-level randomisation, and delivery of text messages to individuals receiving care are notable between these studies and ours. It is also possible that the delayed start, slightly shorter-than-planned mHealth delivery period, and temporary problems in delivering messages to about 40% of registered mobiles at the beginning of the study might have reduced the effectiveness of our intervention.

ing care are notable between these studies and ours. It is also possible that the delayed start, slightly shorter-than-planned mHealth delivery period, and temporary problems in delivering messages to about 40% of registered mobiles at the beginning of the study might have reduced the effectiveness of our intervention. That disease outcome changes were only apparent in the PLA group suggests wider benefits of participatory interventions beyond the provision of information and modelled behaviour employed in our mHealth intervention. This finding builds on existing evidence of the effectiveness of PLA on maternal, neonatal, and child health,13 and ours is the first assessment of this method for a non-communicable disease. Our intervention differs from the recent peer-support lifestyle intervention in Kerala, which improved physical activity and dietary practices but did not affect the incidence of type 2 diabetes among a high-risk cohort.12 First, unlike our measurement of population-level effects, the Kerala outcomes were only measured among the high-risk individuals who, in the intervention group, had been assigned to group support. It is possible that our intervention improved certain secondary outcomes such as diet and physical activity among high-risk sub-groups that are not apparent in general population measurement from only survey questions. Second, the peer-support intervention seems to be education-focused and targeted to a high-risk cohort, unlike our inclusive PLA approach. Finally, better education and literacy indices among the Kerala study population notwithstanding, in the patriarchal context of Bangladesh we implemented separate groups for men and women, whereas the Kerala study implemented mixed-sex groups, which might have affected their effectiveness.

e our inclusive PLA approach. Finally, better education and literacy indices among the Kerala study population notwithstanding, in the patriarchal context of Bangladesh we implemented separate groups for men and women, whereas the Kerala study implemented mixed-sex groups, which might have affected their effectiveness. Our study has several limitations. Although based on WHO STEPS and the Demographic and Health Survey instruments previously applied in Bangladesh and pilot tested before any data collection, large parts of our survey tools, such as for knowledge of diabetes, were developed for this study and were not formally validated in our study population. Our measure of physical activity was crude and based on recall and analysed as a binary measure of at least 150 min per week; this approach might have missed significant changes in the intensity of population exercise between groups. We did not do a detailed food consumption survey, allowing detailed assessment of changes in consumption of certain foods, such as sugar-sweetened beverages. The large effects of PLA on disease without apparent changes in common diabetes risk factors suggests a complex mechanism of action. Perhaps several small changes in lifestyle and diet combined to produce a strong cumulative effect on blood sugar. Dietary and physical activity behaviour change was self-reported in process evaluation data, which showed that rice, oil, salt, and sugar consumption decreased in PLA villages. Process data suggested that intentional exercise increased, particularly among women. One striking effect of the groups that emerged through process evaluation was that women were able to negotiate time for group exercise in communities where individual women were stigmatised if they walked alone. Messages alone might not be able to change this behaviour.30 Furthermore, the removal of stigma about diabetes, the increased solidarity among villagers, and the sharing of information and ideas might have reduced stress levels in the population, which might have contributed to the PLA intervention effect. The differences between process and survey data require further research. The tools we used to measure food consumption and exercise might be too blunt, and participants might be able to more accurately estimate their time spent exercising and quantities of food eaten after the intervention given the clear messaging and emphasis on these behaviours.

and survey data require further research. The tools we used to measure food consumption and exercise might be too blunt, and participants might be able to more accurately estimate their time spent exercising and quantities of food eaten after the intervention given the clear messaging and emphasis on these behaviours. Our approach and some of the assumptions we used for calculating DALYs gives a conservative estimate of the total DALYs averted by PLA. The disability weight that we used, though very small, is for uncomplicated type 2 diabetes. We used the same weight for both type 2 diabetes and intermediate hyperglycaemia, which might slightly overestimate the years lived with disability. However, we did not take into account the likelihood that some of the individuals with type 2 diabetes will develop diabetes-related complications, for which disability weights can range from 0·004 to 0·631.31 Additionally, since we did not have the exact age of diagnosis or onset of diabetes, we used average age of individuals with diagnosed diabetes in our end-of-study survey as age of onset (or diagnosis) of diabetes. Using this age is likely to underestimate years lived with disability and years of life lost, since many individuals were diagnosed before this age.

ct age of diagnosis or onset of diabetes, we used average age of individuals with diagnosed diabetes in our end-of-study survey as age of onset (or diagnosis) of diabetes. Using this age is likely to underestimate years lived with disability and years of life lost, since many individuals were diagnosed before this age. As previously noted, our population-level evaluation means we are unable to estimate QALYs gained. However, the cost-effectiveness ratios of INT$316 to $6518 per case of intermediate hyperglycaemia or type 2 diabetes prevented (or $124 to $2551 per DALY averted) suggests that PLA is highly cost-effective according to the WHO cost-effectiveness threshold,32 considering Bangladesh's gross domestic product per person of $3869 (in 2017). Scale-up of PLA at the national level could prevent about 240 000 cases of type 2 diabetes or intermediate hyperglycaemia each year (equivalent to about $132 million savings in health-care costs per year), assuming a 30% loss in effect through scale-up (appendix). In the absence of prepayment systems in Bangladesh, most health-care expenditures are incurred by patients and their households, potentially leading to financial catastrophes for many.33 Preliminary analyses of our baseline survey data on cost of seeking care suggests that about 60% of people with type 2 diabetes might be at risk of catastrophic and impoverishing health expenditure (ie, spend more than 10% of their household income on diabetes care). Therefore, scaling up cost-effective interventions such as the DMagic PLA has potential to reduce the incidence of impoverishing health expenditure.

of people with type 2 diabetes might be at risk of catastrophic and impoverishing health expenditure (ie, spend more than 10% of their household income on diabetes care). Therefore, scaling up cost-effective interventions such as the DMagic PLA has potential to reduce the incidence of impoverishing health expenditure. The strengths of our study are the large, rural, population-based surveys with high response rates at baseline and the end of the study, and the high 2-year follow-up of the intermediate hyperglycaemia cohort, and assessment of glycaemic status through fasting and 2 h blood glucose tests. Randomisation should eliminate issues of confounding, although baseline differences in wealth were apparent and therefore controlled for, but the possibility of additional unmeasured imbalances between study groups cannot be excluded. Enumerator and participant masking to intervention exposure was infeasible in the end-of-study survey, although the extent to which this could affect the primary outcome measures is likely to be small and we identified no enumerator bias in our sensitivity analysis. Both interventions had good coverage and there was minimal contamination between groups. Although capillary blood glucose concentrations overestimate blood sugar compared with venous samples, the method is feasible and acceptable for epidemiological studies and would not affect the differences we identified between study intervention groups.

good coverage and there was minimal contamination between groups. Although capillary blood glucose concentrations overestimate blood sugar compared with venous samples, the method is feasible and acceptable for epidemiological studies and would not affect the differences we identified between study intervention groups. Further analysis of process evaluation data is likely to provide a more nuanced understanding of intervention mechanisms and effects. Nevertheless, the large, cost-effective impact of PLA in this trial suggests that it might be beneficial in other LMICs with a high burden of type 2 diabetes, and perhaps among high-risk groups in high-income settings. Replication in other populations is an important next step and follow-up of the DMagic study population with mixed-methods approaches will be important to better explain intervention mechanisms of action and long-term impacts. Data sharing De-identified data collected for this study and a data dictionary are available from the corresponding author on reasonable request. Supplementary Material Supplementary appendix

Further analysis of process evaluation data is likely to provide a more nuanced understanding of intervention mechanisms and effects. Nevertheless, the large, cost-effective impact of PLA in this trial suggests that it might be beneficial in other LMICs with a high burden of type 2 diabetes, and perhaps among high-risk groups in high-income settings. Replication in other populations is an important next step and follow-up of the DMagic study population with mixed-methods approaches will be important to better explain intervention mechanisms of action and long-term impacts. Data sharing De-identified data collected for this study and a data dictionary are available from the corresponding author on reasonable request. Supplementary Material Supplementary appendix Acknowledgments This work was funded by the UK Medical Research Council (MR/M016501/1) under the Global Alliance for Chronic Diseases Diabetes Programme. The study team thanks the DMagic trial steering committee (Graham Hitman [Chair], Martin McKee, Dina Balabanova, David Beran, Katherine Fielding, Lou Atkins, and Sophia Wilkinson), the data monitoring committee (Andrew Farmer [Chair], Margaret Thorogood, and Peter Byass) for their valuable contributions to the project. We are grateful to Fatima Zannat and Kazi Faisal Mahmud from mWorld for their contributions to the development of our mHealth intervention and Md Golam Azam for his contributions to field activities.

committee (Andrew Farmer [Chair], Margaret Thorogood, and Peter Byass) for their valuable contributions to the project. We are grateful to Fatima Zannat and Kazi Faisal Mahmud from mWorld for their contributions to the development of our mHealth intervention and Md Golam Azam for his contributions to field activities. Contributors EF was the project principal investigator and led the design of the study, did statistical analyses, and drafted the report. He assumes responsibility for the completeness and integrity of the data and the fidelity of the report to the study protocol and statistical analysis plan. NA contributed to intervention development and survey methods. JM led the process evaluation component of the study and contributed to intervention development. AK was the project manager, overseeing project activities and budgets in Dhaka and Faridpur. SKS coordinated quantitative data collection activities and mHealth intervention development. CK provided technical coordination of the end-of-study survey and data management process. HJ led the formative phase of the study and contributed to intervention development and process evaluation. KAk gathered and analysed qualitative process and implementation data. TN developed and coordinated the implementation of participatory learning and action group activities. HH-B led the economic evaluation component of the project. AKAK and AC provided technical oversight of all project activities and facilitated community engagement and intervention development activities. KAz coordinated project activities in Bangladesh, co-led the project, and contributed to intervention development and study design. All authors contributed to the interpretation of study findings and had the opportunity to review and revise the final manuscript.

unity engagement and intervention development activities. KAz coordinated project activities in Bangladesh, co-led the project, and contributed to intervention development and study design. All authors contributed to the interpretation of study findings and had the opportunity to review and revise the final manuscript. Declaration of interests We declare no competing interests.

Introduction Dementia represents a serious public health challenge because of ageing populations worldwide.1 The prevalence of type 2 diabetes is also rising rapidly around the world, and increases the risk of dementia, including Alzheimer's disease.2–4 Several studies have shown poorer cognitive performance and faster cognitive decline in people with diabetes than in those without the disease.5–9 However, with some notable exceptions,5,7,9 most of the evidence for the effect of type 2 diabetes on cognitive ageing comes from studies in elderly populations.6,8 Typically, such research is based on adults aged 65 years or older at the start of the study, with follow-up measurement of incident dementia. Some researchers believe that diabetes does not necessarily affect cognition before old age,10,11 the implication being that the association between the two exists not because diabetes is a risk factor for dementia, but because of shared risk factors such as hypertension. Since dementia is a progressive disease involving cognitive decline over several years,12,13 investigation is needed to determine whether diabetes affects cognitive decline before old age.

between the two exists not because diabetes is a risk factor for dementia, but because of shared risk factors such as hypertension. Since dementia is a progressive disease involving cognitive decline over several years,12,13 investigation is needed to determine whether diabetes affects cognitive decline before old age. Our aim was to assess whether, compared with normoglycaemia, type 2 diabetes and prediabetes14 are associated with faster cognitive decline from late midlife (age 55 years) to early old age (age 65 years), an age range in which dementia is uncommon. We also aimed to examine the possibility of a dose-response relation by investigating the role of duration of diabetes (the underlying hypothesis being that if diabetes is a risk factor for cognition then longer exposure to diabetes would have a stronger effect on cognitive decline) and to investigate whether glycaemic control (measured by HbA1c), including among individuals with type 2 diabetes, is associated with cognitive decline.

s (the underlying hypothesis being that if diabetes is a risk factor for cognition then longer exposure to diabetes would have a stronger effect on cognitive decline) and to investigate whether glycaemic control (measured by HbA1c), including among individuals with type 2 diabetes, is associated with cognitive decline. Methods Study design and participants The Whitehall II study was established in 1985 with the recruitment of British civil servants aged 35–55 years from 20 London-based departments to investigate determinants of chronic diseases, with the baseline assessment taking place in 1985–88.15 Clinical examinations were also done in 1991–93, 1997–99, 2002–04, and 2007–09. Cognitive testing was introduced to the study in 1997–99 and repeated in 2002–04 and 2007–09. Written informed consent from participants and research ethics approvals (University College London ethics committee) were renewed at each contact; the most recent approval was from the Joint University College London/University College London Hospital Committee on the Ethics of Human Research (Committee Alpha), reference number 85/0938.

nsent from participants and research ethics approvals (University College London ethics committee) were renewed at each contact; the most recent approval was from the Joint University College London/University College London Hospital Committee on the Ethics of Human Research (Committee Alpha), reference number 85/0938. Measurements We ascertained type 2 diabetes status in 1991–93 and 1997–99. We took venous blood samples after a minimum 5 h fast, followed by a 75 g, 2 h oral glucose tolerance test. Blood samples were drawn into fluoride monovette tubes and centrifuged on site. We measured blood glucose using the glucose oxidase method.16 Type 2 diabetes was defined by WHO criteria,17 based on a fasting glucose of 7·0 mmol/L or more, or a 2 h postload glucose of 11·1 mmol/L or more. Participants who met these criteria at the 1991–93 examination or who had known diabetes in 1997–99 (ie, doctor-diagnosed diabetes or use of antidiabetic drugs) were classified as known diabetes. Those without a history of diabetes, but who met the diabetes criteria at the 1997–99 examination were classified as newly diagnosed diabetes. Non-diabetic participants were classified as prediabetic if their fasting plasma glucose concentration was between 6·1 and less than 7·0 mmol/L and their 2 h postload glucose concentration was less than 7·8 mmol/L (impaired fasting glucose), or if their fasting glucose was less than 7·0 mmol/L and their 2 h postload plasma glucose concentration was between 7·8 and less than 11·1 mmol/L (impaired glucose tolerance).17 Others were classified as normoglycaemic.

L and their 2 h postload glucose concentration was less than 7·8 mmol/L (impaired fasting glucose), or if their fasting glucose was less than 7·0 mmol/L and their 2 h postload plasma glucose concentration was between 7·8 and less than 11·1 mmol/L (impaired glucose tolerance).17 Others were classified as normoglycaemic. Glycaemic control was characterised by HbA1c, which was measured in EDTA (edetic acid) whole blood on a calibrated high-performance liquid chromatography system with automated haemolysis before injection. We used mean values from 2002–04 and 2007–09 measurements to represent glycaemic control during follow-up.

L and their 2 h postload glucose concentration was less than 7·8 mmol/L (impaired fasting glucose), or if their fasting glucose was less than 7·0 mmol/L and their 2 h postload plasma glucose concentration was between 7·8 and less than 11·1 mmol/L (impaired glucose tolerance).17 Others were classified as normoglycaemic. Glycaemic control was characterised by HbA1c, which was measured in EDTA (edetic acid) whole blood on a calibrated high-performance liquid chromatography system with automated haemolysis before injection. We used mean values from 2002–04 and 2007–09 measurements to represent glycaemic control during follow-up. We used a comprehensive battery of cognitive tests appropriate for middle-aged individuals.18 Short-term verbal memory was tested with a 20-word free-recall test in which participants were presented a list of 20 one-syllable or two-syllable words at intervals of 2 s and were asked to recall in writing as many of the words as possible, in any order, in 2 min. We used the Alice Heim 4-I test19 to assess inductive reasoning, measuring the ability of participants to identify patterns and infer principles and rules. This test consists of a series of 65 verbal and mathematical reasoning items of increasing difficulty; participants had 10 min to do the test. Verbal fluency was assessed with tests of phonemic fluency (words beginning with s) and semantic fluency (animal names),20 with 1 min allowed for each test. We standardised the results of the four tests to Z scores using the mean and SD from the 1997–99 assessments, and averaged them to create a global cognitive score. The global cognitive score was then re-standardised so that the mean was 0 and the SD was 1. Previous research has used global scores constructed in this way to minimise problems caused by measurement error in the individual tests21 and to allow comparison of findings across studies when effects are not limited to one cognitive domain. For all cognitive tests, including the global score, we used standardised values in the regression analysis to allow comparison of beta coefficients between results for the different tests.

in the individual tests21 and to allow comparison of findings across studies when effects are not limited to one cognitive domain. For all cognitive tests, including the global score, we used standardised values in the regression analysis to allow comparison of beta coefficients between results for the different tests. Covariates included sociodemographic characteristics, health-related behaviours, and chronic diseases. Sociodemographic characteristics were age, sex, marital status (single, divorced, widowed, married, or cohabiting) and education (low [did not complete secondary school], middle [secondary school], and high [university degree or higher]). Health-related behaviours were smoking (current smoker, ex-smoker, or never smoked), alcohol consumption per week (abstainer [zero units], moderate drinker [one to 21 units for men, one to 14 units for women], or heavy drinker [more than 21 units for men, more than 14 units for women]), adequate physical activity (yes or no; as per WHO recommendations,22 assessed via a 20-item questionnaire); and frequency of fruit and vegetable consumption (less than once daily, once daily, or more frequently). Chronic disease covariates (based on self-report of doctor diagnosis and corroborated in medical records) were coronary heart disease, stroke, hypertension (defined as blood pressure of 140/90 mm Hg or higher, or use of antihypertensive drugs), respiratory disease, total cholesterol, obesity (BMI of 30 kg/m2 or greater), use of antidepressants, and use of lipid-lowering drugs.

diagnosis and corroborated in medical records) were coronary heart disease, stroke, hypertension (defined as blood pressure of 140/90 mm Hg or higher, or use of antihypertensive drugs), respiratory disease, total cholesterol, obesity (BMI of 30 kg/m2 or greater), use of antidepressants, and use of lipid-lowering drugs. Statistical analysis We examined associations between baseline characteristics and diabetes status using χ2 squared tests (categorical data) and analysis of variance (continuous data). We did cross-sectional analyses for diabetes status and cognitive measures from 1997–99 using linear regression. We used three sets of adjustments with measures from 1997–99: model 1 was adjusted for sociodemographic measures only; model 2 also included health-related behaviours; and model 3 also included chronic diseases.

id cross-sectional analyses for diabetes status and cognitive measures from 1997–99 using linear regression. We used three sets of adjustments with measures from 1997–99: model 1 was adjusted for sociodemographic measures only; model 2 also included health-related behaviours; and model 3 also included chronic diseases. For longitudinal analyses, we used linear mixed models to examine the association of diabetes status in 1997–99 with cognitive decline over 10 years (1997–99, 2002–04, and 2007–09). In these models, fixed effects were terms for time, main effect terms for diabetes status and all covariates (main effects for covariates allow for adjustment of their cross-sectional effect on cognitive function), and interactions between time and all variables in the model. The interaction between a variable and time represents its effect on cognitive decline, and inclusion of interactions with time for all covariates allows the estimation of diabetes and cognitive decline to be adjusted for the effect of all covariates on cognitive decline. Both the intercept and the slope were fitted as random effects, allowing individuals to have different cognitive scores at baseline and different rates of cognitive decline during follow-up.

tes allows the estimation of diabetes and cognitive decline to be adjusted for the effect of all covariates on cognitive decline. Both the intercept and the slope were fitted as random effects, allowing individuals to have different cognitive scores at baseline and different rates of cognitive decline during follow-up. Age was centred at the median on the basis of the 1997–99 assessment and used as the timescale in the longitudinal analyses. This term was divided by ten; thus, the coefficient associated with unit change in time represents cognitive decline over 10 years to match the 10 year follow-up. These analyses were also adjusted for age at the start of the cognitive follow-up and included covariates from 1997–99 across the three models, as in the cross-sectional analyses. The normoglycaemia group was used as the reference group to calculate differences in cognitive decline in the prediabetes and diabetes groups. To allow interpretation of the cross-sectional and longitudinal estimates, we compared them with the effect of age on cognition by dividing the estimate by the effect of a 1 year increase in age on cognition. We calculated the effect of age by regression of the standardised 1997–99 cognitive score on age.

Age was centred at the median on the basis of the 1997–99 assessment and used as the timescale in the longitudinal analyses. This term was divided by ten; thus, the coefficient associated with unit change in time represents cognitive decline over 10 years to match the 10 year follow-up. These analyses were also adjusted for age at the start of the cognitive follow-up and included covariates from 1997–99 across the three models, as in the cross-sectional analyses. The normoglycaemia group was used as the reference group to calculate differences in cognitive decline in the prediabetes and diabetes groups. To allow interpretation of the cross-sectional and longitudinal estimates, we compared them with the effect of age on cognition by dividing the estimate by the effect of a 1 year increase in age on cognition. We calculated the effect of age by regression of the standardised 1997–99 cognitive score on age. In the final analyses, we examined whether poorer glycaemic control, modelled as a one percentage point increment in HbA1c, was associated with cognitive decline in participants with normoglycaemia, prediabetes, newly diagnosed diabetes, or known diabetes. The normal distribution of the HbA1c measure allowed the use of a one percentage point increment as the exposure in these analyses, which were done by use of linear mixed models with an interaction term between diabetes status, HbA1c, and time of follow-up to estimate cognitive decline associated with a one percentage point increase in HbA1c.

of the HbA1c measure allowed the use of a one percentage point increment as the exposure in these analyses, which were done by use of linear mixed models with an interaction term between diabetes status, HbA1c, and time of follow-up to estimate cognitive decline associated with a one percentage point increase in HbA1c. Non-white participants (n=502) were excluded from the main analyses because a test of interaction showed the association of diabetes with cognition to differ in this subgroup (p=0·0013 for the global cognitive score), and small numbers precluded further analyses. However, to allow comparison, we examined cognitive decline as a function of diabetes status (yes or no) in these participants. We also did sensitivity analyses to assess the robustness of our main findings. These were: replacing the hypertension measure in model 3 with systolic and diastolic blood pressure as continuous variables; removing from the analysis all participants who became diabetic after the clinical examination in 1997–99, based on clinical assessments in 2002–04 and 2007–09; using an alternative classification of diabetes status in which newly diagnosed diabetes and known diabetes were replaced with diabetes diagnosis 0–1·5 years ago and diabetes diagnosis more than 1·5 years ago, respectively, on the basis of age at diagnosis of diabetes and age at the 1997–99 clinical assessment; using all covariates as time-dependent variables; and using a multiple imputation, chained-equations method to replace missing data for cognition and covariates during follow-up, using all available data for exposures, outcomes, and covariates in the analysis. All analyses were done in Stata SE version 12 for Windows (StataCorp, 2011). p values were two sided and p<0·05 was regarded as significant.

, chained-equations method to replace missing data for cognition and covariates during follow-up, using all available data for exposures, outcomes, and covariates in the analysis. All analyses were done in Stata SE version 12 for Windows (StataCorp, 2011). p values were two sided and p<0·05 was regarded as significant. Role of the funding source The sponsors of the study had no role in study 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 the 10 308 participants recruited at the beginning of the study in 1985–88, 8637 (84%) attended the diabetes screening in 1991–93, and 7870 (76%) attended in 1997–99 when the first cognitive assessment took place (figure). The median age of participants was 54·4 years (IQR 50·3–60·3) in 1997–99, 59·9 years (55·8–65·7) in 2002–04, and 64·7 years (60·8–70·6) in 2007–09. No differences in sex (interaction p=0·15–0·97) or age (interaction p=0·39–0·77) were noted in associations of diabetes with cognitive decline, leading us to combine men and women and all age groups in the analyses and to adjust the models for age and sex.

(55·8–65·7) in 2002–04, and 64·7 years (60·8–70·6) in 2007–09. No differences in sex (interaction p=0·15–0·97) or age (interaction p=0·39–0·77) were noted in associations of diabetes with cognitive decline, leading us to combine men and women and all age groups in the analyses and to adjust the models for age and sex. Table 1 shows the characteristics of the study population by diabetes status in 1997–99. The mean duration of diabetes in participants with known diabetes was 4·95 (SD 2·21) years. Of the 5653 people included in the longitudinal analyses, 4073 (72%) had cognitive data recorded at all three assessments, 1000 (18%) at two assessments, and 580 (10%) at one assessment. Compared with individuals not included in the longitudinal analyses, the samples consisted of younger participants (mean age 55·5 vs 57·1 years, p<0·0001), and contained more men (4113 [73%] of 5653 vs 1360 [61%] of 2217, p<0·0001), and more educated individuals (1678 [30%] of 5653 vs 542 [24%] of 2217 with a university degree, p<0·0001). Age was inversely associated with cognition; a 1 year increase in age was associated with a −0·039 SD (95% CI −0·043 to −0·035) decrement in memory, a −0·035 SD (−0·039 to −0·030) decrement in reasoning, a −0·036 SD (−0·041 to −0·032) decrement in phonemic fluency, a −0·040 SD (−0·044 to −0·035) decrement in semantic fluency, and a −0·050 SD (−0·054 to −0·046) decrement in global cognitive score (all p<0·0001).

·039 SD (95% CI −0·043 to −0·035) decrement in memory, a −0·035 SD (−0·039 to −0·030) decrement in reasoning, a −0·036 SD (−0·041 to −0·032) decrement in phonemic fluency, a −0·040 SD (−0·044 to −0·035) decrement in semantic fluency, and a −0·050 SD (−0·054 to −0·046) decrement in global cognitive score (all p<0·0001). Cross-sectional analyses were based on 5183 people with complete cognitive data in 1997–99; 606 (12%) had prediabetes, 110 (2%) had newly diagnosed diabetes, and 146 (3%) had known diabetes. Compared with normoglycaemic individuals, those with known diabetes had a −0·16 SD (95% CI −0·30 to −0·02) lower score in reasoning in the fully adjusted model (model 3; p=0·023; table 2), although the results for the other cognitive measures (including global cognitive score) were not significant. The coefficient from model 3 in individuals with known diabetes corresponds to an age effect of roughly 4·6 years for reasoning. We initially did the longitudinal analyses using a simple binary classification of diabetes status in 1997–99: diabetic versus non-diabetic (normoglycaemia and prediabetes). These results (appendix p 1) show faster declines in reasoning, phonemic fluency, and the global cognitive score in participants with diabetes than in those without diabetes. The effect sizes were larger, albeit with wide CIs, in non-white participants (appendix p 2).

rsus non-diabetic (normoglycaemia and prediabetes). These results (appendix p 1) show faster declines in reasoning, phonemic fluency, and the global cognitive score in participants with diabetes than in those without diabetes. The effect sizes were larger, albeit with wide CIs, in non-white participants (appendix p 2). Estimates for decline in the normoglycaemic group—used as a reference in the analyses—are listed in the appendix (p 10). Compared with normoglycaemic participants, those with known diabetes had a 45% faster decline in memory, a 29% faster decline in reasoning, and a 24% faster decline in global cognition; compared with the reference values, 10 year differences in decline were −0·13 SD (−0·26 to −0·00) for memory; −0·10 SD (−0·19 to −0·01) for reasoning, and −0·11 SD (−0·21 to −0·02) for the global cognitive score in model 3 (table 3). The significant decline in participants with known diabetes was equivalent to an age effect of 3·3 years for memory, 2·9 years for reasoning, and 2·2 years for the global cognitive score. Participants with prediabetes and newly diagnosed diabetes did not show faster cognitive decline than those with normoglycaemia (table 3).

gnificant decline in participants with known diabetes was equivalent to an age effect of 3·3 years for memory, 2·9 years for reasoning, and 2·2 years for the global cognitive score. Participants with prediabetes and newly diagnosed diabetes did not show faster cognitive decline than those with normoglycaemia (table 3). Sensitivity analyses showed that replacing hypertension with systolic and diastolic blood pressure as continuous variables in model 3 had little effect on the estimates (appendix p 3); the results were much the same when participants who became diabetic during the period of cognitive testing were removed from the analysis (appendix p 4). Alternative classification of diabetes status (diagnosed 0–1·5 years ago vs more than 1·5 years ago) showed a faster cognitive decline in participants who had been diagnosed with diabetes more than 1·5 years ago than in those diagnosed more recently (appendix p 5). Use of time-dependent covariates in the longitudinal models showed significantly faster decline in reasoning and the global cognitive score in those with known diabetes (appendix p 6). Associations using imputed data showed stronger cross-sectional effects (appendix p 7), but longitudinal results were similar to those from the main analysis (appendix p 8).

e longitudinal models showed significantly faster decline in reasoning and the global cognitive score in those with known diabetes (appendix p 6). Associations using imputed data showed stronger cross-sectional effects (appendix p 7), but longitudinal results were similar to those from the main analysis (appendix p 8). Mean HbA1c values were highest in participants with known diabetes (6·84%, SD 1·25) and lowest in those with normoglycaemia (5·40%, 0·40); newly diagnosed individuals (6·27%, 1·10) and participants with prediabetes (5·71%, 0·65) had intermediate values. In the fully adjusted analyses (model 3), a one percentage point increment in HbA1c was associated with a significantly faster decline in memory in participants with known diabetes, and a faster decline in reasoning in those with newly diagnosed (significant [p=0·028]) and known diabetes (approaching significance [p=0·052]; table 4). Using time-dependent covariates rather than those drawn from the 1997–99 assessment showed similar associations (appendix p 9).

y in participants with known diabetes, and a faster decline in reasoning in those with newly diagnosed (significant [p=0·028]) and known diabetes (approaching significance [p=0·052]; table 4). Using time-dependent covariates rather than those drawn from the 1997–99 assessment showed similar associations (appendix p 9). Discussion Our results from a large cohort of middle-aged adults show that participants with known diabetes—ie, those who had diabetes at the start of the study—had an increased rate of cognitive decline during the subsequent 10 year period. The effect of diabetes duration cannot be examined when all diabetes cases are analysed together, showing the pertinence of our research design. Faster cognitive decline in people with long-term diabetes, taking into account both cross-sectional and longitudinal analyses, corresponded to an age effect of 7·5 years for reasoning and roughly 4·4 years for the global cognitive score. Cognitive decline in those with newly diagnosed diabetes and prediabetes was not different from that which occurred in normoglycaemic participants. We noted little attenuation of associations after taking into account potential confounding factors. Our results also show that in people with diabetes, those with poorer glycaemic control had faster cognitive decline. These findings suggest that duration of diabetes contributes to faster cognitive decline and that good glycaemic control could decrease this risk (panel).

ing into account potential confounding factors. Our results also show that in people with diabetes, those with poorer glycaemic control had faster cognitive decline. These findings suggest that duration of diabetes contributes to faster cognitive decline and that good glycaemic control could decrease this risk (panel). Substantial evidence suggests that diabetes is a risk factor for cognitive decline and dementia.2,3 Some previous studies have suggested that the faster rate of cognitive decline affects only elderly people with type 2 diabetes,23 with the hypothesis being that type 2 diabetes does not affect cognition before old age.10,11 However, our results, from participants with median ages of 54, 60, and 65 years at the time of cognitive assessments, show that longer duration of diabetes is associated with more rapid cognitive decline. These findings are in line with previous results suggesting that midlife rather than late-life diabetes is a risk factor for dementia.24 These, along with our results, can be interpreted as showing that longer exposure to diabetes is harmful for cognition. A previous case-control study25 in elderly people showed duration and severity of diabetes to be associated with mild cognitive impairment. Another study9 in adults aged 40–83 years, who were followed up for 12 years, showed that the extent of cognitive decline in individuals who developed diabetes during follow-up was between that of individuals without diabetes and those who had diabetes at baseline, albeit not significantly different from either group. Although our finding that cognitive decline was not worse in those with prediabetes is in agreement with that of a previous study26 in elderly women, further research is needed since high glucose concentrations have been associated with an increased risk of dementia in people with glucose concentrations below the clinical threshold of manifest diabetes.27

ecline was not worse in those with prediabetes is in agreement with that of a previous study26 in elderly women, further research is needed since high glucose concentrations have been associated with an increased risk of dementia in people with glucose concentrations below the clinical threshold of manifest diabetes.27 All cognitive tests used in our analyses have been shown previously to be sensitive to age-related changes in cognition in midlife.18 In our study, known diabetes was associated with faster decline in memory, reasoning, and the global cognitive score. Although no significant cross-sectional effects were evident for memory, the memory decline over 10 years was 45% faster in participants with known diabetes. Additionally, poor glycaemic control in diabetes was associated with a faster decline in memory and possibly in reasoning. Thus, various components of cognition seem to be affected by type 2 diabetes, as evident in the robust effects on the global cognitive score. The precise mechanisms that underlie the association of diabetes with cognitive decline and dementia remain unclear. Vascular pathways are considered to be important.3 Diabetes is often associated with common cardiovascular risk factors such as dyslipidaemia, hypertension, and obesity.28 Complications related to microangiopathy have also been implicated.29 Heterogeneous cerebral lesions, which cause cognitive dysfunction, are associated with longer diabetes duration; these lesions include ischaemic stroke, intracerebral haemorrhage, lacunar infarcts, white matter lesions, and cerebral atrophy.30

ity.28 Complications related to microangiopathy have also been implicated.29 Heterogeneous cerebral lesions, which cause cognitive dysfunction, are associated with longer diabetes duration; these lesions include ischaemic stroke, intracerebral haemorrhage, lacunar infarcts, white matter lesions, and cerebral atrophy.30 Our results emphasise the importance of duration of diabetes for cognitive ageing, and suggest that interventions that aim to prevent or delay diabetes onset might have implications for cognitive health. Lifestyle interventions have been shown to reduce the risk of progression from prediabetes to diabetes. A randomised trial31 of adults at high risk of type 2 diabetes showed that an intensive lifestyle modification programme reduced the risk of progression to diabetes by more than use of the antidiabetic drug metformin (58% [95 % CI 48–66] vs 31% [17–43]). Less clear is the effect of tight glycaemic control on those with established disease. In the ACCORD MIND study,32 intervention to reduce HbA1c to less than 6% in people with type 2 diabetes was associated with reduced brain atrophy, although no effect on cognitive decline was evident. By contrast, in the IDEATel trial,33 an HbA1c of 7% or less was associated with slowed cognitive decline. Prevention of microvascular complications is highly dependent on glycaemic control in adults with type 2 diabetes, but no effect on macrovascular disease or mortality was seen in elderly people with longstanding type 2 diabetes.34 Cumulative glycaemic exposure (ie, severity and duration of hyperglycaemia) is important for microvascular complications,35 and increases the risk of more rapid cognitive decline.

ts with type 2 diabetes, but no effect on macrovascular disease or mortality was seen in elderly people with longstanding type 2 diabetes.34 Cumulative glycaemic exposure (ie, severity and duration of hyperglycaemia) is important for microvascular complications,35 and increases the risk of more rapid cognitive decline. The strengths of this analysis from the Whitehall II study are the prospective cohort and the fairly young population—75% of participants were younger than 71 years at the last cognitive assessment. The repeated standardised screening for diabetes before the start of cognitive follow-up allowed us to minimise reverse causation biases. Alternative classification of duration of diabetes gave similar results. We also took into account a range of potential confounders of the association between diabetes and cognition. The major limitation of this study is the issue of generalisability, since the data are for an occupational cohort and the participants are likely to be healthier than the general population. Finally, because of the small numbers of non-white participants, we could not examine the duration-of-diabetes hypothesis in this group, so the extent to which the results apply to non-white populations is unclear. Our results support the hypothesis that the risk of accelerated cognitive decline in people with diabetes depends on how long an individual has had the disease and on the extent to which they can achieve normal carbohydrate metabolism. Further research is needed to determine whether improving management of type 2 diabetes also reduces the risk of dementia.

t the risk of accelerated cognitive decline in people with diabetes depends on how long an individual has had the disease and on the extent to which they can achieve normal carbohydrate metabolism. Further research is needed to determine whether improving management of type 2 diabetes also reduces the risk of dementia. Supplementary Material Supplementary appendix Acknowledgments This research was supported by the US National Institutes of Health (R01AG013196 to AS-M; R01AG034454 to AS-M and MK; R01HL036310 to MK) and the UK Medical Research Council (K013351 to MK), as well as receiving support from the UK Economic and Social Research Council (to MK) and the British Heart Foundation (to EJB). AGT is supported by the TÁMOP 4.2.4.A/1-11-1-2012-0001 National Excellence Programme (research fellowship cofinanced by the European Union and the European Social Fund). We thank all of the participating civil service departments and their welfare, personnel, and establishment officers; the British Occupational Health and Safety Agency; the British Council of Civil Service Unions; all participating civil servants in the Whitehall II study; and all members of the Whitehall II study team. The Whitehall II study team consists of research scientists, statisticians, study coordinators, nurses, data managers, administrative assistants, and data entry staff, all of whom make the study possible.

Unions; all participating civil servants in the Whitehall II study; and all members of the Whitehall II study team. The Whitehall II study team consists of research scientists, statisticians, study coordinators, nurses, data managers, administrative assistants, and data entry staff, all of whom make the study possible. Contributors RHT participated in the preliminary analysis and drafting of the original report. AD supervised the preliminary analysis, did the final analyses, and commented on drafts. AGT, AE, EJB, and MK helped to formulate the research question and commented on drafts. AS-M helped to formulate the research question, searched the published work, supervised the analyses, and wrote successive drafts and the final submitted version of the report. RHT, AD, and AS-M had access to the raw data. Conflicts of interest We declare that we have no conflicts of interest. Figure Study profile Table 1 Characteristics of the study population, by diabetes status in 1997–99

Contributors RHT participated in the preliminary analysis and drafting of the original report. AD supervised the preliminary analysis, did the final analyses, and commented on drafts. AGT, AE, EJB, and MK helped to formulate the research question and commented on drafts. AS-M helped to formulate the research question, searched the published work, supervised the analyses, and wrote successive drafts and the final submitted version of the report. RHT, AD, and AS-M had access to the raw data. Conflicts of interest We declare that we have no conflicts of interest. Figure Study profile Table 1 Characteristics of the study population, by diabetes status in 1997–99 Normoglycaemia (n=4703) Prediabetes*(n=648) Newly diagnosed diabetes (n=115) Known diabetes (n=187) p value Age (years) 55·1 (5·9) 57·5 (6·1) 59·0 (6·1) 57·4 (6·3) <0·0001 Men 3428 (73%) 474 (73%) 81 (70%) 130 (70%) 0·71 Single, divorced, or widowed 1106 (24%) 146 (23%) 32 (28%) 47 (25%) 0·61 Low education 2018 (43%) 292 (45%) 56 (49%) 93 (50%) 0·14 Heavy alcohol consumption† 1270 (27%) 195 (30%) 26 (23%) 52 (28%) 0·26 Current smoker 482 (10%) 48 (7%) 12 (10%) 14 (7%) 0·10 Inadequate physical activity‡ 3522 (75%) 477 (74%) 92 (80%) 153 (82%) 0·08 Fruit and vegetable intake less than once daily 1218 (26%) 144 (22%) 27 (23%) 41 (22%) 0·14 Respiratory illness 351 (7%) 43 (7%) 6 (5%) 20 (11%) 0·23 Total cholesterol (mmol/L) 5·91 (1·05) 6·12 (1·06) 6·12 (1·12) 5·92 (1·03) <0·0001 Obesity (BMI ≥30 kg/m2) 561 (12%) 124 (19%) 20 (17%) 43 (23%) <0·0001 Hypertension 1134 (24%) 257 (40%) 53 (46%) 88 (47%) <0·0001 Stroke 14 (<1%) 1 (<1%) 1 (1%) 2 (1%) 0·17 Coronary heart disease 240 (5%) 44 (7%) 14 (12%) 21 (11%) <0·0001 Use of antidepressant drugs 129 (3%) 12 (2%) 5 (4%) 5 (3%) 0·39 Use of lipid-lowering drugs 122 (3%) 22 (3%) 9 (8%) 14 (7%) <0·0001 Data are n (%) or mean (SD).

ion 1134 (24%) 257 (40%) 53 (46%) 88 (47%) <0·0001 Stroke 14 (<1%) 1 (<1%) 1 (1%) 2 (1%) 0·17 Coronary heart disease 240 (5%) 44 (7%) 14 (12%) 21 (11%) <0·0001 Use of antidepressant drugs 129 (3%) 12 (2%) 5 (4%) 5 (3%) 0·39 Use of lipid-lowering drugs 122 (3%) 22 (3%) 9 (8%) 14 (7%) <0·0001 Data are n (%) or mean (SD). * Prediabetes was defined with a 75 g oral glucose tolerance test as one of two states: impaired fasting glucose, defined as fasting plasma glucose between 6·1 mmol/L and less than 7·0 mmol/L, without impaired glucose tolerance; or impaired glucose tolerance, defined as fasting glucose of less than 7·0 mmol/L and a 2 h postload plasma glucose concentration between 7·8 mmol/L and less than 11·1 mmol/L. † More than 21 units per week for men and more than 14 units per week for women. ‡ Less than WHO recommendations.22 Table 2 Estimated differences in cognitive function, as a function of diabetes status (cross-sectional analysis at 1997–99 assessment)

* Prediabetes was defined with a 75 g oral glucose tolerance test as one of two states: impaired fasting glucose, defined as fasting plasma glucose between 6·1 mmol/L and less than 7·0 mmol/L, without impaired glucose tolerance; or impaired glucose tolerance, defined as fasting glucose of less than 7·0 mmol/L and a 2 h postload plasma glucose concentration between 7·8 mmol/L and less than 11·1 mmol/L. † More than 21 units per week for men and more than 14 units per week for women. ‡ Less than WHO recommendations.22 Table 2 Estimated differences in cognitive function, as a function of diabetes status (cross-sectional analysis at 1997–99 assessment) Model 1 Model 2 Model 3 Memory Normoglycaemia .. .. .. Prediabetes −0·05 (−0·13 to 0·03) −0·05 (−0·13 to 0·03) −0·04 (−0·13 to 0·04) Newly diagnosed diabetes −0·09 (−0·27 to 0·09) −0·08 (−0·26 to 0·10) −0·07 (−0·25 to 0·12) Known diabetes 0·01 (−0·15 to 0·17) 0·02 (−0·14 to 0·17) 0·03 (−0·13 to 0·18) Reasoning Normoglycaemia .. .. .. Prediabetes 0·03 (−0·04 to 0·10) 0·02 (−0·05 to 0·09) 0·03 (−0·04 to 0·10) Newly diagnosed diabetes 0·05 (−0·12 to 0·21) 0·06 (−0·10 to 0·22) 0·07 (−0·09 to 0·23) Known diabetes −0·19 (−0·33 to −0·05)* −0·18 (−0·32 to −0·04)† −0·16 (−0·30 to −0·02)‡ Phonemic fluency Normoglycaemia .. .. .. Prediabetes −0·05 (−0·13 to 0·03) −0·06 (−0·14 to 0·02) −0·05 (−0·13 to 0·03) Newly diagnosed diabetes −0·03 (−0·21 to 0·15) −0·02 (−0·20 to 0·16) −0·01 (−0·18 to 0·17) Known diabetes −0·07 (−0·23 to 0·08) −0·07 (−0·23 to 0·08) −0·05 (−0·21 to 0·10) Semantic fluency Normoglycaemia .. .. .. Prediabetes −0·03 (−0·11 to 0·05) −0·03 (−0·11 to 0·05) −0·03 (−0·11 to 0·05) Newly diagnosed diabetes −0·09 (−0·26 to 0·09) −0·08 (−0·25 to 0·10) −0·07 (−0·24 to 0·11) Known diabetes −0·13 (−0·28 to 0·02) −0·12 (−0·27 to 0·03) −0·11 (−0·27 to 0·04) Global score Normoglycaemia .. .. .. Prediabetes −0·03 (−0·11 to 0·04) −0·04 (−0·12 to 0·03) −0·03 (−0·11 to 0·04) Newly diagnosed diabetes −0·05 (−0·22 to 0·11) −0·04 (−0·21 to 0·12) −0·02 (−0·19 to 0·14) Known diabetes −0·13 (−0·28 to 0·01) −0·12 (−0·26 to 0·02) −0·10 (−0·25 to 0·04) Data are beta coefficients (95% CI) based on standardised cognitive scores (mean =0, SD=1). p values for significant results (p<0·05) compared with reference group are indicated by footnotes. Model 1 is adjusted for age, sex, marital status, and education. Model 2 is adjusted for same parameters as model 1 plus health-related behaviours (smoking, alcohol, physical activity, and fruit and vegetable consumption). Model 3 is adjusted for same parameters as model 2 plus coronary heart disease, stroke, hypertension, respiratory disease, total cholesterol, obesity, use of antidepressants, and use of lipid-lowering drugs.

del 1 plus health-related behaviours (smoking, alcohol, physical activity, and fruit and vegetable consumption). Model 3 is adjusted for same parameters as model 2 plus coronary heart disease, stroke, hypertension, respiratory disease, total cholesterol, obesity, use of antidepressants, and use of lipid-lowering drugs. n for normoglycaemia is 4321; n for prediabetes is 606; n for newly diagnosed diabetes is 110; and n for known diabetes is 146. * p=0·008. † p=0·011. ‡ p=0·023. Table 3 Estimated differences in cognitive decline over 10 years, as a function of diabetes status in 1997–99

del 1 plus health-related behaviours (smoking, alcohol, physical activity, and fruit and vegetable consumption). Model 3 is adjusted for same parameters as model 2 plus coronary heart disease, stroke, hypertension, respiratory disease, total cholesterol, obesity, use of antidepressants, and use of lipid-lowering drugs. n for normoglycaemia is 4321; n for prediabetes is 606; n for newly diagnosed diabetes is 110; and n for known diabetes is 146. * p=0·008. † p=0·011. ‡ p=0·023. Table 3 Estimated differences in cognitive decline over 10 years, as a function of diabetes status in 1997–99 Model 1 Model 2 Model 3 Memory Normoglycaemia .. .. .. Prediabetes 0·02 (−0·05 to 0·09) 0·02 (−0·05 to 0·09) 0·02 (−0·05 to 0·09) Newly diagnosed diabetes 0·07 (−0·09 to 0·22) 0·06 (−0·10 to 0·21) 0·06 (−0·10 to 0·21) Known diabetes −0·13 (−0·26 to −0·01)* −0·13 (−0·26 to −0·00)† −0·13 (−0·26 to −0·00)‡ Reasoning Normoglycaemia .. .. .. Prediabetes −0·02 (−0·07 to 0·02) −0·03 (−0·07 to 0·02) −0·02 (−0·07 to 0·02) Newly diagnosed diabetes −0·04 (−0·14 to 0·06) −0·04 (−0·15 to 0·06) −0·05 (−0·15 to 0·06) Known diabetes −0·10 (−0·19 to −0·01)§ −0·10 (−0·18 to −0·01)¶ −0·10 (−0·19 to −0·01)‖ Phonemic fluency Normoglycaemia .. .. .. Prediabetes 0·02 (−0·04 to 0·08) 0·02 (−0·04 to 0·07) 0·02 (−0·04 to 0·08) Newly diagnosed diabetes −0·10 (−0·23 to 0·04) −0·10 (−0·24 to 0·03) −0·10 (−0·23 to 0·04) Known diabetes −0·09 (−0·20 to 0·02) −0·08 (−0·20 to 0·03) −0·08 (−0·19 to 0·03) Semantic fluency Normoglycaemia .. .. .. Prediabetes 0·00 (−0·05 to 0·06) 0·00 (−0·06 to 0·06) 0·00 (−0·05 to 0·06) Newly diagnosed diabetes 0·03 (−0·10 to 0·16) 0·03 (−0·11 to 0·16) 0·02 (−0·11 to 0·16) Known diabetes −0·06 (−0·17 to 0·05) −0·05 (−0·16 to 0·06) −0·05 (−0·16 to 0·06) Global score Normoglycaemia .. .. .. Prediabetes −0·00 (−0·05 to 0·05) −0·00 (−0·05 to 0·05) −0·00 (−0·05 to 0·05) Newly diagnosed diabetes −0·05 (−0·16 to 0·06) −0·05 (−0·16 to 0·06) −0·05 (−0·16 to 0·06) Known diabetes −0·12 (−0·21 to −0·02)** −0·11 (−0·21 to − 0·02)†† −0·11 (−0·21 to −0·02)‡‡ Data are beta coefficients (95% CI) based on standardised cognitive scores (mean=0, SD=1); −0·00 occurs because of rounding. Longitudinal analyses are based on data for cognitive function from 1997–99, 2002–04, 2007–09. p values for significant results (p<0·05) compared with reference group are indicated by footnotes. Model 1 is adjusted for age, sex, marital status, and education. Model 2 is adjusted for same parameters as model 1 plus health-related behaviours (smoking, alcohol, physical activity, and fruit and vegetable consumption).

values for significant results (p<0·05) compared with reference group are indicated by footnotes. Model 1 is adjusted for age, sex, marital status, and education. Model 2 is adjusted for same parameters as model 1 plus health-related behaviours (smoking, alcohol, physical activity, and fruit and vegetable consumption). Model 3 is adjusted for same parameters as model 2 plus coronary heart disease, stroke, hypertension, respiratory disease, total cholesterol, obesity, use of antidepressants, and use of lipid-lowering drugs. n for normoglycaemia is 4703; n for prediabetes is 648; n for newly diagnosed diabetes is 115; and n for known diabetes is 187. * p=0·039. † p=0·042. ‡ p=0·046. § p=0.028. ¶ p=0·028. ‖ p=0·026. ** p=0·014. †† p=0·015. ‡‡ p=0·014. Table 4 Association of glycaemic control (one percentage point increment in HbA1c) with estimated differences in cognitive decline, by diabetes status in 1997–99

Model 3 is adjusted for same parameters as model 2 plus coronary heart disease, stroke, hypertension, respiratory disease, total cholesterol, obesity, use of antidepressants, and use of lipid-lowering drugs. n for normoglycaemia is 4703; n for prediabetes is 648; n for newly diagnosed diabetes is 115; and n for known diabetes is 187. * p=0·039. † p=0·042. ‡ p=0·046. § p=0.028. ¶ p=0·028. ‖ p=0·026. ** p=0·014. †† p=0·015. ‡‡ p=0·014. Table 4 Association of glycaemic control (one percentage point increment in HbA1c) with estimated differences in cognitive decline, by diabetes status in 1997–99 Model 1 Model 2 Model 3 Memory Normoglycaemia −0·02 (−0·09 to 0·04) −0·03 (−0·09 to 0·03) −0·02 (−0·08 to 0·04) Prediabetes −0·06 (−0·15 to 0·04) −0·05 (−0·15 to 0·04) −0·05 (−0·15 to 0·04) Newly diagnosed diabetes −0·09 (−0·24 to 0·06) −0·09 (−0·23 to 0·06) −0·09 (−0·23 to 0·06) Known diabetes −0·13 (−0·23 to −0·02)* −0·12 (−0·23 to −0·02)† −0·12 (−0·22 to −0·01)‡ Reasoning Normoglycaemia −0·02 (−0·06 to 0·02) −0·01 (−0·05 to 0·03) −0·02 (−0·06 to 0·02) Prediabetes −0·01 (−0·07 to 0·06) −0·00 (−0·07 to 0·06) 0·00 (−0·07 to 0·06) Newly diagnosed diabetes −0·10 (−0·20 to −0·01)§ −0·11 (−0·20 to −0·01)¶ −0·11 (−0·20 to −0·01)‖ Known diabetes −0·07 (−0·15 to −0·00)** −0·08 (−0·15 to −0·00)†† −0·07 (−0·15 to 0·00) Phonemic fluency Normoglycaemia −0·03 (−0·08 to 0·03) −0·02 (−0·07 to 0·03) −0·01 (−0·07 to 0·04) Prediabetes −0·07 (−0·15 to 0·01) −0·07 (−0·15 to 0·02) −0·06 (−0·14 to 0·02) Newly diagnosed diabetes −0·06 (−0·19 to 0·06) −0·07 (−0·19 to 0·06) −0·07 (−0·19 to 0·06) Known diabetes 0·01 (−0·09 to 0·11) 0·00 (−0·09 to 0·10) 0·01 (−0·09 to 0·10) Semantic fluency Normoglycaemia 0·01 (−0·04 to 0·06) 0·01 (−0·04 to 0·07) 0·01 (−0·04 to 0·06) Prediabetes 0·01 (−0·07 to 0·09) 0·01 (−0·07 to 0·09) 0·01 (−0·07 to 0·09) Newly diagnosed diabetes 0·02 (−0·10 to 0·15) 0·02 (−0·11 to 0·14) 0·02 (−0·11 to 0·14) Known diabetes −0·02 (−0·11 to 0·08) −0·02 (−0·11 to 0·07) −0·02 (−0·12 to 0·07) Global score Normoglycaemia −0·01 (−0·06 to 0·03) −0·01 (−0·05 to 0·03) −0·01 (−0·05 to 0·03) Prediabetes −0·05 (−0·11 to 0·02) −0·04 (−0·11 to 0·02) −0·04 (−0·11 to 0·03) Newly diagnosed diabetes −0·08 (−0·18 to 0·02) −0·08 (−0·18 to 0·02) −0·08 (−0·18 to 0·02) Known diabetes −0·06 (−0·14 to 0·02) −0·07 (−0·15 to 0·01) −0·06 (−0·14 to 0·01) Data are beta coefficients (95% CI) based on standardised cognitive scores (mean=0, SD=1); −0·00 occurs because of rounding. Data for cognitive function are from 1997–99, 2002–04, and 2007–09. p values for significant results (p<0·05) compared with reference group are indicated by footnotes. Model 1 is adjusted for age, sex, marital status, and education.

ased on standardised cognitive scores (mean=0, SD=1); −0·00 occurs because of rounding. Data for cognitive function are from 1997–99, 2002–04, and 2007–09. p values for significant results (p<0·05) compared with reference group are indicated by footnotes. Model 1 is adjusted for age, sex, marital status, and education. Model 2 is adjusted for same parameters as model 1 and health-related behaviours (smoking, alcohol, physical activity, and fruit and vegetable consumption). Model 3 is adjusted for same parameters as model 2 and coronary heart disease, stroke, hypertension, respiratory disease, total cholesterol, obesity, use of antidepressants, and use of lipid-lowering drugs. n for normoglycaemia is 4336; n for prediabetes is 572; n for newly diagnosed diabetes is 100; and n for known diabetes is 152. * p=0·020. † p=0·022. ‡ =0·034. § p=0·034. ¶ p=0·027. ‖ p=0·028. ** p=0·049. †† p=0·040. Panel Research in context Systematic review We searched PubMed for research articles and reviews in English published up to Sept 30, 2013. We used the search terms “type 2 diabetes” and “cognitive decline” in the title or abstract. We also searched the reference lists of retrieved articles and identified additional relevant publications on the link between diabetes, cognitive deficits, and dementia through manual search. We found consistent evidence that the risk of dementia is increased in people with type 2 diabetes.2–4 However, previous studies did not closely examine the relation between diabetes duration and cognitive decline, or the effect of glycaemic control. Interpretation

We searched PubMed for research articles and reviews in English published up to Sept 30, 2013. We used the search terms “type 2 diabetes” and “cognitive decline” in the title or abstract. We also searched the reference lists of retrieved articles and identified additional relevant publications on the link between diabetes, cognitive deficits, and dementia through manual search. We found consistent evidence that the risk of dementia is increased in people with type 2 diabetes.2–4 However, previous studies did not closely examine the relation between diabetes duration and cognitive decline, or the effect of glycaemic control. Interpretation Our results show that a longer diabetes duration is associated with faster cognitive decline. Additionally, for people with diabetes, poor glycaemic control was associated with faster cognitive decline. Thus, interventions that delay diabetes onset, as well as tight glycaemic control in those with established disease, might help to prevent some of the deleterious effects of type 2 diabetes on cognitive ageing.

Introduction Metformin is the recommended first-line oral agent for the treatment of hyperglycaemia in patients with type 2 diabetes, with more than 100 million users worldwide. Despite its impressive safety record and efficacy at the population level, the exact mechanism of metformin action is still elusive and patients' glycaemic responses to metformin vary considerably.1–3 Understanding the source of such variation might help to identify patients most likely not to respond to metformin and could help to develop more effective agents by providing insight into the biological mechanism of metformin.

action is still elusive and patients' glycaemic responses to metformin vary considerably.1–3 Understanding the source of such variation might help to identify patients most likely not to respond to metformin and could help to develop more effective agents by providing insight into the biological mechanism of metformin. As with other complex traits, glycaemic response to metformin is probably determined by the interplay between genetic and environmental factors. Clinical variables such as BMI, drug adherence, and dosing only account for part of the variation.3 Pharmacogenetic studies have identified a few variants in genes affecting metformin action or its pharmacokinetics, yet these variants only account for a small fraction of the variation in metformin response.4–8 Two possible explanations have been suggested for why so little genetic contribution to metformin response variability has been identified. First, it might be because the overall genetic contribution to variation in glycaemic response to metformin is low, with variation mainly due to environmental factors; in this case, trying to improve understanding of the genetic and biological variation in metformin response would have little value. A second explanation is that variation in response to metformin does have a large genetic component, but so far most of the variants with small to moderate effects have not been identified in genetic association studies because of inadequate statistical power; in this case, effort and resources should be invested in an effort to discover the genetic contribution to metformin response because it might enable a truly stratified approach to treatment with this drug. Estimation of the extent of genetic contribution to glycaemic response to metformin—often termed heritability—is of key importance to understand which of these explanations is correct.

iscover the genetic contribution to metformin response because it might enable a truly stratified approach to treatment with this drug. Estimation of the extent of genetic contribution to glycaemic response to metformin—often termed heritability—is of key importance to understand which of these explanations is correct. Historically, the heritability of drug response has rarely been established, largely because of the impracticality of applying the traditional twin and family study designs to drug-response phenotypes; assembling sufficient family members with the same diagnosis who have received the same medication and have been assessed using the same treatment outcome is all but impossible. Alternative methods using population-based genome-wide association study (GWAS) data for heritability estimation have been developed.9–11 One of these methods, genome-wide complex trait analysis (GCTA), can estimate the distant genetic relationship between unrelated individuals using GWAS single-nucleotide polymorphism (SNP) data and can correlate the genetic similarity to the phenotypic similarity, thus partitioning the total phenotypic variance into genetic and environmental causes. Since modern GWAS arrays have good coverage of most common variants in the human genome, the genetic variance estimated by the GCTA method—often referred to as chip heritability—is a good indicator of the additive genetic contribution from common SNPs.12 Because of the insufficient coverage of rare variants on GWAS arrays, heritability estimates by the GCTA method are often lower than the narrow-sense heritability derived from traditional twin and family studies. However, the GCTA method offers a more relevant and accurate estimate of drug-response heritability than other approaches that have been done using cell lines or animal models.13 In this study, we apply the GCTA method to GWAS data from the Genetics of Diabetes Audit and Research in Tayside Scotland (GoDARTS) study14 with the aim of establishing the heritability of glycaemic response to metformin.

sponse heritability than other approaches that have been done using cell lines or animal models.13 In this study, we apply the GCTA method to GWAS data from the Genetics of Diabetes Audit and Research in Tayside Scotland (GoDARTS) study14 with the aim of establishing the heritability of glycaemic response to metformin. Methods Samples In this GCTA study, we used a bioresource linked to electronic health record data (GoDARTS) rather than a specific cohort developed to assess metformin pharmacogenetics. As part of the Wellcome Trust UK type 2 diabetes case-control collection, patients with type 2 diabetes in Tayside, Scotland, have been invited to give written informed consent for DNA collection since October, 1997. So far, nearly 10 000 patients with type 2 diabetes have participated in the GoDARTS study.14 All clinical information about these patients can be obtained in an anonymised form from SCI-Diabetes (an electronic medical record for all patients with diabetes in Scotland) and these data are linked to biochemistry records and prescription encashments from 1992 onwards, giving a comprehensive longitudinal record of diabetes-related therapy. Participants consented for their data to be used in research into diabetes and related disorders, and this bioresource was approved by Tayside Regional Ethics Committee. The bioresource is now governed by Tayside Tissue Bank, which has approved the use of the bioresource for the study of metformin pharmacogenetics.

py. Participants consented for their data to be used in research into diabetes and related disorders, and this bioresource was approved by Tayside Regional Ethics Committee. The bioresource is now governed by Tayside Tissue Bank, which has approved the use of the bioresource for the study of metformin pharmacogenetics. Glycaemic response phenotypes We used HbA1c concentration, which is a routinely measured clinical test of glycaemic control in patients with diabetes, to establish glycaemic response to metformin (appendix); fasting glucose or other non-HbA1c measurements of glycaemic control are not available in the GoDARTS study. Pretreatment (baseline) HbA1c was defined as the measurement closest to, and within 6 months of, the metformin start date (index date), whereas on-treatment HbA1c was defined as the minimum recorded HbA1c achieved within 18 months after the index date. We used four different response phenotypes: absolute reduction in HbA1c, which was the difference between baseline and on-treatment HbA1c; proportional reduction in HbA1c, which was the absolute reduction divided by baseline HbA1c; adjusted reduction in HbA1c, which was the residuals of absolute reduction adjusted by known clinical covariates such as baseline HbA1c, adherence, dose, creatinine clearance, and treatment group; and a dichotomous phenotype of whether or not the target on-treatment HbA1c of <7% (53 mmol/mol) was achieved, with adjustment for baseline HbA1c and known clinical covariates.

of absolute reduction adjusted by known clinical covariates such as baseline HbA1c, adherence, dose, creatinine clearance, and treatment group; and a dichotomous phenotype of whether or not the target on-treatment HbA1c of <7% (53 mmol/mol) was achieved, with adjustment for baseline HbA1c and known clinical covariates. Patients who received metformin monotherapy used no other antidiabetic drugs in the 6 months before the index date or during the study period; sulfonylurea treatment was continued throughout the study period in patients who used metformin as an add-on therapy. The sulfonylurea dose was allowed to vary. Details about how the covariates are defined, and the response models, are outlined in the appendix. GWAS data and quality control GWAS data in the GoDARTS cohort were available from two previous studies. The Wellcome Trust Case Control Consortium 2 study (WTCCC2)8 genotyped 4134 patients with the Affymetrix 6.0 microarray (Santa Clara, CA, USA). The SUrrogate markers for Micro- and Macro-vascular hard endpoints for Innovative diabetes Tools (SUMMIT) study genotyped 3499 patients with the Illumina HumanOmniExpress microarray (Illumina, San Diego, CA, USA). Imputation to the HapMap3 panel—a database of common genetic variants that occur in human beings—and a stringent quality control pipeline were used to combine the two datasets and reduce the systematic discrepancy between the genotypes produced by the two microarrays and their corresponding calling algorithms (appendix).

putation to the HapMap3 panel—a database of common genetic variants that occur in human beings—and a stringent quality control pipeline were used to combine the two datasets and reduce the systematic discrepancy between the genotypes produced by the two microarrays and their corresponding calling algorithms (appendix). We did two benchmark analyses with GCTA to validate the combined GWAS dataset. The first analysis showed that the heritability of human height was 46% (SE 6) in this cohort, which was consistent with previous estimates by studies applying the GCTA method.15 The second analysis estimated the heritability of a pseudo case-control phenotype assuming that samples from one genotyping platform were cases and those from the other platform were controls. As expected, the estimated heritability of this dummy platform phenotype was less than 1% (SE 5), confirming that the original GWAS datasets were combined without introduction of artificial heritability. Heritability estimation We used GCTA version 1.11 to calculate the pair-wise genetic relationship between individuals and create the genetic relationship matrix.13 We then applied principal components analysis to all the SNPs to calculate the first ten eigenvectors, which we included as covariates in all the heritability estimation analyses to control for potential population structure. We then estimated univariate heritability of each drug-response phenotype by the restricted maximum likelihood method in GCTA, with sex and age at index date included as covariates.

n eigenvectors, which we included as covariates in all the heritability estimation analyses to control for potential population structure. We then estimated univariate heritability of each drug-response phenotype by the restricted maximum likelihood method in GCTA, with sex and age at index date included as covariates. Additionally, we used a bivariate analysis to jointly estimate the heritability of baseline HbA1c concentrations and the heritability of on-treatment HbA1c concentrations. The most informative parameter estimated from this bivariate analysis was the genetic correlation (rg), which represents the proportion of variance shared between baseline HbA1c and on-treatment HbA1c concentrations that was contributed by common genetic determinants. The correlation between the residual variance is re, which represents, in part, contribution from environmental factors. We established statistical significance using the likelihood-ratio test of specific hypothesis. We report the asymptotic 95% CI, which was calculated as 1·96 times the SE. Because the SEs of the parameter estimates were derived from first-order Taylor series expansions about the likelihood in GCTA, they might be biased for moderate study sample sizes,15 which at borderline levels of significance explains the discrepancy between p value and 95% CI reported.

alculated as 1·96 times the SE. Because the SEs of the parameter estimates were derived from first-order Taylor series expansions about the likelihood in GCTA, they might be biased for moderate study sample sizes,15 which at borderline levels of significance explains the discrepancy between p value and 95% CI reported. Role of the funding source The sponsor 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 The combined dataset included 1 150 943 autosomal SNPs from 6992 patients. After filtering for cryptic relatedness, 5386 independent individuals were included in the final dataset. Of these, 2085 patients had sufficient clinical data to define their glycaemic response to metformin phenotypes. Table 1 summarises the main characteristics of the 2085 patients included in this study (for sample selection procedure see appendix), stratified by 1465 patients on metformin monotherapy and 620 patients who received metformin as add-on therapy to sulfonylureas.

define their glycaemic response to metformin phenotypes. Table 1 summarises the main characteristics of the 2085 patients included in this study (for sample selection procedure see appendix), stratified by 1465 patients on metformin monotherapy and 620 patients who received metformin as add-on therapy to sulfonylureas. Heritability (h2) for baseline HbA1c was 29% (95% CI −1 to 60; p=0·048 for the null hypothesis of being non-heritable), increasing to 42% (10–73; p=0·0052) for on-treatment HbA1c (table 2). Baseline-adjusted on-treatment HbA1c had a heritability of 36% (95% CI 4–69; p=0·011). Of the four drug-response phenotypes, the model-adjusted reduction in HbA1c (h2=34%, 95% CI 1–68; p=0·022) and the ability to reach target HbA1c (h2=32%, −1 to 64; p=0·030) were the most heritable. The heritability estimates for absolute reduction in HbA1c (h2=23%, 95% CI −8 to 54) and proportional reduction in HbA1c (h2=20%, 95% CI −11 to 51) were smaller, and were not statistically significant (table 2).

CI 1–68; p=0·022) and the ability to reach target HbA1c (h2=32%, −1 to 64; p=0·030) were the most heritable. The heritability estimates for absolute reduction in HbA1c (h2=23%, 95% CI −8 to 54) and proportional reduction in HbA1c (h2=20%, 95% CI −11 to 51) were smaller, and were not statistically significant (table 2). To assess whether the genetic contribution to variation in response to metformin is driven by a few loci with a large effect or many loci with small effect, we did univariate heritability estimations for each chromosome separately for the two glycaemic response phenotypes that were significantly heritable. The genetic contribution to variation in response is distributed across several chromosomes (figure). When the proportion of variance in the model-adjusted reduction in HbA1c attributable to each chromosome (chromosome-wise heritability) was regressed against the chromosome length, we noted a significant linear trend (p=0·037) for longer chromosomes to explain larger proportions of the variance. We also noted a similar trend (p=0·034) for achievement of target HbA1c. Bivariate analysis of baseline and on-treatment HbA1c concentrations estimated a moderate genetic correlation (rg) of 0·58 (95% CI 0·06–1·09) between these two traits (table 3). Likelihood ratio tests showed that the genetic correlation was statistically greater than 0 (p=0·053) and marginally less than 1 (p=0·097); where 0 would mean no genetic correlation and 1 would represent 100% genetic correlation.

te genetic correlation (rg) of 0·58 (95% CI 0·06–1·09) between these two traits (table 3). Likelihood ratio tests showed that the genetic correlation was statistically greater than 0 (p=0·053) and marginally less than 1 (p=0·097); where 0 would mean no genetic correlation and 1 would represent 100% genetic correlation. Discussion This study is, to our knowledge, the first to show that genetic differences contribute considerably to the variation noted in patients' glycaemic response to metformin (panel). The heritability estimates for the frequently used definitions of glycaemic response range from 20% to 34%, suggesting that genetic variants are likely to have an important contribution to variation in glycaemic response to metformin in patients with type 2 diabetes. In the context of GCTA estimates for other complex traits with well established heritability by family or twin studies, the point estimates are similar to GCTA estimates for schizophrenia (h2=23% [SE 1]) and Alzheimer's disease (h2=30% [SE 3]),16,17 suggesting that genetic variants contribute to the variation in HbA1c response to metformin to a similar extent.

complex traits with well established heritability by family or twin studies, the point estimates are similar to GCTA estimates for schizophrenia (h2=23% [SE 1]) and Alzheimer's disease (h2=30% [SE 3]),16,17 suggesting that genetic variants contribute to the variation in HbA1c response to metformin to a similar extent. We did the chromosome-wise heritability estimation to provide information about the genetic architecture of glycaemic response to metformin. Clearly, several variants across different chromosomes contribute to the metformin response variation. The finding that the contribution by an individual chromosome is significantly correlated to its length suggests that on each chromosome might be many variants with a small to moderate effect size rather than a few variants with major effect. This hypothesis is also supported by results of the metformin response GWAS, which reported that no individual variant explained a large proportion of the variance.8 Notably, the point estimates of chromosome-wise heritability all have large 95% CIs and the estimates for each chromosome vary between the two different response phenotypes (figure). Thus, individual extreme values, such as the estimate of chromosome 1 in the analysis of model-adjusted reduction in HbA1c, could have had an undue effect on the reported trend.

me-wise heritability all have large 95% CIs and the estimates for each chromosome vary between the two different response phenotypes (figure). Thus, individual extreme values, such as the estimate of chromosome 1 in the analysis of model-adjusted reduction in HbA1c, could have had an undue effect on the reported trend. In the univariate GCTA analysis we were able to assess whether different metformin response phenotypes are heritable. We do not have statistical power to conclude that one phenotype is more heritable than another, although the point estimates for the heritability of the response phenotypes that adjusted for the baseline HbA1c were greater than for the unadjusted models. Because baseline HbA1c has been well documented to have a major effect on the absolute reduction phenotype,18 the higher heritability estimates for baseline-adjusted phenotypes of metformin efficacy are likely to be a result of the successful adjustment for common environmental variance between baseline and on-treatment HbA1c measurements. These adjusted phenotypes probably best address the pharmacogenetics of metformin when considering what factors are associated with the greatest reduction in HbA1c for a given HbA1c concentration before metformin initiation. An adjusted phenotype was used to successfully identify variants near the ATM locus that affect on-treatment HbA1c but not baseline HbA1c.8 However, such definitions adjusted for baseline HbA1c do capture some of the shared genetic component (rg) described above, and thus identified variants might reflect not only the response to metformin, but also the variance in HbA1c per se. When considering what the biological determinants of response to metformin are, a better phenotype might be the unadjusted absolute reduction in HbA1c because this measure does not capture any shared genetic contribution, only the variants with differential genetic effects between the HbA1c before and after initiation of metformin treatment.19 Such an HbA1c reduction model unadjusted for the baseline measure has been used in studies of statin pharmacogenetics.20,21 However, the heritability for the absolute reduction in HbA1c did not achieve statistical significance in our study of response to metformin treatment.

d after initiation of metformin treatment.19 Such an HbA1c reduction model unadjusted for the baseline measure has been used in studies of statin pharmacogenetics.20,21 However, the heritability for the absolute reduction in HbA1c did not achieve statistical significance in our study of response to metformin treatment. We report a new application of the GCTA bivariate analysis in this drug-response study. This approach has advantages over univariate approaches because it does not make assumptions about the response model; rather, it uses a quantitative genetic approach to partition variance into genetic and environmental fractions that are shared and non-shared between two states or traits. The idea behind such an analysis is that intervention with metformin can change the physiological state of a patient. In the pretreatment state, a set of genetic and environmental factors determine the HbA1c variation; a potentially different set of genetic and environmental determinants affect the on-treatment state (appendix). The bivariate analysis can tell us not only how much of the HbA1c variance is genetically determined in each state, but also how much of the genetically determined HbA1c variance is shared between the two physiological states, as estimated by genetic correlation (rg). The shared variants that underlie the genetic correlation have the same effect on HbA1c variation in the two states, and their genetic contribution to HbA1c is not changed by metformin treatment. Thus an rg of 1 would imply that metformin intervention does not change the genetic determinants of HbA1c in the pretreatment and on-treatment state—ie, no pharmacogenetic effect occurs.20 By contrast, a low rg would imply that the genetic determinants of HbA1c are largely different before and after metformin treatment, hence a strong pharmacogenetic effect. In our analysis, the point estimate for the shared genetic contribution was 0·58, suggesting that around half of the genetic determinants contributing to baseline HbA1c and on-treatment HbA1c concentrations were shared, with half the genetic determinants differing between the baseline and metformin treatment states; however, we do acknowledge that the 95% CI precludes a definitive conclusion of this bivariate analysis.

nd half of the genetic determinants contributing to baseline HbA1c and on-treatment HbA1c concentrations were shared, with half the genetic determinants differing between the baseline and metformin treatment states; however, we do acknowledge that the 95% CI precludes a definitive conclusion of this bivariate analysis. A key limitation of this study is the reasonably small sample size. However, the GoDARTS GWAS data used are from the largest metformin pharmacogenomic cohort done so far, including 2085 individuals who received metformin. Yet we still noted considerable 95% CIs for the heritability and genetic correlation estimates due to the limited sample size. Thus, despite having sufficient power to find that glycaemic response to metformin is a heritable trait, we do not have power to establish whether one drug response trait is more heritable than another. To do this, 4450 patients would be needed to statistically differentiate true heritability of 20% and 34%, which correspond to the two extremes of the estimated heritability of our four reported phenotypes.22 This shortfall in available data emphasises the importance of a consortium effort to assemble even more GWAS data, which will enable us to not only achieve more accurate estimates of heritability, but also discover more genetic variants that account for this heritability. The Metformin Genetics Consortium (MetGen) consists of research groups in Europe and the USA that have cohorts available for the study of the genetics of metformin. This consortium currently consists of about 5600 patients who have received metformin, and hopefully in the next 2–3 years additional academic and commercial clinical trial data and observational data might enable a GWAS of about 8000 individuals. Of note, interpretations of the heritability estimates from our current GCTA analyses can only be made in the context of the SNPs captured by the GWAS arrays. Contributions from the rare variants that are poorly covered by the GWAS panels will not form part of the heritability estimated by GCTA, but will remain in the environmental component. The observed environmental (residual) correlation (re) of 0·28 could be contributed by both environmental factors and shared rare genetic variants. Thus sequencing-based genomic studies with an emphasis on the rare drug-response variants are valid irrespective of the heritability estimates from GWAS SNPs.

al component. The observed environmental (residual) correlation (re) of 0·28 could be contributed by both environmental factors and shared rare genetic variants. Thus sequencing-based genomic studies with an emphasis on the rare drug-response variants are valid irrespective of the heritability estimates from GWAS SNPs. In summary, using GWAS data from 2085 patients with type 2 diabetes, our analysis showed that genetic variants contributed to the variation in glycaemic response to metformin, with the heritability of metformin response estimated at up to 34% (95% CI 1–68; p=0·022). This result shows that a moderate proportion of the variance in glycaemic response is genetic, and represents underlying biological differences between individuals. The variants are likely to have a small-to-moderate effect and be scattered across the genome. So far, very little of the genetic contribution to metformin response has been identified; GWAS analyses with larger samples could find more genetic variants that enable better predictions to be made for personalised or stratified medicine, and unravel new mechanisms of metformin action in the reduction of hyperglycaemia. Supplementary Material Supplementary appendix

In summary, using GWAS data from 2085 patients with type 2 diabetes, our analysis showed that genetic variants contributed to the variation in glycaemic response to metformin, with the heritability of metformin response estimated at up to 34% (95% CI 1–68; p=0·022). This result shows that a moderate proportion of the variance in glycaemic response is genetic, and represents underlying biological differences between individuals. The variants are likely to have a small-to-moderate effect and be scattered across the genome. So far, very little of the genetic contribution to metformin response has been identified; GWAS analyses with larger samples could find more genetic variants that enable better predictions to be made for personalised or stratified medicine, and unravel new mechanisms of metformin action in the reduction of hyperglycaemia. Supplementary Material Supplementary appendix Acknowledgments We are grateful to all the participants who took part in this study, the general practitioners, the Scottish School of Primary Care for their help in recruiting the participants, and the whole team, which includes interviewers, computer and laboratory technicians, clerical workers, research scientists, volunteers, managers, receptionists, and nurses. The Wellcome Trust provides support for Wellcome Trust UK type 2 diabetes case-control collection (GoDARTS; 099177/Z/12/12) and informatics support is provided by the Chief Scientist Office. The Wellcome Trust funds the Scottish Health Informatics Programme, provides core support for the Wellcome Trust Centre for Human Genetics in Oxford and funds the Wellcome Trust Case Control Consortium 2 (084726/Z/08/Z). The research leading to these results has received support from the Innovative Medicines Initiative Joint Undertaking under grant agreement number 115006 (IMI-SUMMIT), resources of which are composed of financial contributions from the European Union's Seventh Framework Programme (FP7/2007-2013) and European Federation of Pharmaceutical Industries and Associations companies' kind contributions. PCS is supported by the Hong Kong Research Grants Council General Research Fund project (grants 777511 and 776513), and European Commission Seventh Framework Programme grant for the European Network of National Schizophrenia Networks Studying Gene-Environmental Interactions (EU-GEI). ERP holds a Wellcome Trust New Investigator award. MIMcC holds a National Institute for Health Research Senior Investigator award and a Wellcome Trust Senior Investigator award (098381). This research was specifically funded by the Wellcome Trust (092272/Z/10/Z) for a Henry Wellcome Post-Doctoral Fellowship to KZ.

EI). ERP holds a Wellcome Trust New Investigator award. MIMcC holds a National Institute for Health Research Senior Investigator award and a Wellcome Trust Senior Investigator award (098381). This research was specifically funded by the Wellcome Trust (092272/Z/10/Z) for a Henry Wellcome Post-Doctoral Fellowship to KZ. Contributors KZ, ERP, and CNAP designed the study and wrote the manuscript. HD, NVZ, EA, LG, MIMcC, and HMC undertook the SUMMIT (SUrrogate markers for Micro- and Macro-vascular hard endpoints for Innovative diabetes Tools) genome-wide association study (GWAS), were responsible for its quality control, and supplied the data. The Wellcome Trust Case Control Consortium 2 (WTCC2) and CCS undertook the WTCCC2 GWAS and supplied the data. KZ analysed the data, and ERP and CNAP assisted in data interpretation. LD, JY, ML, and PCS assisted in the analysis and interpretation of the results. LD, JY, ML, PCS, HD, NVZ, EA, LG, MIMcC, HMC, CCS, and ADM edited the manuscript. Declaration of interests We declare that we have no competing interests. Figure Chromosome-wise heritability estimation for glycaemic response to metformin

Contributors KZ, ERP, and CNAP designed the study and wrote the manuscript. HD, NVZ, EA, LG, MIMcC, and HMC undertook the SUMMIT (SUrrogate markers for Micro- and Macro-vascular hard endpoints for Innovative diabetes Tools) genome-wide association study (GWAS), were responsible for its quality control, and supplied the data. The Wellcome Trust Case Control Consortium 2 (WTCC2) and CCS undertook the WTCCC2 GWAS and supplied the data. KZ analysed the data, and ERP and CNAP assisted in data interpretation. LD, JY, ML, and PCS assisted in the analysis and interpretation of the results. LD, JY, ML, PCS, HD, NVZ, EA, LG, MIMcC, HMC, CCS, and ADM edited the manuscript. Declaration of interests We declare that we have no competing interests. Figure Chromosome-wise heritability estimation for glycaemic response to metformin Chromosome-wise heritability plotted for whether or not the target of on-treatment HbA1c<7% (53 mmol/mol) was achieved (A), and for model-adjusted reduction in HbA1c—ie, residuals of absolute reduction adjusted by known clinical covariates (B). The circled numbers show the heritability point estimates of each chromosome (sex chromosomes were not included). The solid lines plot the linear regression of chromosome-wise heritability against chromosome length; the dotted lines show 95% CI. Table 1 Sample characteristics

Chromosome-wise heritability plotted for whether or not the target of on-treatment HbA1c<7% (53 mmol/mol) was achieved (A), and for model-adjusted reduction in HbA1c—ie, residuals of absolute reduction adjusted by known clinical covariates (B). The circled numbers show the heritability point estimates of each chromosome (sex chromosomes were not included). The solid lines plot the linear regression of chromosome-wise heritability against chromosome length; the dotted lines show 95% CI. Table 1 Sample characteristics Metformin monotherapy (n=1465) Metformin plus sulfonylureas (n=620) Age, years 61·4 (10·5) 65·4 (9·4) Men 836 (57%) 390 (63%) BMI, kg/m2 32·6 (5·6) 29·1 (4·9) Baseline HbA1c, % 8·7 (1·3) 9·2 (1·3) Baseline to metformin,* days 18 (29) 21 (30) On-treatment HbA1c, % 7·0 (1·0) 7·4 (1·1) Metformin dose, g/day 1·26 (0·47) 1·29 (0·51) Adherence, % 78·4 (16·6) 78·3 (11·1) Creatinine clearance, mL/min 96·1 (32·7) 79·5 (27·0) HbA1c measurements, n 3·9 (1·8) 4·2 (1·9) Data are mean (SD) or number (%). * Time from baseline measurement of HbA1c to initiation of metformin treatment. Table 2 Univariate heritability estimates of glycaemic response to metformin

Metformin monotherapy (n=1465) Metformin plus sulfonylureas (n=620) Age, years 61·4 (10·5) 65·4 (9·4) Men 836 (57%) 390 (63%) BMI, kg/m2 32·6 (5·6) 29·1 (4·9) Baseline HbA1c, % 8·7 (1·3) 9·2 (1·3) Baseline to metformin,* days 18 (29) 21 (30) On-treatment HbA1c, % 7·0 (1·0) 7·4 (1·1) Metformin dose, g/day 1·26 (0·47) 1·29 (0·51) Adherence, % 78·4 (16·6) 78·3 (11·1) Creatinine clearance, mL/min 96·1 (32·7) 79·5 (27·0) HbA1c measurements, n 3·9 (1·8) 4·2 (1·9) Data are mean (SD) or number (%). * Time from baseline measurement of HbA1c to initiation of metformin treatment. Table 2 Univariate heritability estimates of glycaemic response to metformin n Heritability (h2) 95% CI p value* HbA1cconcentrations Baseline HbA1c 2085 29% −1 to 60 0·048 On-treatment HbA1c 2085 42% 10 to 73 0·0052 Adjusted on-treatment HbA1c 2085 36% 4 to 69 0·011 Response phenotypes Absolute reduction in HbA1c 2085 23% −8 to 54 0·074 Proportional reduction in HbA1c 2085 20% −11 to 51 0·10 Adjusted reduction in HbA1c 2069 34% 1 to 68 0·022 Achieved target HbA1c concentration† 1942 32% −1 to 64 0·030 * p values are from likelihood tests of null hypothesis of heritability being 0. † The sample size was reduced to 1942 because some patients had a baseline HbA1c concentration of 7% or lower. Table 3 Bivariate analysis of baseline and on-treatment HbA1c Point estimate 95% CI Baseline HbA1c 0·29 −0·02 to 0·60 On-treatment HbA1c 0·42 0·11 to 0·73 rg 0·58 0·06 to 1·09 re 0·28 −0·02 to 0·58 The point estimates of baseline and on-treatment HbA1c are for heritability; correlation for rg (genetic) and re (environmental).

Table 3 Bivariate analysis of baseline and on-treatment HbA1c Point estimate 95% CI Baseline HbA1c 0·29 −0·02 to 0·60 On-treatment HbA1c 0·42 0·11 to 0·73 rg 0·58 0·06 to 1·09 re 0·28 −0·02 to 0·58 The point estimates of baseline and on-treatment HbA1c are for heritability; correlation for rg (genetic) and re (environmental). Panel Research in context Systematic review We searched PubMed on Feb 18, 2013, with the search terms “heritability” and “metformin”. We did not apply any publication date or language restrictions. We found no previous reports on the heritability of glycaemic response to metformin. Interpretation This study is, to our knowledge, the first to establish that glycaemic response to metformin is likely to be moderately heritable. Enhanced GWAS studies will identify more variants, enabling better response predictions to be made, and will identify new mechanisms of metformin action in the reduction of hyperglycaemia in the treatment of type 2 diabetes.

Introduction Type 1 diabetes is an autoimmune disorder targeting pancreatic β cells that secrete insulin.1 It is one of the most common chronic diseases of children, and the incidence is increasing worldwide.2 For a time after diagnosis, some β-cell function remains, although not enough to maintain euglycaemia. In this context, even slight protection of residual β-cell function can be expected to have clinically significant benefits.3 Although insulin is life-saving, no treatment is available to address the underlying disease process. Various immune therapies have been shown to slow the progressive loss of pancreatic islet β-cell function and insulin secretion after disease onset in type 1 diabetes.4–11 Unfortunately, none of these immunomodulatory drugs have induced lasting disease remission.

is available to address the underlying disease process. Various immune therapies have been shown to slow the progressive loss of pancreatic islet β-cell function and insulin secretion after disease onset in type 1 diabetes.4–11 Unfortunately, none of these immunomodulatory drugs have induced lasting disease remission. We have shown that combination therapy with the gastrointestinal hormones glucagon-like peptide-1 (GLP-1) and gastrin increased β-cell mass and restored normoglycaemia in non-obese diabetic (NOD) mice.12 Combination therapy with a dipeptidyl peptidase-4 (DPP-4) inhibitor that raised blood concentrations of GLP-1 and a proton-pump inhibitor (PPI) that raised blood concentrations of gastrin restored normoglycaemia in NOD mice.13 In preparation for translating these results to clinical studies, we showed that these drugs had a similar effect on human islets engrafted into immunodeficient diabetic mice: GLP-1 and gastrin induced β-cell neogenesis from adult human pancreatic exocrine duct cells.14 Therapy with a combination of a DPP-4 inhibitor and PPI stimulated human β-cell neogenesis in these mice.15 A DPP-4 inhibitor in NOD mice reduced insulitis and increased CD4+ CD25+ FoxP3+ cells.16

munodeficient diabetic mice: GLP-1 and gastrin induced β-cell neogenesis from adult human pancreatic exocrine duct cells.14 Therapy with a combination of a DPP-4 inhibitor and PPI stimulated human β-cell neogenesis in these mice.15 A DPP-4 inhibitor in NOD mice reduced insulitis and increased CD4+ CD25+ FoxP3+ cells.16 On the basis of these findings, we postulated that combination therapy with a DPP-4 inhibitor and a PPI would increase GLP-1 and gastrin concentrations in patients. In turn, these increases would act through both direct actions on β cells that promote growth and survival, and modulation of the immunological mechanisms that destroy β cells. We also sought to establish whether combination of a DPP-4 inhibitor and a PPI might improve glucose control, while decreasing insulin use, which are outcomes expected if β-cell function were improved by the intervention. Methods Study design and participants REPAIR-T1D is a parallel-group, randomised, double-blind, placebo-controlled, multicentre study. We recruited participants at Sanford Health Systems (Sioux Falls, SD, USA; Fargo, ND, USA), Children's Hospitals and Clinics of Minnesota (St Paul, MN, USA), and Rady Children's Hospital (San Diego, CA, USA).