Browse the corpus

Introduction Tuberculosis (TB) remains a health challenge in most low- and middle-income countries, including those in sub-Saharan Africa. Poverty, smoking, alcoholism, HIV-1 and diabetes mellitus are known drivers of the TB epidemic [1]. Diabetes triples the risk of TB and is associated with adverse outcomes [2–4]. The TB/diabetes association is increasingly important in low- and middle-income countries experiencing a growing burden of noncommunicable disease. According to the 2013 Global Burden of Disease Study, HIV-1, TB and diabetes rank first, third and ninth among the top 10 causes of life-years lost in South Africa [5]. The prevalence of diabetes among TB patients varies by geographic region and background diabetes population in the general population, with a higher prevalence reported in South Asian (25.3–44%) and Latin American (14–39%) populations [6–12]. In sub-Saharan Africa, diabetes prevalence in TB patients ranges from 8.5% to 16.4%, although a large proportion of diabetes in sub-Saharan Africa remains undiagnosed [13–17]. Of note, most studies have used fasting or random blood glucose to diagnose diabetes, which may underestimate the true prevalence. A study in South Africa estimated the general population prevalence of diabetes and impaired glucose tolerance to be 12.3% and 11.2%, respectively [18].

Africa remains undiagnosed [13–17]. Of note, most studies have used fasting or random blood glucose to diagnose diabetes, which may underestimate the true prevalence. A study in South Africa estimated the general population prevalence of diabetes and impaired glucose tolerance to be 12.3% and 11.2%, respectively [18]. The increased risk of TB associated with diabetes is well documented. However, this risk is variable dependent on the population studied. A systematic review that reported a 3-fold increased risk of TB in diabetes patients did not include studies from sub-Saharan Africa where there are competing TB risk factors such as HIV-1 [2]. Recent studies in African settings showed increased odds of TB among diabetes patients with the exception of one study [19], possibly due to the low background prevalence of diabetes. A Ugandan study found HIV-1-infected TB patients to have reduced odds of diabetes compared with HIV-1-uninfected patients [20] and a Tanzanian study showed a stronger TB/diabetes association observed among HIV-1-uninfected (OR 4.23, 95% CI 1.54–11.57) compared with HIV-1-infected (OR 0.14, 95% CI 0.01–1.81) individuals [21]. The reasons for this apparent incongruity in the associations between TB, diabetes and HIV-1 are unclear as there are only a limited number of published studies [22].

betes association observed among HIV-1-uninfected (OR 4.23, 95% CI 1.54–11.57) compared with HIV-1-infected (OR 0.14, 95% CI 0.01–1.81) individuals [21]. The reasons for this apparent incongruity in the associations between TB, diabetes and HIV-1 are unclear as there are only a limited number of published studies [22]. The only South African study that assessed the prevalence of diabetes, using the oral glucose tolerance test (OGTT), among TB patients was conducted in 1980, when South Africa was 42% urbanised (compared with 62% urbanised today), and reported a 2.1% prevalence [16]. South Africa has the largest antiretroviral therapy (ART) programme globally and is ranked as having the sixth highest TB incidence globally [23, 24]. Given this HIV-1/TB burden, reducing the TB burden in South Africa is a priority in the global fight against TB. The strong association between HIV-1 and TB, dysglycaemic effects of some ART drugs, especially protease inhibitors and nonnucleoside reverse transcriptase inhibitors [25, 26], and the emerging diabetes epidemic highlight the importance of investigating the association between TB, diabetes and HIV-1 in this setting, the results of which will have implications for TB control strategies. This study investigated the prevalence of diabetes and impaired glucose regulation (IGR), the association between TB and diabetes/IGR, and the population-attributable risk of TB due to diabetes in South Africa.

B, diabetes and HIV-1 in this setting, the results of which will have implications for TB control strategies. This study investigated the prevalence of diabetes and impaired glucose regulation (IGR), the association between TB and diabetes/IGR, and the population-attributable risk of TB due to diabetes in South Africa. Methods Study setting and population In South Africa, TB patients in the public sector are largely screened and treated in TB clinics. This study was conducted at the largest TB clinic in Khayelitsha, a peri-urban township of around 390 000 predominantly black Africans, in Cape Town, Western Cape province. In this province, diabetes, HIV-1 and TB rank as the first, third and fourth leading causes of death, respectively [27]. The 2012 HIV-1 antenatal prevalence in Khayelitsha was 34% (95% CI 31.0–36.6%) (Western Cape Dept of Health, Cape Town, South Africa; 2012 Antenatal Survey, unpublished data) and the 2015 TB case notification rate was 917 per 100 000 population (V. de Azevedo, City Health Manager, Khayelitsha, Cape Town, South Africa; personal communication), with a 60% HIV-1/TB co-infection rate [28].

1.0–36.6%) (Western Cape Dept of Health, Cape Town, South Africa; 2012 Antenatal Survey, unpublished data) and the 2015 TB case notification rate was 917 per 100 000 population (V. de Azevedo, City Health Manager, Khayelitsha, Cape Town, South Africa; personal communication), with a 60% HIV-1/TB co-infection rate [28]. Study design and sampling We conducted a cross-sectional study on consecutive patients with respiratory symptoms presenting to the clinic from July 2013 from August 2015. Patients were eligible if they provided consent, were aged ≥18 years and had not received >48 h of TB chemotherapy. Those who were critically ill and in need of emergency clinical care were ineligible as they were too physically unwell to give informed consent. The study was approved by the University of Cape Town Human Research Ethics Committee (HREC 403/2011). Assuming a 3-fold higher diabetes prevalence in HIV-1-uninfected TB cases, 7% diabetes prevalence in HIV-1-uninfected individuals, 70% of TB cases HIV-1-co-infected and 80% power, the study aimed to recruit 400 TB and 400 non-TB participants.

the University of Cape Town Human Research Ethics Committee (HREC 403/2011). Assuming a 3-fold higher diabetes prevalence in HIV-1-uninfected TB cases, 7% diabetes prevalence in HIV-1-uninfected individuals, 70% of TB cases HIV-1-co-infected and 80% power, the study aimed to recruit 400 TB and 400 non-TB participants. Study procedures Case definitions TB cases were diagnosed according to South Africa guidelines with the GeneXpert system (Cepheid, Sunnyvale, CA, USA) and analysed in a centralised national health laboratory [29]. Non-TB participants were those with a negative TB diagnosis and resolution of respiratory symptoms. All participants were encouraged to undertake an HIV-1 test and were tested for diabetes using all three of the following: fasting plasma glucose (FPG), a 2-h OGTT and glycated haemoglobin (HbA1c). Diabetes diagnosis was defined as self-reported diabetes, FPG ≥7.0 mmol·L−1, OGTT ≥11.1 mmol·L−1 or HbA1c ≥6.5% [30, 31]. IGR was defined as FPG 5.5–<7.0 mmol·L−1, OGTT 7.7–<11.1 mmol·L−1 or HbA1c 5.7–<6.5%.

tes using all three of the following: fasting plasma glucose (FPG), a 2-h OGTT and glycated haemoglobin (HbA1c). Diabetes diagnosis was defined as self-reported diabetes, FPG ≥7.0 mmol·L−1, OGTT ≥11.1 mmol·L−1 or HbA1c ≥6.5% [30, 31]. IGR was defined as FPG 5.5–<7.0 mmol·L−1, OGTT 7.7–<11.1 mmol·L−1 or HbA1c 5.7–<6.5%. Measurements Venous blood was drawn from the antecubital vein at 0 (after an overnight fast) and 120 min in evacuated fluoride (glucose) and EDTA (HbA1c) tubes. All blood samples were processed on the day of collection at a centralised national health laboratory using standardised operating procedures of the cobas c311 (Roche/Hitachi, Basel, Switzerland) system analyser assay. Weight, height and waist circumference were measured using standardised techniques [32]. Body mass index (BMI) was categorised as underweight <18.5 kg·m−2, normal 18.5–24.9 kg·m−2, overweight 25–29.9 kg·m−2 and obese ≥30 kg·m−2 [32]. The cut-point for high waist circumference was ≥94 cm for males and ≥88 cm for females [32]. Hypertension was defined as a single measured blood pressure of systolic blood pressure >140 mmHg or diastolic blood pressure >90 mmHg [31], or a pre-existing diagnosis. Questionnaire Chronic disease risk factors were ascertained using the STEPs instrument [33]. Information on socioeconomic, demographic and medical history was collected using a researcher-administered questionnaire.

Measurements Venous blood was drawn from the antecubital vein at 0 (after an overnight fast) and 120 min in evacuated fluoride (glucose) and EDTA (HbA1c) tubes. All blood samples were processed on the day of collection at a centralised national health laboratory using standardised operating procedures of the cobas c311 (Roche/Hitachi, Basel, Switzerland) system analyser assay. Weight, height and waist circumference were measured using standardised techniques [32]. Body mass index (BMI) was categorised as underweight <18.5 kg·m−2, normal 18.5–24.9 kg·m−2, overweight 25–29.9 kg·m−2 and obese ≥30 kg·m−2 [32]. The cut-point for high waist circumference was ≥94 cm for males and ≥88 cm for females [32]. Hypertension was defined as a single measured blood pressure of systolic blood pressure >140 mmHg or diastolic blood pressure >90 mmHg [31], or a pre-existing diagnosis. Questionnaire Chronic disease risk factors were ascertained using the STEPs instrument [33]. Information on socioeconomic, demographic and medical history was collected using a researcher-administered questionnaire. Statistical analysis Medians (interquartile ranges (IQRs)) and proportions were used to summarise continuous and categorical variables, respectively. Chi-squared and Fisher's exact tests assessed associations between categorical variables. The Mann–Whitney test was used to compare medians between two groups and the Kruskal–Wallis test for more than two groups. A multivariable logistic regression model for the association between TB and diabetes was built manually, using forward selection, controlling for potential confounding variables. In order to retain statistical power and reduce potential biases, in the regression analysis involving the key variable HIV serostatus, multiple imputation was used to impute HIV serostatus for 50 participants with unknown HIV-1 status. Imputation using the chained equations method was implemented in Stata (StataCorp, College Station, TX, USA) using the “ice” command. Five imputed datasets were created. The “mi” Stata command was used to perform logistic regression analysis on the combined imputed datasets. The base model included a priori confounding variables of age and sex. Variables associated with TB (p<0.10) on univariable analysis were then added, sequentially including variables that improved the model based on the significant lowering of the Akaike Information Criterion in nonnested model comparisons. Potential effect modification between variables was determined by exploring the statistical significance of interaction variables. The fit of the model was assessed using Pearson's goodness-of-fit test with p>0.05 indicating a good fit. The population-attributable risk was calculated as: ARpopdiabetes=p(diabetes)(OR–1)/(1+p(diabetes)(OR–1)), where p(diabetes) is general population diabetes prevalence and OR is the adjusted odds ratio [34]. Statistical significance was set at p<0.05. All data were analysed using Stata version 13.0.

dicating a good fit. The population-attributable risk was calculated as: ARpopdiabetes=p(diabetes)(OR–1)/(1+p(diabetes)(OR–1)), where p(diabetes) is general population diabetes prevalence and OR is the adjusted odds ratio [34]. Statistical significance was set at p<0.05. All data were analysed using Stata version 13.0. Results Baseline characteristics 986 participants were recruited. 48 participants (4.9%) were excluded as they had an undetermined TB status (TB could not be confirmed or excluded). A further 86 participants did not complete diabetes screening at baseline (did not return for fasting bloods). 852 participants were therefore included in the final analyses: 414 TB cases and 438 non-TB participants. The overall median (IQR) age of participants was 38 (31–47) years, with 53% male. The overall prevalence of HIV-1 was 61.2% (95% CI 29.9–36.2%), and was significantly higher in participants with TB (versus non-TB) and in females (table 1). Compared with those without TB, TB cases were younger, with a lower prevalence of obesity and a greater proportion of men. Of the 414 TB cases, nine had rifampicin resistance (2.2%), all of whom did not have diabetes. Table 1 summarises baseline characteristics of participants, stratified by TB status. TABLE 1 Baseline sociodemographic, anthropometric and comorbidity characteristics by tuberculosis (TB) status

Results Baseline characteristics 986 participants were recruited. 48 participants (4.9%) were excluded as they had an undetermined TB status (TB could not be confirmed or excluded). A further 86 participants did not complete diabetes screening at baseline (did not return for fasting bloods). 852 participants were therefore included in the final analyses: 414 TB cases and 438 non-TB participants. The overall median (IQR) age of participants was 38 (31–47) years, with 53% male. The overall prevalence of HIV-1 was 61.2% (95% CI 29.9–36.2%), and was significantly higher in participants with TB (versus non-TB) and in females (table 1). Compared with those without TB, TB cases were younger, with a lower prevalence of obesity and a greater proportion of men. Of the 414 TB cases, nine had rifampicin resistance (2.2%), all of whom did not have diabetes. Table 1 summarises baseline characteristics of participants, stratified by TB status. TABLE 1 Baseline sociodemographic, anthropometric and comorbidity characteristics by tuberculosis (TB) status Non-TB TB Total# p-value Subjects 438 414 852 Age group years <0.001* 18–24 16 (3.7) 43 (10.4) 59 (6.9) 25–34 126 (28.8) 151 (36.5) 277 (32.5) 35–44 126 (28.8) 138 (33.3) 264 (31.0) 45–54 95 (21.7) 53 (12.8) 148 (17.4) ≥55 75 (17.1) 29 (7.0) 104 (12.2) Age years 41 (32–50) 36 (30–43) 38 (31–47) <0.001* (range 20–80) (range 18–80) (range 18–80) Female n=848 224 (51.3) 175 (42.6) 399 (47.0) 0.011* Education level n=827 0.268 Up to primary 130 (30.4) 129 (32.3) 259 (31.3) Up to secondary 289 (67.7) 257 (64.3) 456 (66.0) Higher education 8 (1.9) 14 (3.5) 22 (2.7) Single n=826 276 (64.8) 299 (74.8) 575 (69.6) 0.002* Unemployed n=825 235 (55.0) 213 (53.5) 448 (54.3) 0.662 Household size n=802 0.031* 0–2 individuals 214 (51.7) 230 (59.3) 444 (55.4) >2 individuals 200 (48.3) 158 (40.7) 358 (44.6) Income categories ZAR n=730 <0.001* 0 21 (5.5) 9 (2.6) 30 (4.1) 1–1600 248 (64.8) 160 (46.1) 408 (55.9) 1601–3200 71 (18.5) 103 (29.7) 174 (23.8) 3201–6400 34 (8.9) 62 (17.9) 96 (13.2) 6401–12 800 8 (2.1) 12 (3.5) 20 (2.7) ≥12 801 1 (0.3) 1 (0.3) 2 (0.3) Binge drinking among drinkers 425 (97.0) 403 (97.3) 828 (97.8) 0.784 Current smoker n=828 123 (28.7) 90 (22.5) 213 (25.7) 0.04* Prison history n=837 21 (4.9) 42 (10.3) 63 (7.5) 0.003* Miner (past or present) n=833 17 (4.0) 7 (1.7) 24 (2.9) 0.053 Health care worker 8 (1.9) 7 (1.7) 15 (1.8) 0.889 TB contact n=836 54 (12.6) 51 (12.6) 105 (12.6) 0.999 Previous TB n=830 196 (46.0) 129 (31.9) 325 (39.2) <0.001* Previous diabetes n=838 19 (4.4) 20 (4.9) 39 (4.7) 0.717 HIV-1 status <0.001* Uninfected 160 (36.5) 121 (29.2) 281 (33.0) Infected 242 (55.3) 279 (67.4) 521 (61.2) Unknown 36 (8.2) 14 (3.4) 50 (5.9) ART (among HIV-1-infected) n=521 166 (68.6) 89 (31.9) 255 (48.9) <0.001 Previous gestational diabetes (among females) n=399 4 (1.6) 8 (4.6) 12 (3.0) 0.105 Hypertension 154 (35.2) 75 (18.1) 229 (26.9) <0.001* BMI kg·m−2 n=810 <0.001* <18.5 (underweight) 27 (6.6) 41 (10.3) 68 (8.4) 18.5–24.9 (normal) 224 (54.4) 277 (69.6) 501 (61.9) 25–29.9 (overweight) 69 (16.8) 59 (14.8) 128 (15.8) ≥30 (obese) 92 (22.3) 21 (5.3) 113 (14.0) Wide waist circumference¶ n=765 144 (28.9) 53 (14.3) 167 (21.8) <0.001* Data are presented as n, n (%) or medi

2 n=810 <0.001* <18.5 (underweight) 27 (6.6) 41 (10.3) 68 (8.4) 18.5–24.9 (normal) 224 (54.4) 277 (69.6) 501 (61.9) 25–29.9 (overweight) 69 (16.8) 59 (14.8) 128 (15.8) ≥30 (obese) 92 (22.3) 21 (5.3) 113 (14.0) Wide waist circumference¶ n=765 144 (28.9) 53 (14.3) 167 (21.8) <0.001* Data are presented as n, n (%) or medi an (interquartile range), unless otherwise stated. ART: antiretroviral therapy; BMI: body mass index. #: total n=852, unless otherwise stated in column 1; ¶: ≥94 cm males, ≥88 cm females. *: p<0.05. Prevalence of diabetes and IGR The overall prevalence of diabetes (using any of the three diagnostic criteria) was 11.3% (95% CI 9.3–13.6%); 12.6% among TB cases (95% CI 9.7–16.1%) and 10.1% (95% CI 7.6–13.2%) in non-TB participants (p=0.246). Among diabetes participants, 59.4% (57 out of 96) were not previously diagnosed with diabetes (61.5% in TB cases and 56.8% in non-TB) and all previously diagnosed diabetes patients were on diabetes treatment. Despite treatment, participants with a prior diabetes diagnosis had higher median (IQR) FPG (7.45 (5.2–11.3) versus 5.45 (5.1–7.1) mmol·L−1; p<0.001) and HbA1c (9.7% (7.0–11.4%) versus 6.5% (6.4–6.9%); p<0.001) levels compared with newly diagnosed diabetes patients.

agnosed diabetes patients were on diabetes treatment. Despite treatment, participants with a prior diabetes diagnosis had higher median (IQR) FPG (7.45 (5.2–11.3) versus 5.45 (5.1–7.1) mmol·L−1; p<0.001) and HbA1c (9.7% (7.0–11.4%) versus 6.5% (6.4–6.9%); p<0.001) levels compared with newly diagnosed diabetes patients. The overall statistically significant difference in glycaemic status between participants with and without TB was driven by the prevalence of IGR, which was significantly higher in TB cases than non-TB (65.2% versus 50.0%; p<0.001). The prevalence of diabetes and IGR varied by diagnostic test (table 2). The majority of diabetes diagnoses were made based on the FPG and HbA1c tests, while IGR was largely driven by a positive HbA1c test. TABLE 2 Prevalence of diabetes and impaired glucose regulation (IGR), including previously diagnosed diabetes, by diagnostic test

The overall statistically significant difference in glycaemic status between participants with and without TB was driven by the prevalence of IGR, which was significantly higher in TB cases than non-TB (65.2% versus 50.0%; p<0.001). The prevalence of diabetes and IGR varied by diagnostic test (table 2). The majority of diabetes diagnoses were made based on the FPG and HbA1c tests, while IGR was largely driven by a positive HbA1c test. TABLE 2 Prevalence of diabetes and impaired glucose regulation (IGR), including previously diagnosed diabetes, by diagnostic test Overall Non-TB TB p-value (Chi-squared) Diabetes Any test 11.3 (9.3–13.6) 10.1 (7.6–13.2) 12.6 (9.7–16.1) 0.246 FPG 4.1 (3.0–5.7) 3.9 (2.4–6.2) 4.4 (2.8–6.9) 0.752 OGTT 3.3 (2.3–4.8) 3.5 (2.1–5.8) 3.1 (1.7–5.3) 0.704 HbA1c 8.2 (6.5–10.2) 6.2 (4.3–8.9) 10.2 (7.7–13.6) 0.032* IGR (excluding diabetes) Any test 57.3 (53.7–60.8) 50.0 (45.1–54.9) 65.2 (60.0–70.0) <0.001* FPG 10.6 (8.6–13.1) 12.5 (9.5–16.1) 8.6 (6.1–12.2) 0.089 OGTT 10.6 (8.6–13.0) 4.9 (3.1–7.5) 16.9 (13.3–21.2) <0.001* HbA1c 39.5 (36.1–43.0) 34.1 (29.6–38.9) 45.4 (40.3–50.6) 0.002* Data are presented as % (95% CI), unless otherwise stated. TB: tuberculosis; FPG: fasting plasma glucose; OGTT: oral glucose tolerance test; HbA1c: glycated haemoglobin. *: p<0.05.

6.1–12.2) 0.089 OGTT 10.6 (8.6–13.0) 4.9 (3.1–7.5) 16.9 (13.3–21.2) <0.001* HbA1c 39.5 (36.1–43.0) 34.1 (29.6–38.9) 45.4 (40.3–50.6) 0.002* Data are presented as % (95% CI), unless otherwise stated. TB: tuberculosis; FPG: fasting plasma glucose; OGTT: oral glucose tolerance test; HbA1c: glycated haemoglobin. *: p<0.05. The overall prevalence of diabetes was lower in HIV-1-infected patients compared with those uninfected or those with HIV-1 status unknown (8.9% versus 16.0% or 10.0%, respectively) (table 3). While there was no association between TB and diabetes overall, when stratified by HIV-1 status, the prevalence of diabetes was found to be higher in HIV-1-infected TB cases versus HIV-1-infected participants without TB (11.1% versus 6.2%; p=0.049). This difference was not observed in HIV-1-uninfected participants (table 3). TABLE 3 Diabetes prevalence in tuberculosis (TB) and non-TB participants, stratified by HIV-1 status Overall Non-TB TB p-value Subjects 852 438 414 HIV-1-uninfected n=281 16.0 (12.2–20.8) (n=45/281) 16.9 (11.8–23.6) (n=27/160) 14.9 (9.5–22.5) (n=18/121) 0.651 HIV-1-infected n=521 8.9 (6.7–11.6) (n=46/521) 6.2 (3.8–10.1) (n=15/242) 11.1 (7.9–15.4) (n=31/279) 0.049* HIV-1 status unknown n=50 10.0 (4.1–22.4) (n=5/50) 5.6 (1.3–20.8) (n=2/36) 21.4 (6.0–54.0) (n=3/14) 0.093 Data are presented as n or % (95% CI), unless otherwise stated. *: p<0.05.

14.9 (9.5–22.5) (n=18/121) 0.651 HIV-1-infected n=521 8.9 (6.7–11.6) (n=46/521) 6.2 (3.8–10.1) (n=15/242) 11.1 (7.9–15.4) (n=31/279) 0.049* HIV-1 status unknown n=50 10.0 (4.1–22.4) (n=5/50) 5.6 (1.3–20.8) (n=2/36) 21.4 (6.0–54.0) (n=3/14) 0.093 Data are presented as n or % (95% CI), unless otherwise stated. *: p<0.05. TB/diabetes association and population-attributable risk On univariable analysis, age, sex, marital status, household size, income, employment status, previous TB, HIV-1, smoking, hypertension, waist circumference, a history of time in prison, education, BMI and a history of being a miner were all associated with TB at the 10% level of significance. These variables were used in turn to build the multivariable model. After adjusting for confounding variables, diabetes was associated with a 2.4-fold higher odds of TB (95% CI 1.1–5.2), with this association remaining significant in HIV-1-infected but not HIV-1-uninfected individuals (table 4). Further analysis by diabetes diagnostic test revealed that this significant association was only noted using the HbA1c test. IGR was also associated with a 2.3-fold higher odds of TB (95% CI 1.6–3.3), with the strongest association shown when the OGTT was used as a diagnostic test compared with other tests (table 4). Unlike the diabetes analysis, the association between TB and IGR (using any test) was significant in both HIV-1-infected and -uninfected individuals, albeit driven by different tests: a significant association using the OGTT in HIV-1-infected individuals and the HbA1c test in HIV-1-uninfected individuals. Based on a 12% prevalence of diabetes in the general population and the adjusted OR of 2.4, the population-attributable risk of TB due to diabetes is 14%.

viduals, albeit driven by different tests: a significant association using the OGTT in HIV-1-infected individuals and the HbA1c test in HIV-1-uninfected individuals. Based on a 12% prevalence of diabetes in the general population and the adjusted OR of 2.4, the population-attributable risk of TB due to diabetes is 14%. TABLE 4 Univariable and multivariable analysis of the association between tuberculosis and diabetes/impaired glucose regulation (IGR)

viduals, albeit driven by different tests: a significant association using the OGTT in HIV-1-infected individuals and the HbA1c test in HIV-1-uninfected individuals. Based on a 12% prevalence of diabetes in the general population and the adjusted OR of 2.4, the population-attributable risk of TB due to diabetes is 14%. TABLE 4 Univariable and multivariable analysis of the association between tuberculosis and diabetes/impaired glucose regulation (IGR) Overall HIV-1-infected HIV-1-uninfected Crude OR (95% CI) p-value Adjusted OR# (95% CI) p-value Crude OR (95% CI) p-value Adjusted OR# (95% CI) p-value Crude OR (95% CI) p-value Adjusted OR# (95% CI) p-value Diabetes Overall 1.3 (0.8–2.0) 0.247 2.4 (1.3–4.3) 0.005* 2.0 (1.0–3.8) 0.033 2.4 (1.1–5.2) 0.030* 1.0 (0.5–1.9) 0.995 2.4 (0.9–6.7) 0.081 HbA1c 1.7 (1.0–2.9) 0.034 2.4 (1.2–4.6) 0.012* 2.2 (1.1–4.6) 0.032 2.4 (1.0–5.9) 0.05* 1.5 (0.7–3.2) 0.268 2.2 (0.7–6.4) 0.160 FPG 1.1 (0.6–2.2) 0.725 2.3 (0.9–5.5) 0.068 1.8 (0.5–6.0) 0.336 2.9 (0.7–12.2) 0.148 1.1 (0.5–2.6) 0.800 1.9 (0.6–6.4) 0.295 OGTT 0.9 (0.4–1.9) 0.704 1.2 (0.5–3.3) 0.690 1.8 (0.5–6.2) 0.322 2.5 (0.6–10.3) 0.218 0.5 (0.2–1.7) 0.285 0.5 (0.1–3.1) 0.491 Previously diagnosed diabetes only 1.1 (0.6–2.1) 0.717 3.7 (1.5–9.1) 0.004* 2.1 (0.5–8.3) 0.283 6.3 (1.3–30.8) 0.022* 1.2 (0.6–2.6) 0.652 3.1 (0.9–10.1) 0.066 IGR (excludes diabetes cases) Overall 1.9 (1.4–2.5) <0.001 2.3 (1.6–3.3) <0.001* 2.4 (1.7–3.4) <0.001 2.4 (1.5–3.8) <0.001* 1.2 (0.7–2.0) 0.467 2.3 (1.1–4.7) 0.024* HbA1c 1.4 (1.0–1.8) 0.026 1.6 (1.1–2.3) 0.009* 1.6 (1.1–2.3) 0.012 1.5 (1.0–2.3) 0.084 1.1 (0.7–1.8) 0.569 2.2 (1.1–4.1) 0.017* FPG 0.8 (0.5–1.1) 0.182 0.9 (0.5–1.5) 0.695 1.0 (0.6–1.7) 0.867 1.2 (0.6–2.2) 0.588 0.4 (0.2–0.8) 0.018 0.4 (0.1–1.2) 0.116 OGTT 3.0 (1.9–4.8) <0.001 3.9 (2.1–7.0) <0.001* 5.0 (2.6–9.5) <0.001 5.4 (2.4–12.0) <0.001* 1.4 (0.6–2.9) 0.432 2.8 (1.0–8.0) 0.058 HbA1c: glycated haemoglobin; FPG: fasting plasma glucose; OGTT: oral glucose tolerance test. #: adjusted for sex, age, household size, income, hypertension (baseline), previous miner, previous prisoner, marital status, work status and HIV-1 status. *: p<0.05.

4 (2.4–12.0) <0.001* 1.4 (0.6–2.9) 0.432 2.8 (1.0–8.0) 0.058 HbA1c: glycated haemoglobin; FPG: fasting plasma glucose; OGTT: oral glucose tolerance test. #: adjusted for sex, age, household size, income, hypertension (baseline), previous miner, previous prisoner, marital status, work status and HIV-1 status. *: p<0.05. Discussion The prevalence of diabetes, using any three criteria, was 13% in TB cases in Cape Town, largely contributed to by HbA1c. There was also an alarmingly high prevalence of IGR, with 65% all TB cases having IGR. There is good evidence that TB is associated with transient hyperglycaemia, best measured using FPG and 2-h OGTT (markers of short-term glucose status), unlike HbA1c, which is a marker of glycaemic control over 2–3 months. The finding of the prevalence of diabetes being highest using HbA1c is therefore surprising. One possible explanation is the choice of cut-point. The American Diabetes Association/World Health Organization cut-point is 6.5% [30, 31], but a recent study investigating the diagnostic accuracy of HbA1c against the OGTT in a population survey of black South Africans suggested that the optimal cut-point for detection of diabetes was 6.0% [35]. However, using this lower cut-point would have resulted in a higher, not lower, prevalence. Similarly, anaemia, common in TB patients, may have resulted in lower HbA1c in TB cases, unless there was iron deficiency with or without anaemia which may result in an increase in HbA1c compared with non-TB participants, without a concomitant rise in glucose indices [36]. It may be that use of HbA1c in the context of TB and/or HIV-1 is flawed, or that the cut-point should be revised. Further research is required to better understand factors, such as altered red cell survival, that may influence HbA1c in TB/HIV-1 patients.

participants, without a concomitant rise in glucose indices [36]. It may be that use of HbA1c in the context of TB and/or HIV-1 is flawed, or that the cut-point should be revised. Further research is required to better understand factors, such as altered red cell survival, that may influence HbA1c in TB/HIV-1 patients. We noted a significant association between diabetes and TB, in agreement with the majority of the published literature from other settings, with a 2.4- and 2.3-fold higher odds of TB in diabetes and IGR patients, respectively, and 14% of TB cases attributed to diabetes. Only two of these studies were conducted in a setting with a high HIV-1 burden [21, 37] with ORs ranging from 2.2 to 10.7. The coexistence of a high prevalence of HIV-1, TB and diabetes has significant implications for optimal control of each condition, highlighting the importance of targeting TB control interventions, such as intensified TB screening for diabetes and diabetes/HIV-1 patients, to achieve TB elimination goals.

m 2.2 to 10.7. The coexistence of a high prevalence of HIV-1, TB and diabetes has significant implications for optimal control of each condition, highlighting the importance of targeting TB control interventions, such as intensified TB screening for diabetes and diabetes/HIV-1 patients, to achieve TB elimination goals. There was also a significant TB/IGR association (adjusted OR 2.3), suggesting that a finding of IGR should also prompt TB screening. While the high prevalence of IGR would support the theory of these cases being transient hyperglycaemia, the finding of the high prevalence strongly driven by HbA1c is a conundrum as it would suggest a more sustained state of hyperglycaemia. Negative outcomes have been associated with TB/diabetes [3]. HIV-1-uninfected TB/diabetes patients are at higher risk of death than TB/HIV patients [38, 39], but there is a paucity of research on the association between IGR and TB outcomes to inform the need for interventions to control blood glucose in IGR TB patients.

egative outcomes have been associated with TB/diabetes [3]. HIV-1-uninfected TB/diabetes patients are at higher risk of death than TB/HIV patients [38, 39], but there is a paucity of research on the association between IGR and TB outcomes to inform the need for interventions to control blood glucose in IGR TB patients. The association between diabetes and TB also varied depending on the diagnostic test used. We found that the association was largely driven by the HbA1c test, as did Boillat-Blanco et al. [37] in Tanzania. Using either FPG or OGTT alone as a diabetes diagnostic test did not reveal a statistically significant association. Other studies from Tanzania and Guinea-Bissau also found that IGR/diabetes diagnoses varied when using random blood glucose, FPG and OGTT [19, 21]. Further research is required to adequately define the most appropriate test to identify and monitor diabetes in TB patients, including in HIV-1-co-infected individuals.

on. Other studies from Tanzania and Guinea-Bissau also found that IGR/diabetes diagnoses varied when using random blood glucose, FPG and OGTT [19, 21]. Further research is required to adequately define the most appropriate test to identify and monitor diabetes in TB patients, including in HIV-1-co-infected individuals. On stratification by HIV-1 serostatus, the association between TB and diabetes remained statistically significant only in participants with HIV-1 infection. The reported effect of HIV-1 co-infection on the association between TB and diabetes has been variable. A study using random blood glucose reported reduced odds of diabetes in HIV-1-infected patients [20]. Another study using FPG and OGTT reported similar findings of a stronger TB/diabetes association in HIV-1-uninfected patients [21]. However, a recent study that used all three diabetes diagnostic tests showed a slightly stronger TB/diabetes association among HIV-1-infected individuals, as measured by FPG and OGTT, but not by HbA1c [37]. In our study, the significant TB/diabetes association in HIV-1-infected participants was noted when diabetes was defined using any of the three tests, largely driven by HbA1c. There are limited data on the effect on HIV-1 infection on HbA1c. One study that compared HbA1c values in HIV-1-infected and -uninfected women found slightly lower HbA1c values in HIV-1-infected women after adjustment for fasting glucose values [40]. However, this association was nullified on multivariable analysis, with the differences accounted for by higher red cell mean corpuscular volume in HIV-1-infected individuals. Another study also suggested that HbA1c may underestimate glycaemia in HIV-infected individuals, particularly those with higher mean corpuscular volume, nucleoside reverse transcriptase inhibitor use and lower CD4 count [41].

rences accounted for by higher red cell mean corpuscular volume in HIV-1-infected individuals. Another study also suggested that HbA1c may underestimate glycaemia in HIV-infected individuals, particularly those with higher mean corpuscular volume, nucleoside reverse transcriptase inhibitor use and lower CD4 count [41]. The association between TB and diabetes was stronger in participants with a prior diabetes diagnosis, with a 3.9-fold higher odds of TB. These participants had a higher median FPG and HbA1c, and thus the stronger association may reflect the greater degree of hyperglycaemia in this group. This is in line with published studies that have shown that poorly controlled diabetes is more strongly associated with TB [42]. Of note, HIV-1-infected patients with a previous diabetes diagnosis were at an even greater risk of TB with the adjusted OR increasing from 2.4 to 6.3. In this and other studies [13, 20], a high percentage of diabetes among TB patients was previously undiagnosed. Data from chronic disease audits across more than 100 primary care facilities in the Western Cape Province of South Africa, within which Cape Town is located, have reported that <15% of diabetes patients attending clinics have HbA1c levels <7% (Western Cape Dept of Health, Cape Town, South Africa; unpublished data). Our results therefore strongly support the need to prioritise improved diabetes management as a TB control interventions.

ica, within which Cape Town is located, have reported that <15% of diabetes patients attending clinics have HbA1c levels <7% (Western Cape Dept of Health, Cape Town, South Africa; unpublished data). Our results therefore strongly support the need to prioritise improved diabetes management as a TB control interventions. Strengths and limitations A significant strength of this study was the performance of three different tests for diabetes using venous blood samples with all tests performed in an accredited national laboratory. In addition, multivariable analysis of the association between TB and diabetes/IGR adjusted for the most prevalent and significant known confounders, including a previous history of time in prison, mine work and household income, as well as HIV-1 infection.

with all tests performed in an accredited national laboratory. In addition, multivariable analysis of the association between TB and diabetes/IGR adjusted for the most prevalent and significant known confounders, including a previous history of time in prison, mine work and household income, as well as HIV-1 infection. Our study had a number of limitations. Glycaemia was measured at baseline and it is possible that the hyperglycaemia was transient. We attempted to limit the effect of transient hyperglycaemia on the reported TB/diabetes association by selecting non-TB participants from patients with respiratory symptoms that later resolved. Studies have shown an association between hyperglycaemia at TB diagnosis (even when transient) and TB treatment failure and death [37], highlighting the clinical importance of intervening in patients with baseline hyperglycaemia to improve TB outcomes. Small numbers after stratification by HIV-1 status meant we were unable to investigate the effect of ART or CD4 count. In addition, given that the association between TB and diabetes was significant overall, and that the study was powered to detect this association, the lack of statistical significance noted in HIV-1-uninfected participants may be due to a reduced statistical power after stratification; this study is not able to exclude the possibility of an association in HIV-uninfected individuals. Finally, given the association between diabetes and poorer TB outcomes, the exclusion of critically ill patients due to their inability to give consent may have introduced selection bias with possible exclusion of TB patients with diabetes. Similarly, while the multidrug-resistant (MDR)-TB prevalence found in this study is similar to the national prevalence (2.8%) [43], given the poor outcomes associated with MDR-TB, the exclusion of critically ill patients may also have resulted in a greater proportion of excluded patients with MDR-TB.

nts with diabetes. Similarly, while the multidrug-resistant (MDR)-TB prevalence found in this study is similar to the national prevalence (2.8%) [43], given the poor outcomes associated with MDR-TB, the exclusion of critically ill patients may also have resulted in a greater proportion of excluded patients with MDR-TB. Conclusion Diabetes is a significant contributor to the TB burden in this high TB/HIV-1/diabetes burden setting. Our study utilised international guideline-approved diabetes tests and found a high prevalence of diabetes in TB cases, and a significant association between TB and diabetes, even in the context of HIV-1 co-infection, with greater odds of TB in participants who were previously diagnosed with diabetes. These findings point to the importance of screening for diabetes in those with TB and those with HIV-1, who are already at high risk for developing TB. A surprisingly high prevalence of IGR was also found in this study, emphasising the need for further investigation into the TB outcomes and management of this group. The interpretation of our findings is complex because of the variations in diabetes prevalence and TB/diabetes association by different diagnostic tests. Existing glycaemia measures can be affected by HIV-1, ART, anaemia, acute infection, red cell survival and iron status. This highlights the need for more accurate markers of glycaemia that are independent of these factors in order to inform policy on how best to screen for diabetes at the time of TB diagnosis in (often low-resource and primary care) settings with a high burden of HIV-1 and TB.

fection, red cell survival and iron status. This highlights the need for more accurate markers of glycaemia that are independent of these factors in order to inform policy on how best to screen for diabetes at the time of TB diagnosis in (often low-resource and primary care) settings with a high burden of HIV-1 and TB. Acknowledgements Author contributions: T. Oni designed the study and was principal investigator. T. Oni, N.S. Levitt and R.J. Wilkinson were involved in the design of the study. N. Berkowitz led participant recruitment, retention, phenotypic characterisation of all participants and data management. R. Goliath led study coordination, and contributed to patient recruitment, follow-up and collection of data. Data analysis and interpretation were conducted by M. Kubjane and T. Oni. All authors were involved in the drafting of the manuscript, led by T. Oni. The final manuscript draft was approved by all authors. T. Oni had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Earn CME accreditation by answering questions about this article. You will find these at erj.ersjournals.com/journal/cme

Acknowledgements Author contributions: T. Oni designed the study and was principal investigator. T. Oni, N.S. Levitt and R.J. Wilkinson were involved in the design of the study. N. Berkowitz led participant recruitment, retention, phenotypic characterisation of all participants and data management. R. Goliath led study coordination, and contributed to patient recruitment, follow-up and collection of data. Data analysis and interpretation were conducted by M. Kubjane and T. Oni. All authors were involved in the drafting of the manuscript, led by T. Oni. The final manuscript draft was approved by all authors. T. Oni had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Earn CME accreditation by answering questions about this article. You will find these at erj.ersjournals.com/journal/cme Support statement: This study was funded by the Clinical Infectious Disease Research Initiative, as part of a Wellcome Trust Strategic Grant (WT 084323, 104873), a Carnegie Corporation Postdoctoral Fellowship and a Harry Crossley Senior Clinical Fellowship. The Francis Crick Institute received support from Cancer Research UK, Research Councils UK and the Wellcome Trust. R.J. Wilkinson receives support from the National Research Foundation of South Africa (96841). The funders played no role in the design, conduct or analysis of the study. Funding information for this article has been deposited with the Crossref Funder Registry. Conflicts of interest: None declared.

Introduction The Developmental Origins of Health and Disease (DOHaD) theory proposes that environmental exposures (such as nutritional insults), during critical time “windows” in early life have long-term implications for adult health, particularly non-communicable diseases (NCDs) [1]. Respiratory illness is no exception; there is evidence that early nutritional insults, both prenatally and during early infancy, result in higher risk of chronic obstructive pulmonary disease (COPD) and asthma [2, 3]. The exact mechanisms behind DOHaD are not currently known, but it is hypothesised that disruption of lung development, particularly alveologenesis, could lead to chronic lung disease in adulthood, particularly when compounded by adverse environmental factors [4, 5]. Since the majority of alveologenesis is thought to occur before the age of 2 years [6], nutritional insults during this period could result in long-term impairments. In particular, severe acute malnutrition (SAM), which is most prevalent in the first 2 years of life, might have an important biological link to long-term respiratory disease.

alveologenesis is thought to occur before the age of 2 years [6], nutritional insults during this period could result in long-term impairments. In particular, severe acute malnutrition (SAM), which is most prevalent in the first 2 years of life, might have an important biological link to long-term respiratory disease. Worldwide, SAM affects more than 19 million children under 5 years [7, 8] and is defined as low weight-for-height (below −3 z-scores using the World Health Organization (WHO) growth standards) or a low mid-upper-arm circumference (<115 mm) or presence of oedematous malnutrition (a glossary of nutrition terms and definitions can be found in the supplementary material) [9]. Although some studies have considered the effects of chronic undernutrition (also known as stunting) [10, 11] and micronutrients [12, 13] on lung function, there is very little evidence regarding the long-term effects of SAM on lung function. Understanding whether SAM during infancy and childhood has lasting effects on lung function could facilitate development of interventions aiming to curb the growing burden of NCDs in developing countries and improve survival and long-term outcomes post-SAM. The aim of this study was to quantify the long-term (7-year) effects of SAM on spirometry outcomes, as well as explore predictors of poor lung function, including severity of stunting and wasting, presence of oedema at admission, exposure to household cooking smoke, sex and HIV, in a cohort of Malawian SAM survivors [14, 15].

of this study was to quantify the long-term (7-year) effects of SAM on spirometry outcomes, as well as explore predictors of poor lung function, including severity of stunting and wasting, presence of oedema at admission, exposure to household cooking smoke, sex and HIV, in a cohort of Malawian SAM survivors [14, 15]. Materials and methods Study design This was a longitudinal cohort study that prospectively followed-up ex-SAM “case children”. We estimated that we would locate and recruit 300 case children plus one sibling and one community control per case. This was sufficient for detecting differences in lung function between groups equivalent to 0.5 z-scores (>90% power, 5% significance), generally thought to be the cut-off for clinical significance [16]. Ethical approval for this study was granted by Malawi College of Medicine Research and Ethics Committee (COMREC) (reference P02/13/1342) and University College London Research Ethics Committee (reference 4683/001). Besides lung function, other outcomes were also studied: more complete details of the cohort, as well as additional methods and results have been described elsewhere [17, 18].

ine Research and Ethics Committee (COMREC) (reference P02/13/1342) and University College London Research Ethics Committee (reference 4683/001). Besides lung function, other outcomes were also studied: more complete details of the cohort, as well as additional methods and results have been described elsewhere [17, 18]. Study setting and subjects The cohort originally included all patients admitted to the nutrition ward for treatment of SAM at the Queen Elizabeth Central Hospital, Blantyre, Malawi, from July 12, 2006 to March 9, 2007 (1024 children). The median age of the children at admission was 21.5 months (interquartile range, 15–32 months). Results of survival and anthropometry at the baseline study and the 1-year follow-up have been described previously [14, 15]. 47% (477/1024) of the original cohort were known to be alive 1 year post-discharge; these children formed the “case group” for this study. The sibling control was that closest in age to the case child, between the ages of 6 and 15.9 years. Those younger than 6 years were not eligible for spirometry testing, due to the difficulty in obtaining high-quality spirometry in very young children under field conditions. The community control was defined as a child living in the same community, of the same sex, and within 12 months of age as the case child. This community control was randomly selected by spinning a bottle at the case child's home to select a random direction, then enquiring door-to-door to find the first eligible child. Children who had ever been treated for acute malnutrition were excluded from the control groups. Informed written consent was obtained from the child's parent or guardian; assent was required from the children themselves. Possible sources of missing data included cases who could not be located, those who agreed to participate but did not attend the hospital appointment (more often community controls) and those who failed to achieve spirometry data of acceptable quality.

rent or guardian; assent was required from the children themselves. Possible sources of missing data included cases who could not be located, those who agreed to participate but did not attend the hospital appointment (more often community controls) and those who failed to achieve spirometry data of acceptable quality. Variables Primary lung function outcomes were forced expiratory volume in 1 s (FEV1), forced vital capacity (FVC) and the FEV1/FVC ratio expressed as z-scores, based on the Global Lung Function Initiative (GLI) reference equations for the African–American population, which adjust for standing height, age, sex and ethnicity [16]. Spirometry testing took place in our study room at the hospital using the Easy on-PC spirometer (NDD medical, Zurich, Switzerland), operated by N. Lelijveld or E. Chimwezi, and one of three trained data collectors. Quality control of acceptable spirometry traces was undertaken weekly by N. Lelijveld. J. Kirkby, who was blinded to the study groups, checked 40% of results. The criteria for acceptability and repeatability of the spirometry curves were based on American Thoracic Society/European Respiratory Society recommendations modified slightly for children [19, 20].

ry traces was undertaken weekly by N. Lelijveld. J. Kirkby, who was blinded to the study groups, checked 40% of results. The criteria for acceptability and repeatability of the spirometry curves were based on American Thoracic Society/European Respiratory Society recommendations modified slightly for children [19, 20]. Duplicate measures by independent data collectors were made for height and sitting height to the nearest 0.1 cm (Leicester height stadiometer, HM-250P; Marsden Weighing Group, Rotherham, UK), and weight to the nearest 0.1 kg (MS4202L; Marsden group, Rotherham, UK), following the WHO method [21]. Weight-for-age (WAZ), height-for-age (HAZ) and body mass index (BMI)-for-age (BAZ) z-scores were calculated using WHO 2007 growth standards [22]. HIV status was established from results in health passports; if unknown, an HIV test was offered by a trained counsellor, following Malawi national protocol (using Determine HIV-1/2 (Abbott Laboratories, Irving, TX, USA) and Uni-Gold HIV (Trinity Biotech PLC, Bray, Ireland) tests). Puberty was recorded as a binary variable, as reported by the participant or guardian (onset of menarche in girls, voice change in boys). As birthweight was rarely recorded, mothers were asked whether their child was “small” or “normal/large” at birth, during the baseline data collection; this follows a method widely used in Demographic and Health Surveys (DHS) (http://dhsprogram.com/). Pulse oximetry was used to measure oxygen saturation (SpO2) using the NONIN PalmSAT 2500 device (Nonin Medical Inc., Plymouth, MN, USA) during the iStep exercise test (incremental step test) developed by UCL Institute of Child Health [23]. Body composition, including lean mass, was measured using bioelectrical impedance analysis (BIA) (Quadscan 4000; Bodystat Ltd, Douglas, UK). A questionnaire detailing contra-indications for spirometry and potential confounding factors, including cold symptoms, asthma, cooking fuel use, socioeconomic circumstances (SEC) and history of respiratory illness was also administered. These questions were based on the “American Thoracic Society's ATS-DLD-78C respiratory questionnaire”, UCL's “SLIC” study questionnaire and Malawi DHS [24–26]. History of respiratory illness was assessed by parental reporting. Exposure to household cooking smoke was defined as those reporting use of solid fuels for cooking inside the home.

based on the “American Thoracic Society's ATS-DLD-78C respiratory questionnaire”, UCL's “SLIC” study questionnaire and Malawi DHS [24–26]. History of respiratory illness was assessed by parental reporting. Exposure to household cooking smoke was defined as those reporting use of solid fuels for cooking inside the home. Analysis Primary analysis was based on all children with technically acceptable (grade A–C) spirometry results. Those with grade D results (i.e. not repeatable) [19] and those with reported current upper respiratory (cold) symptoms were also included as they were evenly spread across the sample and made no significant difference to the final conclusions [27]. Basic demographic data for each study group as well as those cases lost to follow-up are presented (table 1). Data for the three main spirometric outcomes and exercise test outcomes were examined by allocation group (figure 1) and statistical differences assessed using simple and multivariable linear regression models (table 2 and table S1). One model was used for each outcome where the predictor was coded zero for cases, one for siblings and two for community controls. HIV status, age, sex, SEC derived from asset scores using DHS questions [26], and puberty were included as a priori potential confounders. For spirometric outcomes, sitting height as a percentage of standing height and leg length were also explored as potential confounders in further analysis. The difference in odds of completing the exercise test between allocation groups was calculated using multivariable logistic regression.

iori potential confounders. For spirometric outcomes, sitting height as a percentage of standing height and leg length were also explored as potential confounders in further analysis. The difference in odds of completing the exercise test between allocation groups was calculated using multivariable logistic regression. TABLE 1 Demographic characteristics for the three study groups and those lost to follow-up

iori potential confounders. For spirometric outcomes, sitting height as a percentage of standing height and leg length were also explored as potential confounders in further analysis. The difference in odds of completing the exercise test between allocation groups was calculated using multivariable logistic regression. TABLE 1 Demographic characteristics for the three study groups and those lost to follow-up Basic demographics Cases Sibling controls Community controls Cases lost to follow-up Subjects n 237 164 131 190 Age (range) years 9.3 (7.6–15.3) 11.5 (4.6–15.6) 9.1 (5.2–15.1) 8 (7–19) Males 128 (54) 76 (46) 68 (52) 108 (57) Birth order median (interquartile range) 2 (1–4) 2 (2–3) 2 (1–3) 2 (1–3) Started puberty 7 (3) 12 (7) 6 (5) NA SEC (asset quintile) 1 (poorest) 50 (21) 36 (22) 22 (17) NA 5 (richest) 44 (19) 28 (17) 28 (21) NA HIV Seropositive 65 (28) 5 (3) 3 (2) 44 (23) Seronegative 155 (65) 100 (61) 75 (57) 121 (64) Status unknown 17 (7) 59 (36) 53 (40) 25 (13) Height-for-age z-score −1.8±1.2 −1.5±1.2 −1.3±1.1 NA Weight-for-age z-score −1.6±1.0 −1.4±1.0 −1.2±1.0 NA BMI-for-age z-score −0.8±0.9 −0.8±0.9 −0.7±0.9 NA Sitting height % 52.2±1.5 51.8±1.7 51.8±1.5 NA Sitting height cm 65.4±4.3 68.2±7.1 66.0±4.7 NA Leg length cm 59.9±5.5 63.0±9.6 61.6±6.0 NA Lean mass index 9.02±1.1 9.20±1.0 9.12±1.0 NA Reported indoor biofuel use (wood/charcoal) 37 (16) ∼16% 18 (14) NA Reported indoor tobacco use 31 (13) ∼13% 19 (15) NA History of TB 11 (5) 1 (0.6) 1 (0.8) 4/173 (2) History of pneumonia admissions 7 (3) 13 (8) 7 (5) NA Ever admitted to hospital (except SAM) 47 (20) 43 (26) 38 (29) NA Results are presented as n (%) or mean±sd, unless otherwise indicated. Height-for-age z-score is based on World Health Organization 2007 growth standards. Lean mass index is calculated from impedance results of bioelectrical impedance analysis and height: a higher value implies more lean mass. Sitting height %=sitting height/standing height×100. History of tuberculosis (TB), pneumonia and hospital admission were self-reported. Cases lost to follow-up are those who were admitted with severe acute malnutrition (SAM) in the original cohort but could not be subsequently located. “∼” for sibling controls indicates that this was not measured but assumed to be the same as for cases with whom they shared a household. NA: not applicable. SEC: socioeconomic circumstances; BMI: body mass index.

admitted with severe acute malnutrition (SAM) in the original cohort but could not be subsequently located. “∼” for sibling controls indicates that this was not measured but assumed to be the same as for cases with whom they shared a household. NA: not applicable. SEC: socioeconomic circumstances; BMI: body mass index. FIGURE 1 Lung function results for cases, siblings and community controls. Solid line with error bars represent mean±sd. Dashed lines indicate the limits of normality as per Global Lung Function Initiative (GLI) spirometry reference data for the African–American population (i.e. mean (0) ±1.96 z-scores). There were no statistically significant differences in any of the spirometry outcomes among the three groups. Most results fall within the normal range; however, mean forced expiratory volume in 1 s (FEV1) and forced vital capacity (FVC) were lower than predicted by the GLI reference for all three groups. TABLE 2 Results of simple and multivariable linear regression analysis for spirometry outcomes across the three study groups

FIGURE 1 Lung function results for cases, siblings and community controls. Solid line with error bars represent mean±sd. Dashed lines indicate the limits of normality as per Global Lung Function Initiative (GLI) spirometry reference data for the African–American population (i.e. mean (0) ±1.96 z-scores). There were no statistically significant differences in any of the spirometry outcomes among the three groups. Most results fall within the normal range; however, mean forced expiratory volume in 1 s (FEV1) and forced vital capacity (FVC) were lower than predicted by the GLI reference for all three groups. TABLE 2 Results of simple and multivariable linear regression analysis for spirometry outcomes across the three study groups Cases (n=201) Mean±sd Sibling (n=143) Community (n=121) Mean±sd Difference case–sibling Mean (sd) Difference case–community Unadjusted (95% CI) Adjusted (95% CI) Unadjusted (95% CI) Adjusted (95% CI) FEV1 z-score −0.47±1.1 −0.48±1.0 0.02 (−0.2 to 0.2) 0.13 (−0.2 to 0.4) −0.34±1.1 −0.13 (−0.4 to 0.1) −0.02 (−0.3 to 0.2) FVC z-score −0.32±1.0 −0.38±1.1 0.06 (−0.2 to 0.3) 0.20 (−0.0 to 0.5) −0.15±1.1 −0.17 (−0.4 to 0.1) −0.05 (−0.3 to 0.2) FEV1/FVC z-score −0.21±0.9 −0.15±0.9 −0.06 (−0.3 to 0.1) −0.10 (−0.3 to 0.1) −0.37±1.0 0.16 (−0.1 to 0.4) 0.15 (−0.1 to 0.4) Adjusted differences include HIV status, socioeconomic circumstances and puberty. When including sitting height % in the model, all p-values remain >0.05. FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity.

−0.15±0.9 −0.06 (−0.3 to 0.1) −0.10 (−0.3 to 0.1) −0.37±1.0 0.16 (−0.1 to 0.4) 0.15 (−0.1 to 0.4) Adjusted differences include HIV status, socioeconomic circumstances and puberty. When including sitting height % in the model, all p-values remain >0.05. FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity. For our secondary analysis, HIV status, SEC, sex, history of pneumonia, history of TB, cooking smoke exposure and body composition (lean/fat mass) were examined for their association with spirometry outcomes across the entire sample using linear regression. A sex-stratified analysis on cooking smoke exposure was also conducted to try to explain significant differences in spirometry found between the sexes. We also used linear regression to explore predictive factors of poor spirometry outcomes in the case group only. Namely, differences associated with severity of wasting and stunting at admission, presence of oedema at admission, HIV status, sex, estimated birth size and age at admission (months) were calculated after adjusting for HIV, SEC and puberty (table 3). For all analyses, a p-value of <0.05 was considered to indicate statistical significance. TABLE 3 Results of linear regression analysis comparing effects of potential predictors of poor long-term spirometric lung function in severe acute malnutrition survivors (case group only)

We also used linear regression to explore predictive factors of poor spirometry outcomes in the case group only. Namely, differences associated with severity of wasting and stunting at admission, presence of oedema at admission, HIV status, sex, estimated birth size and age at admission (months) were calculated after adjusting for HIV, SEC and puberty (table 3). For all analyses, a p-value of <0.05 was considered to indicate statistical significance. TABLE 3 Results of linear regression analysis comparing effects of potential predictors of poor long-term spirometric lung function in severe acute malnutrition survivors (case group only) Potential predictors Adjusted (95% CI) difference z-score FEV1 Adjusted (95% CI) difference z-score FVC Adjusted (95% CI) difference z-score FEV1/FVC Severely stunted at admission (HAZ <−3) (n=92 versus 105) 0.02 (−0.3 to 0.3) 0.25 (−0.1 to 0.6) −0.17 (−0.5 to 0.1) Severely underweight at admission (WAZ ≤−4) (n=92 versus 106) −0.07 (−0.4 to 0.2) 0.20 (−0.1 to 0.5) −0.46* (−0.7 to −0.2) Oedema at admission (n=164 versus 34) 0.05 (−0.4 to 0.5) 0.01 (−0.4 to 0.4) 0.25 (−0.1 to 0.6) HIV positive versus negative (n=53 versus 133) −0.58* (−0.9 to −0.3) −0.49* (−0.8 to −0.2) −0.23 (−0.5 to 0.0) Poorest versus richest SEC (n=118 versus 76) −0.27 (−0.06 to 0.0) −0.22 (−0.5 to 0.1) 0.01 (−0.3 to 0.3) Males versus females (n=108 versus 91) 0.32* (0.0 to 0.6) 0.27 (−0.0 to 0.6) 0.06 (−0.2 to 0.3) Low versus normal birth size (n=18 versus 176) −0.02 (−0.6 to 0.5) −0.04 (−0.6 to 0.5) −0.06 (−0.6 to 0.4) Original admission age ≤2 years versus >2 years (n=106 versus 93) −0.22 (−0.5 to 0.1) −0.21 (−0.5 to 0.1) −0.02 (−0.3 to 0.3) *: indicates p<0.05. Results are adjusted for HIV status, socioeconomic circumstances (SEC) and puberty. n=201 for all cases with spirometry results; n presented for each predictor indicates small numbers of missing data for each variable. FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity; HAZ: height-for-age z-score; WAZ: weight-for-age z-score.

r HIV status, socioeconomic circumstances (SEC) and puberty. n=201 for all cases with spirometry results; n presented for each predictor indicates small numbers of missing data for each variable. FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity; HAZ: height-for-age z-score; WAZ: weight-for-age z-score. Results In all, 398 out of 477 (83%) case children families were contacted; 32 declined to participate, 46 case children had died, and 320 case children were investigated. The study included 721 children in total (320 cases, 217 sibling controls and 184 community controls). Those who attempted spirometry and the final number of technically acceptable results are described in figure 2. The proportion of acceptable results was similar across the study groups (85% for cases, 87% for siblings and 92% for community controls). Although the study aimed to recruit one of each type of control per case, not all cases had an eligible sibling, and some community controls dropped out of the study prior to the hospital appointment. Exploration of important demographic and health characteristics for the three study groups and those cases lost to follow-up indicated that HIV was a key potential confounder in the case group (table 1). All three groups had mean height, weight, and BMI z-scores below global reference norms (table 1); however, cases were significantly more stunted than siblings (adjusted mean (95% CI) height difference: −0.2 (−0.4 to −0.0) z-scores) and community controls (−0.4 (−0.6 to −0.2) z-scores) (see supplementary material for definition of stunting). Further details of growth and survival outcomes are published elsewhere [18]. There was no statistically significant difference in chest circumference, chest depth, or sitting height between cases and controls, but cases had a significantly shorter leg length than either sibling (adjusted difference −1.43 cm, 95% CI −2.3 to −0.5) or community controls (adjusted difference: −1.97 cm, 95% CI −3.0 to −1.0).

no statistically significant difference in chest circumference, chest depth, or sitting height between cases and controls, but cases had a significantly shorter leg length than either sibling (adjusted difference −1.43 cm, 95% CI −2.3 to −0.5) or community controls (adjusted difference: −1.97 cm, 95% CI −3.0 to −1.0). FIGURE 2 Recruitment flow diagram for spirometry results. Recruitment of sibling and community control only commenced at the time of 7-year follow-up; hence, their recruitment history starts after that of cases, who were enrolled at admission. Lung function differences across study groups Although most subjects had spirometry results within the normal range (i.e. within 0±2 z-scores; figure 1), mean FEV1 was significantly lower than predicted in all the groups (by an average of 0.47 z-scores) [16]. Similarly, albeit slightly smaller than for FEV1, deficits were also observed for FVC. There were however no significant differences in FEV1, FVC or FEV1/FVC z-scores between cases and either type of control (table 2; figure 1). This remained true when adjusting for sitting height or leg length in the regression model. Additionally, there were no significant differences in SpO2 either at rest or during the iStep exercise test between the study groups (table S1), although cases were less likely to complete the exercise test than controls (log odds of completing the test adjusted for age, sex, HIV status and SEC: −0.34 (−0.7; 0.0) compared to siblings; −0.53 (−0.9; −0.1) compared to community controls).

rest or during the iStep exercise test between the study groups (table S1), although cases were less likely to complete the exercise test than controls (log odds of completing the test adjusted for age, sex, HIV status and SEC: −0.34 (−0.7; 0.0) compared to siblings; −0.53 (−0.9; −0.1) compared to community controls). Predictors of lung function in the whole sample In the entire sample, FEV1 and FVC z-scores were significantly lower for HIV-positive children than HIV-negative children (0.5 and 0.4 z-scores lower respectively, unadjusted p<0.01 for both) with no significant difference in the FEV1/FVC z-score (p=0.2). There was no difference in spirometry outcomes for those with a previous pneumonia admission, whereas in those with a history of TB (n=13; note small sample size), FEV1 and FVC were lower by mean (95% CI) 0.91 (−1.5 to −0.3) and 1.07 (−1.7 to −0.4) z-scores, respectively (unadjusted p<0.001 for both).

z-score (p=0.2). There was no difference in spirometry outcomes for those with a previous pneumonia admission, whereas in those with a history of TB (n=13; note small sample size), FEV1 and FVC were lower by mean (95% CI) 0.91 (−1.5 to −0.3) and 1.07 (−1.7 to −0.4) z-scores, respectively (unadjusted p<0.001 for both). Although there was no difference in FEV1/FVC, FEV1 and FVC z-scores were both significantly lower in girls than boys for the sample as a whole (unadjusted difference (95% CI): −0.28 (−0.5 to −0.1) z-scores, p=0.004, and −0.25 (−0.4 to −0.1) z-scores, p=0.012, respectively). These statistically significant deficits remained after adjusting for HIV status, SEC and puberty (as well as after adjusting for sitting height percentage and/or leg length). There was no significant difference in age, wealth or HIV status between the sexes for the sample as a whole, although boys were slightly but statistically significantly more stunted than girls (unadjusted difference (95% CI) HAZ: −0.2 (−0.4 to −0.04)).

after adjusting for sitting height percentage and/or leg length). There was no significant difference in age, wealth or HIV status between the sexes for the sample as a whole, although boys were slightly but statistically significantly more stunted than girls (unadjusted difference (95% CI) HAZ: −0.2 (−0.4 to −0.04)). There were no statistically significant differences in spirometry outcomes for those exposed to tobacco smoke at home or for those who lived in homes that cooked with biofuel inside (wood/charcoal). Sex-stratified analysis of the whole sample found no statistically significant difference in FEV1 or FVC z-scores for males exposed and unexposed to household cooking smoke. However, FEV1 was 0.5 (95% CI −0.0 to 1.0) z-scores lower (p=0.059) in females exposed to household smoke than in unexposed females (borderline statistical significance; adjusted for HIV and SEC). There were no significant differences in FVC or FEV1/FVC z-scores. Lean mass index as calculated from BIA was positively associated with FEV1 and FVC z-scores (after adjusting for HIV, SEC, age, sex and puberty: 0.26, p<0.0001 for FEV1 z-score; and 0.28, p<0.0001 for FVC z-score). SAM (i.e. study group) was not included as a potential confounder in this section of analysis, as it was found to have no significant effect on lung function in this sample.

h FEV1 and FVC z-scores (after adjusting for HIV, SEC, age, sex and puberty: 0.26, p<0.0001 for FEV1 z-score; and 0.28, p<0.0001 for FVC z-score). SAM (i.e. study group) was not included as a potential confounder in this section of analysis, as it was found to have no significant effect on lung function in this sample. Predictors of lung function in the case group only We also explored whether factors which differed between the cases at admission could predict poor spirometry outcomes at 7 years post-discharge (table 3). There was no significant difference in spirometry outcomes according to the degree of wasting, degree of stunting, age at admission or presence of oedema at the time of admission, after adjusting for HIV status, SEC and puberty. Reported birth size (i.e. mother's report of her child being “normal or big” versus “small”) also had no association with long-term spirometry outcomes in the case group.

ree of wasting, degree of stunting, age at admission or presence of oedema at the time of admission, after adjusting for HIV status, SEC and puberty. Reported birth size (i.e. mother's report of her child being “normal or big” versus “small”) also had no association with long-term spirometry outcomes in the case group. Discussion When comparing the spirometric lung function of Malawian children who had been admitted 7 years previously with an episode of SAM against “healthy” sibling and community controls, all children (cases and controls) were found to have significant reductions in spirometry as compared to international reference data. There were however no significant differences in spirometric lung function or oxygen saturation during an exercise test between SAM survivors and controls. In the sample as a whole, being HIV positive, of female sex, having less lean mass and having a history of TB were associated with lower spirometric z-scores. Although household smoke was not associated with lung function in the sample as a whole, when stratified by sex, we found that females exposed to indoor cooking smoke had a lower mean FEV1 z-score than unexposed females.

, having less lean mass and having a history of TB were associated with lower spirometric z-scores. Although household smoke was not associated with lung function in the sample as a whole, when stratified by sex, we found that females exposed to indoor cooking smoke had a lower mean FEV1 z-score than unexposed females. Deficits in spirometric z-scores compared to the international reference seen in all study groups likely reflects common environmental factors such as the high background levels of chronic undernutrition and general morbidity found throughout childhood in this region. A recent community survey in urban Malawian adults found the prevalence of spirometric restriction was 38.6% using NHANES (National Health and Nutrition Examination Survey) reference ranges (derived from a healthy Caucasian population in the USA) and 9.0% using local reference ranges [28]. Spirometric restriction was significantly associated with low BMI, which supports the theory that common environmental factors such as chronic undernutrition, combined with high population levels of biofuel exposure and HIV, may be sufficient to cause lung restriction, regardless of early SAM exposures.

ce ranges [28]. Spirometric restriction was significantly associated with low BMI, which supports the theory that common environmental factors such as chronic undernutrition, combined with high population levels of biofuel exposure and HIV, may be sufficient to cause lung restriction, regardless of early SAM exposures. Besides high levels of chronic undernutrition, it is important to consider the complex interaction of lung function, height, sitting height and leg length, as this could also diminish the effects of SAM in these results. Because sitting height as a percentage of standing height was on the hypothesised causal pathway between SAM and spirometry outcomes, our results were not adjusted for this variable. However, sitting height is an important consideration as spirometry z-scores are based on standing height. Although SAM survivors had a significantly lower HAZ, they had a similar sitting height as controls, which could have artificially “improved” their lung function z-scores and hence, potentially masked the long-term effects of SAM. In an attempt to unravel this, regression models including sitting height percentage and leg length were run but had virtually no effect on results; p-values remained insignificant for all spirometry outcomes.

tificially “improved” their lung function z-scores and hence, potentially masked the long-term effects of SAM. In an attempt to unravel this, regression models including sitting height percentage and leg length were run but had virtually no effect on results; p-values remained insignificant for all spirometry outcomes. Owing to this complication, it is important to emphasise that for their standing height, SAM survivors, compared to controls, have preserved lung function. Despite results from both animal models and some human studies in India, which suggest that postnatal malnutrition can cause qualitative changes in lung function beyond merely an effect on lung size [29–31], emerging evidence suggests that nutritional insults in childhood and/or consequential stunting results in smaller lungs that are not necessarily associated with restrictive or obstructive defects [32, 33]. This also seems to be true for our stunted SAM survivors, similar to results recently observed in a population of undernourished Indian children [32]. Another study hypothesised that although poor linear growth often does not affect lung function, poor weight gain does, possibly due to its negative effect on muscle function [34]. Our finding that lean mass was positively associated with FEV1 and FVC z-scores supports this hypothesis that weight, particularly muscle mass, may be more important for lung function than linear growth, particularly if weight loss has been substantial.

does, possibly due to its negative effect on muscle function [34]. Our finding that lean mass was positively associated with FEV1 and FVC z-scores supports this hypothesis that weight, particularly muscle mass, may be more important for lung function than linear growth, particularly if weight loss has been substantial. The finding that lung function in our cohort of SAM survivors appears preserved to the level of control children is particularly interesting considering recently published findings that survivors of SAM have multiple adverse long-term outcomes compared with controls, including low weight- and height-for-age, and weaker hand grip strength [18]. Such observations could be explained at least in part by the thrifty phenotype hypothesis of Hales and Barker [35, 36], which emphasises that poor growth is often associated with selective preservation of vital organs at the expense of other organs and tissues. This may apply to SAM survivors in whom limb growth was compromised during or after the episode of SAM, whereas sitting height and lung function appear preserved for the majority of survivors. This pattern of shorter legs but preserved torso height may allow lung function to be preserved by decreasing the “load” on the lungs, whose primary function is to keep the organs and tissues (especially the brain) supplied with oxygen (see load/capacity theory for more detail [37]). While the thrifty phenotype hypothesis emphasised “brain sparing” [38], the need for preserved lung function in order to support brain energy metabolism is a novel concept hypothesised by these results. It is possible that SAM survivors in this study have been successfully “thrifty” in leg growth, thereby preserving an important vital organ: the lungs.

esis emphasised “brain sparing” [38], the need for preserved lung function in order to support brain energy metabolism is a novel concept hypothesised by these results. It is possible that SAM survivors in this study have been successfully “thrifty” in leg growth, thereby preserving an important vital organ: the lungs. Although these results suggest that lung function is unlikely to be a key area for intervention to curb post-SAM mortality, they do highlight some potentially “high risk” groups. One particularly susceptible population appears to be children with HIV. There is emerging evidence of an increased prevalence of chronic lung disease, including both obstructive and restrictive types, in HIV-positive adults [39, 40]. Despite 85% (62/73) of the HIV-positive children in this cohort reportedly taking antiretroviral therapy, their lung function was relatively impaired when compared to HIV-positive children. Very few studies have considered the effect of HIV on lung function in children; there is one recent study in India that also found significantly lower FEV1 and FVC in children aged 5–12 years with vertically transmitted HIV, the majority of whom were being administered antiretroviral therapy [41]. Repeat infections, such as TB, combined with lower WAZ and lower lean mass could all contribute to the reduced lung function in the HIV-positive group.

significantly lower FEV1 and FVC in children aged 5–12 years with vertically transmitted HIV, the majority of whom were being administered antiretroviral therapy [41]. Repeat infections, such as TB, combined with lower WAZ and lower lean mass could all contribute to the reduced lung function in the HIV-positive group. Besides HIV, females in this cohort also had significantly lower FEV1 and FVC z-scores compared with males. We hypothesised that this could be due to a deviation from the reference values because of delayed puberty in Malawian girls; however, the sex difference is still seen in young girls (<9 years), so this seems to be an unlikely explanation. The sex difference could also reflect differences in body composition or biofuel smoke exposure. Male patients in our cohort had significantly more lean mass than girls (based on BIA); therefore, differences in body composition may have contributed to the observed sex differences in spirometry. Although we did not find any overall association between spirometry outcomes and living with a tobacco smoker or cooking with solid fuel inside homes, differential exposure to cooking smoke between the sexes could also be a contributory factor, especially since biofuel smoke was associated with spirometry outcomes when our data were stratified by sex. However, we would have expected pollution exposure to result in an obstructive pattern rather than the restrictive pattern observed. Interventions such as low-pollution cook stoves, currently being evaluated in Malawi (www.capstudy.org/), could have particular benefits for high-risk groups.

data were stratified by sex. However, we would have expected pollution exposure to result in an obstructive pattern rather than the restrictive pattern observed. Interventions such as low-pollution cook stoves, currently being evaluated in Malawi (www.capstudy.org/), could have particular benefits for high-risk groups. Limitations Possible sources of bias in this study include survivor bias, selection bias with regard to controls, and observer bias. As 46% of the case group are known to have died since admission, survivor bias could have diminished the observed effects of SAM as only the “healthiest” with the fewest long-term impacts are likely to have survived. This survivor bias, as well as high rates of HIV and biomass smoke exposure, means that results may not be generalisable to all settings. In addition, the control group are not necessarily representative of the Malawian population as a whole, as they were recruited in the same communities as survivors of SAM. However, based on mean HAZ and WAZ from Malawi's 2015 DHS, our community controls appear similar to the general population (HAZ −1.3 versus −1.5 in DHS; WAZ −1.2 versus −0.8 in DHS) [42]. Selection bias could have also affected the community control group if those who chose to participate differed systematically from those who declined. This could not be assessed as data were not available on those who declined or dropped-out of the study. Lastly, observer bias could have influenced the results as spirometry is an effort-dependent test, and operators were not blinded to the case/control status of the participants due to risk of data-entry errors. However, this is in part mitigated by the fact that those conducting quality control and selection of final spirometry traces for analysis were blinded.

the results as spirometry is an effort-dependent test, and operators were not blinded to the case/control status of the participants due to risk of data-entry errors. However, this is in part mitigated by the fact that those conducting quality control and selection of final spirometry traces for analysis were blinded. Other limitations of this study include the lack of information on some potential confounders such as the accurate birth weight data. Birth weight is associated with long-term adverse effects on lung function [2] and is also associated with a higher risk of SAM in infancy [43]. However, using the mother's estimate as to whether the child was “small” or “normal” at birth, we found no association with lung function. Finally, it is important to note that our results may be affected by the length of follow-up. It could be that small inter-group differences are not yet clinically apparent. Evidence of “lung sparing” growth could indicate phenotype adaptation, which might not have apparent adverse consequences until much later in life [44]. Future follow-up in adulthood could address this.

affected by the length of follow-up. It could be that small inter-group differences are not yet clinically apparent. Evidence of “lung sparing” growth could indicate phenotype adaptation, which might not have apparent adverse consequences until much later in life [44]. Future follow-up in adulthood could address this. Conclusion Contrary to our initial hypothesis, we found no significant long-term impact of SAM on lung function in surviving children. This could be due to “thrifty” or “lung-sparing” growth preserving sitting height and lung function at the expense of limb length. HIV, despite treatment, was associated with adverse effects on spirometry outcomes, even at these young ages. Females also had significantly poorer lung function than males in this cohort: this could reflect differences in body composition or cooking smoke exposure. These groups could be considered high-risk populations in intervention packages seeking to improve lung function both in survivors of SAM and in the general population. Supplementary material 10.1183/13993003.01301-2016.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-01301-2016_Supplement Disclosures 10.1183/13993003.01301-2016.Supp2J.C. Wells ERJ-01301-2016_Wells

Conclusion Contrary to our initial hypothesis, we found no significant long-term impact of SAM on lung function in surviving children. This could be due to “thrifty” or “lung-sparing” growth preserving sitting height and lung function at the expense of limb length. HIV, despite treatment, was associated with adverse effects on spirometry outcomes, even at these young ages. Females also had significantly poorer lung function than males in this cohort: this could reflect differences in body composition or cooking smoke exposure. These groups could be considered high-risk populations in intervention packages seeking to improve lung function both in survivors of SAM and in the general population. Supplementary material 10.1183/13993003.01301-2016.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-01301-2016_Supplement Disclosures 10.1183/13993003.01301-2016.Supp2J.C. Wells ERJ-01301-2016_Wells Acknowledgements Firstly, we sincerely thank all the children and their families who took part in the study. We also thank those at the Malawi College of Medicine paediatric department who hosted and assisted us with this study. Special thanks go to Sarah Rand, Rachel Bonner and Jane Williams (all UCL) for their help with spirometry, skinfold and iStep training. For acknowledgements of all those who helped in the wider study, see reference [18].

at the Malawi College of Medicine paediatric department who hosted and assisted us with this study. Special thanks go to Sarah Rand, Rachel Bonner and Jane Williams (all UCL) for their help with spirometry, skinfold and iStep training. For acknowledgements of all those who helped in the wider study, see reference [18]. Contributors: conceived and designed the experiments: M. Kerac, N. Lelijveld, M.J. Nyirenda and R.S. Heyderman. Performed the data collection: E. Chimwezi, N. Lelijveld and M. Kerac. Performed data quality control: J. Kirkby. Analysed the data and wrote the first draft of the manuscript: N. Lelijveld. Contributed to the writing of the manuscript and agreed with the manuscript's results and conclusions: N. Lelijveld, M. Kerac, A. Seal, J.C. Wells, M.J. Nyirenda, E. Chimwezi, R.S. Heyderman, J. Stocks and J. Kirkby. All authors have read, and confirm that they meet, ICMJE criteria for authorship. This article has supplementary material available from erj.ersjournals.com Support statement: The Wellcome Trust funded this work with an “Enhancement Award” (grant number 101113/Z/13/A). Funding information for this article has been deposited with the Open Funder Registry. Conflict of interest: Disclosures can be found alongside this article at erj.ersjournals.com

Introduction There has been considerable interest in the role of maternal diet in pregnancy in the aetiology of childhood asthma and atopy [1]. Studies have focused particularly on the potentially beneficial effects of antioxidants, following the hypothesis that a declining intake of antioxidants in Westernised countries has led to a reduction in pulmonary antioxidant defences, and hence to an increase in prevalence of asthma and atopy in recent decades [2]. An alternative hypothesis, which has received less attention, is that the epidemic of asthma and atopy in the West could partly be explained by an increasing dietary intake of foods and constituents which may be harmful. Between 1970 and 2000, there was a 25% increase in the per capita consumption of all refined sugars in the USA, matching a worldwide trend [3]. Current international dietary guidelines advise people to reduce their consumption of sugar, and more particularly free sugars, which comprise sugars (monosaccharides and disaccharides) added to foods or drinks by the manufacturer, cook or consumer, and sugars naturally present in honey, syrups and unsweetened fruit juices [4]. While in children a high consumption of sugar-sweetened beverages [5–7] and fruit juice [7, 8] has been linked to asthma, and particularly atopic asthma [7], the relation between total maternal consumption of free sugar during pregnancy and respiratory and atopic outcomes in the offspring has not been studied. One ecological study reported a correlation between perinatal consumption of sugar and severe childhood asthma symptoms [9], but could not specifically address maternal sugar intake in pregnancy. A recent Danish birth cohort study investigated the relation between soft drink consumption, but not total free sugar intake, during pregnancy and childhood asthma and allergic rhinitis [10].

mption of sugar and severe childhood asthma symptoms [9], but could not specifically address maternal sugar intake in pregnancy. A recent Danish birth cohort study investigated the relation between soft drink consumption, but not total free sugar intake, during pregnancy and childhood asthma and allergic rhinitis [10]. We have investigated whether a high intake of free sugar in pregnancy is associated with adverse respiratory and atopic outcomes in the offspring in a large population-based UK birth cohort. Methods Participants The Avon Longitudinal Study of Parents and Children (ALSPAC) is a population-based birth cohort that recruited 14 541 predominantly white pregnant women resident in Avon, UK with expected dates of delivery from April 1, 1991 to December 31, 1992. These pregnancies resulted in 13 972 singleton or twin children who were alive at 1 year of age. The cohort has been followed since birth with annual questionnaires and, since age 7 years, with objective measures in annual research clinics. The study protocol has been described previously [11, 12] and further information can be found at www.alspac.bris.ac.uk, which contains details of all the data that are available (www.bris.ac.uk/alspac/researchers/data-access/data-dictionary/). Ethics approval was obtained from the ALSPAC Ethics and Law Committee (IRB 00003312) and the Local National Health Service Research Ethics Committees.

mation can be found at www.alspac.bris.ac.uk, which contains details of all the data that are available (www.bris.ac.uk/alspac/researchers/data-access/data-dictionary/). Ethics approval was obtained from the ALSPAC Ethics and Law Committee (IRB 00003312) and the Local National Health Service Research Ethics Committees. Exposure assessment Data on maternal diet in pregnancy were collected by a food frequency questionnaire (FFQ) at 32 weeks gestation, covering all the main foods consumed in Britain [13]. The questionnaire asked about their current weekly frequency of consumption of 43 food groups and food items. More detailed questions were asked about daily consumption of a further eight basic foods (including sugar, coffee and tea). The FFQ was used to estimate total energy intake and daily nutrient intake, by multiplying the daily frequency of consumption of a food by the nutrient content [14] of a standard portion [15] of that food, and summing this for all the foods consumed. In this way free sugar intake was estimated. Free sugar does not include lactose when naturally present in milk and milk products or the sugars contained within the cellular structure of fruits and vegetables. Information on the child's free sugar consumption at age 3 years, as well as maternal and paternal sugar consumption at 4 years post-partum, was collected by a similar FFQ.

ot include lactose when naturally present in milk and milk products or the sugars contained within the cellular structure of fruits and vegetables. Information on the child's free sugar consumption at age 3 years, as well as maternal and paternal sugar consumption at 4 years post-partum, was collected by a similar FFQ. Information from a questionnaire at recruitment and from the obstetric records was used to classify women into four mutually exclusive categories: no evidence of glycosuria or diabetes, existing diabetes mellitus before the pregnancy, gestational diabetes and persistent glycosuria during pregnancy. For the purposes of analysis we combined the last three categories to create a binary maternal “diabetes” variable (see supplementary material for further details). Outcome assessment Binary variables Current doctor-diagnosed asthma was defined in children at age 7.5 years (primary outcome) if mothers responded positively to the question “Has a doctor ever actually said that your study child has asthma?” and to one or both of the questions “Has your child had any of the following in the past 12 months: wheezing with whistling; asthma?” Current wheezing, eczema and hay fever in children at age 7.5 years were defined by a positive answer to the question: “Has your child had any of the following in the past 12 months: wheezing with whistling; eczema; hay fever?”

Outcome assessment Binary variables Current doctor-diagnosed asthma was defined in children at age 7.5 years (primary outcome) if mothers responded positively to the question “Has a doctor ever actually said that your study child has asthma?” and to one or both of the questions “Has your child had any of the following in the past 12 months: wheezing with whistling; asthma?” Current wheezing, eczema and hay fever in children at age 7.5 years were defined by a positive answer to the question: “Has your child had any of the following in the past 12 months: wheezing with whistling; eczema; hay fever?” Atopy at age 7 years was defined as a positive reaction (maximum diameter of any detectable weal) to Dermatophagoides pteronyssinus, cat or grass (after subtracting positive saline reactions from histamine and allergen weals, and excluding children unreactive to 1% histamine). Categorical variables Children were further classified, post hoc, according to their asthmatic/atopic status, thus defining a four-category variable (no atopy or asthma, atopy only, nonatopic asthma and atopic asthma), and according to their number of positive reactions to cat, grass and dust mite allergens (n=0, 1 and ≥2). Data on child's asthma status at age 7 and 14 years were used to derive asthma status phenotypes between age 7 and 14 years (none, remitting, incident and persisting) [16]. Continuous variables Serum total IgE (kU·L−1) was measured by fluoroimmunoassay using the Pharmacia UNICAP system (Pharmacia and Upjohn Diagnostics, Uppsala, Sweden).

Data on child's asthma status at age 7 and 14 years were used to derive asthma status phenotypes between age 7 and 14 years (none, remitting, incident and persisting) [16]. Continuous variables Serum total IgE (kU·L−1) was measured by fluoroimmunoassay using the Pharmacia UNICAP system (Pharmacia and Upjohn Diagnostics, Uppsala, Sweden). Lung function was measured by spirometry (Vitalograph 2120; Vitalograph, Maids Moreton, UK) at age 8.5 years after withholding short-acting bronchodilators for at least 6 h and long-acting bronchodilators and theophyllines for at least 24 h. The best of three reproducible flow–volume curves was used to measure forced expiratory volume in 1 s (FEV1), forced vital capacity (FVC) and maximal mid-expiratory flow (forced expiratory flow at 25–75% of FVC (FEF25–75)), which were further transformed to age-, height- and sex-adjusted standard deviation units [17]. The tests adhered to American Thoracic Society (ATS) criteria for standardisation and reproducibility of flow–volume measurement [18], with the exception of ATS recommendations for duration of expiration [19]; as many children did not fulfil forced expiratory time >6 s end-of-test criteria, a minimal volume change over the final 1 s was used.

rican Thoracic Society (ATS) criteria for standardisation and reproducibility of flow–volume measurement [18], with the exception of ATS recommendations for duration of expiration [19]; as many children did not fulfil forced expiratory time >6 s end-of-test criteria, a minimal volume change over the final 1 s was used. Potential confounders We selected potential confounding factors which are known (from existing literature) to be associated with one or more of the outcomes of interest [20]. These included maternal age at delivery, sex of child, multiple pregnancy, season of birth, maternal history of atopic diseases (hay fever, asthma, eczema, allergies, or attacks of wheezing with whistling on the chest or attacks of breathlessness in the past 2 years), parity, highest educational qualification, housing tenure, financial difficulties, ethnicity, breastfeeding duration and maternal factors during pregnancy (smoking status, anxiety score (Crown–Crisp Experiential Index), paracetamol use, antibiotic use, infections (urinary infection, influenza, rubella, thrush, genital herpes, other), supplement use and total energy intake (kJ·day−1)). Smoking status was categorised as the maximum exposure during pregnancy (never, passive smoking only, 1–9, 10–19 and ≥20 cigarettes per day).

al Index), paracetamol use, antibiotic use, infections (urinary infection, influenza, rubella, thrush, genital herpes, other), supplement use and total energy intake (kJ·day−1)). Smoking status was categorised as the maximum exposure during pregnancy (never, passive smoking only, 1–9, 10–19 and ≥20 cigarettes per day). Statistical analyses We compared the distributions of child and maternal variables across maternal free sugar intake quintiles using F-statistics for differences in continuous variables and Chi-squared tests for differences in categorical variables. Logistic regression, multinomial logistic regression and linear regression were used to analyse relations between maternal free sugar intake in pregnancy and binary, categorical and continuous outcomes, respectively. After log-transforming total IgE, linear regression was used to estimate geometric mean ratios for IgE; confidence limits were calculated using Huber variances. We analysed free sugar intake in quintiles: 1) as a categorical variable using the lowest quintile as reference to allow for a nonlinear pattern of association and 2) as a continuous variable to test for linear trend (i.e. per quintile effect). For all regression analyses, two stages of adjustment were used. In Model 1 we adjusted for total energy intake only. In Model 2 we adjusted additionally for all potential confounders listed above.

or a nonlinear pattern of association and 2) as a continuous variable to test for linear trend (i.e. per quintile effect). For all regression analyses, two stages of adjustment were used. In Model 1 we adjusted for total energy intake only. In Model 2 we adjusted additionally for all potential confounders listed above. When evidence for associations persisted, we considered other factors which can be considered either as potential confounders or potential mediators of associations between maternal free sugar intake in pregnancy and childhood outcomes, i.e. prematurity [21, 22], impaired fetal growth [21, 23], maternal obesity and weight gain [24–26], and offspring obesity [27, 28]. We therefore adjusted additionally for maternal pre-pregnancy body mass index (BMI) (self-reported), gestational age at delivery, birthweight, maternal weight gain during pregnancy (all abstracted from obstetric records) and child's BMI at age 7 years (based on measured height and weight at clinic) (see supplementary figure E1 showing a directed acyclic graph). In order to assess confounding arising from other dimensions of diet, we additionally adjusted separately for maternal intake of vitamin E, zinc, selenium, n-3 polyunsaturated fatty acids (PUFAs), n-6 PUFAs, and total fruits and vegetables in pregnancy [1, 29–31].

plementary figure E1 showing a directed acyclic graph). In order to assess confounding arising from other dimensions of diet, we additionally adjusted separately for maternal intake of vitamin E, zinc, selenium, n-3 polyunsaturated fatty acids (PUFAs), n-6 PUFAs, and total fruits and vegetables in pregnancy [1, 29–31]. To investigate confounding by post-natal sugar intake, we adjusted additionally for child's sugar intake at age 3 years. In order to investigate potential unmeasured confounding by genetic or shared environmental or lifestyle factors, we used a parental comparison approach, whereby effect estimates for maternal sugar intake in pregnancy were compared with effect estimates for maternal and paternal sugar intake after pregnancy. If there is a causal intrauterine effect, one would expect a stronger association with maternal intake in pregnancy than with maternal post-natal intake or paternal intake (the latter two exposures cannot have a direct biological effect on offspring asthma risk) (see further details in the supplementary material) [32, 33].

here is a causal intrauterine effect, one would expect a stronger association with maternal intake in pregnancy than with maternal post-natal intake or paternal intake (the latter two exposures cannot have a direct biological effect on offspring asthma risk) (see further details in the supplementary material) [32, 33]. As sensitivity analyses, we repeated analyses after exclusion of mothers with implausible energy intakes (<2500 or >25 000 kJ·day−1 [34]) and after exclusion of mothers with diabetes (whose offspring will have experienced high fetal exposure to glucose). To correct for potential loss to follow-up bias, we used inverse probability weighting and assigned to each woman a weight that was the inverse of the probability of her selection for given values of covariates (see further details in the supplementary material) [35]. All statistical analyses were carried out using Stata version 12.1 (StataCorp, College Station, TX, USA).

ed inverse probability weighting and assigned to each woman a weight that was the inverse of the probability of her selection for given values of covariates (see further details in the supplementary material) [35]. All statistical analyses were carried out using Stata version 12.1 (StataCorp, College Station, TX, USA). Results Of the 13 972 singleton or twin children alive at 1 year of age, information on maternal diet was available for 12 078, of whom there was information on at least one of the outcomes of interest for 8956 (supplementary figure E2). Characteristics of the 8956 mother–child pairs who were included in the analyses and those of the 3122 mother–child pairs with information on maternal diet who were excluded because of incomplete outcome data are compared in supplementary table E1. Among children with available information, 12.2% had current doctor-diagnosed asthma, 10.7% had current wheezing with whistling, 8.8% had current hay fever, 16.2% had current eczema, 21.5% had atopy and 61.8% did not have any of these five outcomes.

use of incomplete outcome data are compared in supplementary table E1. Among children with available information, 12.2% had current doctor-diagnosed asthma, 10.7% had current wheezing with whistling, 8.8% had current hay fever, 16.2% had current eczema, 21.5% had atopy and 61.8% did not have any of these five outcomes. Maternal characteristics which differed across quintiles of free sugar intake during pregnancy included age, parity, pregnancy size, season of birth, breastfeeding duration, educational level, ethnicity, housing tenure, financial difficulties, anxiety level, tobacco exposure and infection during pregnancy. Women in the highest quintile of total sugar intake during pregnancy had a lower pre-pregnancy BMI, higher total energy intake and gained more weight during pregnancy than women in the lowest quintile. Their offspring were more likely to have weighed less at birth and to have had a lower BMI at age 7 years (table 1). After adjustment for potential confounders, there was weak evidence for positive associations between maternal free sugar intake in pregnancy and childhood doctor-diagnosed asthma and childhood wheeze (OR comparing highest versus lowest quintile 1.31, 95% CI 0.98–1.75; per quintile p-trend=0.09 and 1.42, 95% CI 1.05–1.92; per quintile p-trend=0.08, respectively), and stronger evidence for a positive association with atopy at age 7 years (OR 1.38, 95% CI 1.06–1.78; per quintile p-trend=0.006) (table 2). There was no association with eczema, hay fever, total IgE, FEV1, FVC or FEF25–75 (table 2 and supplementary table E2). Post hoc analysis showed a positive association between maternal intake of free sugar and atopic asthma (OR 2.01, 95% CI 1.23–3.29; per quintile p-trend=0.004) (table 3). The main positive findings of our study are summarised in figure 1.

hay fever, total IgE, FEV1, FVC or FEF25–75 (table 2 and supplementary table E2). Post hoc analysis showed a positive association between maternal intake of free sugar and atopic asthma (OR 2.01, 95% CI 1.23–3.29; per quintile p-trend=0.004) (table 3). The main positive findings of our study are summarised in figure 1. TABLE 1 Characteristics of mothers and offspring who had information on at least one of the outcomes of interest (wheeze, asthma, atopy, eczema, hay fever, total IgE and lung function) by maternal free sugar intake during pregnancy#

hay fever, total IgE, FEV1, FVC or FEF25–75 (table 2 and supplementary table E2). Post hoc analysis showed a positive association between maternal intake of free sugar and atopic asthma (OR 2.01, 95% CI 1.23–3.29; per quintile p-trend=0.004) (table 3). The main positive findings of our study are summarised in figure 1. TABLE 1 Characteristics of mothers and offspring who had information on at least one of the outcomes of interest (wheeze, asthma, atopy, eczema, hay fever, total IgE and lung function) by maternal free sugar intake during pregnancy# Free sugar intake g·day−1 p-value Quintile 1 (1.6–34.0) Quintile 2 (34.0–46.6) Quintile 3 (46.6–60.8) Quintile 4 (60.8–82.4) Quintile 5 (82.4–345.1) Mother's age years 29.4±4.7 29.1±4.5 29.2±4.6 28.9±4.5 27.9±4.7 <0.001 Parity 0 43.2 46.1 48.6 44.9 44.0 0.001 1 35.4 37.2 35.1 36.3 35.9 ≥2 21.4 16.7 16.3 18.8 20.0 Sex of child Male 51.2 52.3 50.5 52.3 49.5 0.39 Female 48.8 47.7 49.5 47.7 50.5 Multiple pregnancy Singleton 97.3 96.7 98.0 98.3 97.3 0.01 Twin 2.7 3.3 2.0 1.7 2.7 Season of birth Winter 15.0 17.6 16.6 15.7 16.0 <0.001 Spring 23.6 25.3 26.8 29.1 30.7 Summer 33.1 29.3 31.8 28.9 27.1 Autumn 28.3 27.8 24.9 26.4 26.2 Breastfeeding duration months Never 21.3 18.9 17.5 22.1 27.1 <0.001 <3 34.0 30.0 30.1 29.2 34.2 3–6 13.1 14.9 14.6 13.3 12.8 ≥6 31.5 36.3 37.4 35.4 25.9 Mother's educational level Certificate of Secondary Education 16.7 12.9 13.4 14.5 20.4 <0.001 Vocational 8.6 9.2 8.1 8.5 10.8 Ordinary level 36.2 35.0 32.8 35.1 38.8 Advanced level 24.6 26.6 26.3 26.0 21.2 Degree 14.0 16.2 19.4 15.9 8.9 Maternal ethnicity White 97.8 97.2 98.5 98.4 98.8 0.003 Non-white 2.2 2.8 1.5 1.6 1.2 Housing tenure Owned/mortgaged 82.6 86.5 86.5 85.1 76.8 <0.001 Council rented 9.9 7.2 6.6 8.8 15.8 Non-council rented 7.5 6.3 7.0 6.1 7.5 Financial difficulties No 82.6 82.8 84.3 85.3 78.8 <0.001 Yes 17.4 17.2 15.7 14.7 21.2 Maternal history of atopic diseases No 32.2 32.0 33.7 31.4 28.9 0.05 Yes 67.8 68.0 66.3 68.6 71.1 Maternal anxiety score during pregnancy 0–9 24.0 23.4 23.1 20.8 14.3 <0.001 10–14 25.1 27.5 25.7 26.1 24.3 15–20 24.4 25.1 27.3 27.1 25.4 ≥20 26.5 24.1 24.0 26.1 36.1 Maximum maternal tobacco exposure cigarettes·day−1 None 27.8 27.6 28.9 28.2 19.2 <0.001 Passive only 45.4 48.5 47.3 46.0 42.3 1–9 7.6 7.7 8.1 8.0 8.1 10–19 12.5 10.1 9.4 10.1 15.2 ≥20 6.7 6.2 6.2 7.7 15.2 Maternal paracetamol use during pregnancy No 38.0 38.0 37.7 38.0 36.4 0.86 Yes 62.0 62.0 62.3 62.0 63.6 Maternal antibiotic use during pregnancy No 83.7 85.4 84.0 83.7 82.9 0.34 Yes 16.3 14.6 16.0 16.3 17.1 Maternal supplement use during pregnancy No 44.5 43.1 44.4 41.9 41.6 0.26 Yes 55.5 56.9 55.6 58.1 58.4 Maternal infections during pregnancy No 57.2 55.1 55.8 53.3 49.3 <0.001 Yes 42.8 44.9 44.

62.0 62.0 62.3 62.0 63.6 Maternal antibiotic use during pregnancy No 83.7 85.4 84.0 83.7 82.9 0.34 Yes 16.3 14.6 16.0 16.3 17.1 Maternal supplement use during pregnancy No 44.5 43.1 44.4 41.9 41.6 0.26 Yes 55.5 56.9 55.6 58.1 58.4 Maternal infections during pregnancy No 57.2 55.1 55.8 53.3 49.3 <0.001 Yes 42.8 44.9 44. 2 46.7 50.7 Total energy intake kJ·day−1 5567±1376 6499±1326 7166±1360 7931±1488 9317±1994 <0.001 Maternal pre-pregnancy BMI kg·m−2 <18.50 2.7 2.8 4.4 4.0 8.1 <0.001 18.50–24.99 70.8 74.3 77.7 76.7 77.4 25.00–29.99 17.7 17.1 13.5 14.8 11.6 ≥30.00 8.8 5.8 4.4 4.5 2.9 Birthweight g <2500 4.2 4.0 3.6 4.3 5.6 0.02 2500–2999 14.2 13.2 13.9 12.8 15.2 3000–3499 34.8 34.9 35.1 36.7 35.6 3500–3999 32.2 33.6 33.5 33.6 33.0 ≥4000 14.7 14.4 13.8 12.6 11.0 Gestational age weeks 39.4±1.8 39.5±1.7 39.5±1.7 39.5±1.8 39.4±1.9 0.28 Child's BMI at age 7 years kg·m−2 <15.00 25.8 26.3 30.3 27.7 30.7 <0.001 15.00–17.49 51.9 52.3 51.2 54.4 52.6 17.50–20.49 17.2 16.6 14.2 14.3 13.6 ≥20.50 5.2 4.8 4.3 3.5 3.1 Maternal weight gain during pregnancy Quartile 1 30.3 23.7 24.5 23.5 24.5 <0.001 Quartile 2 24.3 25.0 25.1 24.7 25.0 Quartile 3 24.2 27.0 26.3 26.4 23.4 Quartile 4 21.3 24.3 24.1 25.4 27.1 Data are presented as mean±sd or %, unless otherwise stated. BMI: body mass index. #: n=8956. TABLE 2 Associations between maternal free sugar intake during pregnancy and asthma, wheeze, eczema, hay fever and atopy in the offspring

2 46.7 50.7 Total energy intake kJ·day−1 5567±1376 6499±1326 7166±1360 7931±1488 9317±1994 <0.001 Maternal pre-pregnancy BMI kg·m−2 <18.50 2.7 2.8 4.4 4.0 8.1 <0.001 18.50–24.99 70.8 74.3 77.7 76.7 77.4 25.00–29.99 17.7 17.1 13.5 14.8 11.6 ≥30.00 8.8 5.8 4.4 4.5 2.9 Birthweight g <2500 4.2 4.0 3.6 4.3 5.6 0.02 2500–2999 14.2 13.2 13.9 12.8 15.2 3000–3499 34.8 34.9 35.1 36.7 35.6 3500–3999 32.2 33.6 33.5 33.6 33.0 ≥4000 14.7 14.4 13.8 12.6 11.0 Gestational age weeks 39.4±1.8 39.5±1.7 39.5±1.7 39.5±1.8 39.4±1.9 0.28 Child's BMI at age 7 years kg·m−2 <15.00 25.8 26.3 30.3 27.7 30.7 <0.001 15.00–17.49 51.9 52.3 51.2 54.4 52.6 17.50–20.49 17.2 16.6 14.2 14.3 13.6 ≥20.50 5.2 4.8 4.3 3.5 3.1 Maternal weight gain during pregnancy Quartile 1 30.3 23.7 24.5 23.5 24.5 <0.001 Quartile 2 24.3 25.0 25.1 24.7 25.0 Quartile 3 24.2 27.0 26.3 26.4 23.4 Quartile 4 21.3 24.3 24.1 25.4 27.1 Data are presented as mean±sd or %, unless otherwise stated. BMI: body mass index. #: n=8956. TABLE 2 Associations between maternal free sugar intake during pregnancy and asthma, wheeze, eczema, hay fever and atopy in the offspring Maternal free sugar intake during pregnancy p-trend Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5 Per quintile Asthma n=7677 Adjusted OR# (95% CI) 1.00 1.14 (0.91–1.41) 0.97 (0.77–1.23) 1.22 (0.96–1.56) 1.37 (1.04–1.81) 1.07 (1.00–1.14) 0.04 Adjusted OR¶ (95% CI) 1.00 1.18 (0.94–1.47) 1.02 (0.80–1.29) 1.24 (0.97–1.58) 1.31 (0.98–1.75) 1.06 (0.99–1.13) 0.09 Wheeze n=7762 Adjusted OR# (95% CI) 1.00 1.21 (0.96–1.52) 1.05 (0.82–1.35) 1.19 (0.92–1.54) 1.38 (1.03–1.85) 1.06 (0.99–1.13) 0.09 Adjusted OR¶ (95% CI) 1.00 1.25 (0.99–1.58) 1.11 (0.86–1.42) 1.22 (0.94–1.59) 1.42 (1.05–1.92) 1.06 (0.99–1.14) 0.08 Eczema n=7748 Adjusted OR# (95% CI) 1.00 1.19 (0.98–1.44) 1.13 (0.93–1.38) 1.05 (0.84–1.30) 0.91 (0.70–1.17) 0.97 (0.92–1.03) 0.33 Adjusted OR¶ (95% CI) 1.00 1.17 (0.96–1.42) 1.15 (0.94–1.41) 1.07 (0.86–1.33) 0.97 (0.74–1.26) 0.99 (0.93–1.05) 0.70 Hay fever n=7728 Adjusted OR# (95% CI) 1.00 1.07 (0.82–1.38) 1.25 (0.96–1.63) 1.23 (0.93–1.63) 1.22 (0.88–1.69) 1.06 (0.98–1.14) 0.13 Adjusted OR¶ (95% CI) 1.00 1.03 (0.79–1.33) 1.24 (0.95–1.62) 1.21 (0.91–1.60) 1.25 (0.89–1.75) 1.07 (0.99–1.15) 0.10 Atopy n=6117 Adjusted OR# (95% CI) 1.00 0.99 (0.81–1.21) 1.09 (0.89–1.34) 1.19 (0.96–1.47) 1.24 (0.97–1.60) 1.06 (1.01–1.13) 0.03 Adjusted OR¶ (95% CI) 1.00 0.98 (0.80–1.20) 1.10 (0.89–1.35) 1.20 (0.96–1.49) 1.38 (1.06–1.78) 1.09 (1.02–1.15) 0.006 #: controlling for energy intake; ¶: controlling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration.

ontrolling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration. TABLE 3 Associations between maternal free sugar intake during pregnancy and atopy without asthma, nonatopic and atopic asthma# in the offspring

ontrolling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration. TABLE 3 Associations between maternal free sugar intake during pregnancy and atopy without asthma, nonatopic and atopic asthma# in the offspring Maternal free sugar intake during pregnancy p-trend Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5 Per quintile Atopy without asthma n=794 Adjusted OR¶ (95% CI) 1.00 0.78 (0.61–1.01) 1.05 (0.82–1.34) 0.96 (0.73–1.25) 1.09 (0.80–1.50) 1.04 (0.97–1.12) 0.30 Adjusted OR+ (95% CI) 1.00 0.77 (0.60–0.99) 1.05 (0.82–1.34) 0.95 (0.72–1.25) 1.17 (0.85–1.62) 1.05 (0.98–1.13) 0.19 Nonatopic asthma n=301 Adjusted OR¶ (95% CI) 1.00 0.82 (0.57–1.16) 0.67 (0.46–0.99) 0.77 (0.52–1.16) 0.83 (0.52–1.34) 0.95 (0.85–1.06) 0.34 Adjusted OR+ (95% CI) 1.00 0.87 (0.61–1.25) 0.73 (0.49–1.08) 0.78 (0.52–1.18) 0.71 (0.43–1.15) 0.92 (0.82–1.03) 0.14 Atopic asthma n=337 Adjusted OR¶ (95% CI) 1.00 1.66 (1.14–2.41) 1.17 (0.78–1.77) 2.09 (1.39–3.14) 1.79 (1.11–2.90) 1.14 (1.03–1.27) 0.01 Adjusted OR+ (95% CI) 1.00 1.75 (1.20–2.56) 1.27 (0.84–1.93) 2.18 (1.45–3.30) 2.01 (1.23–3.29) 1.17 (1.05–1.30) 0.004 #: no atopy or asthma was considered as baseline category (n=3796); ¶: controlling for energy intake; +: controlling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration.

ontrolling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration. FIGURE 1 Summary of the main findings for the associations between maternal free sugar intake during pregnancy and childhood outcomes: a) atopy (n=6117) and b) atopic asthma (n=5228). Q: quintile. #: controlling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration.

ontrolling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration. Further investigation of potential confounding/mediation of main findings and sensitivity analyses Additional separate adjustment for maternal pre-pregnancy BMI, gestational age at delivery, birthweight, maternal weight gain during pregnancy and child's BMI at age 7 years did not substantially alter the main findings and therefore no further formal mediation analysis was conducted (supplementary table E3). Additional separate adjustment for maternal intake of vitamin E, zinc, selenium, n-3 PUFAs, n-6 PUFAs, and total fruits and vegetables in pregnancy did not substantially alter the main findings (data not shown), nor did additional separate adjustment for child's free sugar intake at age 3 years (supplementary table E3). The latter exposure was not associated with any outcome (data not shown).

nium, n-3 PUFAs, n-6 PUFAs, and total fruits and vegetables in pregnancy did not substantially alter the main findings (data not shown), nor did additional separate adjustment for child's free sugar intake at age 3 years (supplementary table E3). The latter exposure was not associated with any outcome (data not shown). In subsets of the cohort with complete data for paternal (respectively, maternal) free sugar intake after pregnancy, no association was found between paternal (respectively, maternal) free sugar intake after pregnancy and childhood atopy or atopic asthma. The significant associations of maternal free sugar intake during pregnancy with childhood atopy and atopic asthma remained, unattenuated, on mutual adjustment for paternal (respectively, maternal) post-natal exposure (table 4 and supplementary table E4, respectively). TABLE 4 Comparison of associations of childhood atopy and atopic asthma with maternal free sugar intake during pregnancy versus paternal intake after pregnancy

In subsets of the cohort with complete data for paternal (respectively, maternal) free sugar intake after pregnancy, no association was found between paternal (respectively, maternal) free sugar intake after pregnancy and childhood atopy or atopic asthma. The significant associations of maternal free sugar intake during pregnancy with childhood atopy and atopic asthma remained, unattenuated, on mutual adjustment for paternal (respectively, maternal) post-natal exposure (table 4 and supplementary table E4, respectively). TABLE 4 Comparison of associations of childhood atopy and atopic asthma with maternal free sugar intake during pregnancy versus paternal intake after pregnancy Free sugar intake p-trend Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5 Per quintile Atopy n=3063 Maternal free sugar intake during pregnancy Adjusted OR# (95% CI) 1.00 1.04 (0.78–1.38) 1.16 (0.87–1.55) 1.33 (0.98–1.82) 1.64 (1.14–2.37) 1.13 (1.04–1.23) 0.004 Adjusted OR¶ (95% CI) 1.00 1.02 (0.77–1.36) 1.15 (0.86–1.53) 1.31 (0.96–1.79) 1.61 (1.11–2.33) 1.13 (1.04–1.22) 0.005 Paternal free sugar intake after pregnancy Adjusted OR# (95% CI) 1.00 1.12 (0.84–1.50) 1.28 (0.95–1.73) 1.36 (0.99–1.86) 1.16 (0.81–1.68) 1.05 (0.97–1.14) 0.22 Adjusted OR¶ (95% CI) 1.00 1.10 (0.82–1.47) 1.25 (0.93–1.69) 1.30 (0.95–1.78) 1.10 (0.76–1.59) 1.04 (0.96–1.13) 0.34 Atopic asthma n=2830 Maternal free sugar intake during pregnancy Adjusted OR# (95% CI) 1.00 2.11 (1.26–3.52) 1.18 (0.66–2.11) 2.82 (1.60–4.96) 2.01 (1.01–4.00) 1.17 (1.02–1.36) 0.03 Adjusted OR¶ (95% CI) 1.00 2.13 (1.27–3.57) 1.17 (0.65–2.11) 2.80 (1.58–4.94) 1.96 (0.98–3.93) 1.17 (1.01–1.35) 0.04 Paternal free sugar intake after pregnancy Adjusted OR# (95% CI) 1.00 0.99 (0.60–1.62) 1.26 (0.75–2.11) 1.47 (0.87–2.48) 0.92 (0.48–1.77) 1.04 (0.90–1.21) 0.55 Adjusted OR¶ (95% CI) 1.00 0.96 (0.59–1.59) 1.23 (0.73–2.06) 1.39 (0.82–2.35) 0.87 (0.45–1.69) 1.03 (0.89–1.19) 0.72 #: controlling only for previously mentioned potential confounders; ¶: mutually adjusting for maternal free sugar intake during pregnancy and paternal free sugar intake after pregnancy, in addition to previously mentioned potential confounders

59) 1.23 (0.73–2.06) 1.39 (0.82–2.35) 0.87 (0.45–1.69) 1.03 (0.89–1.19) 0.72 #: controlling only for previously mentioned potential confounders; ¶: mutually adjusting for maternal free sugar intake during pregnancy and paternal free sugar intake after pregnancy, in addition to previously mentioned potential confounders When we analysed the association between maternal free sugar intake and the number of positive reactions to cat, grass and dust mite allergens, we observed a stronger association for children with two or more positive reactions (table 5). We studied associations between maternal free sugar intake in pregnancy and childhood asthma status phenotypes, and did not observe any association (supplementary table E5). TABLE 5 Associations between maternal free sugar intake in pregnancy and number of positive skin prick tests (SPTs)#

When we analysed the association between maternal free sugar intake and the number of positive reactions to cat, grass and dust mite allergens, we observed a stronger association for children with two or more positive reactions (table 5). We studied associations between maternal free sugar intake in pregnancy and childhood asthma status phenotypes, and did not observe any association (supplementary table E5). TABLE 5 Associations between maternal free sugar intake in pregnancy and number of positive skin prick tests (SPTs)# Free sugar 1 positive SPT¶ ≥2 positive SPTs+ Adjusted OR§ (95% CI) p-trend Adjusted OR§ (95% CI) p-trend Quintile 1 1.00 1.00 Quintile 2 0.87 (0.68–1.12) 1.17 (0.87–1.57) Quintile 3 1.07 (0.83–1.37) 1.14 (0.84–1.56) Quintile 4 1.09 (0.83–1.43) 1.37 (0.99–1.90) Quintile 5 1.19 (0.86–1.63) 1.73 (1.18–2.52) Per quintile 1.06 (0.98–1.14) 0.12 1.13 (1.03–1.23) 0.006 #: no positive SPT was considered as baseline category (n=4797); ¶: n=783; +: n=531; §: controlling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration.

ontrolling for energy intake, smoking, infections, supplements, antibiotics and paracetamol use during pregnancy; maternal educational level, housing tenure, financial difficulties, ethnicity, age, parity, history of atopic diseases, anxiety, sex of child, season of birth, multiple pregnancy and breastfeeding duration. The exclusion of 17 women with implausible energy intake estimates did not alter the main results nor did exclusion of mothers with diabetes. Maternal “diabetes” was not associated with any respiratory or atopic outcome, but was associated with higher birthweight. The inverse probability weighting analysis also produced similar results (data not shown). Discussion In this population-based birth cohort study, we found that a higher maternal intake of free sugar during pregnancy was associated with an increased risk of atopy and atopic asthma in the offspring, independently of sugar intake in early childhood.

The exclusion of 17 women with implausible energy intake estimates did not alter the main results nor did exclusion of mothers with diabetes. Maternal “diabetes” was not associated with any respiratory or atopic outcome, but was associated with higher birthweight. The inverse probability weighting analysis also produced similar results (data not shown). Discussion In this population-based birth cohort study, we found that a higher maternal intake of free sugar during pregnancy was associated with an increased risk of atopy and atopic asthma in the offspring, independently of sugar intake in early childhood. To the best of our knowledge, these are novel findings. While a previous ecological study reported a positive correlation between perinatal consumption of sugar and severe childhood asthma symptoms [9], the limitations of ecological studies for inferring causality are well known (not least because of the high likelihood of confounding) [36]. Furthermore, that study was unable to specifically investigate the potential role of maternal intake of sugar during pregnancy nor the specific role of free sugar. Our longitudinal findings linking maternal free sugar intake in pregnancy to childhood atopy and atopic asthma extend the ecological results and allow stronger causal inference. Interestingly, the findings for atopy became stronger when we examined the association with multiple sensitisation. Although previous cross-sectional studies have reported a positive association between childhood consumption of sugar-containing drinks, including fruit juice, and asthma [5–8], intake of free sugar in early childhood in our study was not associated with any respiratory or atopic outcome.

n with multiple sensitisation. Although previous cross-sectional studies have reported a positive association between childhood consumption of sugar-containing drinks, including fruit juice, and asthma [5–8], intake of free sugar in early childhood in our study was not associated with any respiratory or atopic outcome. Mechanisms We speculate that high maternal fructose consumption may underlie the positive associations between maternal intake of free sugar and childhood atopy and atopic asthma. Fructose, which is a major component of added sugars, and is present naturally in fruit juice and in sweetened drinks as added sucrose (ratio of fructose/glucose 50/50%) or isolated fructose, has been mooted as driving previous cross-sectional findings linking sugar-containing beverage consumption to asthma in children [7, 8]. Fructose consumption, in the form of high fructose corn syrup (ratio of fructose/glucose 60/40%), increased from near 0% to near 30% of per capita consumption of refined sugars in the USA between 1970 and 2000, whereas the consumption of sucrose and glucose declined or remained constant [3].

asthma in children [7, 8]. Fructose consumption, in the form of high fructose corn syrup (ratio of fructose/glucose 60/40%), increased from near 0% to near 30% of per capita consumption of refined sugars in the USA between 1970 and 2000, whereas the consumption of sucrose and glucose declined or remained constant [3]. A prospective randomised controlled trial in adults showed that dietary sugar, and especially fructose, increased levels of C-reactive protein [37]. Fructose also causes generation of uric acid [38], and experimental evidence in mice suggests that uric acid may be an essential initiator and amplifier of T-helper cell type 2 (Th2) immunity and allergic inflammation, through activation of inflammatory dendritic cells [39]. Alternatively, fructose might influence atopic immune responses by conditioning the gut microbiome [40, 41]. The potential of maternal diet in pregnancy to influence inception of offspring allergic airways disease through this mechanism was recently confirmed in a mouse model [42]. We therefore propose that one explanation for our main findings is that high fetal exposure to fructose may cause persistence of Th2 immune responses post-natally and allergic inflammation in the developing lung.

of offspring allergic airways disease through this mechanism was recently confirmed in a mouse model [42]. We therefore propose that one explanation for our main findings is that high fetal exposure to fructose may cause persistence of Th2 immune responses post-natally and allergic inflammation in the developing lung. In contrast to a previous study which reported a link between gestational diabetes and risk of atopic eczema and atopy in early childhood [43], we found no association between maternal diabetes during pregnancy and any outcome in the offspring, although, as expected [44], maternal diabetes was associated with higher birthweight. The lack of a relation with maternal diabetes would suggest that higher fetal exposure to glucose is unlikely to explain our main findings. While high fructose consumption has been proposed as a risk factor for obesity [45], we found no evidence to suggest that the associations between maternal free sugar intake and atopy and atopic asthma in the offspring were mediated by maternal BMI, gestational weight gain or child's BMI, nor by prematurity or low birthweight, assuming key assumptions necessary for mediation analyses were met [46, 47]. Strengths and limitations Strengths of the ALSPAC birth cohort include its size and population-based prospective design, rich information on numerous potential lifestyle and dietary confounders (including information on childhood free sugar intake and parental sugar intake outside of pregnancy), and detailed phenotypic outcome measurements.

ons Strengths of the ALSPAC birth cohort include its size and population-based prospective design, rich information on numerous potential lifestyle and dietary confounders (including information on childhood free sugar intake and parental sugar intake outside of pregnancy), and detailed phenotypic outcome measurements. Although the FFQ that we used had not been formally calibrated against other instruments such as diet diaries, it was based on the one used by Yarnell et al. [48], which has been validated against weighed dietary records and modified in the light of a more recent weighed dietary survey [13]. The FFQ lacked quantitative information on soft drink consumption and this will have led to underestimation of maternal free sugar intake during pregnancy. However, as misclassification of maternal free sugar intake in pregnancy is likely to have been random with respect to childhood outcomes, the strength of associations may have been underestimated. We were unable to assess associations with maternal sugar intake in early pregnancy; however, intakes in early and late pregnancy are likely to be highly correlated. We were unable to assess whether associations between maternal intake of free sugar in pregnancy and childhood atopy and atopic asthma persist beyond the age of 7 years, as no data on atopy (only data on asthma status) have been collected in ALSPAC children after the age of 7 years.

are likely to be highly correlated. We were unable to assess whether associations between maternal intake of free sugar in pregnancy and childhood atopy and atopic asthma persist beyond the age of 7 years, as no data on atopy (only data on asthma status) have been collected in ALSPAC children after the age of 7 years. We think that confounding of the main findings by lifestyle or other aspects of maternal diet in pregnancy is unlikely, as we controlled for numerous potential confounders in the analyses, including nutrients and foods that have been previously linked to childhood asthma and atopy. Importantly, the main findings were not confounded by the offspring's free sugar intake in early childhood. While the possibility of residual confounding cannot be ruled out, the null findings for maternal and paternal free sugar intakes after pregnancy make confounding by unmeasured familial behaviours linked to sugar intake and asthma risk a less likely explanation.

by the offspring's free sugar intake in early childhood. While the possibility of residual confounding cannot be ruled out, the null findings for maternal and paternal free sugar intakes after pregnancy make confounding by unmeasured familial behaviours linked to sugar intake and asthma risk a less likely explanation. As with any longitudinal study, we cannot rule out the possibility that exclusion of mother–child pairs without complete information might have biased our findings. However, it could be argued that, for our results to be totally spurious in those included in our analysis (and for the associations to be truly null in the population as a whole), associations in the excluded mother–child pairs would have to be in the opposite direction and much stronger, compared with the positive associations we reported in the included mother–child pairs, which seems extremely unlikely. Furthermore, loss to follow-up bias has been shown to only slightly modify associations in longitudinal studies, including in ALSPAC [49], and the results of our inverse probability weighting analysis [35] confirmed that loss to follow-up is unlikely to have biased our results. In view of the multiple analyses carried out and the post hoc nature of the findings for atopic asthma, we cannot exclude the possibility that the main findings occurred by chance; hence, they should be interpreted with caution. Given the a priori nature of the hypothesis being tested, and the fact that some outcomes of interest are highly correlated, it did not seem appropriate to correct for multiple testing. However, we plan to re-examine this hypothesis in another birth cohort to see if we can replicate the main findings.

ted with caution. Given the a priori nature of the hypothesis being tested, and the fact that some outcomes of interest are highly correlated, it did not seem appropriate to correct for multiple testing. However, we plan to re-examine this hypothesis in another birth cohort to see if we can replicate the main findings. Conclusions and public health implications We conclude that a higher maternal intake of free sugar during pregnancy may increase the risk of atopy and atopic asthma in the offspring. If these findings are replicated we would design an appropriate intervention study in pregnancy to establish or refute causality. Given the very high levels of sugar consumption currently in the West, where childhood allergy and asthma are so prevalent, confirmation of a causal link would raise exciting prospects for the primary prevention of these disorders. Supplementary material 10.1183/13993003.00073-2017.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-00073-2017_Supplement Disclosures 10.1183/13993003.00073-2017.Supp2A. Bédard ERJ-00073-2017_Bedard A.J. Henderson ERJ-00073-2017_Henderson

Supplementary material 10.1183/13993003.00073-2017.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-00073-2017_Supplement Disclosures 10.1183/13993003.00073-2017.Supp2A. Bédard ERJ-00073-2017_Bedard A.J. Henderson ERJ-00073-2017_Henderson Acknowledgements We are extremely grateful to all the families who took part in this study, the midwives for their help in recruiting them, and the whole Avon Longitudinal Study of Parents and Children team, which includes interviewers, computer and laboratory technicians, clerical workers, research scientists, volunteers, managers, receptionists, and nurses. The authors would like to thank especially Raquel Granell (School of Social and Community Medicine, University of Bristol, Bristol, UK) for preliminary analysis and data collection. This paper is the work of the authors, and A.J. Henderson and S.O. Shaheen will serve as guarantors for its contents. The views expressed are those of the author(s) and not necessarily those of the National Health Service, the National Institute for Health Research or the Dept of Health.

y analysis and data collection. This paper is the work of the authors, and A.J. Henderson and S.O. Shaheen will serve as guarantors for its contents. The views expressed are those of the author(s) and not necessarily those of the National Health Service, the National Institute for Health Research or the Dept of Health. Author contributions: A. Bédard and S.O. Shaheen conceived the study and drafted the manuscript. All authors were involved in the analysis strategy, K. Northstone gave advice on the dietary data and A. Bédard performed the statistical analyses. A.J. Henderson was responsible for all clinical respiratory and allergy data collection. All authors participated in the interpretation of the findings, reviewed the manuscript and revised it critically before submission. All authors have seen and approved the final version of the manuscript. This article has supplementary material available from erj.ersjournals.com Support statement: The UK Medical Research Council, the Wellcome Trust (Grant 102215/2/13/2) and the University of Bristol currently provide core support for the Avon Longitudinal Study of Parents and Children. A. Bédard is funded by a European Respiratory Society Long-Term Research Fellowship (LTRF 2015-5838). K. Northstone is supported by the National Institute for Health Research Collaboration for Leadership in Applied Health Research and Care West at University Hospitals Bristol NHS Foundation Trust. Funding information for this article has been deposited with the Crossref Funder Registry.

rch Fellowship (LTRF 2015-5838). K. Northstone is supported by the National Institute for Health Research Collaboration for Leadership in Applied Health Research and Care West at University Hospitals Bristol NHS Foundation Trust. Funding information for this article has been deposited with the Crossref Funder Registry. Conflict of interest: Disclosures can be found alongside this article at erj.ersjournals.com

Introduction Sarcoidosis is a granulomatous disease of unknown aetiology affecting primarily the pulmonary and lymphatic systems. The disease is characterised by heterogeneity in terms of presentation and disease course [1]. In some patients, chronic fibrosing disease leads to irreversible organ function decline, but for the majority, the disease self-resolves within a few years after diagnosis [1]. The incidence and prevalence of sarcoidosis in Sweden are among the highest worldwide [2]. Sarcoidosis mortality in Sweden has been considered low and in line with that of other Caucasian populations based on studies from the early 1980s [1, 3], but it is unclear whether this holds true today. An evaluation of how sarcoidosis affects individuals is critical for our understanding of the burden of the disease. It is the first step in identifying strategies for preventing complications of sarcoidosis, and for assessing the effectiveness of current therapeutic management.

s unclear whether this holds true today. An evaluation of how sarcoidosis affects individuals is critical for our understanding of the burden of the disease. It is the first step in identifying strategies for preventing complications of sarcoidosis, and for assessing the effectiveness of current therapeutic management. Many estimates of sarcoidosis mortality are derived from death certificate data [4–6]. Interpretations from these studies are unreliable due to changes in classification systems over time, but most importantly, attribution of death to a specific cause is often questionable, especially in chronic diseases with several organ manifestations. Prospectively collected, population-based data with sufficient follow-up are needed to determine whether sarcoidosis is associated with long-term consequences. Previous longitudinal studies have yielded conflicting results [7–10]. Two studies showed that the risk of death is more than two-fold higher in patients with sarcoidosis compared to controls [7, 8]. Excess mortality in pulmonary sarcoidosis may only be observed in those with advanced disease stage at diagnosis, as suggested by one Danish study [9]. In contrast, a small study from the Rochester Epidemiology Project in the USA reported no difference in mortality between individuals with sarcoidosis and the general population [10].

y in pulmonary sarcoidosis may only be observed in those with advanced disease stage at diagnosis, as suggested by one Danish study [9]. In contrast, a small study from the Rochester Epidemiology Project in the USA reported no difference in mortality between individuals with sarcoidosis and the general population [10]. We used nationwide Swedish registers to conduct a large, population-based cohort study. The objective of our study was to compare the mortality experience of individuals recently diagnosed with sarcoidosis to that of the general population, overall, by age, sex and by use of treatment around the time point of diagnosis as a marker for more severe sarcoidosis. Methods Setting and data sources In Sweden, residents are entitled to universal healthcare coverage that is provided by a tax-funded system of hospitals. The National Patient Register (NPR) captures information on hospitalisations and outpatient (non-primary care) visits. The inpatient component has national coverage since 1987 and the outpatient component since 2001. The Cause of Death Register holds information on the date and causes of death (main and contributory) for almost all residents since 1961. Dispensed prescribed medications in pharmacies across Sweden have been recorded in the Prescribed Drug Register since July 2005. Records in population-based registers can be linked using a unique personal identification number.

on the date and causes of death (main and contributory) for almost all residents since 1961. Dispensed prescribed medications in pharmacies across Sweden have been recorded in the Prescribed Drug Register since July 2005. Records in population-based registers can be linked using a unique personal identification number. Study design and participants We conducted a matched cohort study based on data from 2003 to 2013. Information on visits for sarcoidosis was obtained from the NPR using International Classification of Diseases (ICD) codes (ICD-10 D86 including all subcategories, ICD-9/8 135). We identified individuals who had their first ever sarcoidosis visit on January 1, 2003 or later, 2 years after the NPR's outpatient component became available to capture incident sarcoidosis. We required patients to have at least two visits for sarcoidosis >14 days apart and at least one visit at which sarcoidosis was the primary discharge diagnosis. Start of follow-up began on the date they fulfilled the inclusion criteria (index date). To compare sarcoidosis mortality to that of the general population, sarcoidosis cases were individually matched on birth year, sex and county of residence 1:10 to nonsarcoidosis comparators identified from the Total Population Register, which includes all residents. They were required to be resident in Sweden at the time of the case's index date.

lity to that of the general population, sarcoidosis cases were individually matched on birth year, sex and county of residence 1:10 to nonsarcoidosis comparators identified from the Total Population Register, which includes all residents. They were required to be resident in Sweden at the time of the case's index date. To minimise misclassification, the study population was restricted to individuals aged 18‒85 years at start of follow-up. Subjects with a lung cancer or lymphoma diagnosis listed in the Swedish Cancer Register (ICD-7 162, 163, 200‒205) 6 months before or after the first visit for sarcoidosis (or the matched case's first visit for comparators) were excluded. Ethical permission for this study was granted by the regional ethics review board in Stockholm (DNR 2014/230–31). Outcome and follow-up We obtained information on the date of death from the Cause of Death Register. Individuals were followed for all-cause death, first emigration (date obtained from the Total Population Register) or the end of the study (December 31, 2014), whichever occurred first. Following the definition of premature mortality by the Organisation for Economic Co-operation and Development (dx.doi.org/10.1787/193a2829-en), deaths that occurred in individuals aged <70 years were considered premature. During the study period, the average life expectancy at birth in Sweden was 83 years for females and 80 years for males (www.statistikdatabasen.scb.se).

Organisation for Economic Co-operation and Development (dx.doi.org/10.1787/193a2829-en), deaths that occurred in individuals aged <70 years were considered premature. During the study period, the average life expectancy at birth in Sweden was 83 years for females and 80 years for males (www.statistikdatabasen.scb.se). Covariates and other variables The date of birth (to calculate age), sex, county of residence (categorised into southern, middle and northern Sweden), and the country of birth (Nordic, non-Nordic) were collected from the Total Population Register. Attained education was obtained from the Education Register (≤9, 10‒12, ≥13  years or missing). Using information from the NPR, we calculated the Charlson comorbidity index using the original Charlson weights [11]. We searched for each comorbidity occurring 1–3 years before the first sarcoidosis visit (see online supplementary table E1 for ICD codes and weights). By restricting the time window for identifying comorbidity, we attempted to capture the load of multimorbidity that was relevant to mortality at baseline and to avoid detection bias affecting the sarcoidosis group.

3 years before the first sarcoidosis visit (see online supplementary table E1 for ICD codes and weights). By restricting the time window for identifying comorbidity, we attempted to capture the load of multimorbidity that was relevant to mortality at baseline and to avoid detection bias affecting the sarcoidosis group. We hypothesised that individuals in need of treatment at the time of diagnosis represent a group with more severe disease and probably have a different mortality rate. According to national and international guidelines, individuals in need of treatment are those with severe clinical disease (e.g. incapacitating symptoms or involvement of vital organs) [1, 12]. Systemic corticosteroids are the first-line treatment and methotrexate or azathioprine are predominantly used as second-line (steroid-sparing) options [12, 13]. We used information in the Prescribed Drug Register to classify individuals as needing treatment at the time of diagnosis if they dispensed at least one prescription of either systemic corticosteroids (ATC H02AB01/02/04/06/07), methotrexate (L01BA01/L04AX03) or azathioprine (L04AX01) within a period of 3 months before or after their first visit for sarcoidosis. As the register was established in July 2005 information was available for a subset of the population entering the cohort starting October 1, 2005 (n=67 408, 75% of the study population).

hotrexate (L01BA01/L04AX03) or azathioprine (L04AX01) within a period of 3 months before or after their first visit for sarcoidosis. As the register was established in July 2005 information was available for a subset of the population entering the cohort starting October 1, 2005 (n=67 408, 75% of the study population). In addition, using information obtained from the Cause of Death Register, we identified the 10 most common underlying and contributory causes of death in the sarcoidosis group, and calculated their corresponding proportion in the comparator group. Statistical analysis We used Poisson models to estimate age- and sex-adjusted mortality rates and their corresponding 95% confidence intervals in the sarcoidosis and comparator groups. Rate differences were estimated using an additive Poisson model [14] and hazard ratios (HRs) for all-cause death using Cox models adjusted for age, sex and county of residence (model 1), and further for country of birth, education and comorbidity (Charlson index score; model 2). We analysed mortality stratified by age at inclusion, sex and treatment status at diagnosis, and estimated adjusted 1-, 5- and 10-year survival probabilities (model 2) [15]. To estimate the HR for premature mortality we used a fully-adjusted Cox model in which all subjects aged <70 years at inclusion (n=80 686) were followed for all-cause death and right censored at their 70th birthday, first emigration or the end of the study (December 31, 2014).

ear survival probabilities (model 2) [15]. To estimate the HR for premature mortality we used a fully-adjusted Cox model in which all subjects aged <70 years at inclusion (n=80 686) were followed for all-cause death and right censored at their 70th birthday, first emigration or the end of the study (December 31, 2014). We tested the robustness of the HR against the potential confounding of current smoking (data for which we did not have) using probabilistic bias analysis methods [16]. Smoking is associated with a lower risk of developing sarcoidosis and a higher risk of death [17, 18]. In simulations, we made the following informed assumptions for three bias parameters based on Swedish health surveys (The Public Health Agency of Sweden; www.folkhalsomyndigheten.se) and previously published literature: smoking prevalence in sarcoidosis (range 9‒16%) and comparators (21‒27%), and a relative risk of death due to smoking of 2.6 (see online supplementary table E2 for detailed definitions) [17–19]. We simulated another dataset to test possible misclassification of our register-based definition for sarcoidosis. Study participants had their sarcoidosis status reclassified assuming positive predictive values 50‒70% and negative predictive values 98‒100%. The effect of misclassification and confounding on the HR was then tested together in a single simulation. Simulation confidence intervals were obtained using bootstrapping techniques. We used SAS (version 9.4; SAS Institute, Cary, NC, USA) and Stata (version 14.2; StataCorp, College Station, TX, USA) for data management and statistical analysis.

ion and confounding on the HR was then tested together in a single simulation. Simulation confidence intervals were obtained using bootstrapping techniques. We used SAS (version 9.4; SAS Institute, Cary, NC, USA) and Stata (version 14.2; StataCorp, College Station, TX, USA) for data management and statistical analysis. Results We identified 8207 individuals with sarcoidosis and 81 119 matched general population comparators between 2003 and 2013. Table 1 shows the demographic and clinical characteristics of the two groups at baseline. The mean±sd age at inclusion was 49±14.4 years and 56% were male. Individuals with sarcoidosis had more comorbid conditions than comparators (mean±sd Charlson comorbidity index score 0.24±0.86 versus 0.13±0.60, p<0.001 from t-test). 12% of individuals with sarcoidosis and 7% of general population comparators had at least one comorbidity. 6191 (42%) out of 2599 individuals with sarcoidosis entering the cohort starting October 1, 2005 received treatment at the time of diagnosis. Baseline characteristics of the two subgroups of individuals with sarcoidosis and of matched comparators are presented in online supplementary table E3. TABLE 1 Baseline demographic and clinical characteristics of individuals with sarcoidosis and matched general population comparators

Results We identified 8207 individuals with sarcoidosis and 81 119 matched general population comparators between 2003 and 2013. Table 1 shows the demographic and clinical characteristics of the two groups at baseline. The mean±sd age at inclusion was 49±14.4 years and 56% were male. Individuals with sarcoidosis had more comorbid conditions than comparators (mean±sd Charlson comorbidity index score 0.24±0.86 versus 0.13±0.60, p<0.001 from t-test). 12% of individuals with sarcoidosis and 7% of general population comparators had at least one comorbidity. 6191 (42%) out of 2599 individuals with sarcoidosis entering the cohort starting October 1, 2005 received treatment at the time of diagnosis. Baseline characteristics of the two subgroups of individuals with sarcoidosis and of matched comparators are presented in online supplementary table E3. TABLE 1 Baseline demographic and clinical characteristics of individuals with sarcoidosis and matched general population comparators Sarcoidosis General population Subjects 8207 81 119 Age at inclusion years 49±14.4 49±14.4 Sex Female 3613 (44) 35 765 (44) Male 4594 (56) 45 354 (56) Region of residence Northern Sweden 1118 (14) 11 118 (14) Middle Sweden 3285 (40) 32 329 (40) Southern Sweden 3804 (46) 37 672 (46) Country of birth Nordic 7407 (90) 71 268 (88) Non-Nordic 800 (10) 9851 (12) Years of education ≤9 1627 (20) 16 274 (20) 10–12 4058 (50) 37 672 (46) ≥13 2435 (30) 26 256 (32) Missing 87 (1) 917 (1) Charlson comorbidity index score 0.24±0.86 0.13±0.60 Data are presented as n, mean±sd or n (%).

weden 3804 (46) 37 672 (46) Country of birth Nordic 7407 (90) 71 268 (88) Non-Nordic 800 (10) 9851 (12) Years of education ≤9 1627 (20) 16 274 (20) 10–12 4058 (50) 37 672 (46) ≥13 2435 (30) 26 256 (32) Missing 87 (1) 917 (1) Charlson comorbidity index score 0.24±0.86 0.13±0.60 Data are presented as n, mean±sd or n (%). During a similar median follow-up of 5.9 years (interquartile range 3.4‒8.7 years), 528 deaths occurred in the sarcoidosis group and 3204 in the comparator group (table 2). Sarcoidosis was the predominant underlying or contributory cause of death among the 445 individuals with sarcoidosis who died during 2003‒2013 (n=134, 30%; online supplementary table E4). On average, individuals with sarcoidosis died 2 years younger than their general population comparators (mean±sd age at death 69.5±12.6 years versus 71.5±12.5, p=0.001 from t-test). TABLE 2 All-cause mortality in individuals with sarcoidosis versus matched general population comparators

.8)** 1.1 (1–1.3) Exacerbations None 1 (reference) 1 (reference) ≤2 (moderate) 1.1 (1.1–1.2)*** 1.0 (0.9–1.1) 1.0 (1–1.1) 1.0 (1–1.1) ≥3 moderate/≥1 severe 1.6 (1.5–1.6)*** 1.2 (1.1–1.3)*** 1.0 (0.9–1) 1.0 (1–1.1) BMI: body mass index; GOLD: Global Initiative for Obstructive Lung Disease; MRC: Medical Research Council. #: adjusted for all other characteristics in the table. *: p<0.05; **: p<0.01; ***: p<0.0001. In the deceased patients, age ≤65 years, anxiety and not being obese were significantly associated with an increased odds of receiving PCS after adjusting for all other characteristics (table 2). The reduced association of PCS with heart failure may be related to residual confounding from age as it tended towards unity after the multivariate analysis. Lung cancer had the strongest association with receiving PCS, even after adjusting for all other characteristics (adjusted OR 6.5, 95% CI 5.5–6.8). The sensitivity analysis that included GOLD stage in both multivariate models showed little difference in the odds ratios (whole cohort, lung cancer: adjusted OR 14.4, 95% CI 13–15.9; p<0.0001; deceased patients, lung cancer: adjusted OR 6.1, 95% CI 5.4–6.9; p<0.0001). The sensitivity analysis that included the MRC Dyspnoea score in the deceased patients multivariate model showed little difference in the odds ratios (deceased patients, lung cancer: adjusted OR 6.1, 95% CI 5.5–6.7; p<0.0001).

During a similar median follow-up of 5.9 years (interquartile range 3.4‒8.7 years), 528 deaths occurred in the sarcoidosis group and 3204 in the comparator group (table 2). Sarcoidosis was the predominant underlying or contributory cause of death among the 445 individuals with sarcoidosis who died during 2003‒2013 (n=134, 30%; online supplementary table E4). On average, individuals with sarcoidosis died 2 years younger than their general population comparators (mean±sd age at death 69.5±12.6 years versus 71.5±12.5, p=0.001 from t-test). TABLE 2 All-cause mortality in individuals with sarcoidosis versus matched general population comparators Sarcoidosis General population Adjusted mortality rate difference (95% CI)¶ Hazard ratio (95% CI) Deaths per person-years Adjusted mortality rate (95% CI)# Deaths per person-years Adjusted mortality rate (95% CI)# Model 1+ Model 2§ Subjects n 8207 81 119 Overall 528/47 810 11.0 (10.1–12.0) 3204/479 200 6.7 (6.5–6.9) 2.7 (2.3–3.1) 1.74 (1.59–1.91) 1.61 (1.47–1.76) Age at inclusion years 18–29 2/4049 0.9 (0.6–1.3) 24/40 689 0.5 (0.3–0.8) -0.1 (-0.8–0.6) 0.84 (0.20–3.56) 0.69 (0.16–2.96) 30–39 25/12 464 1.6 (1.3–1.9) 109/121 763 0.9 (0.8–1.1) 1.1 (0.3–1.9) 2.24 (1.45–3.46) 1.62 (1.02–2.56) 40–49 41/11 080 3.0 (2.5–3.5) 192/110 249 1.8 (1.5–2.0) 1.9 (1.5–2.2) 2.13 (1.52–2.98) 2.03 (1.45–2.85) 50–59 80/9606 8.6 (7.6–9.6) 476/96 636 5.0 (4.6–5.5) 2.8 (2.0–3.6) 1.70 (1.34–2.15) 1.54 (1.21–1.96) 60–69 156/6999 21.0 (18.8–23.1) 838/71 388 12.3 (11.5–13.0) 9.0 (6.8–11.3) 1.92 (1.62–2.28) 1.65 (1.38–1.96) 70–85 224/3614 72.5 (65.8–79.1) 1565/38 476 42.5 (40.3–44.5) 18.0 (11.2–25.4) 1.61 (1.39–1.84) 1.52 (1.32–1.75) Sex Female 260/20 973 11.5 (10.5–12.5) 1659/211 050 6.7 (6.4–7.0) 3.2 (2.5–3.9) 1.64 (1.44–1.87) 1.55 (1.36–1.77) Male 268/26 838 15.2 (13.9–16.6) 1545/268 150 8.9 (8.4–9.3) 2.6 (2.1–3.1) 1.85 (1.64–2.11) 1.68 (1.47–1.91) #: per 1000 person-years, estimated using Poisson models mutually adjusted for age and sex; ¶: estimated using additive Poisson models adjusted for exponential age and sex, region of residence, country of birth, education and comorbidity (Charlson comorbidity index score; continuous); the model did not converge for age groups 18–29 and 30–39 years, hence estimates are unadjusted rate differences; +: estimated using Cox models adjusted for age (continuous), sex and region of residence; §: estimated using Cox models further adjusted for country of birth, education and comorbidity (Charlson comorbidity index score; continuous).

for age groups 18–29 and 30–39 years, hence estimates are unadjusted rate differences; +: estimated using Cox models adjusted for age (continuous), sex and region of residence; §: estimated using Cox models further adjusted for country of birth, education and comorbidity (Charlson comorbidity index score; continuous). The age- and sex-adjusted mortality rate was 11.0 per 1000 person-years in sarcoidosis (95% CI 10.1‒12.0) and 6.7 per 1000 person-years in comparators (95% CI 6.5‒6.9; table 2). There were three more deaths per 1000 person-years in the sarcoidosis group compared to the comparators (95% CI 2.4‒3.2). Mortality rates were higher in individuals with sarcoidosis compared to comparators irrespective of age or sex, except in the youngest individuals (18‒39 years), for whom death was uncommon. The overall adjusted 1-, 5- and 10-year survival probabilities were 98.9%, 95.4% and 89.4% for sarcoidosis, respectively, and 99.6%, 96.9% and 92.9% for the comparators (figure 1a). FIGURE 1 Adjusted survival functions of individuals with incident sarcoidosis and general population comparators a) overall and b) by age at inclusion.

The age- and sex-adjusted mortality rate was 11.0 per 1000 person-years in sarcoidosis (95% CI 10.1‒12.0) and 6.7 per 1000 person-years in comparators (95% CI 6.5‒6.9; table 2). There were three more deaths per 1000 person-years in the sarcoidosis group compared to the comparators (95% CI 2.4‒3.2). Mortality rates were higher in individuals with sarcoidosis compared to comparators irrespective of age or sex, except in the youngest individuals (18‒39 years), for whom death was uncommon. The overall adjusted 1-, 5- and 10-year survival probabilities were 98.9%, 95.4% and 89.4% for sarcoidosis, respectively, and 99.6%, 96.9% and 92.9% for the comparators (figure 1a). FIGURE 1 Adjusted survival functions of individuals with incident sarcoidosis and general population comparators a) overall and b) by age at inclusion. The fully adjusted HR for all-cause death was 1.61 (95% CI 1.47‒1.76; table 2). The risk for premature death was 64% higher in sarcoidosis compared to comparators (HR 1.64, 95% CI 1.43‒1.89). No large variation in the relative risk of death by age or sex was observed. For individuals treated at diagnosis the HR was 2.34 (95% CI 1.99‒2.75) versus 1.13 (95% CI 0.94‒1.35) for those who did not receive treatment at diagnosis (online supplementary table E5). Survival functions stratified by age and treatment status at diagnosis are depicted in figure 1b and figure 2, respectively. FIGURE 2 Adjusted survival functions of individuals with incident sarcoidosis compared to the general population, stratified by treatment status.

The fully adjusted HR for all-cause death was 1.61 (95% CI 1.47‒1.76; table 2). The risk for premature death was 64% higher in sarcoidosis compared to comparators (HR 1.64, 95% CI 1.43‒1.89). No large variation in the relative risk of death by age or sex was observed. For individuals treated at diagnosis the HR was 2.34 (95% CI 1.99‒2.75) versus 1.13 (95% CI 0.94‒1.35) for those who did not receive treatment at diagnosis (online supplementary table E5). Survival functions stratified by age and treatment status at diagnosis are depicted in figure 1b and figure 2, respectively. FIGURE 2 Adjusted survival functions of individuals with incident sarcoidosis compared to the general population, stratified by treatment status. The bias analysis showed that the conventional HR (from the main analysis) was relatively robust even under extreme bias assumptions (table 3). The HR for death did not differ greatly (1.66, 95% bootstrapped interval 1.40‒1.93) when we accounted for potential confounding by current smoking and for misclassification of our sarcoidosis definition in a single simulation. TABLE 3 Probabilistic bias analysis accounting for unmeasured confounding by current smoking and sarcoidosis misclassification

3–15.9; p<0.0001; deceased patients, lung cancer: adjusted OR 6.1, 95% CI 5.4–6.9; p<0.0001). The sensitivity analysis that included the MRC Dyspnoea score in the deceased patients multivariate model showed little difference in the odds ratios (deceased patients, lung cancer: adjusted OR 6.1, 95% CI 5.5–6.7; p<0.0001). Timing of PCS in respect to the patient's death Only 21.4% of deceased COPD patients received PCS (5595 out of 26 135), of whom 31.5% had a co-diagnosis of lung cancer (1764 out of 5595). Of those that had received PCS, 48.4% received it only within 6 months of their death (of whom 29.1% had lung cancer) (figure 3). FIGURE 3 Cumulative proportion of deceased chronic obstructive pulmonary disease (COPD) patients that first received palliative care support (PCS) in relation to the time of their death and the proportion of those that had a lung cancer co-diagnosis.

The bias analysis showed that the conventional HR (from the main analysis) was relatively robust even under extreme bias assumptions (table 3). The HR for death did not differ greatly (1.66, 95% bootstrapped interval 1.40‒1.93) when we accounted for potential confounding by current smoking and for misclassification of our sarcoidosis definition in a single simulation. TABLE 3 Probabilistic bias analysis accounting for unmeasured confounding by current smoking and sarcoidosis misclassification Hazard ratio (95% CI) Conventional analysis Random error 1.61 (1.47–1.76) Sensitivity analyses# Unmeasured confounding by current smoking Systematic error 1.74 (1.71–1.78) Systematic error and random error 1.74 (1.60–1.93) Misclassification of sarcoidosis definition Systematic error 1.54 (1.42–1.66) Systematic error and random error 1.54 (1.31–1.79) Unmeasured confounding by current smoking and misclassification of sarcoidosis definition Systematic error 1.66 (1.52–1.78) Systematic error and random error 1.66 (1.40–1.93) #: estimated by simulations based on predetermined assumptions for bias parameters (see online supplementary material for their specification). Confidence intervals are simulation intervals for systematic error analyses and bootstrapped intervals for systematic error and random error analyses.

om error 1.66 (1.40–1.93) #: estimated by simulations based on predetermined assumptions for bias parameters (see online supplementary material for their specification). Confidence intervals are simulation intervals for systematic error analyses and bootstrapped intervals for systematic error and random error analyses. Discussion In this study, individuals with sarcoidosis had a higher mortality rate compared to matched comparators from the general population. After adjusting for relevant confounders, individuals with sarcoidosis had a 62% higher risk for all-cause death compared to the general population and an excess of three deaths per 1000 person-years were related to sarcoidosis. In addition, sarcoidosis was associated with a greater risk of premature death (death at <70 years of age). Despite the overall differences in mortality, our results suggest that individuals with sarcoidosis who did not receive treatment at diagnosis had only a marginally greater risk of death. However, the risk was increased more than two-fold for those treated, representing more severe disease at diagnosis.

years of age). Despite the overall differences in mortality, our results suggest that individuals with sarcoidosis who did not receive treatment at diagnosis had only a marginally greater risk of death. However, the risk was increased more than two-fold for those treated, representing more severe disease at diagnosis. The results of this study are in line with most, but not all previous investigations. A small study from the USA of mainly Caucasian incident cases reported no association between sarcoidosis and mortality (standardised mortality ratio 0.90) [10]. However, it should be noted that the mortality rate in the sarcoidosis group was similar to the one in our study (13.3 per 1000 person-years) [10], hence the difference on the ratio scale may be attributable to the choice of comparators. Nonetheless, our results were similar to a study from the UK (1991‒2003; mortality rate 14 per 1000 person-years, HR 2.09) despite differences in case ascertainment (they used primary care data to identify sarcoidosis) [7].

[10], hence the difference on the ratio scale may be attributable to the choice of comparators. Nonetheless, our results were similar to a study from the UK (1991‒2003; mortality rate 14 per 1000 person-years, HR 2.09) despite differences in case ascertainment (they used primary care data to identify sarcoidosis) [7]. In the report by Tukey et al. [8], sarcoidosis in the Black Women's Health Study was associated with a 2.4-fold higher risk of death, which is higher than our overall estimate. Two factors are likely to have contributed to the observed difference. Tukey et al. included cases at various disease stages in their cohort (incident, prevalent and deaths due to sarcoidosis), whereas we limited inclusion to newly diagnosed cases. Another explanation could be that individuals of African American descent have worse prognosis, with previous studies suggesting that sarcoidosis is diagnosed earlier, more extrapulmonary organs are involved and the rate of sarcoidosis-related hospitalisations is higher in black individuals compared to white individuals [20, 21].

lanation could be that individuals of African American descent have worse prognosis, with previous studies suggesting that sarcoidosis is diagnosed earlier, more extrapulmonary organs are involved and the rate of sarcoidosis-related hospitalisations is higher in black individuals compared to white individuals [20, 21]. Mortality estimates in this study, both absolute and relative, are generally lower than those reported in previous investigations for sarcoidosis [1]. This might reflect improved diagnosis and survival compared to previous decades, or might be a feature of our population, in which prognosis might be better [1]. Moreover, despite the later (10 years) peak in disease incidence in females [2], the HR for death was similar for both sexes. This observation is consistent with previous longitudinal studies [7, 9] and contradicts some studies utilising death certificate datasets that indicated prominently higher mortality in females [4, 5, 22]. Our study showed that sarcoidosis is associated with an excess of three deaths per 1000 person-years in our population, an estimate comparable to other inflammatory diseases such as rheumatoid arthritis [23].

studies utilising death certificate datasets that indicated prominently higher mortality in females [4, 5, 22]. Our study showed that sarcoidosis is associated with an excess of three deaths per 1000 person-years in our population, an estimate comparable to other inflammatory diseases such as rheumatoid arthritis [23]. To address our concern that any potential greater risk for death in sarcoidosis compared to the general population was possibly driven by a smaller group of patients with more severe disease, we defined a proxy for severe disease based on dispensation of sarcoidosis-related treatments around the time of disease diagnosis. We showed that the risk of death in individuals treated was more than twice that of the general population, whereas the risk was only slightly increased for patients who did not require such treatment. This finding, while in line with a single-centre study from the 1990s in pulmonary sarcoidosis patients [9], should be cautiously interpreted, as short-term corticosteroid treatment courses are not free from adverse effects and the role of early administration of such therapies remains controversial for some patient subgroups [24, 25]. In addition, there might be other factors affecting the decision to initiate treatment, which may have had some influence on the observed results.

steroid treatment courses are not free from adverse effects and the role of early administration of such therapies remains controversial for some patient subgroups [24, 25]. In addition, there might be other factors affecting the decision to initiate treatment, which may have had some influence on the observed results. Our analysis of death certificate data of individuals with sarcoidosis in our cohort showed that the disease is the most commonly mentioned underlying and/or contributing cause of death in these individuals. However, sarcoidosis was listed in only 30% of the deaths in the sarcoidosis cohort. Factors such as organ engagement (e.g. cardiac sarcoidosis), time since diagnosis, reporting practices or sarcoidosis misclassification in this study might influence this observation. Therefore, it should be noted that while studies utilising death certificate datasets provide an interesting snapshot of disease burden at death, they are methodologically limited to account for the whole picture of disease course, as they capture only a subset of sarcoidosis patients [26, 27].

luence this observation. Therefore, it should be noted that while studies utilising death certificate datasets provide an interesting snapshot of disease burden at death, they are methodologically limited to account for the whole picture of disease course, as they capture only a subset of sarcoidosis patients [26, 27]. Our study has some limitations. A register-based definition was used to identify incident sarcoidosis cases and therefore some misclassification is expected. We did not have any information on primary care visits. However, the majority of patients with sarcoidosis in Sweden are diagnosed in specialist care, therefore we do not believe we have missed very many cases. Furthermore, the HR for death did not change significantly when subjected to testing for misclassification under extreme terms. Lastly, due to the lack of clinical patient information, we were unable to investigate mortality by disease phenotype or contrast our medication-based proxy for severity with other disease severity indices. The use of treatment as a proxy for disease severity might have also led to some misclassification. Future studies are warranted to investigate mortality in well-defined patient subgroups with different disease phenotypes.

henotype or contrast our medication-based proxy for severity with other disease severity indices. The use of treatment as a proxy for disease severity might have also led to some misclassification. Future studies are warranted to investigate mortality in well-defined patient subgroups with different disease phenotypes. A major strength of our study is the use of high-quality registers with excellent coverage of the entire Swedish population. They provided a population-based sample and the largest study to date to examine mortality longitudinally. Interpretations from this study are generalisable to populations with similar standards of healthcare. Complete follow-up was possible and relevant confounding factors (except lifestyle factors such as smoking) were available for all individuals. Moreover, two potential sources of bias were quantitatively addressed using advanced analysis techniques, indicating the robustness of our results.

lar standards of healthcare. Complete follow-up was possible and relevant confounding factors (except lifestyle factors such as smoking) were available for all individuals. Moreover, two potential sources of bias were quantitatively addressed using advanced analysis techniques, indicating the robustness of our results. In conclusion, we have demonstrated that the overall risk of death in individuals with sarcoidosis is greater compared to the general population. However, it varies considerably with disease severity. It was only slightly increased in the majority of patients who did not receive treatment at the time of diagnosis, but was increased two-fold for individuals who received corticosteroids, methotrexate or azathioprine at the time of diagnosis, probably due to more severe disease. This subset of patients with a higher risk of poor outcomes should be prioritised for future interventions aiming to reduce the burden of sarcoidosis, and the effectiveness of treatments should be evaluated. Supplementary material 10.1183/13993003.01815-2017.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary tables ERJ-01815-2017_Supplement Disclosures 10.1183/13993003.01815-2017.Supp2E.V. Arkema ERJ-01815-2017_Arkema Acknowledgements The authors would like to thank all patients and the personnel who contributed information to the databases used in this study.

Supplementary material 10.1183/13993003.01815-2017.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary tables ERJ-01815-2017_Supplement Disclosures 10.1183/13993003.01815-2017.Supp2E.V. Arkema ERJ-01815-2017_Arkema Acknowledgements The authors would like to thank all patients and the personnel who contributed information to the databases used in this study. Author contributions are as follows. Conception and design of the study: M. Rossides and E.V. Arkema; statistical analysis: M. Rossides; interpretation of the data: all authors; drafting or revising the manuscript critically for important intellectual content: all authors; approved the final version for submission: all authors. This article has supplementary material available from erj.ersjournals.com Support statement: The study received funding from the Swedish Society of Medicine and the Swedish Heart-Lung Foundation. Sarcoidosis research at Karolinska Institutet is supported by the Swedish Research Council, the regional agreement on medical training and clinical research (ALF) between Stockholm County Council and Karolinska Institutet, King Oscar's II Jubilee Foundation, the Stockholm County Council and Karolinska Institutet's Research Foundation. Funding information for this article has been deposited with the Crossref Funder Registry. Conflict of interest: Disclosures can be found alongside this article at erj.ersjournals.com

To the Editor: We wholly applaud the move by Johnson et al. [1] to improve awareness of breathlessness and to raise its profile as a subject for focussed clinical research. We consider their research and the ensuing proposal to recognise breathlessness via a new medical term, “chronic breathlessness syndrome”, as important and justified. We share their goal, which is to direct attention to this neglected, undertreated and under-researched symptom. There are two important caveats to be made in response to this article, however. First, there is a need to involve those who live with chronic breathlessness and are thus “experts by experience” in discussions about the framework proposed here, rather than bringing them into the conversation once consensus has been achieved. Second, further medicalisation of breathlessness via the term “syndrome” may not be the best way forward. Research into patients' experience of breathlessness shows that the ways in which breathlessness is spoken about (medicalised and otherwise) not only reflect their experiences but also helps to shape how breathlessness is lived [2–6].

calisation of breathlessness via the term “syndrome” may not be the best way forward. Research into patients' experience of breathlessness shows that the ways in which breathlessness is spoken about (medicalised and otherwise) not only reflect their experiences but also helps to shape how breathlessness is lived [2–6]. O.K. Faull and colleagues, in their response to Johnson et al. [1], comment on the individuality of responses to breathlessness that rely on prior experiences and bodily awareness (interoception). Context and culture play an important role in shaping the understanding and perception of breathlessness [2–4, 6]. For example, among African American communities across the USA, the last words of Eric Garner, “I can't breathe”, as he suffocated in a tussle with police officers, have become a slogan for the Black Lives Matter movement and a metaphor for the lives of those living under other kinds of oppression [2]. Başoğlu [7] suggests that asphyxiation is the most traumatic form of torture and that persistent breathlessness because of an underlying medical condition may be even worse due to the duration of the suffering involved.

es Matter movement and a metaphor for the lives of those living under other kinds of oppression [2]. Başoğlu [7] suggests that asphyxiation is the most traumatic form of torture and that persistent breathlessness because of an underlying medical condition may be even worse due to the duration of the suffering involved. Sufferers of respiratory illness vary in relation to the intensity, affect, ideation and meaning they attribute to their breathlessness [3–6]. It affects every aspect of the life of a breathless person in ways that description of it as a medical symptom cannot capture in full [8]. There is a need to legitimise a range of attitudes towards breathlessness in order for them to inform the clinical encounter. Collecting such experiences under the umbrella term “syndrome” may not be sufficient to enable full expression of the variability and multiple meanings of the experience of breathlessness, and may carry unexpected cognitive and affective “baggage” that detracts from its utility as a proxy for experience.

cal encounter. Collecting such experiences under the umbrella term “syndrome” may not be sufficient to enable full expression of the variability and multiple meanings of the experience of breathlessness, and may carry unexpected cognitive and affective “baggage” that detracts from its utility as a proxy for experience. In view of the highly contextualised experience of breathlessness, it is critical to think about whose views are part of the debate. Discussions with experts by experience and first-person reports of experiences of breathlessness [3, 5] have revealed how powerful language and context are in determining how people with breathlessness think about and experience their problem [4, 6], and how this influences what they might do. Words such as “pulmonary” and “rehabilitation”, for example, may negatively impact upon the uptake of one of the most effective interventions for breathlessness [6].

and context are in determining how people with breathlessness think about and experience their problem [4, 6], and how this influences what they might do. Words such as “pulmonary” and “rehabilitation”, for example, may negatively impact upon the uptake of one of the most effective interventions for breathlessness [6]. There is a further stage necessary in the research process of Johnson et al. [1] in order to validate the claim made in the paper that “a recognised syndrome would […] give permission for patients to discuss their ongoing breathlessness with their clinicians”. As Johnson et al. [1] suggest, patients and their families need to be involved in the discussion, but they should be able to critique the framework suggested by the paper, rather than be presented with it as a fait accompli. Otherwise there is a danger that the words “chronic” and “syndrome” will drive people with breathlessness further underground, in part because they have not been involved in the process of describing their own condition [9]. We encourage Johnson et al. [1] to take this research on to its next logical stage, that of developing a truly consensual terminology that considers the critical role language, metaphor and meaning play in both living with and treating breathlessness. This could be done using the Delphi technique within a more participatory paradigm [10]. Such an approach offers the chance of empowering patients and caregivers in ways that would result in real changes to both their experience and treatment. Disclosures 10.1183/13993003.02331-2017.Supp1H. Carel ERJ-02331-2017_Carel

There is a further stage necessary in the research process of Johnson et al. [1] in order to validate the claim made in the paper that “a recognised syndrome would […] give permission for patients to discuss their ongoing breathlessness with their clinicians”. As Johnson et al. [1] suggest, patients and their families need to be involved in the discussion, but they should be able to critique the framework suggested by the paper, rather than be presented with it as a fait accompli. Otherwise there is a danger that the words “chronic” and “syndrome” will drive people with breathlessness further underground, in part because they have not been involved in the process of describing their own condition [9]. We encourage Johnson et al. [1] to take this research on to its next logical stage, that of developing a truly consensual terminology that considers the critical role language, metaphor and meaning play in both living with and treating breathlessness. This could be done using the Delphi technique within a more participatory paradigm [10]. Such an approach offers the chance of empowering patients and caregivers in ways that would result in real changes to both their experience and treatment. Disclosures 10.1183/13993003.02331-2017.Supp1H. Carel ERJ-02331-2017_Carel J.W. Dodd ERJ-02331-2017_Dodd J. Macnaughton ERJ-02331-2017_MacNaughton R. Oxley ERJ-02331-2017_Oxley A. Rose ERJ-02331-2017_Rose A. Russell ERJ-02331-2017_Russell Conflict of interest: Disclosures can be found alongside this article at erj.ersjournals.com

Introduction Bronchiectasis is a chronic, progressive respiratory disease associated with irreversible widening of the bronchi [1]. Recent data suggest that in the UK, incidence rates in females and males have risen to 35.2 and 26.9 per 100 000 person-years, respectively [2]. In the USA, the prevalence of adult bronchiectasis has been estimated at 52 in 100 000 people, with higher prevalence among females and older individuals [3]. Persistent Pseudomonas aeruginosa lung infections of bronchiectasis patients, occurring in ∼30% of cases, are associated with poorer outcomes and premature mortality [4, 5]. The study of chronic P. aeruginosa lung infections has focused on cystic fibrosis (CF)-associated bronchiectasis, where patients are diagnosed, monitored and subjected to antibiotic therapy from a very early age. This contrasts with non-CF bronchiectasis patients, who present at a much older age and often have a shorter history of therapeutic interventions. Hence, bacterial isolates from non-CF bronchiectasis patients exhibit less resistance to antibiotics compared to isolates from adult CF patients [6]. Previous studies have characterised the evolution of P. aeruginosa during chronic lung infections in CF patients [7, 8]. High-resolution analyses have revealed extensive heterogeneity within P. aeruginosa populations in the CF lung [9–12], including the co-existence of multiple divergent lineages [13].

CF patients [6]. Previous studies have characterised the evolution of P. aeruginosa during chronic lung infections in CF patients [7, 8]. High-resolution analyses have revealed extensive heterogeneity within P. aeruginosa populations in the CF lung [9–12], including the co-existence of multiple divergent lineages [13]. In CF, a number of transmissible strains of P. aeruginosa have been identified, leading to the introduction of measures to control cross-infection [14]. The study of P. aeruginosa in relation to non-CF bronchiectasis is less advanced. In our single-centre study of 50 P. aeruginosa isolates from 40 bronchiectasis patients using molecular typing, there was no compelling evidence for cross-infection or a dominant clone [15]. However, whole-genome sequence analysis of multiple bronchiectasis isolates has not been performed. Here, we report the use of genomics to assess the diversity of P. aeruginosa strains causing infections in non-CF bronchiectasis across multiple UK centres, to identify multistrain infections, and to look for evidence for cross-infection or common source acquisition. In addition, we characterise adaptive mutations and present evidence for within-population divergence during P. aeruginosa chronic lung infections of bronchiectasis patients.

ctasis across multiple UK centres, to identify multistrain infections, and to look for evidence for cross-infection or common source acquisition. In addition, we characterise adaptive mutations and present evidence for within-population divergence during P. aeruginosa chronic lung infections of bronchiectasis patients. Methods Patients and bacterial isolates The 189 P. aeruginosa isolates used in this study (online supplementary table S1) were isolated from sputum samples obtained from 93 patients with bronchiectasis and chronic P. aeruginosa infection (defined as two or more positive respiratory tract cultures in the preceding 12 months) attending 16 adult bronchiectasis centres throughout England and Wales. These included isolates collected as part of a multicentre nebulised antibiotic trial [16], where patients were enrolled within 21 days of completing a course of antipseudomonal antibiotics for an exacerbation. Additional isolates from Newcastle (n=8) and Liverpool (n=53) were collected during observational studies. The methodology used for isolating P. aeruginosa from patient sputum samples is described in the online supplementary material.

in 21 days of completing a course of antipseudomonal antibiotics for an exacerbation. Additional isolates from Newcastle (n=8) and Liverpool (n=53) were collected during observational studies. The methodology used for isolating P. aeruginosa from patient sputum samples is described in the online supplementary material. For 24 patients, sets of isolates (two or more) from the same sample were analysed to look for evidence of multilineage infections. For three of these patients (patients 147–149), sets of 14 or 15 isolates from a single sample were sequenced for higher resolution analysis of within-population heterogeneity. For some analyses, to avoid biases arising from inclusion of multiple clonal genomes from the same patient, a subset of 99 genomes from 91 patients was used. This subset consisted of one randomly selected genome per clonal lineage per patient (online supplementary table S1). We use the term “clonal lineage” to describe isolates with shared multilocus sequence type (MLST) profile and clustering according to core genome single nucleotide polymorphism (SNP)-based phylogeny. DNA preparation and whole-genome sequencing Details of the extraction of genomic DNA from P. aeruginosa isolates, library preparation and whole-genome shotgun sequencing using Illumina (San Diego, CA, USA) short-read sequencing technology are given in the online supplementary material. The European Nucleotide Archive accession number for the study is PRJEB14952.

Details of the extraction of genomic DNA from P. aeruginosa isolates, library preparation and whole-genome shotgun sequencing using Illumina (San Diego, CA, USA) short-read sequencing technology are given in the online supplementary material. The European Nucleotide Archive accession number for the study is PRJEB14952. Methods used for genome sequence assembly, extraction of MLST data, phylogenetic reconstruction using the core genome and variant calling by mapping to the genome of PAO1 [17] to identify SNPs or small insertions or deletions are described in the online supplementary material. Identification of large deletions and virulence factor genes Genome sequences were aligned to the reference genomes P. aeruginosa PAO1 (NC_002516 [17]) and P. aeruginosa LESB58 (FM209186 [18]) and large clone-specific deletions (≥10 kb) were identified using the BLAST Ring Image Generator (BRIG) [19]. The boundaries of deletions were determined by aligning the genome sequences with the P. aeruginosa PAO1 genome using Mauve [20], implemented as part of the Geneious package (www.geneious.com). The presence and absence of virulence factor genes in genome assemblies was determined using Blastable (www.github.com/bawee/blastable). The Pseudomonas genome database (beta.pseudomonas.com) [21] was used to facilitate analysis of gene function.

uve [20], implemented as part of the Geneious package (www.geneious.com). The presence and absence of virulence factor genes in genome assemblies was determined using Blastable (www.github.com/bawee/blastable). The Pseudomonas genome database (beta.pseudomonas.com) [21] was used to facilitate analysis of gene function. Results Diversity of P. aeruginosa non-CF bronchiectasis isolates and evidence for P. aeruginosa multilineage co-infections Core genome SNP phylogenetic analysis alongside a collection of 331 P. aeruginosa isolate genomes from diverse clinical sources [22] indicated that the bronchiectasis isolates were widely distributed (online supplementary figure S1). From the 189 isolates, it was possible to extract complete MLST profiles for 160 (online supplementary tables S1 and S2), with the most widespread sequence types (STs) being ST-253 (PA14-like [23]; 14 patients, eight centres), ST-179 (seven patients, four centres), ST-17 (clone C [23]; five patients, three centres), ST-252 (four patients, four centres) and ST-260 (four patients, three centres). Using core genome SNP phylogeny, previous studies have subdivided the wider P. aeruginosa population into two major groups (group I, which includes strain PAO1, and group II, which includes strain PA14) and one minor group of mostly unrelated clonal lineages [24, 25]. Of a subset of 99 genomes consisting of one randomly selected genome per clonal lineage per patient, 71 were located in group I and 27 in group II (figure 1). Based on a combination of MLST genotype and core genome SNP phylogeny, of the 24 patients from whose samples multiple isolates were examined, there were seven examples of multilineage infections. In one patient (patient 92), three distinct clonal lineages of P. aeruginosa were identified. In patients 42, 72, 73, 84, 85 and 148 there were two co-existing lineages (figure 1).

SNP phylogeny, of the 24 patients from whose samples multiple isolates were examined, there were seven examples of multilineage infections. In one patient (patient 92), three distinct clonal lineages of P. aeruginosa were identified. In patients 42, 72, 73, 84, 85 and 148 there were two co-existing lineages (figure 1). FIGURE 1 Evidence for multilineage co-infections in seven patients. A core genome single nucleotide polymorphism phylogeny is shown for the subset of 99 isolates, confirming that all but one isolate (B113) clusters into one of two major groups. Each bronchiectasis centre is represented by a different colour. Arrows sharing the same colour indicate isolates that were obtained from the same patient. The three isolates from the same patient 92 sample are numbered 1–3.

99 isolates, confirming that all but one isolate (B113) clusters into one of two major groups. Each bronchiectasis centre is represented by a different colour. Arrows sharing the same colour indicate isolates that were obtained from the same patient. The three isolates from the same patient 92 sample are numbered 1–3. Evidence of shared lineages causing infections in different patients attending the same centre The core genome SNP phylogeny identified a number of examples where closely related clonal lineages were isolated from more than one patient attending the same centre (online supplementary table S3). In order to obtain a higher resolution comparison, these isolates were analysed using pairwise comparisons across their entire genomes (online supplementary table S3), identifying five instances where the genomes of isolates from different patients attending the same centre varied at <200 sites (C6/C7, C29/C30, C105/109, C139/C141 and C156/C159; figure 2). This level of genome similarity is greater than in some pairwise comparisons of contemporary isolates of the same lineage from the same sputum sample (online supplementary table S3; from 184 variant sites (C110/C111) to >750 variant sites (C125/C126)).

ites (C6/C7, C29/C30, C105/109, C139/C141 and C156/C159; figure 2). This level of genome similarity is greater than in some pairwise comparisons of contemporary isolates of the same lineage from the same sputum sample (online supplementary table S3; from 184 variant sites (C110/C111) to >750 variant sites (C125/C126)). FIGURE 2 Example pairwise comparisons between isolates sharing the same clonal lineage that were isolated from more than one patient attending the same centre. The number of single nucleotide polymorphism variations are indicated, with the number of small insertion and deletion variations shown in brackets. Full details are shown in online supplementary table S3. The five examples where isolates shared <200 variant sites are highlighted in green. All isolates of ST-244 from patients attending centre 4 were compared, with similarity graded according to variant sites.

all insertion and deletion variations shown in brackets. Full details are shown in online supplementary table S3. The five examples where isolates shared <200 variant sites are highlighted in green. All isolates of ST-244 from patients attending centre 4 were compared, with similarity graded according to variant sites. The draft genome sequences of the subset of 99 bronchiectasis isolates were examined for the presence of large (>10 kb) deletions (figure 3). A total of 36 different deletions (25 >100 kb), ranging in size from 11 to 300 kb and representing independent genetic events, were identified (online supplementary table S4). These were distributed across 28 genomes in the 99-member genome subset. Most genomes had only one deletion, although two (C54 and C164) had three deletions and four (A119, C4, C85 and C119) had two. In most cases, isolates of the same clonal lineages from the same patient shared the same deletions. However, in patients 45, 55, 79 and 92 not all isolates of the same lineage had the same deletion. The genomes of isolate pairs C6/C7, C29/C30, C105/109, C139/C141 and C156/C159, which are from different patients but vary at <200 sites (table 1), were indistinguishable by BRIG analysis (example shown in figure 3b).

ns. However, in patients 45, 55, 79 and 92 not all isolates of the same lineage had the same deletion. The genomes of isolate pairs C6/C7, C29/C30, C105/109, C139/C141 and C156/C159, which are from different patients but vary at <200 sites (table 1), were indistinguishable by BRIG analysis (example shown in figure 3b). FIGURE 3 Examples of alignment of genomes of bronchiectasis strains with that of reference strain Pseudomonas aeruginosa PAO1. Sequences identified as present (dark grey) or absent (white) in the genome of PAO1 are indicated. a) Isolates of the same lineage (ST-253) from the same patient. From innermost to outermost: C95, C97, C98, C99 and C96. A deletion present in isolate C96 only is highlighted (arrow). b) Pairs of isolates (from innermost to outermost: C6 and C7; and C156 and C159) that both share the same clonal lineage but are from different patients attending the same hospital. Isolates C6 and C7 share a large deletion and isolates C156 and C159 share a smaller overlapping deletion (online supplementary table S4), as indicated (arrow). c) Isolates of different lineages from the same patient. From innermost to outermost: A77, A80 and A85 (all ST-175); and A78, A81 and A82 (all ST-17). A large deletion present in the ST-17 isolates is indicated by an arrow. The figures were generated using the BLAST Ring Image Generator [19]. TABLE 1 Summary of genomic diversity observed within the same clonal lineage of Pseudomonas aeruginosa in individual patients

FIGURE 3 Examples of alignment of genomes of bronchiectasis strains with that of reference strain Pseudomonas aeruginosa PAO1. Sequences identified as present (dark grey) or absent (white) in the genome of PAO1 are indicated. a) Isolates of the same lineage (ST-253) from the same patient. From innermost to outermost: C95, C97, C98, C99 and C96. A deletion present in isolate C96 only is highlighted (arrow). b) Pairs of isolates (from innermost to outermost: C6 and C7; and C156 and C159) that both share the same clonal lineage but are from different patients attending the same hospital. Isolates C6 and C7 share a large deletion and isolates C156 and C159 share a smaller overlapping deletion (online supplementary table S4), as indicated (arrow). c) Isolates of different lineages from the same patient. From innermost to outermost: A77, A80 and A85 (all ST-175); and A78, A81 and A82 (all ST-17). A large deletion present in the ST-17 isolates is indicated by an arrow. The figures were generated using the BLAST Ring Image Generator [19]. TABLE 1 Summary of genomic diversity observed within the same clonal lineage of Pseudomonas aeruginosa in individual patients Isolates SNPs INDELs Patient 147 15 336.35; 261.00 (88–640) 15.20; 14.00 (0–35) Patient 148 (ST17) 4 451.50; 482.50 (159–654) 23.83; 25.50 (6–34) Patient 148 (ST175) 11 195.45; 179.00 (79–403) 9.27; 5.00 (0–36) Patient 149 14 209.01; 206.00 (68–327) 11.40; 10.00 (3–28) Data are presented as n or mean; median (range). The number of single nucleotide polymorphisms (SNPs) and small insertion and deletion (INDEL) differences between the genomes of contemporary isolates from single sputum samples are presented.

) 9.27; 5.00 (0–36) Patient 149 14 209.01; 206.00 (68–327) 11.40; 10.00 (3–28) Data are presented as n or mean; median (range). The number of single nucleotide polymorphisms (SNPs) and small insertion and deletion (INDEL) differences between the genomes of contemporary isolates from single sputum samples are presented. Genomic diversity of isolates within patients can be similar to diversity between patients In order to further assess the within-patient diversification exhibited by P. aeruginosa populations, larger sets of isolates from single sputum samples were analysed for three patients: 147 (15 isolates), 148 (15 isolates) and 149 (14 isolates) (table 1). For two of these patients, the P. aeruginosa population comprised a single clonal lineage. For patient 148, two distinct clonal lineages were identified and these two sets of isolates were analysed separately. In all four isolate sets analysed, the maximum pairwise SNP variations between two isolates of the same lineage was >300, with a median of ≥79 (table 1), indicating the occurrence of within-patient diversification.

48, two distinct clonal lineages were identified and these two sets of isolates were analysed separately. In all four isolate sets analysed, the maximum pairwise SNP variations between two isolates of the same lineage was >300, with a median of ≥79 (table 1), indicating the occurrence of within-patient diversification. Loss-of-function mutations and deletions identified in multiple isolates We used variant calling approaches to identify independent occurrences of loss-of-function mutations within the subset of 99 bronchiectasis isolate genomes. This yielded a number of examples of genes with known functions carrying independent loss-of-function mutations in multiple isolates (table 2 and online supplementary table S5). These include genes linked to mucoidy, virulence, osmoprotection, biofilm formation, motility, DNA repair and antimicrobial resistance (table 2). The genes encoding all three components of the MexAB-OprM efflux pump appear among the most common loss-of-function mutations. In addition, multiple isolates carried loss-of-function mutations in genes encoding regulators (including lasR, algU, fleR and vfr). Among the 99 bronchiectasis isolates, the number of genes with loss-of-function mutations as listed in table 2 ranged from zero to six (online supplementary figure S2 and table S6). TABLE 2 Loss-of-function mutations occurring in multiple isolates

Loss-of-function mutations and deletions identified in multiple isolates We used variant calling approaches to identify independent occurrences of loss-of-function mutations within the subset of 99 bronchiectasis isolate genomes. This yielded a number of examples of genes with known functions carrying independent loss-of-function mutations in multiple isolates (table 2 and online supplementary table S5). These include genes linked to mucoidy, virulence, osmoprotection, biofilm formation, motility, DNA repair and antimicrobial resistance (table 2). The genes encoding all three components of the MexAB-OprM efflux pump appear among the most common loss-of-function mutations. In addition, multiple isolates carried loss-of-function mutations in genes encoding regulators (including lasR, algU, fleR and vfr). Among the 99 bronchiectasis isolates, the number of genes with loss-of-function mutations as listed in table 2 ranged from zero to six (online supplementary figure S2 and table S6). TABLE 2 Loss-of-function mutations occurring in multiple isolates Gene PAO1 gene number Independent occurrences of a mutation n Function/comment mexB PA0426 16 Transporter from MexAB-OprM efflux pump, antibiotic resistance, virulence mucA PA0763 13 Anti-σ factor, mutations can lead to mucoidy betT2 PA5291 9 Transporter, uptake of small molecules such as choline and glycine betaine, contributing to growth via phosphatidyl choline metabolism and osmoprotection bifA PA4367 7 Cyclic-di-GMP phosphodiesterase, inversely regulates biofilm formation mexA PA0425 7 Membrane fusion protein from MexAB-OprM efflux pump, antibiotic resistance, virulence pcoA PA2065 7 Copper resistance PA4469 PA4469 7 Hypothetical protein encoded by a gene in same operon as and upstream of sodM (superoxide dismutase; response to oxidative stress) rbdA PA0861 7 Cyclic-di-GMP phosphodiesterase, modulation of biofilm dispersal, negative regulation of Pel production pilJ PA0411 6 Methyl-accepting chemotaxis receptor-like protein involved in twitching motility and biofilm formation oprM PA0427 6 Outer membrane protein from MexAB-OprM efflux pump, antibiotic resistance, virulence oprF PA1777 6 Major porin, biofilm formation chpA PA0413 5 Chemotaxis-like chemosensory protein involved in twitching motility fimV PA3115 5 Peptidoglycan-binding protein, promotes type IV pilin assembly, twitching motility ladS PA3974 5 Sensor kinase, implicated in switch between acute and chronic infection mutL PA4946 5 Mismatch repair system, DNA repair, mutation can lead to mutator phenotype gmd PA5453 5 GDP-mannose 4,6-dehydratase, mexS PA2491 5 Mutations promote MexT-dependent mexEF-oprN expression and multidrug resistance pchE PA4226 5 Pyochelin synthesis PA0054 PA0054 5 Hypothetical protein Only mutations predicted to lead to loss-of-function were included (i.e. introduction of a stop codon, or a frame-shift mutation). The number of independent mutations indicates the number of isolates carrying unique mutations in the listed gene. Those genes where the number of independent occurrences of a mutation was equal to or greater than five are shown.

-of-function were included (i.e. introduction of a stop codon, or a frame-shift mutation). The number of independent mutations indicates the number of isolates carrying unique mutations in the listed gene. Those genes where the number of independent occurrences of a mutation was equal to or greater than five are shown. Hypermutability is a common trait among CF isolates of P. aeruginosa. Of the 99 panel isolates, 11 carried loss-of-function mutations in the DNA mismatch repair genes mutS or mutL (online supplementary table S1). All but two of these were confirmed as having the hypermutable phenotype. An alignment of all of the genomes containing deletions >10 kb relative to the genome of strain PAO1 revealed a strikingly nonrandom distribution, with 30 of the 36 deletions lying within the 1.9- to 2.8-Mb portion of the strain PAO1 genome. Genes within this region include the psl genes, encoding an extracellular polysaccharide [26], genes encoding the siderophore pyoverdine and genes encoding a type VI secretion apparatus [27].

trikingly nonrandom distribution, with 30 of the 36 deletions lying within the 1.9- to 2.8-Mb portion of the strain PAO1 genome. Genes within this region include the psl genes, encoding an extracellular polysaccharide [26], genes encoding the siderophore pyoverdine and genes encoding a type VI secretion apparatus [27]. Next, we specifically examined one representative of each of the 99 clonal lineages for the presence or absence of genes associated with pathogenicity (online supplementary table S6). 23 of these genomes lacked one or more of the psl genes. In contrast, all of the genomes contained all of the alg genes required for making alginate and the pel genes required for making Pel exopolysaccharide. 11 of the genomes lacked genes required for synthesis of pyoverdine, with nine of these also lacking an fpvA receptor gene for uptake of ferripyoverdine, although the genes required for synthesis of an alternative siderophore, pyochelin, were present in all cases. In addition, 11 of the genomes lacked two or more genes of the type VI secretion system (PA2360 (hsiA3)–PA2373 (vgrG3)) (online supplementary table S6). These findings are consistent with the occurrence of deletions of the region of the genome containing Psl, pyoverdine and type VI secretion genes in multiple isolates, although in some isolates smaller deletions (<10 kb) were detected.

VI secretion system (PA2360 (hsiA3)–PA2373 (vgrG3)) (online supplementary table S6). These findings are consistent with the occurrence of deletions of the region of the genome containing Psl, pyoverdine and type VI secretion genes in multiple isolates, although in some isolates smaller deletions (<10 kb) were detected. Discussion We used whole-genome sequencing to obtain a cross-section of the diversity of P. aeruginosa strains causing infections in bronchiectasis in the UK. Our data suggest that the distribution of P. aeruginosa lineages found among the bronchiectasis isolate collection broadly represents what is present in the global P. aeruginosa population. In contrast to CF [14], we found no data to suggest that there is a widespread transmissible strain among the UK non-CF bronchiectasis community. However, our study did not include large numbers of patients from individual centres. Lineages such as PA14-like and clone C, that are naturally more abundant in nature [23], were among the most abundant in the bronchiectasis collection. Because some lineages are naturally more abundant, their occurrence (based on MLST) in multiple patients is not necessarily indicative of common source or cross-infection. Whole-genome sequencing offers higher resolution than methods such as MLST, allowing us to address this issue.

ant in the bronchiectasis collection. Because some lineages are naturally more abundant, their occurrence (based on MLST) in multiple patients is not necessarily indicative of common source or cross-infection. Whole-genome sequencing offers higher resolution than methods such as MLST, allowing us to address this issue. In a previous comparison of paired isolates from patients within the same bronchiectasis centre, in most patients (nine out of 10) the two isolates shared a common genotype, with one patient found to be infected with two strains simultaneously [15]. In this study, of 24 patients from whose samples multiple isolates were examined, seven had multilineage infections. Similar multilineage infections have also been reported in CF, generally associated with children [28]. In addition, a number of studies in CF have also demonstrated the phenotypic [9, 11, 12] and genomic [10, 13, 29] diversification of single-lineage P. aeruginosa populations in the CF lung. Here, we show for the first time that similar diversification occurs during infections of non-CF bronchiectasis patients. Both the prevalence of multilineage infections and the diversification that occurs during the infection process emphasise the need to be cautious when interpreting the analysis of sputum samples based on single isolates of P. aeruginosa.

imilar diversification occurs during infections of non-CF bronchiectasis patients. Both the prevalence of multilineage infections and the diversification that occurs during the infection process emphasise the need to be cautious when interpreting the analysis of sputum samples based on single isolates of P. aeruginosa. We found several examples of isolates from patients attending the same centre that not only shared the same clonal lineage, but also differed genomically by <200 sites. Genomic variations between isolates from the same patient sample revealed similar, and in some cases higher, levels of variation. The occurrence of isolates with very high genetic relatedness in different patients strongly implies that there has been common source acquisition or cross-infection. The extent of the nucleotide variations differentiating two isolates will be dependent upon 1) the length of time since the transmission event and 2) the rate of mutation of the P. aeruginosa population during the infection. Further studies will be needed to better define the role of cross-infection or common source acquisitions in this patient group.

ations differentiating two isolates will be dependent upon 1) the length of time since the transmission event and 2) the rate of mutation of the P. aeruginosa population during the infection. Further studies will be needed to better define the role of cross-infection or common source acquisitions in this patient group. There was clear evidence for bacterial adaptation to the lung environment by the accumulation of mutations and deletions, including loss-of-function mutations in genes identified previously as being commonly mutated in CF, such as mucA (mucoidy) and lasR (quorum sensing). However, it is worth noting that mutations in genes encoding some of the regulators highlighted in previous CF studies (mexT, retS, exsD and ampR) were observed either infrequently (two mexT and two ampR mutants) or not at all (online supplementary table S5). Mutations in global regulators potentially affect numerous processes. In CF, the pathoadaptive genes identified in different studies have varied, suggesting that there are multiple routes to adaptation to the CF lung [7, 8], a scenario which is likely to apply also to non-CF bronchiectasis.

ine supplementary table S5). Mutations in global regulators potentially affect numerous processes. In CF, the pathoadaptive genes identified in different studies have varied, suggesting that there are multiple routes to adaptation to the CF lung [7, 8], a scenario which is likely to apply also to non-CF bronchiectasis. Loss-of-function mutations in genes encoding the MexAB-OprM efflux pump were common among the bronchiectasis isolates. Although generally thought of as a multidrug efflux system important for antibiotic resistance, this system has been implicated in virulence [30]. Hence, although it may seem counterintuitive that P. aeruginosa should adapt by losing an antibiotic resistance-related efflux pump, it may be that the driver for selection is related to a function other than antibiotic efflux. In contrast, the loss-of-function mutations in mexS can be linked directly to antibiotic resistance, since mutations in mexS promote upregulation of the MexEF-OprN MDR efflux pump [31].

ibiotic resistance-related efflux pump, it may be that the driver for selection is related to a function other than antibiotic efflux. In contrast, the loss-of-function mutations in mexS can be linked directly to antibiotic resistance, since mutations in mexS promote upregulation of the MexEF-OprN MDR efflux pump [31]. The prevalence among non-CF bronchiectasis isolates of deletions in a specific genomic region encoding pyoverdine and Psl polysaccharide was higher than in a dataset of 331 P. aeruginosa clinical isolate genomes [22], where 22 genomes lacked one or more psl genes, only three lacked one or more of the pyoverdine synthesis genes and only one did not have an fpvA receptor gene. P. aeruginosa can utilise multiple pathways for iron acquisition [32]. During chronic lung infections in CF, P. aeruginosa adapts by favouring the heme utilisation route for iron acquisition rather than the pyoverdine siderophore system [33]. Our observations suggest a similar adaptation in non-CF bronchiectasis.

gene. P. aeruginosa can utilise multiple pathways for iron acquisition [32]. During chronic lung infections in CF, P. aeruginosa adapts by favouring the heme utilisation route for iron acquisition rather than the pyoverdine siderophore system [33]. Our observations suggest a similar adaptation in non-CF bronchiectasis. In order to protect itself from hostile environmental conditions or host defences P. aeruginosa can produce three exopolysaccharides contributing to biofilm formation: alginate, Psl and Pel [26]. It has been suggested that Psl is a key surface attachment determinant [34], whereas in the CF lung free-floating biofilm structures may be more important [35]. Other mutations favouring the production of Pel rather than Psl include mutations in bifA [36], rbdA [37], oprF [38] and ladS [39]. Hence, overall our observations indicate that in non-CF BE chronic lung infections, the Pel and alginate exopolysaccharides are favoured over Psl. Other common loss-of-function mutations (in pilJ, chpA and fimV) are implicated in lost or amended twitching motility, an adaptation also seen both in CF [8] and in an artificial sputum biofilm model [40], suggesting that this may be an adaptation related to the viscosity of the sputum environment.

In order to protect itself from hostile environmental conditions or host defences P. aeruginosa can produce three exopolysaccharides contributing to biofilm formation: alginate, Psl and Pel [26]. It has been suggested that Psl is a key surface attachment determinant [34], whereas in the CF lung free-floating biofilm structures may be more important [35]. Other mutations favouring the production of Pel rather than Psl include mutations in bifA [36], rbdA [37], oprF [38] and ladS [39]. Hence, overall our observations indicate that in non-CF BE chronic lung infections, the Pel and alginate exopolysaccharides are favoured over Psl. Other common loss-of-function mutations (in pilJ, chpA and fimV) are implicated in lost or amended twitching motility, an adaptation also seen both in CF [8] and in an artificial sputum biofilm model [40], suggesting that this may be an adaptation related to the viscosity of the sputum environment. Our study represents the first comparative genomics analysis of multiple P. aeruginosa isolates associated with chronic lung infections of non-CF bronchiectasis patients. Although a larger, more targeted study, analysing greater numbers of isolates per sample, would be needed to determine the true prevalence of multilineage infections, this observation does suggest that it is common for multiple P. aeruginosa lineages to coexist in bronchiectasis infections. Our study also demonstrates that within-sample diversity can be comparable in scale to the genetic variations that occur between isolates from different patients attending the same centre. These observations suggest that there is an urgent need for more detailed and larger scale longitudinal studies in non-CF patients, and for surveillance that captures the diversity within centres and would identify cross-infection or common source acquisition events earlier, allowing measures to be taken in order to minimise the spread of this important pathogen.

rgent need for more detailed and larger scale longitudinal studies in non-CF patients, and for surveillance that captures the diversity within centres and would identify cross-infection or common source acquisition events earlier, allowing measures to be taken in order to minimise the spread of this important pathogen. Supplementary material 10.1183/13993003.02108-2016.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-02108-2016_Supplement Figure S1. Core genome SNP phylogeny showing the distribution of bronchiectasis isolates. The figure shows analysis of the genomes of all bronchiectasis isolates used in this study (highlighted in blue) alongside 331 genomes from Kos et al. [14] and the genomes of commonly studied strains PAO1 (labelled PAO1107), PA14 (UCBPPPA14109), PA7 and LESB58. Line colours indicate the two major clusters of P. aeruginosa (I, green; II, blue) as well as those isolates not clustering in the two main groups (red). ERJ-02108-2016_Figure_S1 Table S4. Clone-specific deletions, relative to PAO1. ERJ-02108-2016_Table_S4 Acknowledgements We would like to thank the investigators and patients involved in the PROMIS trial (ISRCTN 49790596). We also thank Paul Roberts (Royal Liverpool and Broadgreen University Hospitals NHS Trust, Liverpool, UK) for technical assistance. This article has supplementary material available from erj.ersjournals.com Earn CME accreditation by answering questions about this article. You will find these at erj.ersjournals.com/journal/cme

Acknowledgements We would like to thank the investigators and patients involved in the PROMIS trial (ISRCTN 49790596). We also thank Paul Roberts (Royal Liverpool and Broadgreen University Hospitals NHS Trust, Liverpool, UK) for technical assistance. This article has supplementary material available from erj.ersjournals.com Earn CME accreditation by answering questions about this article. You will find these at erj.ersjournals.com/journal/cme Support statement: This work was supported by the UK Medical Research Council as part of the BRONCH-UK grant (reference MR/L011263/1) awarded to A. De Soyza as principal investigator, with C. Winstanley among the co-investigators. I.L. Lamont would like to acknowledge funding support from Cystic Fibrosis New Zealand, the New Zealand Lotteries Board (Health) and the National Health and Medical Research Council (Australia). M.P. Moore and C. Winstanley would like to acknowledge the support of the Cystic Fibrosis Trust (grant RS34). Funding information for this article has been deposited with the Open Funder Registry. Conflict of interest: None declared.

Timing of PCS in respect to the patient's death Only 21.4% of deceased COPD patients received PCS (5595 out of 26 135), of whom 31.5% had a co-diagnosis of lung cancer (1764 out of 5595). Of those that had received PCS, 48.4% received it only within 6 months of their death (of whom 29.1% had lung cancer) (figure 3). FIGURE 3 Cumulative proportion of deceased chronic obstructive pulmonary disease (COPD) patients that first received palliative care support (PCS) in relation to the time of their death and the proportion of those that had a lung cancer co-diagnosis. Using multinomial regression, it can be seen that patients with COPD and lung cancer were more likely to receive PCS earlier in their last 12 months of life (≤11 months pre-death) than patients who only had a COPD diagnosis (figure 4). There was a 40% increased chance of receiving support within 1–6 months pre-death and 30% increased chance of receiving PCS within 6–12 months pre-death, compared with receiving palliative care at a late time period (≤1 month pre-death), for patients with both COPD and lung cancer compared with not having lung cancer (middle versus late for “COPD+lung cancer” versus “COPD alone”: risk ratio 1.4, 95% CI 1.3–1.7; p<0.0001; early versus late for “COPD+lung cancer” versus “COPD alone”: risk ratio 1.3, 95% CI 1.1–1.6; p<0.01).

e period (≤1 month pre-death), for patients with both COPD and lung cancer compared with not having lung cancer (middle versus late for “COPD+lung cancer” versus “COPD alone”: risk ratio 1.4, 95% CI 1.3–1.7; p<0.0001; early versus late for “COPD+lung cancer” versus “COPD alone”: risk ratio 1.3, 95% CI 1.1–1.6; p<0.01). FIGURE 4 Patients with a dual diagnosis, i.e. chronic obstructive pulmonary disease (COPD) and lung cancer, were more likely to receive palliative care support (PCS) earlier before their death than patients with COPD alone. RR: risk ratio (“COPD+lung cancer” versus “COPD”). Includes only deceased COPD patients that received PCS within 12 months of their death.

.e. chronic obstructive pulmonary disease (COPD) and lung cancer, were more likely to receive palliative care support (PCS) earlier before their death than patients with COPD alone. RR: risk ratio (“COPD+lung cancer” versus “COPD”). Includes only deceased COPD patients that received PCS within 12 months of their death. Discussion From a large primary care cohort of 92 365 COPD patients, only 7.8% of the whole cohort and 21.4% of deceased patients received PCS during the study follow-up. However, just under a third of these patients also had lung cancer; it is highly probable that many patients were receiving PCS in respect of their cancer diagnosis rather than their airways disease. Indeed, only 6% of COPD patients without lung cancer were provided with PCS compared with 50% of the lung cancer patients in the whole cohort. A co-diagnosis of lung cancer had the strongest association with PCS compared with any of the other measured patient characteristics, in both the whole cohort (15 times increased odds of receiving PCS) and in only the deceased patients (6 times increased odds of receiving PCS). Therefore, the largest influence on PCS for COPD patients was not related to the advancement of their airways disease but to having lung cancer. However, it is encouraging that the proportion of COPD patients who received PCS steadily increased between 2005 and 2014, in particular for the deceased COPD patients, both with and without lung cancer. The rise was steepest from 2009 onwards, which could be related to the introduction of the national End of Life Care Strategy, a programme developed to specifically reach patients with nonmalignant life-limiting diseases, and is in keeping with data showing a reduction in undesirable hospital deaths in COPD patients over the same time period [26].

Introduction Globally, chronic obstructive pulmonary disease (COPD) prevalence is on the rise and is the only disease that continues to have an increasing age-adjusted mortality rate [1]. In the UK, 5.2% of all deaths are secondary to COPD, which is approaching the proportion (6.2%) of deaths that are due to lung cancer [2, 3]. Patients with advanced COPD have a burden of disabling physical symptoms that are often compounded by multiple comorbidities, psychological distress and social isolation. Currently, there is a lack of palliative care support (PCS) for patients with end-stage COPD, despite evidence that it improves their quality of life [4, 5]. The natural course of physical decline for patients with COPD can be variable, but overall it is characterised by a long-term steady deterioration [6]. However, most healthcare resources are dedicated towards management and prevention of acute events, with significantly less emphasis from physicians and researchers on palliative and supportive care [7].

ine for patients with COPD can be variable, but overall it is characterised by a long-term steady deterioration [6]. However, most healthcare resources are dedicated towards management and prevention of acute events, with significantly less emphasis from physicians and researchers on palliative and supportive care [7]. The 2010 National Institute for Health and Clinical Excellence (NICE) guidelines state that patients with end-stage COPD should have access to the full range of services offered by palliative care teams [8]. These include palliation for breathlessness and other symptoms, advance care planning, addressing emotional and social needs, and end-of-life care. However, compared with patients with lung cancer who suffer from a similar burden of disabling symptoms and psychological distress, small studies or studies using select COPD cohorts have found over the past decade that COPD patients often have limited access to palliative care services [9–13], even when compared with patients with other chronic terminal diseases [14]. It is likely that a major reason for this is the difficulty in predicting when to initiate PCS; current recommendations are to deliver PCS during the last year of life, but prognosticating the life expectancy of individuals with COPD has proven to be extremely difficult [1, 15, 16]. Consequentially, there is no commonly accepted definition of “end-stage COPD”. In addition, it is well recognised that communication between COPD patients with advanced disease and physicians remains poor regarding disease prognosis and end-of-life planning, which is likely to be related to a lack of training and guidance [17, 18].

ly, there is no commonly accepted definition of “end-stage COPD”. In addition, it is well recognised that communication between COPD patients with advanced disease and physicians remains poor regarding disease prognosis and end-of-life planning, which is likely to be related to a lack of training and guidance [17, 18]. Previous studies have investigated the planning of PCS from acute UK hospital trusts; unfortunately, these have consistently found poor access for COPD patients to palliative care services, alongside limited provision of patient information [19, 20]. This is the first nationally representative study to describe the uptake of PCS in COPD patients, the characteristics that are associated with receiving PCS and how this has changed over the past decade. Methods Data sources The UK Clinical Practice Research Datalink (CPRD) includes 674 general practitioner (GP) practices and has current coverage of over 11.3 million patients, and represents the UK's population with respect of age, sex, body mass index (BMI) and ethnicity [21]. Approximately 60% of CPRD practices have patient-level linkage to Hospital Episode Statistics (HES) data and Office of National Statistics (ONS) mortality data.

and has current coverage of over 11.3 million patients, and represents the UK's population with respect of age, sex, body mass index (BMI) and ethnicity [21]. Approximately 60% of CPRD practices have patient-level linkage to Hospital Episode Statistics (HES) data and Office of National Statistics (ONS) mortality data. Study design and population We conducted an open cohort study of COPD patients that contribute towards routinely collected CPRD–HES–ONS linked electronic healthcare records (figure 1). Patients had to have a COPD diagnosis as determined using a previous validated algorithm [22]. Patients entered the study 1 year after their latest COPD diagnosis date, the date the practice began recording research quality data, their continuous CPRD-GP registration date or study start date (December 31, 2004). Patients were censored at date of death, the end of the study period (March 31, 2015), the GP practice last data collection date or the date of transfer out of a CPRD-linked practice. FIGURE 1 Flowchart of patient inclusion in the study. COPD: chronic obstructive pulmonary disease; CPRD: Clinical Practice Research Datalink; HES: Hospital Episode Statistics; ONS: Office of National Statistics; PCS: palliative care support.

Study design and population We conducted an open cohort study of COPD patients that contribute towards routinely collected CPRD–HES–ONS linked electronic healthcare records (figure 1). Patients had to have a COPD diagnosis as determined using a previous validated algorithm [22]. Patients entered the study 1 year after their latest COPD diagnosis date, the date the practice began recording research quality data, their continuous CPRD-GP registration date or study start date (December 31, 2004). Patients were censored at date of death, the end of the study period (March 31, 2015), the GP practice last data collection date or the date of transfer out of a CPRD-linked practice. FIGURE 1 Flowchart of patient inclusion in the study. COPD: chronic obstructive pulmonary disease; CPRD: Clinical Practice Research Datalink; HES: Hospital Episode Statistics; ONS: Office of National Statistics; PCS: palliative care support. Outcome and variables The primary outcome was documentation of PCS, as recorded in the patients' electronic healthcare records using relevant Read Codes (the clinical terminology system used by CPRD); two-thirds of codes were for “terminal care”, “palliative care”, “on gold standards palliative care framework” or “palliative care plan review”; other commonly used codes included “palliative care treatment”, “palliative care specialist” or “seen by palliative care physician” (supplementary table S1). A history of smoking (current or ex-smoker), myocardial infarction, lung cancer, heart failure, stroke, anxiety and depression was identified using appropriate Read Codes. BMI (kg·m−2) was categorised as underweight (<19 kg·m−2), normal (19–25 kg·m−2), overweight (25–30 kg·m−2) or obese (≥30 kg·m−2). Exacerbations were identified using a validated algorithm [23]; the baseline year was defined as the year before study entry. A moderate exacerbation was defined as having been treated within the GP practice, while a severe exacerbation was defined as requiring hospitalisation; exacerbations recorded within 14 days of the index exacerbation were considered as the same event. COPD severity was classified using Global Initiative for Chronic Obstructive Lung Disease (GOLD) staging classification and the Medical Research Council (MRC) Dyspnoea scale [24, 25]. Death was defined as death from any cause in the analysis; however, in a sensitivity analysis, we investigated a subgroup of patients for whom it was highly likely COPD led directly to the cause of death based upon ONS data.

e (GOLD) staging classification and the Medical Research Council (MRC) Dyspnoea scale [24, 25]. Death was defined as death from any cause in the analysis; however, in a sensitivity analysis, we investigated a subgroup of patients for whom it was highly likely COPD led directly to the cause of death based upon ONS data. Statistical analysis Logistic regression was used to calculate the association between patient characteristics and PCS, for the whole cohort and for the deceased patients. Multivariate logistic regression was conducted using the following variables: sex, age at study entry, history of myocardial infarction, stroke, lung cancer, anxiety, depression, heart failure, BMI, GOLD stage, MRC Dyspnoea score, smoking status and number of exacerbations in the baseline year. These demographic and clinical variables, taking into account disease severity and comorbidity, were selected as known or suspected risk factors for receiving PCS [11–13]. Where data was <25% missing, a complete case analysis was undertaken; for variables with >25% missing, the variable was excluded from the multivariate regression model (GOLD stage in the whole cohort and deceased subpopulation, and MRC Dyspnoea score in the deceased subpopulation) and a sensitivity analysis was then carried out to include that variable in the model. Multinomial regression was applied to estimate the relative effect of a coexisting lung cancer diagnosis on the timing of the PCS in relation to time of death. Two time periods, i.e. “early period” (6–12 months before death) and “medium period” (1–6 months before death), were compared to the “late period” (≤1 month before death). Statistical analysis was conducted using Stata version 14.1 (StataCorp, College Station, TX, USA).

e timing of the PCS in relation to time of death. Two time periods, i.e. “early period” (6–12 months before death) and “medium period” (1–6 months before death), were compared to the “late period” (≤1 month before death). Statistical analysis was conducted using Stata version 14.1 (StataCorp, College Station, TX, USA). Ethics The protocol for this research was approved by the Independent Scientific Advisory Committee (ISAC) for Medicines and Healthcare Products Regulatory Agency Database Research (protocol 17_083); the approved protocol was made available during peer review. Generic ethical approval for observational research using the CPRD with approval from the ISAC was granted by a Health Research Authority Research Ethics Committee (East Midlands – Derby; 05/MRE04/87). Results Characteristics of the whole cohort There were 92 365 eligible COPD patients in the cohort (figure 1). At study entry, mean age was 67.8 years, 49 801 (53.9%) were male, 18 962 (20.6%) had cardiovascular disease (myocardial infarction, stroke or heart failure), 31 260 (33.8%) had anxiety or depression, 30 656 (34.1%) had a normal BMI, 46 900 (52.3%) were current smokers, 19 841 (35.2%) had GOLD stage ≥3, 18 295 (25.5%) had MRC Dyspnoea score ≥4, 46 898 (50.8%) had no exacerbations in their baseline year and 26 135 (28.3%) died during study follow-up (table 1). TABLE 1 Demographic and clinical characteristics of the whole chronic obstructive pulmonary disease (COPD) cohort and the deceased COPD patients from 2005 to 2014

Results Characteristics of the whole cohort There were 92 365 eligible COPD patients in the cohort (figure 1). At study entry, mean age was 67.8 years, 49 801 (53.9%) were male, 18 962 (20.6%) had cardiovascular disease (myocardial infarction, stroke or heart failure), 31 260 (33.8%) had anxiety or depression, 30 656 (34.1%) had a normal BMI, 46 900 (52.3%) were current smokers, 19 841 (35.2%) had GOLD stage ≥3, 18 295 (25.5%) had MRC Dyspnoea score ≥4, 46 898 (50.8%) had no exacerbations in their baseline year and 26 135 (28.3%) died during study follow-up (table 1). TABLE 1 Demographic and clinical characteristics of the whole chronic obstructive pulmonary disease (COPD) cohort and the deceased COPD patients from 2005 to 2014 Whole cohort Deceased patients Whole cohort Received PCS No PCS Deceased patients Received PCS No PCS Total (% of total) n=92 365 (100) n=7198 (7.8) n=85 167 (92.2) n=26 135 (100) n=5595 (21.4) n=20 540 (78.6) Follow-up years 4.2 (1.9–7.2) 4.3 (2.2–7.0) 4.1 (1.9–7.3) 3.3 (1.5–5.6) 3.9 (2–6.3) 3.1 (1.4–5.3) Age years 67.8±11.4 71.8±9.7 67.5±11.6 74±9.6 72.3±9.3 74.5±9.5 Age category years ≤65 35 994 (39) 1703 (23.7) 34 291 (40.3) 4628 (17.7) 3408 (16.6) 1220 (21.8) 65–75 29 424 (31.9) 2624 (36.5) 26 800 (31.5) 8339 (31.9) 6296 (30.7) 2043 (36.5) ≥75 26 947 (29.2) 2871 (39.9) 24 076 (28.3) 13 168 (50.4) 10 836 (52.8) 2332 (41.7) Male 49 801 (53.9) 4134 (57.4) 45 667 (53.6) 15 215 (58.2) 3256 (58.2) 11 959 (58.2) Myocardial infarction 6232 (7.4) 726 (10.1) 6309 (7.4) 3004 (11.5) 576 (10.3) 2428 (11.8) Stroke 5526 (6.1) 573 (8) 5062 (5.9) 2390 (9.1) 451 (8.1) 1939 (9.4) Heart failure 9673 (10.6) 1140 (15.8) 8747 (10.3) 5335 (20.4) 903 (16.1) 4432 (21.6) Lung cancer 4138 (4.5) 2071 (28.8) 2380 (2.8) 3121 (11.9) 1764 (31.5) 1357 (6.6) Anxiety 19 990 (21.9) 1765 (24.5) 18 536 (21.8) 5425 (20.8) 1356 (24.2) 4069 (19.8) Depression 21 836 (23.9) 1693 (23.5) 20 459 (24.0) 5375 (20.6) 1268 (22.7) 4107 (20) BMI kg·m−2 n=86 789 n=6 486 n=89 789 n=24 686 n=5382 n=19 304 <19 (underweight) 8144 (9.1) 1060 (15.3) 7084 (8.6) 4075 (16.5) 843 (15.7) 3232 (16.7) 19–25 (normal) 30 656 (34.1) 2780 (40) 27 876 (33.7) 9842 (39.9) 2207 (41) 7635 (39.6) 25–30 (overweight) 27 738 (30.9) 1854 (26.7) 25 884 (31.3) 6272 (25.4) 1417 (26.3) 4855 (25.2) ≥30 (obese) 23 251 (25.9) 1253 (18) 21 998 (26.7) 4497 (18.2) 915 (17) 3582 (18.6) Smoking status n=89 789 n=6726 n=84 392 n=24 686 n=5382 n=19 304 Current smoker 46 900 (52.3) 3440 (49.5) 43 460 (52.5) 11 744 (47.6) 2646 (49.2) 9098 (47.1) Ex-smoker 42 889 (47.8) 3507 (50.5) 39 382 (47.5) 12 942 (52.4) 2736 (50.8) 10 206 (52.9) GOLD stage n=56 159 n=5017 n=59 362 n=16 118 n=3614 n=12 504 1 12 996 (23.1) 917 (19.7) 12 049 (23.4) 3411 (21.2) 723 (20) 2688 (21.5) 2 23 352 (41.6) 1651 (35.5) 21 701 (42.1) 4989 (31) 1249 (34.6) 3740 (29.9) 3 15 259 (27.2) 1442 (31) 13 817 (26.8) 5360 (33.3) 1132 (31.3) 4228 (33.8) 4 4582 (8.2) 643 (13.8) 3939 (7.7) 23

stage n=56 159 n=5017 n=59 362 n=16 118 n=3614 n=12 504 1 12 996 (23.1) 917 (19.7) 12 049 (23.4) 3411 (21.2) 723 (20) 2688 (21.5) 2 23 352 (41.6) 1651 (35.5) 21 701 (42.1) 4989 (31) 1249 (34.6) 3740 (29.9) 3 15 259 (27.2) 1442 (31) 13 817 (26.8) 5360 (33.3) 1132 (31.3) 4228 (33.8) 4 4582 (8.2) 643 (13.8) 3939 (7.7) 23 58 (14.6) 510 (14.1) 1848 (14.8) MRC Dyspnoea score n=71 985 n=4656 n=66 177 n=15 104 n=3668 n=11 436 1 11 168 (15.5) 393 (7.7) 10 775 (16.1) 1156 (7.7) 296 (8.1) 860 (7.5) 2 23 856 (33.1) 1099 (21.6) 22 757 (34) 3097 (20.5) 774 (21.1) 2323 (20.3) 3 18 666 (25.9) 1233 (24.2) 17 433 (26) 3899 (25.8) 877 (23.9) 3 022 (26.4) 4 13 644 (19) 1421 (27.9) 12 223 (18.3) 4431 (29.3) 1011 (27.6) 3420 (29.9) 5 4651 (6.5) 943 (18.5) 3708 (5.5) 2521 (16.7) 710 (19.4) 1811 (15.8) Exacerbations in baseline year n=92 365 n=6726 n=91 201 n=26 135 n=5595 n=20 540 0 46 898 (50.8) 3227 (44.8) 43 671 (51.3) 11 762 (45) 2516 (45) 9246 (45) ≤2 (moderate) 27 991 (30.3) 2178 (30.3) 25 813 (30.3) 7690 (29.4) 1695 (30.3) 5995 (29.9) ≥3 moderate/≥1 severe 17 476 (18.9) 1793 (24.9) 15 683 (18.4) 6683 (25.6) 1384 (24.7) 5299 (25.8) Data are presented as median (interquartile range), mean±sd or n (%), unless otherwise stated. PCS: palliative care support; BMI: body mass index; GOLD: Global Initiative for Obstructive Lung Disease; MRC: Medical Research Council.

58 (14.6) 510 (14.1) 1848 (14.8) MRC Dyspnoea score n=71 985 n=4656 n=66 177 n=15 104 n=3668 n=11 436 1 11 168 (15.5) 393 (7.7) 10 775 (16.1) 1156 (7.7) 296 (8.1) 860 (7.5) 2 23 856 (33.1) 1099 (21.6) 22 757 (34) 3097 (20.5) 774 (21.1) 2323 (20.3) 3 18 666 (25.9) 1233 (24.2) 17 433 (26) 3899 (25.8) 877 (23.9) 3 022 (26.4) 4 13 644 (19) 1421 (27.9) 12 223 (18.3) 4431 (29.3) 1011 (27.6) 3420 (29.9) 5 4651 (6.5) 943 (18.5) 3708 (5.5) 2521 (16.7) 710 (19.4) 1811 (15.8) Exacerbations in baseline year n=92 365 n=6726 n=91 201 n=26 135 n=5595 n=20 540 0 46 898 (50.8) 3227 (44.8) 43 671 (51.3) 11 762 (45) 2516 (45) 9246 (45) ≤2 (moderate) 27 991 (30.3) 2178 (30.3) 25 813 (30.3) 7690 (29.4) 1695 (30.3) 5995 (29.9) ≥3 moderate/≥1 severe 17 476 (18.9) 1793 (24.9) 15 683 (18.4) 6683 (25.6) 1384 (24.7) 5299 (25.8) Data are presented as median (interquartile range), mean±sd or n (%), unless otherwise stated. PCS: palliative care support; BMI: body mass index; GOLD: Global Initiative for Obstructive Lung Disease; MRC: Medical Research Council. During follow-up (median (interquartile range (IQR)) 4.2 (1.9–7.2) years), only 7198 (7.8%) received PCS through primary care. 4138 (4.5%) of the whole cohort were diagnosed with lung cancer by the end of follow-up. In patients without a co-diagnosis of lung cancer, only 5.8% (5127 out of 88 227) received PCS; in contrast, in the subgroup of patients with COPD and lung cancer, 50% (2071 out of 4138) received PCS.

ived PCS through primary care. 4138 (4.5%) of the whole cohort were diagnosed with lung cancer by the end of follow-up. In patients without a co-diagnosis of lung cancer, only 5.8% (5127 out of 88 227) received PCS; in contrast, in the subgroup of patients with COPD and lung cancer, 50% (2071 out of 4138) received PCS. Characteristics of the deceased patients 26 135 patients died during the cohort follow-up. Compared with the whole cohort, deceased patients had a slightly shorter follow-up time (median (IQR) 3.3 (1.5–5.6) years), were slightly older (mean age 74 years), but had a similar percentage of males (table 1). The proportion of patients with a history of myocardial infarction, stroke, heart failure or lung cancer was higher in deceased COPD patients than the cohort as a whole; anxiety and depression were of similar proportions. There was a higher proportion of deceased COPD patients when compared with the whole cohort, who were underweight, ex-smokers, had a higher GOLD stage, higher MRC Dyspnoea scores and more exacerbations in the baseline year (table 1). During follow-up (median (IQR) 3.7 (1.5–5.6) years), only 4655 (22.5%) received or accessed PCS through primary care. In patients without a co-diagnosis of lung cancer, only 16.7% (3831 out of 23 014) received PCS; in contrast, in the subgroup of patients with COPD and lung cancer, 56.5% (1764 out of 3121) received PCS. In the sensitivity analysis, including only patients known to have died from their COPD, there were 15 745 patients, only 15.4% (1641) of whom had received PCS.

ancer, only 16.7% (3831 out of 23 014) received PCS; in contrast, in the subgroup of patients with COPD and lung cancer, 56.5% (1764 out of 3121) received PCS. In the sensitivity analysis, including only patients known to have died from their COPD, there were 15 745 patients, only 15.4% (1641) of whom had received PCS. Longitudinal changes in the proportion of patients receiving PCS from 2005 to 2014 From 2005 to 2014, there was a steady rise in the proportion of patients who received PCS (figure 2). In 2005, 0.5% (216 out of the 39 685 patients in the cohort in 2005) had received support; by 2014, 2.0% (907 out of the 44 493 patients in the cohort in 2014) had received PCS. The proportion of deceased COPD patients that received PCS followed a similar pattern, overall increasing from 1.2% in 2005 to 16.5% in 2014 (figure 2). The rate of increase over the years appeared to be the same for COPD patients with or without lung cancer (supplementary figure S1). FIGURE 2 Proportion of chronic obstructive pulmonary disease patients that received palliative care support (PCS) in each year during study follow-up.

Longitudinal changes in the proportion of patients receiving PCS from 2005 to 2014 From 2005 to 2014, there was a steady rise in the proportion of patients who received PCS (figure 2). In 2005, 0.5% (216 out of the 39 685 patients in the cohort in 2005) had received support; by 2014, 2.0% (907 out of the 44 493 patients in the cohort in 2014) had received PCS. The proportion of deceased COPD patients that received PCS followed a similar pattern, overall increasing from 1.2% in 2005 to 16.5% in 2014 (figure 2). The rate of increase over the years appeared to be the same for COPD patients with or without lung cancer (supplementary figure S1). FIGURE 2 Proportion of chronic obstructive pulmonary disease patients that received palliative care support (PCS) in each year during study follow-up. Association between PCS and patient characteristics In the whole COPD cohort, age ≥75 years, being female, cardiovascular comorbidities (myocardial infarction, stroke or heart failure), anxiety, underweight BMI, MRC Dyspnoea score ≥2 and a history of four or more moderate exacerbations or one or more severe exacerbation in the baseline year were significantly associated with an increased odds of receiving PCS after adjusting for all other characteristics (table 2); being overweight or obese was significantly associated with a reduced odds of receiving PCS (table 2). Lung cancer had the strongest association with receiving PCS, even after adjusting for all other characteristics (adjusted OR 14.7, 95% CI 13.5–16).

ceiving PCS after adjusting for all other characteristics (table 2); being overweight or obese was significantly associated with a reduced odds of receiving PCS (table 2). Lung cancer had the strongest association with receiving PCS, even after adjusting for all other characteristics (adjusted OR 14.7, 95% CI 13.5–16). TABLE 2 Association between lung cancer and other patient characteristics and receiving palliative care support for the whole chronic obstructive pulmonary disease (COPD) cohort and the deceased COPD patients

ceiving PCS after adjusting for all other characteristics (table 2); being overweight or obese was significantly associated with a reduced odds of receiving PCS (table 2). Lung cancer had the strongest association with receiving PCS, even after adjusting for all other characteristics (adjusted OR 14.7, 95% CI 13.5–16). TABLE 2 Association between lung cancer and other patient characteristics and receiving palliative care support for the whole chronic obstructive pulmonary disease (COPD) cohort and the deceased COPD patients Whole cohort Deceased patients Crude OR (95% CI) Adjusted OR# (95% CI) Crude OR (95% CI) Adjusted OR# (95% CI) Lung cancer No 1 (reference) 1 (reference) Yes 14 (13.0–15.0)*** 14.7 (13.5–16.0)*** 6.5 (6–7)*** 6.1 (5.6–6.6)*** Age years ≤65 1 (reference) 1 (reference) 65–75 2.0 (1.9–2.1)*** 1.6 (1.5–1.7) 0.9 (0.8–1)* 0.9 (0.9–1.1) ≥75 2.4 (2.3–2.6)*** 1.9 (1.8–2.1)*** 0.6 (0.6–0.7)*** 0.7 (0.7–0.8)*** Sex Female 1 (reference) 1 (reference) Male 0.9 (0.8–0.9)*** 0.8 (0.8–0.9)*** 1.0 (0.9–1.1) 1.0 (0.9–1.1) Myocardial infarction No 1 (reference) 1 (reference) Yes 1.4 (1.3–1.5)*** 1.2 (1.0–1.3)** 0.9 (0.8–0.9)** 0.9 (0.8–1.0) Stroke No 1 (reference) 1 (reference) Yes 1.4 (1.3–1.5)*** 1.1 (1.0–1.3)* 0.8 (0.8–0.9)** 1 (0.8–1.1) Heart failure No 1 (reference) 1 (reference) Yes 1.6 (1.5–1.8)*** 1.5 (1.3–1.6)*** 0.7 (0.7–0.8)*** 0.9 (0.8–0.9)** Anxiety No 1 (reference) 1 (reference) Yes 1.2 (1.1–1.2)*** 1.3 (1.2–1.4)*** 1.3 (1.2–1.4)*** 1.2 (1.1–1.3)*** Depression No 1 (reference) 1 (reference) Yes 1.0 (0.9–1.0)*** 1.0 (0.9–1.1) 1.2 (1.1–1.3)*** 1.1 (1–1.2) BMI <19 (underweight) 1 (reference) 1 (reference) 19–25 (normal) 1.5 (1.4–1.6)*** 1.4 (1.3–1.6)*** 0.9 (0.8–1)* 0.9 (0.8–1) 25–30 (overweight) 0.7 (0.7–0.8)*** 0.7 (0.7–0.8)*** 1 (0.9–1.1) 1.0 (0.9–1.1) ≥30 (obese) 0.6 (0.5–0.6)*** 0.6 (0.6–0.7)*** 0.9 (0.8–1.0)** 0.9 (0.8–1.0)** Smoking Ex-smoker 1 (reference) 1 (reference) Current smoker 0.9 (0.8–0.9)*** 1 (0.9–1.1) 1.1 (1.0–1.2)** 1.0 (0.9–1.0) GOLD stage 1 1 (reference) 1 (reference) 2 1.0 (0.9–1.1) 1.2 (1.1–1.4)*** 3 1.4 (1.3–1.5)*** 1.0 (0.9–1.1) 4 2.1 (1.9–2.4)*** 1.0 (0.9–1.2) MRC Dyspnoea score 1 1 (reference) 1 (reference) 2 1.3 (1.2–1.5)*** 1.2 (1.1–1.4)** 1 (0.8–1.1) 3 1.9 (1.7–2.2)*** 1.6 (1.4–1.8)*** 0.8 (0.7–1)* 4 3.2 (2.8–3.6)*** 2.4 (2.1–2.7)*** 0.9 (0.7–1)* 5 7.0 (6.2–7.9)*** 5.1 (4.5–5.8)** 1.1 (1–1.3) Exacerbations None 1 (reference) 1 (reference) ≤2 (moderate) 1.1 (1.1–1.2)*** 1.0 (0.9–1.1) 1.0 (1–1.1) 1.0 (1–1.1) ≥3 moderate/≥1 severe 1.6 (1.5–1.6)*** 1.2 (1.1–1.3)*** 1.0 (0.9–1) 1.0 (1–1.1) BMI: body mass index; GOLD: Global Initiative for Obstructive Lung Disease; MRC: Medical Research Council.

09 onwards, which could be related to the introduction of the national End of Life Care Strategy, a programme developed to specifically reach patients with nonmalignant life-limiting diseases, and is in keeping with data showing a reduction in undesirable hospital deaths in COPD patients over the same time period [26]. Palliative care is not just synonymous with end-of-life care, but also centres on symptom management, improving a patient's quality of life, and psychological support for the patient and their family. Indeed, modern palliative care approaches are more needs based rather than prognosis based, appropriate for COPD patient's whose life expectancy is difficult to predict [4]. Two other factors, i.e. anxiety and MRC Dyspnoea score, indicative of requiring support for psychological and symptom management, respectively, were also significantly associated with PCS in the whole cohort, suggesting PCS was considered for symptom control, but to a far lesser degree than a coexisting diagnosis of lung cancer. Having a high BMI was associated with a reduced chance of receiving PCS; this may reflect a lack of an understanding/willingness to consider PCS in obese COPD patients who may suffer with a relatively increased sensation of breathlessness [27]. Interestingly, older age (≥75 years) was associated with an increased odds of receiving PCS in the whole cohort, but younger age was significantly associated with an increased odds of receiving support in the deceased patients. The deceased patients were a frailer cohort with more comorbidities and more severe clinical features; it could be that in this setting a younger patient is more likely to be considered for PCS than an older patient and it is noteworthy that a similar ageist phenomenon has been recognised in studies addressing referral of general cancer patients to palliative care services [28].

dities and more severe clinical features; it could be that in this setting a younger patient is more likely to be considered for PCS than an older patient and it is noteworthy that a similar ageist phenomenon has been recognised in studies addressing referral of general cancer patients to palliative care services [28]. It is also imperative that provision of PCS occurs in a timely manner, not just a few weeks before death [29]. Sadly, this study found that a third of COPD patients only received PCS in the last month of their life. In addition, there was again an inequality between patients with or without a lung cancer diagnosis. Those with lung cancer had an increased chance of receiving PCS earlier in the year before death, whereas patients without lung cancer were much more likely to receive PCS during their final month of life. The inequality between PCS for COPD patients and lung cancer patients has previously also been noted in other countries [17], yet the two diseases are comparable in their manifestation and severity of symptoms. Moreover, several studies have shown that end-stage COPD patients have at least as many, if not more, physical and psychological needs as patients with end-stage lung cancer [30]. It appears that often it was only the terminal care element of PCS that was provided for COPD patients without lung cancer. However, it has been shown previously that even this aspect of PCS is relatively lacking as COPD patients are more likely to die in hospital than lung cancer patients, which is not in keeping with the general preference to die at home [31, 32].

are element of PCS that was provided for COPD patients without lung cancer. However, it has been shown previously that even this aspect of PCS is relatively lacking as COPD patients are more likely to die in hospital than lung cancer patients, which is not in keeping with the general preference to die at home [31, 32]. It has been nearly two decades since a small UK study suggested that PCS was lacking in COPD patients compared with those with lung cancer [13]. Despite an increased recognition during this time that patients with COPD would benefit from a palliative care approach, including implementation of NICE guidelines and the national End of Life Care Programme and Strategy, our data have shown that there has been only a limited increase in PCS uptake. This lack of deliverance from within primary care, including absence of referral to specialist services, has been proposed to be multifactorial and related to difficulties in predicting prognosis, inadequate expertise to engage in palliative care discussions, physician concern in using opioids in chronic respiratory disease, fear of diminishing patients’ hopes and bias against patients with smoking-related lung diseases [27]. Additional strategies are therefore required to help improve access to PCS, including defining and sharing good models of care between the respiratory and palliative care communities, enhanced training in identifying patients who will benefit, communication skills and symptom management, as well as increasing the availability of evidence-based pharmacological and nonpharmacological intervention [4, 33].

defining and sharing good models of care between the respiratory and palliative care communities, enhanced training in identifying patients who will benefit, communication skills and symptom management, as well as increasing the availability of evidence-based pharmacological and nonpharmacological intervention [4, 33]. Study limitations Recording of palliative care provision in primary care is one of the NICE Quality and Outcomes Framework indicators and therefore, fortuitously, this incentive payment programme ensures high recording of PCS. However, it is possible some instances of PCS were not recorded, such as a prescription of opioids/benzodiazepines to palliate breathlessness, but it is likely most GPs would have documented the reason for such prescriptions, especially given they are incentivised to use Quality and Outcomes Framework Read Codes. In this study we only addressed PCS from within primary care, including referrals to specialist services, but not including PCS provided during hospital admissions. Therefore, this study may have excluded some patients who received PCS only through secondary care within the last couple of weeks of their life and were not seen back in primary care after hospital/hospice discharge.

including referrals to specialist services, but not including PCS provided during hospital admissions. Therefore, this study may have excluded some patients who received PCS only through secondary care within the last couple of weeks of their life and were not seen back in primary care after hospital/hospice discharge. It is very difficult to identify which patients have “end-stage” COPD; therefore, as a proxy, we carried out much of our analysis only on patients who had died. It is possible that we included some patients who did not have end-stage COPD as they may have died from unrelated causes. However, using ONS data we also identified a subgroup of patients whose deaths were highly likely to have been related to their COPD and in that analysis the proportion of patients that received PCS in this subgroup was even lower than for the total deceased population. There was missing data for the MRC Dyspnoea score and, as we carried out a complete case analysis for the whole cohort, it is possible there was some selection bias as these patients may have had more severe disease (fewer patients had missing MRC Dyspnoea scores in GOLD stage 4 than stages 1–3); however, this should not have affected the findings except to pull the estimate further from unity. Lastly, it is possible that there were other factors (e.g. social and cultural preferences) which were not included in our analysis that could have affected our findings; however, due to the strength of the association with lung cancer that relationship is unlikely to have changed considerably.

mate further from unity. Lastly, it is possible that there were other factors (e.g. social and cultural preferences) which were not included in our analysis that could have affected our findings; however, due to the strength of the association with lung cancer that relationship is unlikely to have changed considerably. Conclusions This is the first nationally representative study to investigate the provision of PCS for COPD patients. We have shown that the majority of COPD patients within primary care were not provided PCS in their last year of life; when PCS was provided, it appeared to be related to their co-diagnosis of lung cancer rather than their airways disease, including their disease severity (GOLD score and exacerbation frequency) or their perceived respiratory disability (MRC Dyspnoea score). Furthermore, PCS was frequently only given within the last few weeks of life, whereas patients with a lung cancer co-diagnosis were far more likely to receive PCS in a more timely manner. Many COPD patients suffer from symptom-related distress even when medical treatment has been optimised. The low proportion of patients receiving PCS, not immediately before death, suggests it is an especially poorly provided service with regard to its capacity to reduce symptom burden, improve quality of life and reduce psychological distress. This study advocates an urgent need to improve all aspects of PCS, not just terminal care, for COPD patients within the UK.

not immediately before death, suggests it is an especially poorly provided service with regard to its capacity to reduce symptom burden, improve quality of life and reduce psychological distress. This study advocates an urgent need to improve all aspects of PCS, not just terminal care, for COPD patients within the UK. Supplementary material 10.1183/13993003.01879-2017.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary table S1 ERJ-01879-2017_Table_S1 Supplementary figure S1 ERJ-01879-2017_Figure_S1 Disclosures 10.1183/13993003.01879-2017.Supp2J.K. Quint ERJ-01879-2017_Quint L. Smeeth ERJ-01879-2017_Smeeth This article has supplementary material available from erj.ersjournals.com Support statement: This study was funded by Wellcome (WT107183). P. Stone is supported by funding from Marie Curie (MCCC-FCO-16-U and MCCC-HMT-17-U). Funding information for this article has been deposited with the Crossref Funder Registry. Conflict of interest: Disclosures can be found alongside this article at erj.ersjournals.com

Introduction In 2015, only 20% of the 580 000 people eligible for multidrug-resistant tuberculosis (MDR-TB) treatment received an appropriate drug regimen [1]. Treatment of MDR-TB is long, expensive and toxic; errors in the design of the regimen are associated with increased rates of failure and death [2, 3]. Drug-resistant tuberculosis regimens need to include a sufficient number of effective drugs, a significant challenge for clinicians worldwide, as most are forced to make therapy decisions without any drug susceptibility testing (DST) information. Additionally, the World Health Organization (WHO) policy guidance for the use of novel antituberculosis drugs (bedaquiline and delamanid) and newly developed shorter regimens for the treatment of drug-resistant TB requires rapid diagnosis and triaging of patients to identify those who are most likely to benefit from the new treatment options [4–8]. Phenotypic DST is not suitable for this purpose, as it takes weeks to complete due to the slow growth rate of Mycobacterium tuberculosis complex (MTBC) strains, and requires both expensive infrastructure and considerable technical expertise [9]. As targeted genotypic DST assays (that provide results within hours to days) have been shown to be accurate, increasingly automated and cost-effective, they are proving to be a viable alternative or effective complement to phenotypic DST [9]. However, continuing technical constraints of the current molecular assays restrict the number of resistance determinants and genomic regions that can be evaluated, which limit the clinical value of the assays. By contrast, whole-genome sequencing (WGS) has the potential to provide near-complete information as it includes almost the entire genetic repertoire of a given clinical MTBC strain. However, the data analysis is more complex, and in order to maximise clinical utility of WGS, healthcare workers need clear rules to interpret the clinical relevance of genetic changes that are detected [10]. A high-quality and comprehensive catalogue of genetic markers of resistance (i.e.

a given clinical MTBC strain. However, the data analysis is more complex, and in order to maximise clinical utility of WGS, healthcare workers need clear rules to interpret the clinical relevance of genetic changes that are detected [10]. A high-quality and comprehensive catalogue of genetic markers of resistance (i.e. mutations that either cause resistance or compensate for resistance) is needed to distinguish significant resistant variants from those that are not. For some drugs this requires a precise understanding of the level of resistance conferred by the mutation in question, which is expressed as a range of minimum inhibitory concentrations (MICs) found in strains harbouring a specific mutation, as well as an understanding of the degree of cross-resistance conferred for antibiotics with shared modes of action [11, 12]. Although in vitro allelic exchange experiments prove conclusively that a particular mutation is both necessary and sufficient to confer phenotypic resistance, these approaches are expensive, slow and technically demanding, and they are only suitable to investigate the function of novel resistance genes or, at best, a limited number of resistance mutations per gene [13]. Consequently, in silico association studies are indispensable to investigate the vast majority of suspected resistance mutations, particularly in nonessential genes, where hundreds of loss-of-function mutations can result in resistance [14].

nce genes or, at best, a limited number of resistance mutations per gene [13]. Consequently, in silico association studies are indispensable to investigate the vast majority of suspected resistance mutations, particularly in nonessential genes, where hundreds of loss-of-function mutations can result in resistance [14]. Attempts have been made to combine disparate datasets to collect the necessary evidence for these associations. Unfortunately, most of these databases are not actively curated or lack significant clinical metadata [15]. Moreover, these databases focus mainly on collecting and presenting published data, leaving the final interpretation of the genotype–phenotype correlation to the user [16, 17]. More fundamentally, there is no consensus regarding the threshold of evidence required to classify a mutation as a valid marker for phenotypic resistance.

hese databases focus mainly on collecting and presenting published data, leaving the final interpretation of the genotype–phenotype correlation to the user [16, 17]. More fundamentally, there is no consensus regarding the threshold of evidence required to classify a mutation as a valid marker for phenotypic resistance. In this study we 1) describe a standardised analytical approach for assessing and quantifying the strength of the association of a particular mutation or group of mutations with phenotypic antibiotic resistance; 2) demonstrate how this approach can be used by applying it to data from the most comprehensive systematic review of MTBC drug resistance mutations conducted to date; and 3) apply the resulting graded mutation list to the interpretation guidelines for the WHO-endorsed, targeted genotypic DST assays Hain GenoType MTBDRplus v2.0 and MTBDRsl v2.0 [9]. This study has implications for the clinical interpretation of both targeted molecular and WGS-based diagnostics for drug-resistant TB, and is intended to provide clarity and build confidence regarding the genetic basis of resistance in MTBC for both molecular assay developers and the clinicians interpreting those assays.

s study has implications for the clinical interpretation of both targeted molecular and WGS-based diagnostics for drug-resistant TB, and is intended to provide clarity and build confidence regarding the genetic basis of resistance in MTBC for both molecular assay developers and the clinicians interpreting those assays. Materials and methods Data collection A systematic literature review on the association of sequencing and phenotypic DST data for MTBC was undertaken for selected anti-TB drugs and resistance genes (table 1). Expert consensus from the global ReSeqTB Data Sharing Platform was utilised to define the loci with the highest likelihood of association with resistance, and the review was limited to those loci (online supplementary material 1) [18]. A comprehensive search of the National Center for Biotechnology Information (NCBI) PubMed database for relevant citations was performed; the list of search terms and the data collection form are available as online supplementary material 2 and 3. The quality of the studies included was appraised using a modified Quality Assessment of Diagnostic Accuracy Studies (QUADAS)-2 tool (online supplementary material 4) [19]. The sensitivity and specificity of predicting phenotypic ofloxacin (OFX) and levofloxacin (LFX) resistance by sequencing were found to be independent of the phenotypic method used, whereas there were substantial differences in specificity for moxifloxacin (MFX) resistance prediction, depending on whether liquid or solid DST was used as the reference method (data not shown). Results for OFX and LFX from both testing methods were therefore pooled, whereas MFX results were analysed separately for each DST method. To maximise the number of isolates studied and thus increase statistical power, results for ethionamide (ETO) and prothionamide (PTO) were also pooled.

ence method (data not shown). Results for OFX and LFX from both testing methods were therefore pooled, whereas MFX results were analysed separately for each DST method. To maximise the number of isolates studied and thus increase statistical power, results for ethionamide (ETO) and prothionamide (PTO) were also pooled. TABLE 1 Overview of the data included in the study Collected data Studies Loci of interest Total isolates Isolation time frame years Countries represented Screened Included Rifampicin (R) rpoB 13 424 1999–2014 37 459 95 Isoniazid (H) katG 11 847 1992–2014 42 650 127 inhA-mabA 9407 furA 361 mshA 288 Ethionamide and prothionamide (ETO/PTO) inhA-mabA 346 ethA 181 mshA 117 Ofloxacin (OFX) gyrA 5911 1991–2013 36 243 75 gyrB 3078 Moxifloxacin (MFX) gyrA 1019 gyrB 735 Levofloxacin (LFX) gyrA 449 gyrB 218 Pyrazinamide (Z) pncA 4949 1990–2014 36 378 81 Streptomycin (S) rpsL 3263 1985–2013 43 423 104 tap 0 rrs 2598 whiB7 0 gidB 812 Amikacin (AM) rrs 2105 Capreomycin (CM) rrs 2533 tlyA 1854 Kanamycin (KM) rrs 1727 eis 2029 whiB7 56 Data are presented as n. Inclusion and exclusion criteria for individual studies are reported in online supplementary material 2.

36 378 81 Streptomycin (S) rpsL 3263 1985–2013 43 423 104 tap 0 rrs 2598 whiB7 0 gidB 812 Amikacin (AM) rrs 2105 Capreomycin (CM) rrs 2533 tlyA 1854 Kanamycin (KM) rrs 1727 eis 2029 whiB7 56 Data are presented as n. Inclusion and exclusion criteria for individual studies are reported in online supplementary material 2. Development of a standardised methodology for the statistical validation of the association of a mutation with resistance An expert, consensus-driven approach was used to develop a standardised procedure for grading drug resistance-associated mutations. The collated data were used to calculate the frequency of each mutation in resistant and susceptible MTBC isolates and to derive a likelihood ratio. In this approach, likelihood ratios were used for objectively evaluating whether mutations were positively or negatively associated with phenotypic resistance. Moreover, odds ratios were considered when evaluating the association of the genotypic and phenotypic data. Using this rationale, the thresholds commonly adopted in evidence-based medicine were adapted to grade the MTBC mutations [20–22]. Details of the statistical analysis are provided in online supplementary material 5. Mutations were classified as having either high, moderate or minimal confidence for being associated with resistance, or as indeterminate or “not associated with resistance” (see table 2 for the definitions of each category). The procedure used for nonsense mutations, insertions/deletions and silent mutations is described in detail in online supplementary material 5. Results from the three types of phenotypic reference standards used (liquid, solid and combined-media DST) were compared using a series of rules that are outlined in online supplementary material 5 to yield confidence values for individual mutations (“individual confidence values” (ICVs)), associations with a specific medium (“medium confidence values” (MCVs)) and an overall, best confidence value (BCV). Moreover, interpretive confidence values for each of the aforementioned categories (i.e. iICVs, iMCVs and iBCV, respectively) were calculated to extrapolate the confidence values of individual mutations and the pooled results for insertions/deletions and nonsense mutations.

and an overall, best confidence value (BCV). Moreover, interpretive confidence values for each of the aforementioned categories (i.e. iICVs, iMCVs and iBCV, respectively) were calculated to extrapolate the confidence values of individual mutations and the pooled results for insertions/deletions and nonsense mutations. TABLE 2 Overview of proposed confidence levels for grading mutations associated with phenotypic resistance Symbol LR+ and OR p-value value High (Hi) confidence for association with resistance Strong association of the mutation with phenotypic drug resistance; sufficient evidence that the mutation confers or is strongly associated with drug resistance #• <0.05 >10 Moderate (Mo) confidence for association with resistance Moderate association of the mutation with phenotypic drug resistance; additional data desirable for improved evidence that the mutation confers or is strongly associated with drug resistance #• <0.05 5< … ≤10 Minimal (Mi) confidence for association with resistance Weak association of the mutation with phenotypic drug resistance; inconclusive evidence that the mutation confers or is strongly associated with drug resistance. Substantial additional data required #• <0.05 1< … ≤5 No association with resistance No evidence of association between the mutation and drug resistance #• <0.05 <1 Indeterminate

Moderate association of the mutation with phenotypic drug resistance; additional data desirable for improved evidence that the mutation confers or is strongly associated with drug resistance #• <0.05 5< … ≤10 Minimal (Mi) confidence for association with resistance Weak association of the mutation with phenotypic drug resistance; inconclusive evidence that the mutation confers or is strongly associated with drug resistance. Substantial additional data required #• <0.05 1< … ≤5 No association with resistance No evidence of association between the mutation and drug resistance #• <0.05 <1 Indeterminate No statistically significant threshold reached; additional data required Indeter ≥0.05 The table shows the thresholds applied to likelihood ratios (LR) and odds ratios (OR) to grade the association of mutations with phenotypic drug resistance. LR+: positive likelihood ratio. “Additional data” is defined as a requirement for 1) more phenotypically drug resistant and susceptible isolates tested with the mutation in question; and/or 2) better understanding of the mechanism of drug resistance (e.g. to investigate epistasis, or the interactions between drug-resistance conferring mutations, lineage-specific genetic factors and compensatory mutations [23, 24] or synergistic factors when more than one mutation is required to confer resistance [25]).

or 2) better understanding of the mechanism of drug resistance (e.g. to investigate epistasis, or the interactions between drug-resistance conferring mutations, lineage-specific genetic factors and compensatory mutations [23, 24] or synergistic factors when more than one mutation is required to confer resistance [25]). Results were stratified according to the following reference standards for each target-drug combination, given that systematic differences between media have been observed previously. 1) Liquid-media phenotypic DST performed according to WHO guidelines; 2) solid-media phenotypic DST performed according to WHO guidelines (for pyrazinamide (Z), the Wayne enzymatic assay, which is not WHO-endorsed, was also considered in this category); and 3) liquid- and solid-media DST combined (this category included DST results for which the medium used was unclear, which meant that the number in this category was sometimes larger than the sum of categories 1 and 2). Results Overview of the datasets included in the systematic review Table 1 summarises the main features of the datasets considered. Data from up to 43 countries and up to 13 424 isolates per gene locus or antimycobacterial drug were included. PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) diagrams for each drug and a detailed breakdown of exclusion criteria, types of studies included, global representativeness of datasets and phenotypic DST methods can be found in online supplementary material 6.

ycobacterial drug were included. PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) diagrams for each drug and a detailed breakdown of exclusion criteria, types of studies included, global representativeness of datasets and phenotypic DST methods can be found in online supplementary material 6. The majority of the phenotypically resistant isolates harboured either single point mutations or, more rarely, insertions/deletions in the resistance genes/loci that were studied (mean 82.6%, range 69.8–95.6%; online supplementary material 6, figure S6.20A). Conversely, phenotypically susceptible isolates had mostly wild-type results (mean 85.6%, range 70.9–97.3%; online supplementary material 6, figure S6.20B). However, both frequencies probably represent an underestimate, given that some studies only reported the variants that were thought to be responsible for the resistant phenotype. Well-known nonsynonymous polymorphisms that do not confer resistance, such as katG R463L, or synonymous mutations were not always reported [26] (online supplementary material 4).

present an underestimate, given that some studies only reported the variants that were thought to be responsible for the resistant phenotype. Well-known nonsynonymous polymorphisms that do not confer resistance, such as katG R463L, or synonymous mutations were not always reported [26] (online supplementary material 4). Overview of confidence graded mutations A full list of confidence values for associations of ICVs, MCVs or the overall BCVs and their corresponding interpretative confidence values can be found in online supplementary material 7 and 9. Figure 1 provides an overview of the proportion of isolates with different confidence levels for MCVs. For all drugs, data for the majority of variants were only available from one medium (solid or liquid). Where mutations were tested on both media, the MCVs were usually identical. However, there were also variants, for which the confidence levels differed (online supplementary material 9). The number of these discrepancies varied from just one variant for drugs such as amikacin (AM), to 56 for pyrazinamide (Z). Some of these discrepancies were minor (e.g. the gyrA D94N variant had minimal confidence on solid medium for OFX/LFX, but high confidence for liquid medium), whereas others were major. Important differences included 1) gyrA A90V, which was not associated with resistance when the phenotypic reference standard was liquid medium but had a high-confidence MCV for MFX resistance when associated with phenotypic resistance determined in solid media; and 2) rpoB L511P, which was not associated with resistance in liquid medium, but had minimal confidence on solid medium. In order to be conservative in the interpretation of these cases, the lower statistical confidence was overruled by the higher MCV to yield a BCV (see online supplementary material 5 for more details). Similarly, MCVs for which evidence was available on one medium only were used as the BCVs.

confidence on solid medium. In order to be conservative in the interpretation of these cases, the lower statistical confidence was overruled by the higher MCV to yield a BCV (see online supplementary material 5 for more details). Similarly, MCVs for which evidence was available on one medium only were used as the BCVs. FIGURE 1 Medium confidence values (MCVs) stratified by confidence value, drug susceptibility testing medium and antibiotic-resistance gene combination. In the three rows above the graph we show variants that were concordant on both media, the number of variants that had different confidence levels on liquid and solid (these are marked as “discrepant variants” and are listed in full in online supplementary material 9) and unique variants for which confidence levels were available on only one of the two media.

variants that were concordant on both media, the number of variants that had different confidence levels on liquid and solid (these are marked as “discrepant variants” and are listed in full in online supplementary material 9) and unique variants for which confidence levels were available on only one of the two media. Overview of BCVs Table 3 lists all of the 394 MTBC genetic variants with high, moderate (192 mutations plus 202 frameshifts and premature stop codons) or minimal BCVs, as well as 40 changes that were found not to be associated with resistance according to nominal p-values (online supplementary material 5). Six of these variants identified with our association method had to be graded as not associated with phenotypic resistance manually based on expert knowledge. For example, the a514c and c517t mutations in rrs had high-confidence BCVs for predicting kanamycin (KM) resistance, but were excluded from further analysis because there was no known causative link between these mutations and KM resistance [27, 28]. Other mutations (e.g. inhA g-102a) were excluded as they are known markers for particular MTBC genotypes (lineage or sublineages) and do not confer resistance [26, 29]. We highlighted the 286 variants with high, moderate or minimal BCVs (111 mutations plus 150 frameshifts and 25 premature stop codons), as well as 18 changes that were found to be “not associated” with resistance (likelihood ratio <1) that remained statistically significant after correcting the p-value for the false discovery rate (see online supplementary material 5 for details). The resulting subset of 304 BCVs is referred to as the corrected BCVs for the remainder of this article. Overall, we identified 286 confidence-graded (high + moderate + minimal (Hi+Mo+Mi)) mutations associated with phenotypic resistance.

ue for the false discovery rate (see online supplementary material 5 for details). The resulting subset of 304 BCVs is referred to as the corrected BCVs for the remainder of this article. Overall, we identified 286 confidence-graded (high + moderate + minimal (Hi+Mo+Mi)) mutations associated with phenotypic resistance. TABLE 3 List of confidence-graded mutations associated with phenotypic drug resistance as determined by best confidence values

ue for the false discovery rate (see online supplementary material 5 for details). The resulting subset of 304 BCVs is referred to as the corrected BCVs for the remainder of this article. Overall, we identified 286 confidence-graded (high + moderate + minimal (Hi+Mo+Mi)) mutations associated with phenotypic resistance. TABLE 3 List of confidence-graded mutations associated with phenotypic drug resistance as determined by best confidence values Drug (phenotypic testing) Gene High-confidence mutations Moderate-confidence mutations Minimal-confidence mutations No association with resistance First-line Rifampicin (R) rpoB F505V+D516Y, S512T, Q513H+L533P, Q513-F514ins, Q513K, Q513L, Q513P, F514dupl, M515I+D516Y, D516A, D516F, D516G, D516G+L533P, D516ins, D516N, D516V, Del N518, S522Q, H526C, H526D, H526F, H526G, H526L, H526R, H526Y S531F, S531L, S531Q, S531W, S531Y, D626E D516Y, S522L, H526P, L533P L511P, H526N, I572F Isoniazid (H) inhA-mabA g-102a#,¶ c-15t g-102a#,¶, t-80g, g-47c, T4I katG S315I, S315N, S315T, pooled frameshifts and premature stop codons A110V, R463L, L499M furA L68F mshA A187V#,¶ N111S Second-line (group A) Moxifloxacin (MFX) gyrA G88C, A90V, S91P, D94A, D94G, D94N, D94Y E21Q, S95T, G247S, G668D, V712L Ofloxacin (OFX)/levofloxacin (LFX) gyrA G88A, G88C, S91P, A90V, D94A, D94G, D94H, D94N, D94Y D89N E21Q, T80A, S95T, G247S, G668D, V712L gyrB E459K, A504V Second-line (group B) Amikacin (AM) rrs a1401g, g1484t Kanamycin (KM) eis c-14t, g-10a g-37t, c-12t a1338c rrs a514c#, a1401g, c1402t, g1484t rrs+eis rrs c517t# + eis g-37t Capreomycin (CM) rrs a1401g, c1402t, g1484t c517t tlyA N236K, pooled frameshifts and premature stop codons D149H Streptomycin (S) rpsL K43R, K43T, K88Q, K88R, T40I rrs a1401g#, a514c, a514t, c462t, c513t, c517t gidB E92D#,¶ L16R, V110G, pooled frameshifts and premature stop codons Second-line (group C) Ethionamide and prothionamide (ETO/PTO)  inhA c-15t+I194T, c-15t+S49A c-15t ethA Q347Stop Second-line (group D) Pyrazinamide (Z) pncA t-12c, a-11g, t-7c, A3E, L4S, I6T, V7G, D8E, D8G, D8N, Q10P, D12A, D12N, C14R, G17D, L19P, G24D, Y34D, A46V, K48T, D49G, D49N, H51Q, H51R, P54S, H57D¶, H57P, H57R, H57Y, S59P, P62L, P62Q, D63G, S66P, S67P, W68C, W68R, H71D, H71Q, H71Y, C72R, T76P, H82R, L85P, L85R, F94L, F94S, K96N, K96R, G97C, G97D, G97S, Y103H, S104R, G108R, L116P, L116R, L120P, R123P, V125F, V125G, V128G, G132A, G132D, G132S, A134V, T135N, T135P, H137P, C138Y, V139G, V139L, Q141P, T142A, T142K, T142M, indel - R148ins (inframe), L151S, V155G, L159P, T160P, G162D, T168P, L172P, M175T, M175V, V180F, V180G, Pooled frameshifts and premature stop codons V7G, Q10R, P54L, W68G,

, L116P, L116R, L120P, R123P, V125F, V125G, V128G, G132A, G132D, G132S, A134V, T135N, T135P, H137P, C138Y, V139G, V139L, Q141P, T142A, T142K, T142M, indel - R148ins (inframe), L151S, V155G, L159P, T160P, G162D, T168P, L172P, M175T, M175V, V180F, V180G, Pooled frameshifts and premature stop codons V7G, Q10R, P54L, W68G, K96E, K96T, A171E, M175I D12G, F58L, H71R, I133T, V139A indel - c-125del, I31T, L35R, T47A, I6L, K48T, T114M The table includes all the mutations graded according to the proposed standardised approach for providing confidence levels to their association with phenotypic drug resistance. Standard type represents associations based on nominal p-values (putative); bold type represents associations based on corrected p-values. The rationale for pooling insertions/deletions and nonsense mutations can be found in online supplementary material 5. Tables 1 and 2 provide the details of the data included in the grading system and the definitions for the confidence categories. Indeterminate mutations were not included in the table and can be found in online supplementary material 8. Drugs were classified based on the updated guidelines for short and individualised regimens [4]. #: six associations were not considered for further analysis as there was probably no causative relationship between these genetic changes and the resistance to the antibiotic in question; ¶: genotype-specific mutation.

8. Drugs were classified based on the updated guidelines for short and individualised regimens [4]. #: six associations were not considered for further analysis as there was probably no causative relationship between these genetic changes and the resistance to the antibiotic in question; ¶: genotype-specific mutation. Diagnostic performance of corrected iBCVs Online supplementary material 10 provides a comprehensive overview of the performance characteristics (sensitivity, specificity and diagnostic accuracy) for different categories of corrected iBCVs (see online supplementary material 5 for a detailed explanation of the differences between BCV and iBCV): 1) high, moderate and minimal confidence mutations individually; 2) Hi+Mo+Mi confidence mutations combined; 3) indeterminate (I) mutations; and 4) a combination of Hi+Mo+Mi+I confidence mutations as well as mutations that are “not associated with phenotypic resistance” (referred herewith as “all mutations”).

and iBCV): 1) high, moderate and minimal confidence mutations individually; 2) Hi+Mo+Mi confidence mutations combined; 3) indeterminate (I) mutations; and 4) a combination of Hi+Mo+Mi+I confidence mutations as well as mutations that are “not associated with phenotypic resistance” (referred herewith as “all mutations”). The sensitivities (95% CI) of mutations with high-confidence corrected iBCVs compared with phenotypic DST (i.e. the observed resistant phenotype) ranged from 0.0% (0.0–0.01%) for ETO/PTO to 88.2% (85.1–90.9%) for MFX (figure 2a). Specificities (95% CI) varied from 95.6% (94.7–96.4%) for capreomycin (CM) to 99.5% (99.0–99.8%) for AM. The inclusion of Mo and Mo+Mi confidence mutations resulted in a gain in sensitivity of 0–47.3% with only marginal decreases in specificity (i.e. 0–3.8%). The performance of the Hi+Mo+Mi confidence mutations identified in this study performed as well or better than a set of diagnostic mutations recently proposed by Farhat et al. [30] that were based on detecting resistance-associated mutations using random forest modelling on a set of 1400 MTBC isolates (online supplementary material 11, table S11.1).

confidence mutations identified in this study performed as well or better than a set of diagnostic mutations recently proposed by Farhat et al. [30] that were based on detecting resistance-associated mutations using random forest modelling on a set of 1400 MTBC isolates (online supplementary material 11, table S11.1). FIGURE 2 Comparison of the sensitivity and specificity of different groups of mutations. For each drug, two types of comparison were performed. First, the a) sensitivities and b) specificities were calculated with the associated 95% confidence levels for the “observed” phenotypic result. Specifically, the figures for high (Hi), high and moderate (Hi+Mo) and high and moderate and minimal (Hi+Mo+Mi) confidence interpretative best confidence values (iBCVs) were compared with using all mutations observed in the study. Moreover, genetic variants were included from a recent study by Farhat et al. [30]. Second, we conducted the same comparison using a “corrected” phenotype as reference (i.e. where we assumed that strains that were phenotypically susceptible but harboured either a Hi, Hi+Mo or H+Mo+Mi confidence iBCV mutation or mutations by Farhat et al. were false-susceptible results). For some drugs, such as capreomycin (CM), both percentages remained unchanged as all mutations were high confidence. Minimal target product profile (TPP) thresholds set by the World Health Organization for new molecular-based diagnostic tools compared to phenotypic drug susceptibility testing [31] are shown. These were intended for rifampicin (R), isoniazid (H), fluoroquinolones, kanamycin (KM), amikacin (AM) and CM only. However, in addition we included the threshold for the remaining drugs and overall results for comparison (for additional details see online supplementary material 11). MFX: moxifloxacin; OFX: ofloxacin; LFX: levofloxacin; S: streptomycin; ETO/PTO: ethionamide and prothionamide; Z: pyrazinamide.

kacin (AM) and CM only. However, in addition we included the threshold for the remaining drugs and overall results for comparison (for additional details see online supplementary material 11). MFX: moxifloxacin; OFX: ofloxacin; LFX: levofloxacin; S: streptomycin; ETO/PTO: ethionamide and prothionamide; Z: pyrazinamide. Assuming that mutations with high, moderate and minimal confidence corrected iBCVs are true markers or resistance, a “corrected phenotype” was calculated for each drug (i.e. the sum of the phenotypically resistant isolates and phenotypically susceptible isolates with one of the aforementioned mutations). Accordingly, the proportion of resistance missed by phenotypic DST ranged from 0.7% (95% CI 0.3–1.6%) for AM to 11.9% (95% CI 9.8–14.1%) for CM, which resulted in a difference of 0.2% and 3.9%, respectively, between the sensitivities using the “corrected phenotype” as the reference standard versus phenotypic DST (the specificities became, by definition, 100%) (table 4 and figure 2b). TABLE 4 Overview of phenotypically susceptible isolates with high (Hi), moderate (Mo) or minimal (Mi) confidence corrected interpretative best confidence values (iBCVs)

Assuming that mutations with high, moderate and minimal confidence corrected iBCVs are true markers or resistance, a “corrected phenotype” was calculated for each drug (i.e. the sum of the phenotypically resistant isolates and phenotypically susceptible isolates with one of the aforementioned mutations). Accordingly, the proportion of resistance missed by phenotypic DST ranged from 0.7% (95% CI 0.3–1.6%) for AM to 11.9% (95% CI 9.8–14.1%) for CM, which resulted in a difference of 0.2% and 3.9%, respectively, between the sensitivities using the “corrected phenotype” as the reference standard versus phenotypic DST (the specificities became, by definition, 100%) (table 4 and figure 2b). TABLE 4 Overview of phenotypically susceptible isolates with high (Hi), moderate (Mo) or minimal (Mi) confidence corrected interpretative best confidence values (iBCVs) Drug Drug-resistant phenotype Hi confidence iBCVs Hi+Mo confidence iBCVs Hi+Mo+Mi confidence iBCVs False-susceptible Resistance missed % (95% CI) Difference in sensitivity % False-susceptible Resistance missed % (95% CI) Difference in sensitivity % False-susceptible Resistance missed % (95% CI) Difference in sensitivity % Rifampicin (R) 8294 55 0.7 (0.5–0.9) 0.1 124 1.5 (1.2–1.8) 0.2 192 2.3 (2.0–2.6) 0.2 Isoniazid (H) 11001 55 0.5 (0.4–0.7) 0.2 81 0.7 (0.6–0.9) 0.2 81 0.7 (0.6–0.9) 0.2 Moxifloxacin (MFX) 517 50 8.8 (6.6–11.5) 1.0 50 8.8 (6.6–11.5) 1.0 50 8.8 (6.6–11.5) 1.0 Ofloxacin (OFX) /levofloxacin (LFX) 3809 93 2.4 (1.9–2.9) 0.5 94 2.4 (2.0–2.9) 0.5 94 2.4 (2.0–3.0) 0.5 Amikacin (AM) 809 6 0.7 (0.3–1.6) 0.2 6 0.7 (0.3–1.6) 0.2 6 0.7 (0.3–1.6) 0.2 Kanamycin (KM) 943 25 2.6 (1.7–3.8) 0.8 25 2.6 (1.7–3.8) 0.8 25 2.6 (1.7–3.8) 0.8 Capreomycin (CM) 810 109 11.9 (9.8–14.1) 3.9 109 11.9 (9.8–14.1) 3.9 109 11.9 (9.8–14.1) 3.9 Streptomycin (S) 2204 16 0.7 (0.4–1.2) 0.3 16 0.7 (0.4–1.2) 0.3 16 0.7 (0.4–1.2) 0.3 Ethionamide and prothionamide (ETO/PTO) 298 0 0.0 (0.0–1.2) 0.0 7 2.3 (0.9–4.7) 1.2 7 2.3 (0.9–4.7) 1.2 Pyrazinamide (Z) 2595 59 2.2 (1.7–2.9) 0.8 67 2.5 (2.0–3.2) 0.9 83 3.1 (2.5–3.8) 1.0 The resistance missed corresponds to the false-susceptible isolates divided by the “corrected phenotype”, consisting of the sum of phenotypically resistant isolates and false-susceptible isolates (the probable smallest and largest figures in each category are shown in bold). The difference in sensitivity was calculated by subtracting the sensitivity for the “observed phenotype” from the sensitivity of the “corrected phenotype” (both sensitivities are plotted in figure 2a).

solates and false-susceptible isolates (the probable smallest and largest figures in each category are shown in bold). The difference in sensitivity was calculated by subtracting the sensitivity for the “observed phenotype” from the sensitivity of the “corrected phenotype” (both sensitivities are plotted in figure 2a). Some genotypic diagnostic tests, such as the Xpert MTB/RIF (Cepheid, Sunnyvale, CA, USA) and Hain line-probe assays (Hain Lifescience, Nehren, Germany), infer the presence of clinically relevant mutations when wild-type assay probes do not anneal to MTBC sequence in a clinical sample. When they do this, they are effectively using an indiscriminate “all mutations” approach of predicting phenotypic resistance. We looked at the potential effect of including all observed mutations in the entire gene on our prediction estimates, rather than just the graded mutation set. The gains in sensitivity compared with the Hi+Mo+Mi set was marked in some cases (mean 11.5%, range 0.9–32.6%), such as ETO/PTO and Z, but also resulted in a large decrease in specificity (mean −10.8%, range −26.5–−1.5%) (figure 2a and b). Notably, these decreases in specificity were probably an underestimate, since synonymous mutations that can cause systematic false-resistance results were excluded from this study [32].

2.6%), such as ETO/PTO and Z, but also resulted in a large decrease in specificity (mean −10.8%, range −26.5–−1.5%) (figure 2a and b). Notably, these decreases in specificity were probably an underestimate, since synonymous mutations that can cause systematic false-resistance results were excluded from this study [32]. Assessment of the interpretation guidelines of the Hain GenoType MTBDRplus v2.0 and MTBDRsl v2.0 Based on the package inserts of the Hain GenoType MTBDRplus v2.0, 32 mutations are identified as mutations that confer resistance to isoniazid (H) or rifampicin (R) [33]. 18 of these mutations had either high, moderate or minimal confidence iBCVs, whereas the remaining 14 where either indeterminate, occurred only in combination with other mutations or were not evaluated in this review (online supplementary material 12, table S12.1). Of the 19 genetic markers identified as predictors of resistance to fluoroquinolones, KM, AM or CM, as defined by the package insert for the GenoType MTBDRsl v2.0 [34], 10 mutations were found to be indeterminate in our study, were only found in combination with other mutations or were located in a region not considered for at least one antibiotic in this review (online supplementary material 12, table S12.1).

or CM, as defined by the package insert for the GenoType MTBDRsl v2.0 [34], 10 mutations were found to be indeterminate in our study, were only found in combination with other mutations or were located in a region not considered for at least one antibiotic in this review (online supplementary material 12, table S12.1). Discussion Rapid evidence-based triaging of patients with drug resistant MTBC strains to appropriate drug-resistant TB treatment regimens can only be achieved using genotypic DST methods. Yet, in practice, our understanding of the consequences of classifying patient MTBC strains as “resistant” based on the detection of certain mutations is biased by subjective methods and limited datasets. A case point is the rapid detection of R resistance in clinical MTBC samples. This is now largely achieved using molecular tests, but the emergence of data on discrepancies between genotypic and phenotypic DST, and some systematic false-positive results have created some uncertainty regarding the use of molecular data for early management of patients [35–39]. Using an expert, consensus-driven approach, we developed and verified a standardised procedure to assess the level of confidence in the association between individual mutations and clinically relevant phenotypic drug resistance in MTBC. Our comprehensive approach provides clear, objective and quantitative estimates of the correlation of genotype with phenotypic resistance that is consistent with methods previously established for evidence-based medicine. These findings have immediate implications for molecular and WGS diagnostic assays currently under development, as well as for the interpretation of existing commercially available molecular DST assays. For example, to our knowledge, the eis c-2a mutation, which is interpreted as conferring KM resistance in the package insert of the MTBDRsl v2.0 assay, has only ever been observed in two KM-resistant strains that also harbour the high-confidence eis c-14t mutation (online supplementary material 12) [34, 40]. Consequently, there is currently no convincing evidence that the eis c-2a mutation alone is a valid marker for phenotypic KM resistance and the interpretation of the assay should probably be changed to remove this mutation from consideration or require co-occurrence of the eis c-14t mutation for clinically relevant interpretation.

, there is currently no convincing evidence that the eis c-2a mutation alone is a valid marker for phenotypic KM resistance and the interpretation of the assay should probably be changed to remove this mutation from consideration or require co-occurrence of the eis c-14t mutation for clinically relevant interpretation. This example illustrates the potential value of our findings to guide molecular diagnostics developers in terms of which mutations to include and exclude in their assays and interpretation guides, as well as to help regulators evaluating manufacturer claims and clinicians to minimise systematic false-positive and false-negative results [32]. In addition, our results confirmed previous findings that some potentially clinically relevant resistance mutations could be systematically overlooked if certain phenotypic methods are used for DST [35].

evaluating manufacturer claims and clinicians to minimise systematic false-positive and false-negative results [32]. In addition, our results confirmed previous findings that some potentially clinically relevant resistance mutations could be systematically overlooked if certain phenotypic methods are used for DST [35]. Likelihood ratios are not only useful in computing the (post-test) probability of a diagnosis, but can also be used to evaluate the association between a mutation and a given phenotype of interest, in this case drug resistance [20, 41]. Using likelihood ratio thresholds, we classified observed variants as high, moderate or minimal confidence resistance mutations (table 2). While likelihood ratios are commonly used to refine clinical judgements and pretest probabilities, they have not been used previously in this manner for predicting phenotypic TB drug resistance. This approach has two main strengths. First, the likelihood ratio is a universal measure of association in diagnostics that is not affected by local or regional prevalence of drug resistance [42]. Second, unlike sensitivity and specificity, which are often used to assess resistance mutations as predictors of phenotypic resistance, likelihood ratios do not lead to an exaggeration of the benefits of a test or the strength of an association [43], since they simply provide a multiplier for the pretest probability of resistance. In particular, a high sensitivity and a high specificity do not ensure that the positive result of a diagnosis is correct if the underlying condition is exceedingly rare. However, the grading system actually does not take into account the uncertainty (95% confidence intervals) around the likelihood ratio estimate for mutations that are positively associated with resistance. This means that mutations can be graded as high confidence, despite having been observed in only few resistant isolates (e.g. the tlyA N236K mutation was assigned a high confidence ICV despite occurring in just three resistant and one susceptible isolates; online supplementary material S7). However, the confidence level in the grading of each mutation from this study must necessarily be regarded as provisional, since it could change in either direction as more data are accumulated. This is an inevitable attribute of any evidence-based approach, as any conclusion is open to revision when new evidence comes to light [44].

nfidence level in the grading of each mutation from this study must necessarily be regarded as provisional, since it could change in either direction as more data are accumulated. This is an inevitable attribute of any evidence-based approach, as any conclusion is open to revision when new evidence comes to light [44]. Applying our grading scheme to a large, systematically collected set of MTBC sequencing and phenotypic data, we were able to identify a total of 286 high, moderate and minimal confidence corrected BCVs (table 3). The resulting diagnostic sensitivities and specificities compared with phenotypic DST can be found in figure 2a and b.

r grading scheme to a large, systematically collected set of MTBC sequencing and phenotypic data, we were able to identify a total of 286 high, moderate and minimal confidence corrected BCVs (table 3). The resulting diagnostic sensitivities and specificities compared with phenotypic DST can be found in figure 2a and b. The WHO has defined a specificity of ≥98% and a sensitivity of >95% (for R) or >90% (for H, fluoroquinolones, KM, AM and CM) compared to phenotypic reference standards as a requirement for diagnosing drug-resistant TB [31]. While the diagnostic sensitivities we observed in this study were lower than the WHO thresholds using only the graded mutations to predict resistant phenotypes, this is probably an underestimate of maximum potential sensitivity of genotypic prediction of resistance due to a combination of five important limitations. First, and most fundamental, are the genes and mutations considered. In this systematic review, we were limited to including only those genes and mutations previously documented to be associated with resistance and included in the published literature. While to our knowledge, our review is one of the most comprehensive yet completed, the global knowledge base on all genes associated with resistance is still growing, and we know that certain genes, for example, ahpC were not included as potential predictors of H resistance. Additionally, not all studies included data on all known resistance associated genes, which limited the sensitivity. Second, current sequencing technologies have varying capabilities to detect low frequencies (<20%) of resistant strains mixed with susceptible stains relative to phenotypic testing that can detect resistant strains making up only 1% of the total population [45]. This can be a major source of discordance between the detected genotype (apparently wild-type) and a resistant phenotype for some drugs, particularly the fluoroquinolones [9]. Third, breakpoint artefacts (i.e. inappropriately high critical concentrations) can be a major source of misclassification of phenotypes. This is well illustrated for CM, for which 11.9% (95% CI 9.8–14.1%) of strains harbouring markers of resistance were missed by phenotypic DST (table 4) [28]. Fourth, the specific biology and genetics of some resistance mechanisms occasionally limited the sensitivity of our method. For resistance caused by loss-of-function mutation in a nonessential gene (e.g.

h 11.9% (95% CI 9.8–14.1%) of strains harbouring markers of resistance were missed by phenotypic DST (table 4) [28]. Fourth, the specific biology and genetics of some resistance mechanisms occasionally limited the sensitivity of our method. For resistance caused by loss-of-function mutation in a nonessential gene (e.g. pncA), the number of different resistance mutations was very large and, consequently individual mutations were infrequent [11]. Our grading scheme scored such mutations as indeterminate until sufficient evidence can be gathered. Additionally, resistance mutations with MIC distributions that overlap substantially with the MIC distribution of susceptible strains are inherently difficult to distinguish from mutations not associated with resistance. This is because the MIC distributions of these mutations are truncated by the CC, which means that a mutant strain will not consistently test resistant due to the inherent variation in phenotypic testing. This phenomenon was most noticeable for the eis g-37t and c-12t mutations, which reduced confidence in the association and did not meet statistical significance for an association with KM resistance after the more conservative p-value correction was applied (table 3) [46]. Fifth, synonymous mutations were excluded from this analysis because these are not reported routinely in association studies. However, it is known that these mutations can sometimes confer resistance [47, 48].

r an association with KM resistance after the more conservative p-value correction was applied (table 3) [46]. Fifth, synonymous mutations were excluded from this analysis because these are not reported routinely in association studies. However, it is known that these mutations can sometimes confer resistance [47, 48]. The specificities of the corrected iBCVs were usually superior to the sensitivities and would be 100% if an expert rule was adopted for the genotype to overrule the phenotype whenever a high, moderate or minimal confidence mutation was detected (figure 2a and b). This is especially relevant for mutations affected by breakpoint artefacts, mutations that confer modest MIC increases, such as the “disputed” rpoB mutations [49], and drugs for which resistance is currently defined inconsistently on the phenotypic level. The latter point is best illustrated with the gyrA A90V mutation, which confers low-level resistance to MFX [50]. Consequently, it has a high specificity as predictor of phenotypic resistance when the phenotypic standard is the Clinical and Laboratory Standards Institute critical concentration of 0.5 mg·L−1 with 7H10, but not when the WHO critical concentration of 2 mg·L−1 on 7H10 medium is used as the reference [51]. Until critical concentrations are harmonised (which is particularly important for fluoroquinolones, where the evidence is mounting that strains with slightly elevated MICs might still be treatable [52]), and genotypic interpretations adjusted accordingly, this mutation and related mutations will continue to pose diagnostic challenges.

oncentrations are harmonised (which is particularly important for fluoroquinolones, where the evidence is mounting that strains with slightly elevated MICs might still be treatable [52]), and genotypic interpretations adjusted accordingly, this mutation and related mutations will continue to pose diagnostic challenges. Additional improvements to the graded mutation list could be implemented based on lessons learned from limitations of our study. Our study was limited by the fact that it mostly relied on amplicon-based sequencing data, which meant that the underlying population structure could not be taken into account. If WGS data had been available for all strains, resistance mutations that are currently classified as indeterminate or potentially even as not associated with resistance because they only confer modest MIC increases, as was the case for eis g-37t and c-12t, could be identified as resistance associated by the virtue of them being homoplastic (i.e. arising in unrelated isolates independently [26]). Such observations could help focus future MIC testing and/or allelic exchange experiments to clarify their MIC ranges and confirm or refute an association with resistance [46]. Conversely, mutations that are not homoplastic, such as gidB E92D, that are known lineage markers for particular genotypes and not markers for resistance could be excluded [26]. However, even using WGS, some manual curation based on an assessment of the mode of action of the antibiotic may still be required to remove spurious associations, as was the case for the rrs a1401g mutation.

that are known lineage markers for particular genotypes and not markers for resistance could be excluded [26]. However, even using WGS, some manual curation based on an assessment of the mode of action of the antibiotic may still be required to remove spurious associations, as was the case for the rrs a1401g mutation. The methods presented in this study will be used as a standardised analytical approach for assessing potential resistance mutations in the Relational Sequencing TB Data Sharing Platform currently available at https://platform.reseqtb.org/. The ReSeqTB platform serves as a globally harmonised knowledge base for the curation, validation and interpretation of existing and newly created genotypic and phenotypic data for TB drug resistance correlations [18]. In this context, the grading system presented here will be refined further by taking the following criteria into consideration: phylogenetic information, laboratory evidence (e.g. MIC values, epidemiological cut-offs and/or pharmacokinetics/pharmacodynamics-driven thresholds, biochemical assays and site-directed mutagenesis) and clinical evidence.

system presented here will be refined further by taking the following criteria into consideration: phylogenetic information, laboratory evidence (e.g. MIC values, epidemiological cut-offs and/or pharmacokinetics/pharmacodynamics-driven thresholds, biochemical assays and site-directed mutagenesis) and clinical evidence. This study establishes the first confidence-graded list of mutations for predicting drug resistance, and as such should serve as a gene target guide for developing new molecular diagnostics, and as a tool for supporting the clinical interpretation of existing molecular diagnostics such as the Hain GenoType assays. Once incorporated into the ReSeqTB knowledge base, we expect the confidence-graded list of mutations to improve in precision iteratively as data are accumulated, and it will be revised annually through an expert review process similar to the methods established for the Stanford HIV Drug Resistance Database (https://hivdb.stanford.edu). This will be of particular value for the interpretation of WGS-based in vitro diagnostics that are currently being piloted as decision support tools for rapid and comprehensive characterisation of clinically relevant resistance to guide individualised treatment regimens containing the most effective, least toxic drug combinations. Ultimately, we aim to provide a comprehensive and user friendly tool to assist clinicians with the interpretation of resistance mutations in MTBC.

and comprehensive characterisation of clinically relevant resistance to guide individualised treatment regimens containing the most effective, least toxic drug combinations. Ultimately, we aim to provide a comprehensive and user friendly tool to assist clinicians with the interpretation of resistance mutations in MTBC. Supplementary material 10.1183/13993003.01354-2017.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary materials 1, 4, 5, 6 and 11 ERJ-01354-2017_Suppl_1-4-5-6-11 Supplementary materials 2, 3, 7, 8, 9, 10 and 12 ERJ-01354-2017_Suppl_2-3-7-8-9-10-12 Disclosures 10.1183/13993003.01354-2017.Supp2D. Alland ERJ-01354-2017_Alland I. Comas ERJ-01354-2017_Comas C.M. Denkinger ERJ-01354-2017_Denkinger K. Dheda ERJ-01354-2017_Dheda D. Hanna ERJ-01354-2017_Hanna C.U. Koser ERJ-01354-2017_Koser C. Lange ERJ-01354-2017_Lange R. Liwski ERJ-01354-2017_Liwski S. Niemann ERJ-01354-2017_Niemann L. Rigouts ERJ-01354-2017_Rigouts M. Schito ERJ-01354-2017_Schito Acknowledgements We acknowledge Naomi Hillery (Health Services Research Center, Dept of Family Medicine and Public Health University of California, San Diego, CA, USA) for help with data management.

C. Lange ERJ-01354-2017_Lange R. Liwski ERJ-01354-2017_Liwski S. Niemann ERJ-01354-2017_Niemann L. Rigouts ERJ-01354-2017_Rigouts M. Schito ERJ-01354-2017_Schito Acknowledgements We acknowledge Naomi Hillery (Health Services Research Center, Dept of Family Medicine and Public Health University of California, San Diego, CA, USA) for help with data management. Author contributions were as follows. P. Miotto: literature search, study design, data collection, data analysis, data interpretation, manuscript writing; C.U. Köser and T.C. Rodwell: study design, data interpretation, manuscript writing; L. Chindelevitch: study design, data analysis, data interpretation, manuscript writing; D.M. Cirillo, M. Schito and K. Dheda: study design, manuscript writing; B. Tessema and E. Tagliani: literature search, data collection, critical revision of manuscript; A.M. Starks, C. Emerson, D. Hanna, P.S. Kim, R. Liwski, M. Zignol, C. Gilpin, S. Niemann, C.M. Denkinger, J. Fleming, R.M. Warren, D. Crook, J. Posey, S. Gagneux, S. Hoffner, C. Rodrigues, I. Comas, D.M. Engelthaler, M. Murray, D. Alland, L. Rigouts, C. Lange, R. Hasan, U.D.K. Ranganathan, R. McNerney and M. Ezewudo: study design, critical revision of manuscript. This article has supplementary material available from erj.ersjournals.com

Author contributions were as follows. P. Miotto: literature search, study design, data collection, data analysis, data interpretation, manuscript writing; C.U. Köser and T.C. Rodwell: study design, data interpretation, manuscript writing; L. Chindelevitch: study design, data analysis, data interpretation, manuscript writing; D.M. Cirillo, M. Schito and K. Dheda: study design, manuscript writing; B. Tessema and E. Tagliani: literature search, data collection, critical revision of manuscript; A.M. Starks, C. Emerson, D. Hanna, P.S. Kim, R. Liwski, M. Zignol, C. Gilpin, S. Niemann, C.M. Denkinger, J. Fleming, R.M. Warren, D. Crook, J. Posey, S. Gagneux, S. Hoffner, C. Rodrigues, I. Comas, D.M. Engelthaler, M. Murray, D. Alland, L. Rigouts, C. Lange, R. Hasan, U.D.K. Ranganathan, R. McNerney and M. Ezewudo: study design, critical revision of manuscript. This article has supplementary material available from erj.ersjournals.com Use of trade names is for identification only and does not constitute endorsement by the US Department of Health and Human Services, the US Public Health Service, or the Centers for Disease Control and Prevention. The findings and conclusions in this report are those of the authors and do not necessarily represent the views of the funding agency.

entification only and does not constitute endorsement by the US Department of Health and Human Services, the US Public Health Service, or the Centers for Disease Control and Prevention. The findings and conclusions in this report are those of the authors and do not necessarily represent the views of the funding agency. Support statement: This study was supported by the Bill and Melinda Gates Foundation under grant agreement FIND OPP1115209 to address how to score mutations in the ReSeqTB data sharing platform initiative. The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. K. Dheda is supported by the South African MRC and the EDCTP. I. Comas is supported by the Ministerio de Economía y Competitividad (Spanish Government) research grant SAF2016–77346-R and the European Research Council (ERC) (638553-TB-ACCELERATE). L. Chindelevitch is supported by a Sloan Fellowship. S. Niemann is supported by grants of the German Center for Infection Research. Funding information for this article has been deposited with the Crossref Funder Registry. Conflict of interest: Disclosures can be found alongside this article at erj.ersjournals.com

Introduction Community-acquired pneumonia (CAP) is an increasingly common cause of admission to hospital and is potentially fatal [1]. However, 80% of hospitalised patients survive the acute illness and are discharged [2]. Patients, clinicians and researchers have an interest in accurately describing symptom recovery among those who survive the initial pneumonia insult. Understanding the factors associated with symptom recovery time would not only enable us to prognosticate for patients but also to address modifiable risk factors for delayed symptom recovery. However, the pneumonia recovery literature is sparse. As a consequence, it is difficult for clinicians faced with simple questions such as “How long will it take me to feel better?” or “Will I ever get back to normal?” to provide anything more than a very general answer. Several cross-sectional studies have used scoring systems to summarise the level of symptoms within a cohort at fixed time-points following CAP [3, 4]. However, our understanding of which factors influence recovery has been hampered by a lack of longitudinal studies. Owing to comorbidities whose symptoms overlap with pneumonia, those who have a higher level of symptoms pre-pneumonia are likely to have a relatively high level at maximum recovery. Moreover, the perception of symptoms is unique, and the way an individual patient scores symptoms at a particular time point will be correlated with their previous and future scores; cross-sectional studies do not take into account this longitudinal correlation.

kely to have a relatively high level at maximum recovery. Moreover, the perception of symptoms is unique, and the way an individual patient scores symptoms at a particular time point will be correlated with their previous and future scores; cross-sectional studies do not take into account this longitudinal correlation. Statistical modelling produces a function that can explain the pattern of variation in a series of observed responses using as few input variables as possible [5]. The input variables that produce the best-fitting model may also give clues as to the mechanisms underlying the phenomenon being studied [6]. We modelled symptom scores from a prospective, longitudinal, observational study of symptom recovery from CAP. Our aims were to address three fundamental questions relating to the symptoms of pneumonia: Do patients completely recover from pneumonia symptoms? How long does this recovery take? Which factors influence symptomatic recovery? Methods Ethics statement This work was approved by the UK NHS Research Ethics Committee (NHS REC 10/WNo03/40), was sponsored by Aintree University Hospital NHS Foundation Trust and was listed on the NIHR Clinical Research Network portfolio. All subjects provided informed consent to join the study. A consultee provided assent on behalf of those who lacked capacity as a consequence of CAP related delirium, with consent being retrospectively obtained upon recovery of capacity.

S Foundation Trust and was listed on the NIHR Clinical Research Network portfolio. All subjects provided informed consent to join the study. A consultee provided assent on behalf of those who lacked capacity as a consequence of CAP related delirium, with consent being retrospectively obtained upon recovery of capacity. Study subjects Eligible subjects aged 16 years or older were recruited from two hospitals in Liverpool, UK, between February 2011 and March 2013. Subjects with CAP (British Thoracic Society definition) were recruited within 24 h of their first dose of in-hospital antibiotic [7]. We excluded patients admitted within the last 14 days, with cystic fibrosis (CF) or non-CF bronchiectasis, who: were immunocompromised or mechanically ventilated, required renal replacement therapy, had thoracic malignancy or advanced cancer of any type, were receiving palliative treatment, or had chronic cognitive impairment preventing completion of a symptom questionnaire. In-hospital management and study procedures The study team had no role in the clinical management of study subjects. Both hospitals had similar pneumonia protocols and are part of a robust regional pneumonia performance audit [8]. At enrolment, demographics and clinical data were recorded and the subjects provided clinical samples. Follow-up was 48 h following enrolment, on the day of discharge, and at clinic visits 1, 6 and 12 months following admission.

ilar pneumonia protocols and are part of a robust regional pneumonia performance audit [8]. At enrolment, demographics and clinical data were recorded and the subjects provided clinical samples. Follow-up was 48 h following enrolment, on the day of discharge, and at clinic visits 1, 6 and 12 months following admission. The CAP-sym questionnaire We measured symptoms using the CAP-sym (community acquired pneumonia-symptom) questionnaire, which is a validated patient-based tool for measuring CAP symptoms (see supplementary figure 1) [9]. At the time of enrolment, subjects conducted the CAP-sym questionnaire twice – the first iteration representing their symptoms at recruitment and the second thinking back 30 days prior to admission representing how they felt pre-pneumonia. The CAP-sym questionnaire was then repeated at all subsequent visits. FIGURE 1 Schematic representation of a pneumonia symptom profile; δ represents the pre-pneumonia level, β is the peak symptom level, γ is the symptom decay after admission and α is the residual symptom score after follow-up.

The CAP-sym questionnaire We measured symptoms using the CAP-sym (community acquired pneumonia-symptom) questionnaire, which is a validated patient-based tool for measuring CAP symptoms (see supplementary figure 1) [9]. At the time of enrolment, subjects conducted the CAP-sym questionnaire twice – the first iteration representing their symptoms at recruitment and the second thinking back 30 days prior to admission representing how they felt pre-pneumonia. The CAP-sym questionnaire was then repeated at all subsequent visits. FIGURE 1 Schematic representation of a pneumonia symptom profile; δ represents the pre-pneumonia level, β is the peak symptom level, γ is the symptom decay after admission and α is the residual symptom score after follow-up. The choice of covariate predictors of recovery We selected the following potential predictors of recovery as covariates for our model: advancing age, which is associated with short-term death following CAP [10]; sex, men have previously been shown to have adverse outcomes following CAP [11]; the CURB65 (confusion, urea, respiratory rate, blood pressure, age >65 years) score, which predicts 30-day mortality following CAP [12]; the Charlson comorbidity index, which estimates the risk of death attributable to comorbidity during a hospital admission [13]; smoking, which impairs recovery mechanisms [14]; COPD, which is associated with an increased risk of pneumonia [15]; pro-calcitonin (PCT) level, which has been suggested to have prognostic value [16]; C-reactive protein (CRP) level, which has been associated with risk of complications following CAP [17]; prior statin use, which has been associated with improved outcomes following CAP [18]; and socioeconomic deprivation, which is associated with increased risk of CAP [11].

een suggested to have prognostic value [16]; C-reactive protein (CRP) level, which has been associated with risk of complications following CAP [17]; prior statin use, which has been associated with improved outcomes following CAP [18]; and socioeconomic deprivation, which is associated with increased risk of CAP [11]. Statistical modelling From an initial exploratory analysis of the CAP-sym data, a statistical function was derived, which included the four parameters represented schematically in figure 1. Non-linear mixed effects modelling (NONMEM, version 7.3, ICON, Dublin, Ireland) was applied to the CAP-sym data, including inter-individual variability (IIV) as a multiplicative, log-normally distributed random effect on each of the four model parameters. Graphical plots of model fit were generated. See supplementary material for a full algebraic derivation of the model.

.3, ICON, Dublin, Ireland) was applied to the CAP-sym data, including inter-individual variability (IIV) as a multiplicative, log-normally distributed random effect on each of the four model parameters. Graphical plots of model fit were generated. See supplementary material for a full algebraic derivation of the model. Covariate effects on each parameter were assessed univariably; we used linear functions for continuous variables and fitted separate parameters for each level of each categorical variable. We then carried out backward eliminations, with covariates being retained if their removal from the model produced a statistically significant increase in objective function value. If covariate values were missing, the median was used for continuous variables if less than 10% were not recorded. An additional parameter was estimated for missing categorical variables. CAP-sym data-points provided by patients who subsequently died during follow-up were included. Our model was fitted by maximum likelihood, which automatically corrects for selection bias and which depends on a patient's observed CAP-sym measurements prior to death, although not on any additional dependence on unmeasured features of their CAP-sym trajectory. An alternative approach using “informative dropout modelling” would have been inappropriate given the small number of observed deaths.

lection bias and which depends on a patient's observed CAP-sym measurements prior to death, although not on any additional dependence on unmeasured features of their CAP-sym trajectory. An alternative approach using “informative dropout modelling” would have been inappropriate given the small number of observed deaths. Results Cohort characteristics Of 169 patients, 792 CAP-sym data-points were recorded. The cohort's characteristics are presented in table 1. The median age was 68 years. Patients frequently had comorbid conditions; 39% of subjects were active smokers and 53.4% had prior-pulmonary disease. The highest Charlson comorbidity level was 6 out of 24, and modelling subjects were grouped into levels 1–3 and 4–6, which were compared to those with score 0. Supplementary figure 2 shows the distribution of socioeconomic deprivation, with 43% of subjects being drawn from the most deprived centile of the population of England. During the first 24 h of hospital admission, 53% of patients were pyrexial. Fifty-four per cent of patients had a procalcitonin (PCT) level >0.5 ng·mL­1, a threshold above which antibiotics for bacterial infection are advised. Pyrexia, raised neutrophils or raised PCT was found in 82.8% of patients, demonstrating that the majority had evidence of systemic inflammation. Figure 2a and b show PCT and CRP values on admission and during follow-up. The distribution of CURB65 scores across low risk (scores 0 and 1), intermediate (score 2) and high-risk (score ≥3) categories was similar to those found in UK CAP audits [2]. Median length of hospital stay was 6 days and in-patient mortality was 7.7%.

2a and b show PCT and CRP values on admission and during follow-up. The distribution of CURB65 scores across low risk (scores 0 and 1), intermediate (score 2) and high-risk (score ≥3) categories was similar to those found in UK CAP audits [2]. Median length of hospital stay was 6 days and in-patient mortality was 7.7%. TABLE 1 Cohort characteristics

2a and b show PCT and CRP values on admission and during follow-up. The distribution of CURB65 scores across low risk (scores 0 and 1), intermediate (score 2) and high-risk (score ≥3) categories was similar to those found in UK CAP audits [2]. Median length of hospital stay was 6 days and in-patient mortality was 7.7%. TABLE 1 Cohort characteristics Subjects n 169 Age years 68, 16–98 (18) Males n (%) 88 (52.0%) BMI# kg·m­2 26 (22–30) Ethnicity White British 166 (98.2%) White other 2 (1.2%) Black African 1 (0.6%) Comorbidities COPD 70 (41.0%) Chronic lung disease other than COPD 21 (12.4%) Congestive cardiac failure 23 (13.6%) Dementia 2 (1.2%) Diabetes# 28 (16.7%) Hepatic disease 5 (3.0%) Renal disease 14 (8.3%) Lived in nursing/residential care 8 (4.7%) Smoking status# Active smoker 63 (39%) Ex-smoker 66 (41%) Never-smoker 32 (20%) Charlson comorbidity index 0 56 (33.1%) 1 69 (40.8%) 2 18 (10.7%) 3 17 (10.1%) 4 6 (3.6%) 5 2 (1.2%) 6 1 (0.6%) >6 0 Influenza infection# 18 (16.8%) CURB65 score 0–1 79 (46.7%) 2 50 (29.6%) 3–5 40 (23.7%) Infection markers Pyrexial 90 (53.0%) Neutrophil count ×109 per L 9.9 (7.1–14.8) CRP mg·mL­1 145 (61–248) Pro-calcitonin# ng·mL­1 0.70 (0.1–3.9) >0.25 ng·mL­1 98 (64.5%) >0.5 ng·mL­1 83 (54.6%) Antibiotic regimen¶ Appropriate 107/159 (67.3%) Over treated 41/159 (25.8%) Under treated 11/159 (6.9%) Received macrolide 133/159 (83.6%) Outcome Length of stay days 6, 0–58 (7.8) Readmission within 30 days of discharge 16/135 (11.8%) In-hospital mortality 13 (7.7%) Death within 30 days of discharge 1/135 (0.7%) Death post discharge 13/135 (9.6%) Total 1-year mortality 26 (15.4%) Cause of in-hospital death CAP 8 (61.5) Sepsis 2 (15.4) Myocardial infarction 1 (7.7) Respiratory failure 1 (7.7) Unknown 1 (7.7) Cause of death post discharge CAP 2 (15.4%) HAP 1 (7.7%) Gastric cancer 1 (7.7%) Lung cancer 3 (23.1%) Interstitial lung disease 1 (7.7%) COPD 2 (15.4%) Unknown 3 (23.1%) Data are presented as median, range (sd), or median (interquartile range), unless otherwise stated. BMI: body mass index; COPD: chronic obstructive pulmonary disease; CURB65: confusion, urea, respiratory rate, blood pressure, age >65 years; CRP: C-reactive protein; CAP: community-acquired pneumonia; HAP: hospital-acquired pneumonia. #: incomplete data for diabetes (n=168), BMI (n=126), smoking status (n=161) and pro-calcitonin (n=166). ¶: initial empirical antibiotic choice was deemed appropriate if it was consistent with that stated in the local guidelines.

: C-reactive protein; CAP: community-acquired pneumonia; HAP: hospital-acquired pneumonia. #: incomplete data for diabetes (n=168), BMI (n=126), smoking status (n=161) and pro-calcitonin (n=166). ¶: initial empirical antibiotic choice was deemed appropriate if it was consistent with that stated in the local guidelines. Local guidelines are based on the British Thoracic Society guidelines and are based around CURB65 score on admission. Over-treatment was therefore a treatment regime ordinarily reserved for a higher CURB65 score and under-treatment was a regime aimed at lower risk patients based on the CURB65 score. Here, the assessment of appropriateness does not take into account factors other than CURB65 score, such as treatment duration. Patients were recorded as having received a macrolide if at any point in their pneumonia treatment they received a macrolide of any sort and for any duration.

tients based on the CURB65 score. Here, the assessment of appropriateness does not take into account factors other than CURB65 score, such as treatment duration. Patients were recorded as having received a macrolide if at any point in their pneumonia treatment they received a macrolide of any sort and for any duration. FIGURE 2 a) Distribution of pro-calcitonin (PCT) levels on admission and during follow-up. At all time points, the distribution of PCT values was wide. Using accepted respiratory tract infection treatment thresholds, on admission 64.5% of values were >25 ng·mL­1 and 54.6% of values were >0.5 ng·mL­1. 48 h after in-hospital treatment, the median PCT level had fallen when compared with the median on admission, but many subjects had levels in the treatment range. By 1 month, the median PCT level had fallen below the lower treatment threshold of 0.25 ng·mL­1; however, at 1 month and 6 months, some subjects had high PCT levels. b) Distribution of C-reactive protein (CRP) levels on admission and during follow-up. The pattern of CRP level was very similar to that of pro-calcitonin. Values were high at presentation, had begun to fall by 48 h, had fallen substantially by 1 month, and had changed very little between 1 month and 6 months. The 2014 NICE pneumonia guidelines suggest that antibiotic treatment should be offered to all patients diagnosed with pneumonia. If a diagnosis of pneumonia cannot be made, but a lower respiratory tract infection has been diagnosed, then the decision to treat with antibiotics can be assisted by the CRP level. If the CRP level is >100 mg·L­1, antibiotics are recommended; antibiotics are considered if the level is between 20 and 100 mg·L­1; antibiotics are withheld if the level is <20 mg·L­1.

t a lower respiratory tract infection has been diagnosed, then the decision to treat with antibiotics can be assisted by the CRP level. If the CRP level is >100 mg·L­1, antibiotics are recommended; antibiotics are considered if the level is between 20 and 100 mg·L­1; antibiotics are withheld if the level is <20 mg·L­1. General symptom trends for the cohort Table 2 compares data from our cohort with data from the multicentre study with which the CAP-sym questionnaire was validated; mean CAP-sym values were similar at equivalent time points [19]. Figure 3 displays the range of CAP-sym scores at each time point and reveals a highly skewed distribution, suggesting that summarising the score for a whole cohort with a mean value may be misleading. To explore possible causes for different levels of symptoms, we plotted the summary CAP-sym scores over time grouped by various covariates, e.g. smoking status (figure 4). At every time point, smokers reported higher symptom scores than ex-smokers, who in turn reported higher scores than never-smokers. TABLE 2 A comparison of community-acquired pneumonia symptom (CAP-sym) values with the CAP-sym validation cohort

General symptom trends for the cohort Table 2 compares data from our cohort with data from the multicentre study with which the CAP-sym questionnaire was validated; mean CAP-sym values were similar at equivalent time points [19]. Figure 3 displays the range of CAP-sym scores at each time point and reveals a highly skewed distribution, suggesting that summarising the score for a whole cohort with a mean value may be misleading. To explore possible causes for different levels of symptoms, we plotted the summary CAP-sym scores over time grouped by various covariates, e.g. smoking status (figure 4). At every time point, smokers reported higher symptom scores than ex-smokers, who in turn reported higher scores than never-smokers. TABLE 2 A comparison of community-acquired pneumonia symptom (CAP-sym) values with the CAP-sym validation cohort Clinical stage CAP-sym score mean±sd This study (n=169) Torreset al. [19] Standard treatment (n=244) Moxifloxacin (n=233) Pre-pneumonia 13.6±14.5 NA NA Enrolment 32.8±14.6 33.9±13.6 34.3±13.2 Mid-treatment 23.8±15.1 20.6±11.0 20.9±11.8 Discharge 15.3±10.6 12.0±10.3 13.5±11.5 Early follow-up 13.6±11.8 9.6±10.8 10.1±10.9 Medium-term follow-up 12.6±11.8 NA NA Late follow-up 13.3±12.7 NA NA Mid-treatment in this study was 48 h after enrolment, while in the study by Torres et al. [19] it was between days 3 and 5. Discharge in this study is equivalent to the 7–10 day “test of cure” time point in the study by Torres et al. [19]. Early follow-up in this study was at 28 days, while in the study by Torres et al. [19] it was between days 28 and day 35. NA: not available.

in the study by Torres et al. [19] it was between days 3 and 5. Discharge in this study is equivalent to the 7–10 day “test of cure” time point in the study by Torres et al. [19]. Early follow-up in this study was at 28 days, while in the study by Torres et al. [19] it was between days 28 and day 35. NA: not available. FIGURE 3 Distribution of community-acquired pneumonia symptom (CAP-sym) scores at each time point, and the median trend. FIGURE 4 The association between smoking status and community-acquired pneumonia symptom (CAP-sym) scores. Bars represent the inter-quartile range (IQR) and whiskers extend to 1.5× the IQR. The line within the box represents the median value for that group.

FIGURE 3 Distribution of community-acquired pneumonia symptom (CAP-sym) scores at each time point, and the median trend. FIGURE 4 The association between smoking status and community-acquired pneumonia symptom (CAP-sym) scores. Bars represent the inter-quartile range (IQR) and whiskers extend to 1.5× the IQR. The line within the box represents the median value for that group. Non-linear mixed effects modelling of CAP-sym scores Figure 5 shows the longitudinal profile of each patient's symptoms from which we explored the kinetics of individual patient recovery. Following univariable covariate analysis, the Charlson comorbidity index, sex, smoking status and COPD all had significant associations with the level of pre-pneumonia symptoms (δ). Age (centred on the median 68 years), PCT, CRP level, sex and CURB65 score had significant associations with the peak level of symptoms (β) (see supplementary table 1 for the univariable analysis). There was little variability in the pattern of symptom resolution (γ) or in the extent of recovery to baseline (α), and therefore, covariate effects on these parameters were not supported. Following backward elimination, the final model included two covariate effects: the effect of Charlson comorbidity index on pre-pneumonia symptoms (δ), and the effect of age on the peak level of symptoms (β); (see supplementary table 1 for the multivariable analysis). The effect of age was such that, for every year older than 68 years, patients described a lower level of symptoms, with symptoms increasing in patients aged <68 years. See supplementary table 2 for effect estimates produced by the final model and supplementary figures 3a, 3b and 4 for plots of model fit.

ble analysis). The effect of age was such that, for every year older than 68 years, patients described a lower level of symptoms, with symptoms increasing in patients aged <68 years. See supplementary table 2 for effect estimates produced by the final model and supplementary figures 3a, 3b and 4 for plots of model fit. FIGURE 5 Symptom kinetics of individual patients. Each line joins community-acquired pneumonia symptom (CAP-sym) scores recorded on an individual patient at multiple time points. Recovery time and extent of recovery The modelling revealed that most patients made a near-complete recovery to baseline. The pattern of recovery (γ) was exponential and the time taken to recover was dependent on the initial severity of symptoms. Exponential processes have constant half-lives; this enabled us to calculate that, on average, patients' symptoms would have reduced by approximately 97% of their initial value by 9.8 days (95% CI 7.3–12.2 days) (see supplementary material for half-life calculation). Discussion Principal findings This is the first study to longitudinally model the determinants of symptom patterns among patients who had been hospitalised, but not ventilated, for CAP. The time taken for an individual patient to recover to their baseline was dependent on the severity of their acute symptoms, which in turn was influenced by the patient's age and comorbidity. Our model estimated that, on average, patients had recovered 97% of their CAP symptoms by 10 days.

but not ventilated, for CAP. The time taken for an individual patient to recover to their baseline was dependent on the severity of their acute symptoms, which in turn was influenced by the patient's age and comorbidity. Our model estimated that, on average, patients had recovered 97% of their CAP symptoms by 10 days. Strengths and weaknesses The CAP-sym questionnaire is one of two psychometrically validated pneumonia questionnaires available, the other being the CAP-SCORE; we believe that the CAP-sym score is the better tool for assessing symptoms [20]. The CAP-SCORE gathers information about only three symptoms (shortness of breath, cough and sputum production) and since 50% of patients presenting with CAP produce no sputum, it has very limited scope to capture the spectrum of symptoms described by patients with pneumonia. In contrast, the CAP-sym questionnaire gathers information about 18 symptoms and has the advantage of having been translated and validated in 13 different languages, enabling its use in international multicentre studies. Two randomised clinical trials have used CAP-sym as an outcome measure, and have demonstrated its responsiveness to treatment and consistency in scoring average symptoms at similar time points [19, 21]. In our study the univariable associations between the presenting CAP-sym score and other markers of severity (PCT, CRP levels and CURB65 score) provided further evidence of the questionnaire's responsiveness and convergent validity regarding hospitalised CAP patients not requiring ventilation.

lar time points [19, 21]. In our study the univariable associations between the presenting CAP-sym score and other markers of severity (PCT, CRP levels and CURB65 score) provided further evidence of the questionnaire's responsiveness and convergent validity regarding hospitalised CAP patients not requiring ventilation. During validation, the CAP-sym questionnaire was found to be more responsive to pneumonia treatment than the widely used generic questionnaire SF36 [9]. However, a criticism of all pneumonia symptom questionnaires is that, since there are no symptoms unique to pneumonia, the scores they produce are the sum of chronic comorbidity plus acute pneumonia. As a consequence of this relative lack of specificity, when prior studies have presented an average symptom score for a group of patients, it is not possible to determine what proportion of that average is derived from the pneumonia as opposed to comorbidities. Separating acute from chronic symptoms involves studying how an individual changes from his/her own baseline. This approach will make intuitive sense to clinicians who are used to asking patients how they feel “now” compared to “their normal”. Our model was specified in just this way; it took into account the effects of multiple clinical variables, including comorbidity and socioeconomic status, and determined how these affected CAP-sym scores before, during and after pneumonia. Linking repeated measurements on the same patient is a fundamental tenet of longitudinal analysis, but previous studies of symptomatic recovery from pneumonia have not taken this into account [5].

comorbidity and socioeconomic status, and determined how these affected CAP-sym scores before, during and after pneumonia. Linking repeated measurements on the same patient is a fundamental tenet of longitudinal analysis, but previous studies of symptomatic recovery from pneumonia have not taken this into account [5]. There are some limitations to this work. The results from this hospitalised cohort cannot be directly extrapolated to primary care, where most CAP is managed. Patients who require mechanical ventilation represent a numerically small but important fraction of all CAP patients. We excluded these individuals as the effect of ventilation on recovery from pneumonia would have introduced a bias that could only be accounted for by a larger study where ventilation was an explanatory variable. In-patient mortality and median age were low when compared with both contemporary UK CAP audits and a large German cohort study [2, 10]. The observed differences were related to the prospective nature of our study and its rigorous inclusion criteria; another UK prospective CAP cohort study reported very similar patient characteristics and outcomes to ours [18].

w when compared with both contemporary UK CAP audits and a large German cohort study [2, 10]. The observed differences were related to the prospective nature of our study and its rigorous inclusion criteria; another UK prospective CAP cohort study reported very similar patient characteristics and outcomes to ours [18]. Our multivariable analysis revealed that age explained variation in peak symptoms. A previous study of acute pneumonia symptoms in a cohort of mean age 56 years compared the mean level of symptoms described by three age groups of patients at presentation. This study revealed that older people, on average, described a lower level of symptoms than younger people [22]. Previous work has suggested that the elderly report symptoms differently to younger patients [23], and others have argued that pneumonia in the elderly is a distinct entity [24, 25]. We have described a method for incorporating covariates such as age in the analysis of longitudinal symptom data and have provided evidence to power future studies to account for age. Another possible source of bias was the effect of age on the decisions made by the admitting clinicians. It is possible that younger patients who were very symptomatic but had little comorbidity may have been admitted less frequently than older patients with fewer symptoms but social reasons for admission. If these effects were pronounced, they could have influenced the age effect seen in the acute CAP-sym scores. Millet et al. [26] have shown that, among those presenting to hospital with CAP, admission is more likely as age increases from 65 to 85 years, and in the presence of comorbidities such as COPD. However, their group were unable to assess the impact of severity on admission and they did not investigate symptoms. To further evaluate these effects, future studies of recovery would record qualitative data on decisions to admit and discharge.

s from 65 to 85 years, and in the presence of comorbidities such as COPD. However, their group were unable to assess the impact of severity on admission and they did not investigate symptoms. To further evaluate these effects, future studies of recovery would record qualitative data on decisions to admit and discharge. Prior to developing pneumonia, many patients have symptoms which originate from comorbidities, and in order to determine recovery it is necessary to establish this baseline symptom level. We did this by patient recall (model parameter delta (δ)), though we acknowledge that this may be subject to recall bias. However, we observed that most patients’ CAP-sym score was very similar at the end of follow-up to the pre-pneumonia level they recalled – in fact, the mean difference between pre-pneumonia scores and post-pneumonia scores was smaller than the detection limit of the CAP-sym score, i.e. <1 CAP-sym point. This suggests that the effect of recall bias was small and that patients' recall of pre-pneumonia symptoms was accurate. Another note of caution is that the chosen functional form of the model we used is but one of many that would have given a tolerable fit to the data, and our estimate of the half-life for symptom recovery was determined by our choice. Validation of this model would require a larger study.

eumonia symptoms was accurate. Another note of caution is that the chosen functional form of the model we used is but one of many that would have given a tolerable fit to the data, and our estimate of the half-life for symptom recovery was determined by our choice. Validation of this model would require a larger study. Meaning of the study: implications for clinicians and policymakers Patient-based measures are now regarded as essential to the comprehensive assessment of clinical quality, and as important measures of outcome in research studies. This is a recent development; as recently as the 2011 European Respiratory Society guidelines for the management of lower respiratory tract infections no reference was made to symptomatic recovery from pneumonia [27]. The most recent comprehensive pneumonia guideline is the 2014 UK National Institute of Clinical Excellence (NICE) guidelines, which set out a number of recovery milestones, although they caution that the advice is based on low-grade evidence [28]. NICE suggests that clinicians tell patients that “the rate of improvement will vary with the severity of the pneumonia”. Our results showed that resolution of symptoms was constant across the cohort and that what determined the time to recovery was the initial severity of symptoms.

is based on low-grade evidence [28]. NICE suggests that clinicians tell patients that “the rate of improvement will vary with the severity of the pneumonia”. Our results showed that resolution of symptoms was constant across the cohort and that what determined the time to recovery was the initial severity of symptoms. Another NICE milestone suggests: “By 6 months: most people will feel back to normal”. We calculated that, on average, patients recovered to within 3% of baseline symptoms by 10 days. These findings should be treated with caution since covariate effects were not included in the analysis of α of γ owing to the limited range of variation in those parameters. A larger study with more frequent sampling during the first 6 months of recovery would be required for validation. Another note of caution is that the CAP-sym questionnaire has been validated to assess pneumonia symptoms but not functional status. A patient may have had a near-complete resolution of pneumonia symptoms, as measured by the CAP-sym score, but not be able to return home owing to functional decline. This study was not designed to assess this, but other works have shown that functional recovery may be influenced by severity, making the 6-month milestone for complete recovery a more realistic goal for some patients [29].

s, as measured by the CAP-sym score, but not be able to return home owing to functional decline. This study was not designed to assess this, but other works have shown that functional recovery may be influenced by severity, making the 6-month milestone for complete recovery a more realistic goal for some patients [29]. Several stereotypical patterns have been proposed to describe recovery from acute illness such as pneumonia [30]. CAP was traditionally conceptualised as a “big hit” illness that, if survived, was followed by complete recovery. However, pointing to the poor long-term survival of those who recover from CAP, some investigators have suggested that pneumonia may induce a “slow-burn” pattern of deterioration more akin to a chronic disease [31, 32]. Our data suggest that the acute symptoms of pneumonia resolve completely in most patients; however, if pneumonia were to destabilise comorbidities, then the recovery pattern may resemble the “relapsing recurrences” paradigm. Recent work suggests that it may soon be possible to trial therapies aimed at enhancing the resolution of inflammation caused by pneumonia [33]. Measuring outcome in these trials will be challenging since most people survive an episode of CAP, and studies aimed at detecting changes in mortality would need to be very large since the effect size is likely to be small. Barlow et al. have argued that the ideal pneumonia intervention study would include patient-related outcomes such as the CAP-sym score, and we have demonstrated a robust methodology for this [34].

nd studies aimed at detecting changes in mortality would need to be very large since the effect size is likely to be small. Barlow et al. have argued that the ideal pneumonia intervention study would include patient-related outcomes such as the CAP-sym score, and we have demonstrated a robust methodology for this [34]. Supplementary material 10.1183/13993003.02170-2016.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-02170-2016_supplementary

nd studies aimed at detecting changes in mortality would need to be very large since the effect size is likely to be small. Barlow et al. have argued that the ideal pneumonia intervention study would include patient-related outcomes such as the CAP-sym score, and we have demonstrated a robust methodology for this [34]. Supplementary material 10.1183/13993003.02170-2016.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-02170-2016_supplementary Acknowledgements We would like to thank the patients who volunteered for the study, the NIHR for funding the work, and the NIHR Clinical Local Research Network (CLRN) for providing strategic support for recruitment. We would also like to thank John Blakey for his critical appraisal of the manuscript. D.G. Wootton conceived the study, wrote the grant to obtain the funding, obtained ethical approval, participated in recruitment, followed up the subjects, analysed the results and wrote the manuscript. L. Dickinson and H. Pertinez performed the modelling and contributed to the manuscript. J. Court, L. Keogan, L. Macfarlane and S. Wilks helped recruit the patients and collate the data, and reviewed the manuscript. In addition, L. Keogan assisted with patient follow-up. J. Gallagher is a patient representative who contributed to the grant proposal, reviewed the study protocol and contributed to the manuscript. P.J. Diggle and S.B. Gordon helped design the study and, along with M. Woodhead, supervised the fellowship project, including discussion of recruitment, data analysis and preparation of the manuscript. P.J. Diggle supervised the statistical plan and analyses and derived the statistical model structure.

the manuscript. P.J. Diggle and S.B. Gordon helped design the study and, along with M. Woodhead, supervised the fellowship project, including discussion of recruitment, data analysis and preparation of the manuscript. P.J. Diggle supervised the statistical plan and analyses and derived the statistical model structure. Patient involvement: Patients told us that research into recovery was important to them and worked with us to develop the study protocol. A patient representative is a co-author on this manuscript. Patient involvement was remunerated as per INVOLVE guidelines (www.invo.org.uk/posttypepublication/payment-for-involvement/). This article has supplementary material available from erj.ersjournals.com Earn CME accreditation by answering questions about this article. You will find these at erj.ersjournals.com/journal/cme Support statement: The study was funded by a National Institute of Health Research (NIHR) Doctoral Research Fellowship awarded to D.G. Wootton, with support from the North West CLRN. Funding information for this article has been deposited with the Crossref Funder Registry. Conflict of interest: None declared.

Introduction Chronic airflow obstruction (CAO) is the primary characteristic of patients with chronic obstructive pulmonary disease (COPD) and affects up to one in five adults, depending on where they live, according to data from the Burden of Obstructive Lung Disease (BOLD) study [1]. COPD is expected to keep its position as the third most important cause of death worldwide [2], and imposes a substantial burden on quality of life [3] and healthcare utilisation [4]. So far, data on productivity-related burden of CAO or COPD have been scant [4]. Only three population-based studies have provided employment rates in CAO [5–7]. Erdal et al. [5] showed that 55% of individuals with CAO from a general Norwegian population were in a paid job, versus 87% of controls without CAO. However, controls were younger and had higher levels of education and the authors did not examine employment in multivariate analyses. Jansson et al. [6] examined CAO-specific disability in northern Sweden, but did not include a control group and did not report employment rates. In the PLATINO (Latin American Project for Research in Pulmonary Obstruction) study undertaken in five Latin-American countries, Montes de Oca et al. [7] showed that the workforce participation among subjects with CAO was lower than in healthy subjects (41.8% versus 57.1%). However, in multivariable analyses they found that higher age, dyspnoea, number of comorbid conditions, female sex and lower education were associated with unemployment, whereas CAO was only of borderline significance.

kforce participation among subjects with CAO was lower than in healthy subjects (41.8% versus 57.1%). However, in multivariable analyses they found that higher age, dyspnoea, number of comorbid conditions, female sex and lower education were associated with unemployment, whereas CAO was only of borderline significance. The BOLD study is a large international study providing population-based estimates of the prevalence and burden of CAO. One of the primary objectives of the BOLD study is to estimate disease burden in terms of activity limitation and economic impact [8]. In the current analysis, we have compared estimates of employment status in BOLD participants with and without CAO across the world. Methods The BOLD protocol has been published previously [8]. It was written in compliance with the Helsinki declaration and is approved by local ethics committees at all sites. All participants provided written consent.

The BOLD study is a large international study providing population-based estimates of the prevalence and burden of CAO. One of the primary objectives of the BOLD study is to estimate disease burden in terms of activity limitation and economic impact [8]. In the current analysis, we have compared estimates of employment status in BOLD participants with and without CAO across the world. Methods The BOLD protocol has been published previously [8]. It was written in compliance with the Helsinki declaration and is approved by local ethics committees at all sites. All participants provided written consent. Population All participating sites were recruited from well-defined administrative areas with the goal of providing representative samples of the local population of ≥600 non-institutionalised persons aged ≥40 years. The current report includes participants from 26 sites (online supplementary material). Out of 22 118 participants providing interview data, 18 710 performed acceptable post-bronchodilator spirometry and were included in the descriptive part of the current analysis. However, when analysing risk for unemployment as outcome, all subjects aged ≥65 years (defined here as retirees) and homemakers/caregivers were excluded. After excluding these subjects, there were no CAO cases left in Tirana (Albania), so this centre is not part of the analyses assessing the effect of CAO on unemployment. Online supplementary table S1 lists sampling strategy and response rates for all sites.

ned here as retirees) and homemakers/caregivers were excluded. After excluding these subjects, there were no CAO cases left in Tirana (Albania), so this centre is not part of the analyses assessing the effect of CAO on unemployment. Online supplementary table S1 lists sampling strategy and response rates for all sites. Data collection The BOLD study is a cross-sectional study based on a structured, face-to-face interview using standardised questionnaires and pre-/post-bronchodilator spirometry. All study coworkers were trained and certified by BOLD coordinating centres. The interviews gathered information on smoking habits, education, job status, self-reported comorbidities (hypertension, heart disease, diabetes, stroke and lung cancer) and respiratory symptoms (dyspnoea, wheezing and chronic bronchitis). Participants indicated whether they had worked for income at any time in the preceding year or if they served as full-time homemakers/caregivers during that time frame. Since retirement was not formally captured under occupation, we excluded anyone aged ≥65 years from analyses involving employment status. All other individuals not being categorised as working, homemakers/caregivers or retirees were defined as unemployed. The main outcome for the current study was a dichotomous employment status where retirees (≥65 years) and homemakers/caregivers were excluded.

e aged ≥65 years from analyses involving employment status. All other individuals not being categorised as working, homemakers/caregivers or retirees were defined as unemployed. The main outcome for the current study was a dichotomous employment status where retirees (≥65 years) and homemakers/caregivers were excluded. Never-smokers were individuals who had smoked <20 packs of cigarettes during their lifetime, or less than one cigarette daily for a year. Ex-smokers were those who reported an age at which they had stopped smoking. Education was categorised according to the highest level of completed schooling and divided into no schooling, primary school, middle school, high school, some college and completed college/university. Dyspnoea was defined using the modified Medical Research Council (mMRC) questions (grades 0–4, see online supplementary material for details) [9]. Subjects reporting being unable to walk for reasons other than breathing problems were excluded from the dyspnoea variable. Wheezing was defined as attacks of wheezing associated with dyspnoea in the past 12 months. Chronic bronchitis was defined as productive cough on most days in 3 months per year for at least two consecutive years. Post-bronchodilator spirometry was performed using a hand-held spirometer (EasyOne; ndd Medizintechnik, Zürich, Switzerland) according to American Thoracic Society standards [10], before and ≥15 min after inhalation of 200 μg salbutamol through a large-volume spacer. For quality control, all individual manoeuvers were reviewed by a pulmonary function reading centre.

ng a hand-held spirometer (EasyOne; ndd Medizintechnik, Zürich, Switzerland) according to American Thoracic Society standards [10], before and ≥15 min after inhalation of 200 μg salbutamol through a large-volume spacer. For quality control, all individual manoeuvers were reviewed by a pulmonary function reading centre. Predicted values of forced expiratory volume in 1 s (FEV1), forced vital capacity (FVC) and FEV1/FVC ratio were estimated from equations for caucasians from the third National Health and Nutrition Examination Survey (NHANES-III) [11]. Spirometric CAO was defined as post-bronchodilator FEV1/FVC below lower limit of normal (LLN). Analysis The sample size of the BOLD study was set to be able to provide robust CAO prevalence estimates at the individual sites [8]. No power calculations were performed a priori for employment status.

Predicted values of forced expiratory volume in 1 s (FEV1), forced vital capacity (FVC) and FEV1/FVC ratio were estimated from equations for caucasians from the third National Health and Nutrition Examination Survey (NHANES-III) [11]. Spirometric CAO was defined as post-bronchodilator FEV1/FVC below lower limit of normal (LLN). Analysis The sample size of the BOLD study was set to be able to provide robust CAO prevalence estimates at the individual sites [8]. No power calculations were performed a priori for employment status. For unadjusted comparisons of individuals with and without CAO, we used Pearson Chi-squared (categorical variables) and t-tests (continuous variables). To illustrate differences in unemployment and CAO in different parts of the world, we stratified descriptive analyses by high-income sites (Sydney (Australia), Salzburg (Austria), Vancouver (Canada), London (UK), Tartu (Estonia), Hannover (Germany), Reykjavik (Iceland), Maastricht (the Netherlands), Bergen (Norway), Krakow (Poland), Lisbon (Portugal), Uppsala (Sweden), Lexington (KY, USA)) and low- to middle-income sites (Guangzhou (China), Mumbai (India), Pune (India), Manila (Philippines), Nampicuan Talugtug (Philippines), Annaba (Algeria), Cape Town (South Africa), Adana (Turkey), Kashmir (India), Sousse (Tunisia), Ile-Ife (Nigeria) and Fes (Morocco)). Income categories were based on the gross national income per capita (GNIPC) of the country, with the cut point between low-to-middle income and high income being GNIPC 10 000 US$. [12]. We also calculated a risk ratio for unemployment associated with CAO as the prevalence of unemployment in CAO subjects divided by the prevalence of unemployment in non-CAO subjects, using a log-binomial generalised linear model (in Stata (StataCorp, College Station, TX, USA) specified as glm with fam(bin) and link(log)). A risk ratio >1 indicates higher risk of unemployment associated with CAO, while a ratio <1 indicates lower risk of unemployment associated with CAO. To illustrate sex differences in CAO status across sites, we tabulated study sites and CAO status, stratified by sex.

USA) specified as glm with fam(bin) and link(log)). A risk ratio >1 indicates higher risk of unemployment associated with CAO, while a ratio <1 indicates lower risk of unemployment associated with CAO. To illustrate sex differences in CAO status across sites, we tabulated study sites and CAO status, stratified by sex. Multivariable analyses for the pooled dataset were conducted using a multilevel mixed-effects generalised linear model (online supplementary material). An alternative approach would be a fixed-effect model. The difference to the chosen mixed-effects approach would be that the latter treats the sites as a random sample of all possible sites, whereas the former would tend to focus more exclusively on the sites that were included in the study. The BOLD sites are in some sense a random sample of broader sites to which we wish to make an inference.

chosen mixed-effects approach would be that the latter treats the sites as a random sample of all possible sites, whereas the former would tend to focus more exclusively on the sites that were included in the study. The BOLD sites are in some sense a random sample of broader sites to which we wish to make an inference. The main predictor variable was spirometric CAO. We fitted five mixed models, all adjusting for site as a cluster level variable. We first identified the total effect of CAO on unemployment in a model with no additional covariates included (model 1). Model 2 added demographic variables (age, sex and education) and smoking habits. Model 2 was extended into model 3 by adding comorbidities. Model 4 included FVC in addition to model 3 covariates. As our multivariable analyses include height, age and sex, which are the main components when using % predicted values, we thus chose to analyse lung function in terms of absolute values. In addition, a recent publication from the European Community Respiratory Health Survey III study has indicated that FVC in absolute values (lung size) is able to explain most of the difference in symptom burden between males and females [13]. FVC is a robust indicator of lung disease, especially when obstruction is already taken into account. In addition, we included model 5 with respiratory symptoms in addition to the model 4 covariates (online supplementary material). Details of comorbidities and symptoms are presented in the online supplementary material. Covariates added in each model were added not as independent risk factors for unemployment, but as potential confounding or mediating factors influencing the effect of CAO on unemployment. In addition, models 2–5 were performed separately for high-income sites and for low- to middle-income sites.

upplementary material. Covariates added in each model were added not as independent risk factors for unemployment, but as potential confounding or mediating factors influencing the effect of CAO on unemployment. In addition, models 2–5 were performed separately for high-income sites and for low- to middle-income sites. In individual participant data meta-analyses, we estimated site-specific and overall odds ratios for CAO on unemployment in forest plots, with increasing adjustment corresponding to models 1–5 (except for the site adjustment). The Stata command used was ipdmetan which performs a two-stage individual participant data meta-analysis using the inverse-variance method. Unlike traditional meta-analysis, the individual participant data meta-analysis in ipdmetan fits a specified model to the data of one site at a time, making use of all individual participants within the sites. The two-stage approach derives aggregate data in each site separately and then combines these in a traditional meta-analysis model. The I2 statistic was reported to display the percentage of total variation across sites which was due to true site-by-site heterogeneity rather than what would be expected by chance alone (see the online supplementary material for more details) [14]. All analyses were performed using Stata SE version 14 for Macintosh OSX.

In individual participant data meta-analyses, we estimated site-specific and overall odds ratios for CAO on unemployment in forest plots, with increasing adjustment corresponding to models 1–5 (except for the site adjustment). The Stata command used was ipdmetan which performs a two-stage individual participant data meta-analysis using the inverse-variance method. Unlike traditional meta-analysis, the individual participant data meta-analysis in ipdmetan fits a specified model to the data of one site at a time, making use of all individual participants within the sites. The two-stage approach derives aggregate data in each site separately and then combines these in a traditional meta-analysis model. The I2 statistic was reported to display the percentage of total variation across sites which was due to true site-by-site heterogeneity rather than what would be expected by chance alone (see the online supplementary material for more details) [14]. All analyses were performed using Stata SE version 14 for Macintosh OSX. Results Out of 18 710 participants, 2123 (11.3%) had CAO. Compared to subjects without CAO, those with CAO were older, had lower education levels, more smoking exposure, more comorbidities, more respiratory symptoms and substantially lower FEV1 (table 1). Overall, CAO was more common in males than in females. However, site-specific prevalence estimates stratified by sex showed that for some centres the sex ratio was reversed (online supplementary table S2). Excluding those aged ≥65 years, 36.7% of participants with CAO reported paid work during the preceding year, whereas 53.2% of participants without CAO had undertaken paid work during the preceding year.

estimates stratified by sex showed that for some centres the sex ratio was reversed (online supplementary table S2). Excluding those aged ≥65 years, 36.7% of participants with CAO reported paid work during the preceding year, whereas 53.2% of participants without CAO had undertaken paid work during the preceding year. TABLE 1 Study participant characteristics in the Burden of Obstructive Lung Disease (BOLD) study by chronic airflow obstruction (CAO)

estimates stratified by sex showed that for some centres the sex ratio was reversed (online supplementary table S2). Excluding those aged ≥65 years, 36.7% of participants with CAO reported paid work during the preceding year, whereas 53.2% of participants without CAO had undertaken paid work during the preceding year. TABLE 1 Study participant characteristics in the Burden of Obstructive Lung Disease (BOLD) study by chronic airflow obstruction (CAO) Spirometric CAO No spirometric CAO Total Subjects 2123 16 587 18 710 Female 46.4 (44.2–48.5) 51.9 (51.1–52.6) 51.3 (50.5–52.0) Age years 60.7±11.9 55.2±11.0 55.8±11.3 Smoking Never-smoker 33.9 (32.0–36.0) 57.2 (56.4–58.0) 54.6 (53.8–55.3) Ex-smoker 31.0 (29.1–33.0) 23.9 (23.2–24.5) 24.7 (24.1–25.3) Current smoker 35.1 (33.1–37.1) 18.9 (18.4–19.5) 20.8 (20.2–21.4) Education None 14.7 (13. 2–16.3) 12.1 (11.6–12.6) 12.4 (11.9–12.9) Primary school 21.7 (20.0–23.5) 15.7 (15.2–16.3) 16.4 (15.9–16.9) Middle school 17.0 (15.5–18.7) 16.0 (15.5–16.6) 16.1 (15.6–16.7) High school 24.7 (22.9–26.6) 26.2 (25.5–26.8) 26.0 (25.4–26.6) Some college 11.1 (9.8–12.5) 12.8 (12.3–13.4) 12.6 (12.2–13.1) College/university 10.9 (9.6–12.3) 17.2 (16.6–17.8) 16.5 (15.9–17.0) Job status Paid work 36.7 (34.7–38.8) 53.2 (52.4–53.9) 51.3 (50.6–52.0) Homemaker/caregiver 14.8 (13.3–16.4) 13.5 (13.0–14.0) 13.7 (13.2–14.1) Unemployed 19.6 (18.0–21.4) 16.2 (15.6–16.8) 16.6 (16.1–17.1) Above retirement age 28.9 (27.0–30.8) 17.1 (16.6–17.7) 18.5 (17.9–19.0) Lung function FVC % pred 89.2±21.8 90.3±16.1 90.2±16.9 FEV1 % pred 69.2±21.4 92.0±16.7 89.4±18.7 Self-reported doctor's diagnosis COPD 15.3 (13.8–16.9) 2.4 (2.2–2.6) 3.9 (3.6–4.2) Hypertension 32.9 (30.9–34.9) 26.2 (25.6–26.9) 27.0 (26.3–27.6) Heart disease 14.3 (12.9–15.9) 10.0 (9.5–10.4) 10.5 (10.0–10.9) Diabetes 7.2 (6.1–8.3) 7.5 (7.1–7.9) 7.5 (7.1–7.9) Stroke 3.1 (2.4–3.9) 1.9 (1.7–2.1) 2.0 (1.8–2.2) Lung cancer 0.7 (0.4– 1.1) 0.2 (0.1– 0.3) 0.3 (0.2– 0.3) Dyspnoea mMRC 0 55.8 (53.5–58.0) 78.8 (78.1–79.4) 76.2 (75.6–76.9) mMRC 1 17.1 (15.5–18.9) 12.1 (11.6–12.6) 12.7 (12.2–13.2) mMRC 2 13.2 (11.7–14.8) 5.6 (5.2–5.9) 6.4 (6.1–6.8) mMRC 3 8.5 (7.3–9.8) 2.7 (2.5–3.0) 3.3 (3.1–3.6) mMRC 4 5.5 (4.5–6.6) 0.9 (0.7–1.0) 1.4 (1.2–1.6) Attack of wheezing in past 12 months 22.3 (20.6–24.1) 6.3 (5.9–6.6) 8.1 (7.7–8.5) Chronic bronchitis 15.4 (13.9–17.0) 5.1 (4.8–5.5) 6.3 (5.9–6.6) Data are presented as n, % (95% CI) or mean±sd. n=18 710 subjects from 26 study sites. All comparisons between CAO and non-CAO were significant (p<0.01, Pearson Chi-squared test for categorical variables, t-test for continuous variables) except for self-reported diabetes.

is 15.4 (13.9–17.0) 5.1 (4.8–5.5) 6.3 (5.9–6.6) Data are presented as n, % (95% CI) or mean±sd. n=18 710 subjects from 26 study sites. All comparisons between CAO and non-CAO were significant (p<0.01, Pearson Chi-squared test for categorical variables, t-test for continuous variables) except for self-reported diabetes. Missing data: smoking habits n=11 (n=1 CAO, n=10 non-CAO); education n=25 (n=4 CAO, n=21 non-CAO); hypertension, diabetes, stroke, lung cancer and heart disease n=1 (n=1 CAO); dyspnoea n=1834 (n=258 CAO, n=1576 non-CAO), mostly because of other reasons for having trouble walking; wheezing n=11 (n=2 CAO, n=9 non-CAO); chronic bronchitis n=0. FVC: forced vital capacity; FEV1: forced expiratory volume in 1 s; COPD: chronic obstructive pulmonary disease; mMRC: modified Medical Research Council scale. Figure 1 shows that more males than females reported current paid employment in both high- and low- to middle-income countries, but the difference was larger in low- to middle-income countries. This appeared to be explained by a substantial proportion of female unpaid homemakers/caregivers in low- to middle-income countries. More details on sex differences in employment status are given in online supplementary table S3. FIGURE 1 Distribution of job status by sex for participants in a) high-, and b) low- to middle-income countries. n=18 710.

Figure 1 shows that more males than females reported current paid employment in both high- and low- to middle-income countries, but the difference was larger in low- to middle-income countries. This appeared to be explained by a substantial proportion of female unpaid homemakers/caregivers in low- to middle-income countries. More details on sex differences in employment status are given in online supplementary table S3. FIGURE 1 Distribution of job status by sex for participants in a) high-, and b) low- to middle-income countries. n=18 710. Table 2 shows unemployment by CAO status in each study site, excluding homemakers, caregivers and retirees (subjects aged ≥65 years). Despite a wide variation in unemployment rates by site, there was a fairly consistent pattern of higher unemployment among individuals with CAO in high-income sites. This pattern was less clear in the low- to middle-income sites. TABLE 2 Unemployment rates: prevalence of unemployment by site and spirometric chronic airflow obstruction (CAO) status

Table 2 shows unemployment by CAO status in each study site, excluding homemakers, caregivers and retirees (subjects aged ≥65 years). Despite a wide variation in unemployment rates by site, there was a fairly consistent pattern of higher unemployment among individuals with CAO in high-income sites. This pattern was less clear in the low- to middle-income sites. TABLE 2 Unemployment rates: prevalence of unemployment by site and spirometric chronic airflow obstruction (CAO) status Subjects# n Unemployment % Crude OR (95% CI)¶ CAO No CAO Total 11 675 High-income Bergen, Norway 397 20.0 9.5 2.1 (1.0–4.2) Hannover, Germany 361 25.0 20.8 1.2 (0.6–2.5) Krakow, Poland 350 57.9 41.4 1.4 (1.0–1.9) Lexington, USA 305 61.0 27.7 2.2 (1.6–3.0) Lisbon, Portugal 320 53.9 39.8 1.4 (0.9–2.0) London, UK 427 40.4 24.3 1.7 (1.1–2.4) Maastricht, the Netherlands 396 31.3 20.4 1.5 (1.0–2.3) Reykjavik, Iceland 557 14.0 3.3 4.2 (1.8–10.1) Salzburg, Austria 860 35.2 25.4 1.4 (1.1–1.8) Sydney, Australia 339 20.0 15.3 1.3 (0.6–3.0) Tartu, Estonia 348 20.0 7.8 2.6 (0.9–7.5) Uppsala, Sweden 371 23.8 6.0 4.0 (1.7–9.5) Vancouver, Canada 594 21.8 11.5 1.9 (1.1–3.3) Low- to middle-income Adana, Turkey 487 41.1 45.4 0.9 (0.7–1.2) Annaba, Algeria 408 50.0 24.6 2.0 (1.2–3.3) Cape Town, South Africa 510 52.2 33.5 1.6 (1.2–2.0) Fes, Morocco 335 41.7 53.7 0.8 (0.5–1.3) Guangzhou, China 359 35.7 49.9 0.7 (0.4–1.5) Ile-Ife, Nigeria 667 5.1 7.6 0.7 (0.2–2.7) Kashmir, India 366 7.6 1.3 5.8 (1.5–22.9) Manila, Philippines 594 10.3 19.5 0.5 (0.2–1.4) Mumbai, India 250 17.7 10.3 1.7 (0.6–5.1) Nampicuan Talugtug, Philippines 493 23.2 14.7 1.6 (0.9–2.7) Pune, India 671 6.5 4.1 1.6 (0.4–6.4) Sousse, Tunisia 390 53.3 46.1 1.2 (0.7–1.9) Tirana, Albania 520 0.0 5.0 #: retirees (age limit defined as ≥65 years) and homemakers/caregivers were excluded from the analysis. ¶: calculated based on prevalence of unemployment among subjects with CAO divided by prevalence of unemployment among subjects without CAO. A ratio >1 indicates higher unemployment prevalence among CAO subjects than among non-CAO subjects, while a ratio <1 indicates lower unemployment prevalence among CAO subjects.

is. ¶: calculated based on prevalence of unemployment among subjects with CAO divided by prevalence of unemployment among subjects without CAO. A ratio >1 indicates higher unemployment prevalence among CAO subjects than among non-CAO subjects, while a ratio <1 indicates lower unemployment prevalence among CAO subjects. In multivariable analyses, we assessed the odds ratio of being unemployed by CAO status and an increasing number of covariates (table 3). The first model showed that when we adjusted for site, the odds ratio (95% CI) of being unemployed was 1.79 (1.41–2.27) for participants with CAO. Adding the traditional confounders sex, age, smoking habits and education in model 2 decreased the odds ratio for unemployment in participants with CAO (OR reduction from 1.79 to 1.44), but the effect remained statistically significant. Further addition of comorbidities (model 3) and FVC (model 4) had little effect on the association between unemployment and CAO, even when these variables themselves were significantly associated with unemployment: the presence of comorbidities and declining FVC were all associated with increased odds of being unemployed. Table 3 shows that excess unemployment among those with CAO is partially explained by sex, age, smoking and education, but not explained additionally by comorbidities and FVC. When respiratory symptoms were added (online supplementary table S4), these were also significantly associated with unemployment and appeared to explain some of the effects of CAO. In this model, the odds ratio for CAO independent of reported symptoms fell to 1.26 (95% CI 1.00–1.57). Substituting self-reported COPD for LLN-defined CAO in our analyses increased the odds ratio of not being in paid work from 1.43 (95% CI 1.14–1.79) to 3.31 (95% CI 2.17–5.05) (additional analysis, data not shown). However, while the prevalence of spirometry-defined CAO was 11.3% in BOLD, the prevalence of self-reported COPD was only 1.2%, and while 36.7% of the spirometry-defined participants with CAO were in paid employment, the corresponding figure for the self-reported COPD cases was only 25% (results not shown).

not shown). However, while the prevalence of spirometry-defined CAO was 11.3% in BOLD, the prevalence of self-reported COPD was only 1.2%, and while 36.7% of the spirometry-defined participants with CAO were in paid employment, the corresponding figure for the self-reported COPD cases was only 25% (results not shown). TABLE 3 OR (95% CI) for unemployment for lower limit of normal-defined chronic airflow obstruction (CAO) and other risk factors, with an increasing degree of adjustment (demographic characteristics, comorbidities and forced vital capacity (FVC))

not shown). However, while the prevalence of spirometry-defined CAO was 11.3% in BOLD, the prevalence of self-reported COPD was only 1.2%, and while 36.7% of the spirometry-defined participants with CAO were in paid employment, the corresponding figure for the self-reported COPD cases was only 25% (results not shown). TABLE 3 OR (95% CI) for unemployment for lower limit of normal-defined chronic airflow obstruction (CAO) and other risk factors, with an increasing degree of adjustment (demographic characteristics, comorbidities and forced vital capacity (FVC)) Model 1 Model 2 Model 3 Model 4 Spirometric CAO 1.79 (1.41–2.27) 1.44 (1.15–1.81) 1.45 (1.15–1.82) 1.43 (1.14–1.79) FVC 10 percentage points decrease in % pred 1.08 (1.04–1.12) Female 2.07 (1.85–2.32) 2.10 (1.87–2.36) 2.10 (1.87–2.35) Age 10-year increment 3.09 (2.85–3.35) 2.91 (2.68–3.15) 2.90 (2.67–3.15) Smoking status Current smoker 0.96 (0.83–1.10) 0.98 (0.85–1.13) 0.98 (0.85–1.13) Ex-smoker 1.15 (1.01–1.32) 1.13 (0.99–1.30) 1.14 (0.99–1.30) Education Some college 1.51 (1.23–1.85) 1.49 (1.22–1.84) 1.49 (1.21–1.83) High school 2.03 (1.71–2.42) 2.02 (1.69–2.41) 2.01 (1.68–2.39) Middle school 2.24 (1.83–2.73) 2.20 (1.80–2.69) 2.18 (1.79–2.67) Primary school 2.78 (2.27–3.41) 2.76 (2.25–3.39) 2.72 (2.22–3.35) No education 2.73 (2.09–3.57) 2.69 (2.05–3.51) 2.66 (2.03–3.49) Comorbidities Hypertension 1.29 (1.13–1.46) 1.26 (1.10–1.43) Heart disease 1.54 (1.27–1.86) 1.51 (1.25–1.83) Diabetes 1.51 (1.23–1.85) 1.47 (1.19–1.80) Stroke 1.82 (1.16–2.86) 1.80 (1.15–2.83) Lung cancer 2.34 (0.81–6.76) 2.38 (0.82–6.93) Adjustment variables: no fixed effects (model 1); age, sex, education and smoking (model 2); model 2 adjustment + comorbidities (model 3); model 3 adjustment + FVC (model 4). All five models were fit using a multilevel mixed-effects generalised linear model with study site included as random effect to account for within-site clustering. Reference values for categorical variables: no CAO, male, never-smoker, university education, no hypertension, no heart disease, no diabetes, no stroke and no lung cancer. n=11 675. Retirees (age limit defined as ≥65 years) and homemakers/caregivers were excluded from the analysis.

o account for within-site clustering. Reference values for categorical variables: no CAO, male, never-smoker, university education, no hypertension, no heart disease, no diabetes, no stroke and no lung cancer. n=11 675. Retirees (age limit defined as ≥65 years) and homemakers/caregivers were excluded from the analysis. To examine how the observed associations varied by country income, we performed multivariable analyses separately for high-income and low- to middle-income sites (table 4). CAO was a significant risk factor for unemployment in all models in high-income sites, but not in low- to middle-income sites. While age and lower education level were important risk factors for unemployment in high-income sites, female sex was the most pronounced risk factor for unemployment in low- to middle-income sites. To further depict the sex variation in job status, we created online supplementary table S3, which shows the prevalence of job status categories among males and females in each site. This table illustrates that almost no sites had more females than males in paid work (with Lexington, Lisbon and Ile-Ife as the only three exceptions). Further on, focusing on the low- to middle-income sites, this table demonstrates that the difference in “unemployed” job status between the sexes were very high in some sites, with the mean difference being 46.1% more unemployed females than males. There were some sites that had more unemployed males than females, but these were few (Annaba, Cape Town, Kashmir, Mumbai and Pune), and the mean difference was low (5.5%). In model 5, dyspnoea was an additional important risk factor for unemployment in high-income sites, together with age and education (online supplementary table S4).

ites that had more unemployed males than females, but these were few (Annaba, Cape Town, Kashmir, Mumbai and Pune), and the mean difference was low (5.5%). In model 5, dyspnoea was an additional important risk factor for unemployment in high-income sites, together with age and education (online supplementary table S4). TABLE 4 OR (95% CI) for unemployment for chronic airflow obstruction (CAO) and other risk factors, stratified by country income category, with increasing degree of adjustment (demographic characteristics, comorbidities and forced vital capacity (FVC))

ites that had more unemployed males than females, but these were few (Annaba, Cape Town, Kashmir, Mumbai and Pune), and the mean difference was low (5.5%). In model 5, dyspnoea was an additional important risk factor for unemployment in high-income sites, together with age and education (online supplementary table S4). TABLE 4 OR (95% CI) for unemployment for chronic airflow obstruction (CAO) and other risk factors, stratified by country income category, with increasing degree of adjustment (demographic characteristics, comorbidities and forced vital capacity (FVC)) Model 2 Model 3 Model 4 High income Low to middle income High income Low to middle income High income Low to middle income Spirometric CAO 1.71 (1.17–2.49) 1.16 (0.78–1.73) 1.63 (1.16–2.28) 1.18 (0.79–1.76) 1.68 (1.16–2.45) 1.15 (0.77–1.71) FVC 10 percentage points decrease in % pred 1.09 (1.03–1.15) 1.08 (1.02–1.14) Female 1.36 (1.16–1.59) 3.34 (2.76–4.04) 1.43 (1.23–1.68) 3.25 (2.68–3.94) 1.44 (1.23–1.68) 3.23 (2.66–3.91) Age 10-year increment 4.28 (3.77–4.86) 2.31 (2.06–2.59) 4.04 (3.55–4.59) 2.21 (1.96–2.48) 4.02 (3.53–4.57) 2.20 (1.96–2.47) Smoking status Current smoker 1.34 (1.09–1.65) 0.88 (0.72–1.09) 1.36 (1.11–1.68) 0.89 (0.72–1.10) 1.36 (1.10–1.67) 0.89 (0.72–1.10) Ex-smoker 1.31 (1.09–1.56) 1.15 (0.90–1.47) 1.28 (1.07–1.53) 1.13 (0.88–1.44) 1.29 (1.08–1.54) 1.12 (0.88–1.43) Education Some college 1.85 (1.45–2.38) 0.96 (0.60–1.52) 1.83 (1.43–2.36) 0.95 (0.60–1.52) 1.83 (1.42–2.35) 0.97 (0.61–1.54) High school 2.30 (1.84–2.88) 1.26 (0.92–1.73) 2.27 (1.81–2.84) 1.28 (0.93–1.76) 2.24 (1.79–2.81) 1.30 (0.94–1.78) Middle school 3.65 (2.73–4.87) 1.23 (0.89–1.70) 3.54 (2.64–4.74) 1.26 (0.91–1.74) 3.49 (2.60–4.67) 1.26 (0.91–1.75) Primary school 4.14 (3.02–5.66) 1.52 (1.10–2.10) 3.90 (2.84–5.37) 1.56 (1.13–2.16) 3.86 (2.80–5.30) 1.56 (1.13–2.16) No education 1.96 (0.61–6.32) 1.61 (1.13–2.30) 2.01 (0.62–6.43) 1.64 (1.15–2.34) 1.98 (0.61–6.43) 1.65 (1.15–2.35) Comorbidities Hypertension 1.25 (1.05–1.49) 1.25 (1.02–1.53) 1.22 (1.02–1.45) 1.22 (0.99–1.50) Heart disease 1.56 (1.22–2.00) 1.20 (0.86–1.67) 1.53 (1.19–1.96) 1.18 (0.85–1.64) Diabetes 1.53 (1.14–2.04) 1.39 (1.01–1.92) 1.46 (1.09–1.95) 1.38 (1.00–1.91) Stroke 2.17 (1.14–4.13) 1.58 (0.83–3.02) 2.15 (1.13–4.10) 1.56 (0.82–2.99) Lung cancer 2.87 (0.89–9.30) 1.01 (0.04–28.54) 2.86 (0.89–9.26) 1.03 (0.04–28.37) Adjustment variables: no fixed effects (model 1); age, sex, education and smoking (model 2); model 2 adjustment + comorbidities (model 3); model 3 adjustment + FVC (model 4). All five models were fit using a multilevel mixed-effects generalised linear model with study site included as random effect to account for within-site clustering.

es: no fixed effects (model 1); age, sex, education and smoking (model 2); model 2 adjustment + comorbidities (model 3); model 3 adjustment + FVC (model 4). All five models were fit using a multilevel mixed-effects generalised linear model with study site included as random effect to account for within-site clustering. Separate analyses were performed for high-income countries and low- to middle-income countries. Reference values for categorical variables: no CAO, male, never-smoker, university education, no hypertension, no heart disease, no diabetes, no stroke and no lung cancer. n=11 675. Retirees (age limit defined as ≥65 years) and homemakers/caregivers were excluded from the analysis.

and low- to middle-income countries. Reference values for categorical variables: no CAO, male, never-smoker, university education, no hypertension, no heart disease, no diabetes, no stroke and no lung cancer. n=11 675. Retirees (age limit defined as ≥65 years) and homemakers/caregivers were excluded from the analysis. To present the association between CAO and unemployment by site, and to examine site heterogeneity, we performed individual participant data meta-analyses with forest plots of odds ratios and overall I2 statistics (figure 2 and online supplementary figures S1–S4). The overall odds ratio (95% CI) for unemployment among CAO subjects after adjusting for sex, age, smoking, education, comorbidities and FVC (i.e. the equivalent of model 4, but without site adjustment) was 1.41 (1.18–1.69) with site-by-site heterogeneity (I2) of 12.9% (p=0.279). Meta-analyses with covariates corresponding to models 1, 2, 3 and 5 are shown in online supplementary figures S1–S4, and show that there is no significant site heterogeneity in the association between airflow obstruction and unemployment when adjusting for the covariates in models 2, 3 and 5. However, in crude analysis (model 1), there is significant site heterogeneity (I2 49.1%, p=0.003).

shown in online supplementary figures S1–S4, and show that there is no significant site heterogeneity in the association between airflow obstruction and unemployment when adjusting for the covariates in models 2, 3 and 5. However, in crude analysis (model 1), there is significant site heterogeneity (I2 49.1%, p=0.003). FIGURE 2 Odds ratios (95% CI) for unemployment for lower limit of normal-defined chronic airflow obstruction, adjusted for demographic characteristics, comorbidities and forced vital capacity (FVC). Adjustment variables: sex, age, smoking, education, hypertension, heart disease, diabetes, stroke, lung cancer and FVC. n=11 675, meta-analysis with results across sites and overall. Retirees (age limit defined as 65 years) and homemakers/caregivers excluded. BOLD: Burden of Obstructive Lung Disease study.

acity (FVC). Adjustment variables: sex, age, smoking, education, hypertension, heart disease, diabetes, stroke, lung cancer and FVC. n=11 675, meta-analysis with results across sites and overall. Retirees (age limit defined as 65 years) and homemakers/caregivers excluded. BOLD: Burden of Obstructive Lung Disease study. Discussion The unweighted prevalence of spirometry-defined CAO was 11.3% in this sample of almost 19 000 participants from the global BOLD study. The association between CAO and unemployment varied across sites in crude analyses, but the site heterogeneity lost significance after adjustment for relevant covariates: CAO was an overall important risk factor for unemployment after adjusting for sex, age, smoking, education, comorbidities and even FVC. When looking at high-income and low- to middle-income sites separately, this association was only statistically significant in high-income sites. Regarding other covariates, age and education were important risk factors for unemployment in high-income sites, while female sex was important for unemployment in low- to middle-income sites.

gh-income and low- to middle-income sites separately, this association was only statistically significant in high-income sites. Regarding other covariates, age and education were important risk factors for unemployment in high-income sites, while female sex was important for unemployment in low- to middle-income sites. Comparable population-based studies have previously observed similar prevalence rates of COPD as the CAO rates found in the present study. The PLATINO (Latin-American Pulmonary Obstruction Investigation Project) study found the prevalence to be within the range of 7.8–19.7% [15], Hansen et al. [16] found the overall COPD prevalence in a Danish general population to be 12%, and the systematic review by Adeloye et al. [17] found the global prevalence of population-based spirometrically defined COPD to be 11.7%.

tigation Project) study found the prevalence to be within the range of 7.8–19.7% [15], Hansen et al. [16] found the overall COPD prevalence in a Danish general population to be 12%, and the systematic review by Adeloye et al. [17] found the global prevalence of population-based spirometrically defined COPD to be 11.7%. Only one multicentre study has previously provided population-based estimates of unemployment in CAO, identifying CAO using spirometry. In accordance with our findings, the PLATINO study, performed in five Latin-American countries, estimated that 41.8% of participants with CAO and 57.1% of those without CAO had a paid job the preceding year [7]. In the multivariable analysis of the PLATINO study they found a borderline lower probability of paid work (OR 0.83, 95% CI 0.69–1.00) for CAO patients, and, as in our study, they found significant effects of age, sex, education, dyspnoea and comorbidities. However, the PLATINO study researchers adjusted for dyspnoea in their main model, and this has probably reduced the effect of spirometry-defined CAO on the probability of having paid work. We observed the same pattern in our study; while CAO was significantly associated with unemployment in our main model with OR 1.43 (adjusting for sex, age, smoking, education, comorbidities and FVC), the odds ratio decreased to 1.26 (although still remaining significant, with 95% CI 1.00–1.57) after adding reported dyspnoea and other respiratory symptoms. In line with this, we speculate that symptoms and severity of CAO would probably explain the bulk of unemployment, and that it would be better to study these disease aspects than merely spirometry measurements. However, even after adjusting for mMRC, wheezing with dyspnoea and symptoms of chronic bronchitis in our study, the effect of spirometry-defined CAO on unemployment was still significant (model 5; online supplementary material). This suggests that there are properties other than the burden of current wheezing, dyspnoea and bronchitis that lead to unemployment, and adding objectively measured CAO identifies the magnitude of these. For instance, the patient might experience other symptoms (e.g. asthenia), be a frequent exacerbator or there might be some degree of reporting bias.

rties other than the burden of current wheezing, dyspnoea and bronchitis that lead to unemployment, and adding objectively measured CAO identifies the magnitude of these. For instance, the patient might experience other symptoms (e.g. asthenia), be a frequent exacerbator or there might be some degree of reporting bias. Other studies on workforce participation of CAO patients have been based on self-reported COPD diagnosis and not spirometry [18–22]. Studies of self-reported COPD observe stronger associations between the disease and participation in the workplace than the current study. This difference might be due to a bias towards more severely affected patients in studies based on self-reports [23]. Lamprecht et al. [24] showed that >80% of subjects with post-bronchodilator FEV1/FVC <LLN were undiagnosed, and that less severe airflow obstruction was an important predictor lack of diagnosis.

study. This difference might be due to a bias towards more severely affected patients in studies based on self-reports [23]. Lamprecht et al. [24] showed that >80% of subjects with post-bronchodilator FEV1/FVC <LLN were undiagnosed, and that less severe airflow obstruction was an important predictor lack of diagnosis. The inclusion of undiagnosed CAO patients by state-of-the-art spirometric case detection in representative population-based samples is the main strength of the current study. All epidemiological studies are subject to selection bias to some degree, and the use of representative samples and mostly high cooperation rates (over half >70%) reduce the likelihood of strong biases from selection. Furthermore, our main outcome is categorical and objective, and less prone to bias [25, 26] than reports of diagnoses, although some of the covariates may be more prone to recall bias. In addition, we have used post-bronchodilator measurements, in accordance with international guidelines, and we have a large sample size from a general global population with standardised data collection across sites. In addition, we have built regression models based on a priori hypotheses of associations, rather than including all variables that were significant in bivariable analyses or by an automated stepwise approach.

d we have a large sample size from a general global population with standardised data collection across sites. In addition, we have built regression models based on a priori hypotheses of associations, rather than including all variables that were significant in bivariable analyses or by an automated stepwise approach. Some limitations deserve to be mentioned. First of all, the BOLD study is a cross-sectional study, and as such we cannot infer temporality and we have no direct evidence that the CAO was directly responsible for the unemployment. It is not unthinkable that some of the association between CAO and unemployment is a result of unemployed participants being more susceptible to the disease, even if we have adjusted for education, age and smoking habits. Economic hardship in the form of unemployment can worsen individual unhealthy behaviours including smoking [27]. Second, the employment question is based on any paid work in the past year, and does not differentiate between full-time and part-time work. In other words, subjects who have needed to reduce their work participation due to CAO from full-time to part-time will still be defined as in paid work in our analysis. This may lead to an underestimation of associations between CAO and employment. Being able to present absolute rates of disease-related unemployment standardised at the site population level would have been an advantage, but as our data did not include census information with age distribution details from each site this was not feasible. Future research should preferably include such data for this purpose. Furthermore, lack of a direct question on retirement means that we may have underestimated the problem of unemployment above 65 years of age. Our chosen cut-off of 65 years as retirement age may have affected results in both directions. Third, our spirometry-derived variables were calculated from the NHANES III reference equation for caucasians. This is relatively uncontroversial for measures of FEV1/FVC in the age group 40–65 years, as normal values are not strongly associated with ethnicity. However, overall, the prevalence of spirometry-defined CAO (FEV1/FVC <LLN) will be slightly lower with NHANES reference values than with the recently recommended Global Lung Function Initiative reference values [28]. The difference would not be large enough for us to expect substantial differences in the associations observed in the present study.

nce of spirometry-defined CAO (FEV1/FVC <LLN) will be slightly lower with NHANES reference values than with the recently recommended Global Lung Function Initiative reference values [28]. The difference would not be large enough for us to expect substantial differences in the associations observed in the present study. If anything, a higher CAO prevalence would lead to larger effects of CAO on unemployment, including more individuals with less severe obstruction. The use of NHANES may be more controversial for the measures of FVC than for the ratio measures. In this case, we have used FVC as a continuous variable so that the “lower limit of normal” is not an issue, and, as we have allowed a separate baseline in each centre and as most centres are ethnically homogeneous, this should not present a problem [29, 30]. Since the main focus of the present study was on associations rather than prevalences, we chose to implement the same reference values for the whole study population. This may allow for possible factors that might have affected the lung function at a national level to become apparent, instead of being lost with the use of different reference equations at each site. Fourth, regarding study limitations, the registration of never-smokers may have been somewhat exaggerated if there were participants who started smoking recently before study inclusion, but who had not yet reached 20 lifetime packs of cigarettes. However, the risk of this would seem small given that the youngest participants included in the study are aged 40 years. Lastly, there might be a bias toward more females responding as unemployed in low- to middle-income sites due to cultural differences where females might not have formal employment, although they attend work and have an informal income. This information bias might make the sex difference in the risk of being unemployed somewhat higher than the actual risk in these sites, but unfortunately it is beyond the potential of our dataset to disentangle this possible female misclassification. Online supplementary table S3 shows that the differences between males and females applied to almost all sites.

fference in the risk of being unemployed somewhat higher than the actual risk in these sites, but unfortunately it is beyond the potential of our dataset to disentangle this possible female misclassification. Online supplementary table S3 shows that the differences between males and females applied to almost all sites. The association between CAO and unemployment was significant in overall analyses, but in stratified analyses we observed that the association was probably driven by high-income sites. There may be several reasons for this. First, subjects in low- to middle-income countries may have more prevalent diseases than CAO that render them vulnerable to unemployment. Second, there may be more heterogeneity in low- to middle-income sites than in high-income sites. Our analyses showed consistent results across the high-income sites that seemed to be more homogeneous than the low- to middle-income sites, where CAO was a risk factor for unemployment in some sites and almost a protective factor against unemployment in other sites. The suspicion was further strengthened by crude meta-analysis, showing significant site heterogeneity in the univariate association between CAO and unemployment. However, when other covariates were accounted for, the site heterogeneity lost significance. Third, other factors may be more important than health factors for unemployment risk in low- to middle-income countries. We observed that female sex was an important risk factor for unemployment in these sites, while age and education were important for the high-income sites. Traditional male/female roles in low- to middle-income countries may affect work-life participation to such a degree that they blur the association between health-related factors and unemployment. Such large sex differences in work participation were illustrated in online supplementary table S3 in the present study. And last, but not least, our results may be an indication of how disease burden act differently in high-versus low- to middle-income sites, due to a strictly economic component. In high-income sites those most severely affected are given the possibility to be economically sustained by the corresponding social security systems, while in low- to middle-income sites such alternatives are few or nonexistent. While in high-income sites, the welfare system bears the economic burden of disease, in low- to middle-income sites the people affected both bear the personal and the economic burden of disease.

by the corresponding social security systems, while in low- to middle-income sites such alternatives are few or nonexistent. While in high-income sites, the welfare system bears the economic burden of disease, in low- to middle-income sites the people affected both bear the personal and the economic burden of disease. In conclusion, we have found that work-life participation of subjects with CAO is overall lower than work-life participation of subjects without CAO, and that CAO is associated with unemployment after adjusting for sex, age, smoking, education, comorbidities and even FVC. There was no significant heterogeneity between sites, although stratified analyses showed that CAO may be of greater importance for unemployment in high-income sites. Our study shows the risk of unemployment among people with this prevalent respiratory disease, and illustrates how CAO considerably impacts productivity and social security systems worldwide. Supplementary material 10.1183/13993003.00499-2017.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-00499-2017_Supplement Figure S1. Crude odds ratios (OR) with 95% confidence intervals (95% CI) for unemployment for LLN-defined CAO. N #x003D; 11675* subjects, meta-analysis with results across sites and overall. Legend: *Retirees (age limit defined as 65 years old) and homemakers/caregivers excluded. Corresponds to model 1 in Table 4 (but without site adjustment). ERJ-00499-2017_Figure_S1

nce intervals (95% CI) for unemployment for LLN-defined CAO. N #x003D; 11675* subjects, meta-analysis with results across sites and overall. Legend: *Retirees (age limit defined as 65 years old) and homemakers/caregivers excluded. Corresponds to model 1 in Table 4 (but without site adjustment). ERJ-00499-2017_Figure_S1 Figure S2. Odds ratios (OR) with 95% confidence intervals (95% CI) for unemployment for LLN-defined CAO, adjusted for demographic characteristics*. N = 11675** subjects, meta-analysis with results across sites and overall. Legend: *Adjustment variables: gender, age, smoking, education. Corresponds to model 2 in Table 4 (but without site adjustment). **Retirees (age limit defined as 65 years old) and homemakers/caregivers excluded. ERJ-00499-2017_Figure_S2 Figure S3. Odds ratios (OR) with 95% confidence intervals (95% CI) for unemployment for LLN-defined CAO, adjusted for demographic characteristics, and comorbidities*. N #x003D; 11675** subjects, meta-analysis with results across sites and overall. Legend: *Adjustment variables: gender, age, smoking, education, hypertension, heart disease, diabetes, stroke, lung cancer. Corresponds to model 3 in table 4 (but without site adjustments). **Retirees (age limit 65 years old) and homemakers/caregivers excluded. ERJ-00499-2017_Figure_S3

results across sites and overall. Legend: *Adjustment variables: gender, age, smoking, education, hypertension, heart disease, diabetes, stroke, lung cancer. Corresponds to model 3 in table 4 (but without site adjustments). **Retirees (age limit 65 years old) and homemakers/caregivers excluded. ERJ-00499-2017_Figure_S3 Figure S4. Odds ratios (OR) with 95% confidence intervals (95% CI) for unemployment for LLN-defined CAO, adjusted for demographic characteristics, comorbidities, FVC, and respiratory symptoms*. N #x003D; 11675** subjects, meta-analysis with results across sites and overall. Legend: *Adjustment variables: gender, age, smoking, education, hypertension, heart disease, diabetes, stroke, lung cancer, FVC, dyspnea, dyspnoea and wheezing, chronic bronchitis symptoms. Corresponds to model 5 in table S3 (but without site adjustment). **Retirees (age limit 65 years old) and homemakers/caregivers excluded. ERJ-00499-2017_Figure_S4 Disclosures 10.1183/13993003.00499-2017.Supp2P. Burney ERJ-00499-2017_Burney M. Erdal ERJ-00499-2017_Erdal R. Grønseth ERJ-00499-2017_Gronseth W.C. Tan ERJ-00499-2017_Tan W.M. Vollmer ERJ-00499-2017_Vollmer Acknowledgements We would like to thank all study participants and researchers at all sites for contributing to the successful execution of this study.

Disclosures 10.1183/13993003.00499-2017.Supp2P. Burney ERJ-00499-2017_Burney M. Erdal ERJ-00499-2017_Erdal R. Grønseth ERJ-00499-2017_Gronseth W.C. Tan ERJ-00499-2017_Tan W.M. Vollmer ERJ-00499-2017_Vollmer Acknowledgements We would like to thank all study participants and researchers at all sites for contributing to the successful execution of this study. Collaborators: Research teams at centres: NanShan Zhong (principal investigator (PI)), Shengming Liu, Jiachun Lu, Pixin Ran, Dali Wang, Jingping Zheng and Yumin Zhou (Guangzhou Institute of Respiratory Diseases, Guangzhou Medical College, Guangzhou, China); Ali Kocabaş (PI), Attila Hancioglu, Ismail Hanta, Sedat Kuleci, Ahmet Sinan Turkyilmaz, Sema Umut and Turgay Unalan (Cukurova University School of Medicine, Department of Chest Diseases, Adana, Turkey); Michael Studnicka (PI), Torkil Dawes, Bernd Lamprecht and Lea Schirhofer (Paracelsus Medical University, Department of Pulmonary Medicine, Salzburg Austria); Eric Bateman (PI), Anamika Jithoo (PI), Desiree Adams, Edward Barnes, Jasper Freeman, Anton Hayes, Sipho Hlengwa, Christine Johannisen, Mariana Koopman, Innocentia Louw, Ina Ludick, Alta Olckers, Johanna Ryck and Janita Storbeck (University of Cape Town Lung Institute, Cape Town, South Africa); Thorarinn Gislason (PI), Bryndis Benedikdtsdottir, Kristin Jörundsdottir, Lovisa Gudmundsdottir, Sigrun Gudmundsdottir and Gunnar Gundmundsson (Landspitali University Hospital, Dept of Allergy, Respiratory Medicine and Sleep, Reykjavik, Iceland); Ewa Nizankowska-Mogilnicka (PI), Jakub Frey, Rafal Harat, Filip Mejza, Pawel Nastalek, Andrzej Pajak, Wojciech Skucha, Andrzej Szczeklik and Magda Twardowska (Division of Pulmonary Diseases, Department of Medicine, Jagiellonian University School of Medicine, Cracow, Poland); Tobias Welte (PI), Isabelle Bodemann, Henning Geldmacher and Alexandra Schweda-Linow (Hannover Medical School, Hannover, Germany); Amund Gulsvik (PI), Tina Endresen and Lene Svendsen (Department of Thoracic Medicine, Institute of Medicine, University of Bergen, Bergen, Norway); Wan C. Tan (PI) and Wen Wang (iCapture Center for Cardiovascular and Pulmonary Research, University of British Columbia, Vancouver, BC, Canada); David M. Mannino (PI), John Cain, Rebecca Copeland, Dana Hazen and Jennifer Methvin (University of Kentucky, Lexington, KY, USA); Renato B. Dantes (PI), Lourdes Amarillo, Lakan U. Berratio, Lenora C. Fernandez, Norberto A. Francisco, Gerard S. Garcia, Teresita S. de Guia, Luisito F. Idolor, Sullian S. Naval, Thessa Reyes, Camilo C.

ino (PI), John Cain, Rebecca Copeland, Dana Hazen and Jennifer Methvin (University of Kentucky, Lexington, KY, USA); Renato B. Dantes (PI), Lourdes Amarillo, Lakan U. Berratio, Lenora C. Fernandez, Norberto A. Francisco, Gerard S. Garcia, Teresita S. de Guia, Luisito F. Idolor, Sullian S. Naval, Thessa Reyes, Camilo C. Roa Jr, Ma. Flordeliza Sanchez and Leander P. Simpao (Philippine College of Chest Physicians, Manila, Philippines); Christine Jenkins (PI), Guy Marks (PI), Tessa Bird, Paola Espinel, Kate Hardaker and Brett Toelle (Woolcock Institute of Medical Research, Sydney, Australia), Peter G.J. Burney (PI), Caron Amor, James Potts, Michael Tumilty and Fiona McLean (National Heart and Lung Institute, Imperial College, London, UK); E.F.M. Wouters and G.J. Wesseling (Maastricht University Medical Center, Maastricht, the Netherlands); Cristina Bárbara (PI), Fátima Rodrigues, Hermínia Dias, João Cardoso, João Almeida, Maria João Matos, Paula Simão, Moutinho Santos and Reis Ferreira (Portuguese Society of Pneumology, Lisbon, Portugal); Christer Janson (PI), Inga Sif Olafsdottir, Katarina Nisser, Ulrike Spetz-Nyström, Gunilla Hägg and Gun-Marie Lund (Department of Medical Sciences: Respiratory Medicine and Allergology, Uppsala University, Sweden); Rain Jõgi (PI), Hendrik Laja, Katrin Ulst, Vappu Zobel and Toomas-Julius Lill (Lung Clinic, Tartu University Hospital, Tartu, Estonia); Parvaiz A. Koul (PI), Sajjad Malik, Nissar A. Hakim and Umar Hafiz Khan (Sher-i-Kashmir Institute of Medical Sciences, Srinagar, India); Rohini Chowgule (PI), Vasant Shetye, Jonelle Raphael, Rosel Almeda, Mahesh Tawde, Rafiq Tadvi, Sunil Katkar, Milind Kadam, Rupesh Dhanawade and Umesh Ghurup (Indian Institute of Environmental Medicine, Mumbai, India); Imed Harrabi (PI), Myriam Denguezli, Zouhair Tabka, Hager Daldoul, Zaki Boukheroufa, Firas Chouikha and Wahbi Belhaj Khalifa (Faculté de Médecine, Sousse, Tunisia); Luisito F. Idolor (PI), Teresita S. de Guia, Norberto A. Francisco, Camilo C. Roa, Fernando G. Ayuyao, Cecil Z. Tady, Daniel T. Tan, Sylvia Banal-Yang, Vincent M. Balanag Jr, Maria Teresita N. Reyes and Renato. B.

Daldoul, Zaki Boukheroufa, Firas Chouikha and Wahbi Belhaj Khalifa (Faculté de Médecine, Sousse, Tunisia); Luisito F. Idolor (PI), Teresita S. de Guia, Norberto A. Francisco, Camilo C. Roa, Fernando G. Ayuyao, Cecil Z. Tady, Daniel T. Tan, Sylvia Banal-Yang, Vincent M. Balanag Jr, Maria Teresita N. Reyes and Renato. B. Dantes (Lung Centre of the Philippines, Philippine General Hospital, Nampicuan and Talugtug, Philippines); Sanjay Juvekar (PI), Siddhi Hirve, Somnath Sambhudas, Bharat Chaidhary, Meera Tambe, Savita Pingale, Arati Umap, Archana Umap, Nitin Shelar, Sampada Devchakke, Sharda Chaudhary, Suvarna Bondre, Savita Walke, Ashleshsa Gawhane, Anil Sapkal, Rupali Argade and Vijay Gaikwad (Vadu HDSS, KEM Hospital Research Centre Pune, Pune, India); Sundeep Salvi (PI), Bill Brashier, Jyoti Londhe and Sapna Madas (Chest Research Foundation, Pune, India); Mohamed C. Benjelloun (PI), Chakib Nejjari, Mohamed Elbiaze and Karima El Rhazi (Laboratoire d’épidémiologie, Recherche Clinique et Santé Communautaire, Fès, Morroco); Daniel Obaseki (PI), Gregory Erhabor, Olayemi Awopeju and Olufemi Adewole (Obafemi Awolowo University, Ile-Ife, Nigeria); Mohamed Al Ghobain (PI), Hassan Alorainy (PI), Esam El-Hamad, Mohamed Al Hajjaj, Hashi Ayan, Rowena Dela, Rofel Fanuncio, Elizabeth Doloriel, Imelda Marciano and Lyla Safia Thoracic Society, Riyadh, Saudi Arabia); Talant M. Sooronbaev (PI), Bermet M. Estebesova, Meerim Akmatalieva, Saadat Usenbaeva, Jypara Kydyrova, Eliza Bostonova, Ulan Sheraliev, Nuridin Marajapov, Nurgul Toktogulova, Berik Emilov, Toktogul Azilova, Gulnara Beishekeeva, Nasyikat Dononbaeva and Aijamal Tabyshova (Pulmunology and Allergology Department, National Centre of Cardiology and Internal Medicine, Bishkek, Kyrgyzstan); Kevin Mortimer (PI), Wezzie Nyapigoti, Ernest Mwangoka, Mayamiko Kambwili, Martha Chipeta, Gloria Banda, Suzgo Mkandawire and Justice Banda (Malawi Liverpool Wellcome Trust, Blantyre, Malawi); Asma Elsony (PI), Hana A.

nd Allergology Department, National Centre of Cardiology and Internal Medicine, Bishkek, Kyrgyzstan); Kevin Mortimer (PI), Wezzie Nyapigoti, Ernest Mwangoka, Mayamiko Kambwili, Martha Chipeta, Gloria Banda, Suzgo Mkandawire and Justice Banda (Malawi Liverpool Wellcome Trust, Blantyre, Malawi); Asma Elsony (PI), Hana A. Elsadig, Nada Bakery Osman, Bandar Salah Noory, Monjda Awad Mohamed, Hasab Alrasoul Akasha Ahmed Osman, Namarig Moham ed Elhassan, Abdel Muis El‫ Zain, Marwa Mohamed Mohamaden, Suhaiba Khalifa, Mahmoud Elhadi, Mohand Hassan and Dalia Abdelmonam (Epidemiological Laboratory, Khartoum, Sudan); Hasan Hafizi (PI), Anila Aliko, Donika Bardhi, Holta Tafa, Natasha Thanasi, Arian Mezini, Alma Teferici, Dafina Todri, Jolanda Nikolla and Rezarta Kazasi (Tirana University Hospital, Shefqet Ndroqi, Albania); Hamid Hacene Cherkaski (PI), Amira Bengrait, Tabarek Haddad, Ibtissem Zgaoula, Maamar Ghit, Abdelhamid Roubhia, Soumaya Boudra, Feryal Atoui, Randa Yakoubi, Rachid Benali, Abdelghani Bencheikh and Nadia Ait-Khaled (Faculté de Médecine Annaba, SEMEP Elhadjar, Algeria); Akramul Islam (PI), Syed Masud Ahmed (co-PI), Shayla Islam, Qazi Shafayetul Islam, Mesbah-Ul-Haque, Tridib Roy Chowdhury, Sukantha Kumar Chatterjee, Dulal Mia, Shyamal Chandra Das, Mizanur Rahman, Nazrul Islam, Shahaz Uddin, Nurul Islam, Luiza Khatun, Monira Parvin, Abdul Awal Khan and Maidul Islam (James P. Grant School of Public Health, BIGH/BRAC University, Bangladesh); Li-Cher Loh (PI), Abdul Rashid and Siti Sholehah (Penang Medical College, Penang, Malaysia); Herve Lawin (PI), Arsene Kpangon, Karl Kpossou, Gildas Agodokpessi, Paul Ayelo and Benjamin Fayomi (Unit of Teaching and Research in Occupational and Environmental Health, Cotonou, Benin).

th, BIGH/BRAC University, Bangladesh); Li-Cher Loh (PI), Abdul Rashid and Siti Sholehah (Penang Medical College, Penang, Malaysia); Herve Lawin (PI), Arsene Kpangon, Karl Kpossou, Gildas Agodokpessi, Paul Ayelo and Benjamin Fayomi (Unit of Teaching and Research in Occupational and Environmental Health, Cotonou, Benin). Author contributions are as follows. Design, planning and data collection: R. Grønseth, P. Burney and A. Johannessen. Data management and quality control: R. Grønseth, A. Johannessen, M. Erdal and W.M. Vollmer. Statistical analyses: R. Grønseth, A. Johannessen and W.M. Vollmer. Analysis plan: R. Grønseth, A. Johannessen, P. Burney, W.M. Vollmer and M. Erdal. Drafting: R. Grønseth, M. Erdal and A. Johannessen. Revision and approval of drafts: R. Grønseth, M. Erdal, W.C. Tan, D.O. Obaseki, A.F.S. Amaral, T. Gislason, S.K. Juvekar, P.A. Koul, M. Studnicka, S. Salvi, P. Burney, A.S. Buist, W.M. Vollmer and A. Johannessen. All authors contributed toward data analysis, drafting and critically revising the paper and agree to be accountable for all aspects of the work. This article has supplementary material available erj.ersjournals.com Support statement: This study was funded by the Wellcome Trust (grant number 085790/2/08/2). Funding information for this article has been deposited with the Crossref Funder Registry. Conflict of interest: Disclosures can be found alongside this article at erj.ersjournals.com

Introduction Obesity is a well-recognised independent cardiovascular risk factor and is a comorbidity to pulmonary arterial hypertension (PAH) [1]. In the REVEAL (Registry to EValuate Early And Long-term PAH disease management) registry, 32% of patients with PAH were classified as obese at enrolment [2]. This registry reports a higher prevalence of overweight and obese individuals among those with the idiopathic form of PAH [1]. Obesity-related pulmonary hypertension can occur as a result of hypoventilation and hypoxia due to the increased mechanical load of excess body fat [3]. However, the metabolic and inflammatory disturbances that occur during obesity may play a role in the development of PAH. Adipose tissue expresses high levels of the oestrogen-synthesising enzyme aromatase and in obesity can become a major source of oestrogen production [4]. Interestingly, adipose tissue expresses high levels of cytochrome P450 1B1 (CYP1B1), an enzyme involved in oestrogen metabolism [5]. CYP1B1 plays a role in clinical PAH, with CYP1B1 highly expressed in pulmonary artery lesions of PAH patients [6–8]. Furthermore, CYP1B1 single nucleotide polymorphisms have been identified that have significant association with right ventricular ejection fraction and oestrogen metabolism [9]. These are in tight linkage disequilibrium with single nucleotide polymorphisms associated with pulmonary hypertension and oncogenesis [9].

8]. Furthermore, CYP1B1 single nucleotide polymorphisms have been identified that have significant association with right ventricular ejection fraction and oestrogen metabolism [9]. These are in tight linkage disequilibrium with single nucleotide polymorphisms associated with pulmonary hypertension and oncogenesis [9]. The oestrogen metabolite 16α-hydroxyestrone (16αOHE1) is formed via CYP1B1 and has been implicated in PAH. 16αOHE1 is a potent mitogen in pulmonary artery smooth muscle cells (PASMCs), acting in a reactive oxygen species (ROS)-dependent manner with little effect on smooth muscle cells from the systemic circulation [10]. Furthermore, when administered to mice, 16αOHE1 results in the development of a pulmonary hypertension phenotype [8]. Growing evidence suggests that endogenous oestrogen and its metabolites play a role in the development of PAH. High oestradiol (E2) and low dehydroepiandrosterone-sulfate (DHEA-S) levels have been identified as risk factors for PAH in males. More recently, high E2 and low DHEA-S levels have also been associated with the risk and severity of PAH in post-menopausal females [11]. It has been shown clinically that E2 is associated with PAH and correlates inversely with 6-min walk distance in males and post-menopausal females [12]. Additionally, anastrozole (ANA) improved the 6-min walk distance in a small-scale clinical trial of post-menopausal female and male PAH patients [13].

al females [11]. It has been shown clinically that E2 is associated with PAH and correlates inversely with 6-min walk distance in males and post-menopausal females [12]. Additionally, anastrozole (ANA) improved the 6-min walk distance in a small-scale clinical trial of post-menopausal female and male PAH patients [13]. Our hypothesis was that obesity may induce changes in oestrogen metabolism, and that this could play a role in the development of pulmonary hypertension in obese males and females. To characterise the effects of obesity on endogenous oestrogen, and its contribution to PAH, the effects of an aromatase inhibitor, ANA, and the CYP1B1 inhibitor, 2,2′,4,6′-tetramethoxystilbene (TMS), were studied in obese mice. We focused on the leptin-deficient ob/ob mouse but also verified observations in mice fed a high-fat diet (HFD). Materials and methods Detailed descriptions are provided in the supplementary material. ob/ob mice B6.V-Lepob/Lepob/OlaHsd (ob/ob) mice aged 6–10 weeks and their lean littermates (B6.V-(lean)/OlaHsd) were obtained from Envigo (Huntingdon, UK). HFD mice C57BL/6JOlaHsd mice aged 6–8 weeks old (Envigo) were maintained on a normal diet or a HFD (percentage calories from fat 42%, protein 15%, carbohydrate 43%; Special Diet Services, Witham, UK) for 20 weeks. The HFD mice gained 40% more body weight than their normal diet controls (supplementary figure S1).

ob/ob mice B6.V-Lepob/Lepob/OlaHsd (ob/ob) mice aged 6–10 weeks and their lean littermates (B6.V-(lean)/OlaHsd) were obtained from Envigo (Huntingdon, UK). HFD mice C57BL/6JOlaHsd mice aged 6–8 weeks old (Envigo) were maintained on a normal diet or a HFD (percentage calories from fat 42%, protein 15%, carbohydrate 43%; Special Diet Services, Witham, UK) for 20 weeks. The HFD mice gained 40% more body weight than their normal diet controls (supplementary figure S1). Chronic hypoxic studies The development of pulmonary hypertension in ob/ob and HFD mice was achieved with 14 days hypoxia as described previously [14]. Mice were administered with the aromatase inhibitor ANA (3 mg·kg−1·day−1 for 14 days) or vehicle (1% carboxymethylcellulose) daily (intraperitoneally). Mice housed in normoxic conditions were studied as controls. Pharmacological inhibition of CYP1B1 in male ob/ob mice Male ob/ob mice and their lean littermates were injected with TMS, a CYP1B1 inhibitor, 3 mg·kg−1·day−1, or vehicle (4% ethanol), once daily for 14 days (i.p.). Haemodynamic and right ventricular measurements Right ventricular systolic pressure (RVSP) and systemic arterial pressure were measured using a PVR-1045 Millar pressure–conductance catheter (Millar Instruments, Houston, TX, USA). Right ventricular hypertrophy (RVH) was assessed as described previously [14].

Pharmacological inhibition of CYP1B1 in male ob/ob mice Male ob/ob mice and their lean littermates were injected with TMS, a CYP1B1 inhibitor, 3 mg·kg−1·day−1, or vehicle (4% ethanol), once daily for 14 days (i.p.). Haemodynamic and right ventricular measurements Right ventricular systolic pressure (RVSP) and systemic arterial pressure were measured using a PVR-1045 Millar pressure–conductance catheter (Millar Instruments, Houston, TX, USA). Right ventricular hypertrophy (RVH) was assessed as described previously [14]. Lung histopathology Lung sagittal sections (5 μm) were stained with Elastin–Van Gieson. Pulmonary arteries <80 μm external diameter were then microscopically assessed in a blinded fashion to assess vascular remodelling, as described in the supplementary material and previously [14]. Measurement of E2, DHEA-S, testosterone and 16αOHE1 E2, DHEA-S, testosterone and 16αOHE1 levels were determined by ELISA as described in the supplementary material. Isolation and culture of mouse PASMCs Mouse PASMCs (mPASMCs) were isolated from third-order pulmonary arteries of male lean and ob/ob mice, and used between passage 2 and 5. Preparation of visceral adipose tissue conditioned media Cell culture media was incubated with visceral adipose tissue (VAT; 100 mg·mL−1 media) harvested from male ob/ob mice for 24 h at 37°C in the presence or absence of ANA or TMS. VAT conditioned media (VAT-CM) was diluted prior to use. Assessment of cell proliferation Proliferation was assessed using a Countess II FL cell counter (Life Technologies, Loughborough, UK) as described in the supplementary material.

Preparation of visceral adipose tissue conditioned media Cell culture media was incubated with visceral adipose tissue (VAT; 100 mg·mL−1 media) harvested from male ob/ob mice for 24 h at 37°C in the presence or absence of ANA or TMS. VAT conditioned media (VAT-CM) was diluted prior to use. Assessment of cell proliferation Proliferation was assessed using a Countess II FL cell counter (Life Technologies, Loughborough, UK) as described in the supplementary material. Immunoblotting Proteins of interest were assessed by immunoblotting whole lung or mPASMC lysates as described in the supplementary material. Quantitative reverse transcriptase-PCR mRNA expression was assessed by quantitative reverse transcriptase-PCR as described in the supplementary material. Amplex Red assay Hydrogen peroxide (H2O2) was assessed in mPASMC lysates using an Amplex Red Hydrogen Peroxide/Peroxidase Assay Kit (Thermo Fischer Scientific, Loughborough, UK) according to the manufacturer's instructions. Protein tyrosine phosphatase oxidation assessment Briefly, irreversible oxidation of protein tyrosine phosphatases (PTPs) was assessed by immunoblotting using an oxidised PTP antibody that specifically recognises the sulfonic acid form of PTP cysteine residues. ROS determination in lung sections by immunofluorescence Immunohistochemistry of the ROS marker 8-hydroxyguanosine (8-OHG) was determined in whole lung sections as described previously [10].

Protein tyrosine phosphatase oxidation assessment Briefly, irreversible oxidation of protein tyrosine phosphatases (PTPs) was assessed by immunoblotting using an oxidised PTP antibody that specifically recognises the sulfonic acid form of PTP cysteine residues. ROS determination in lung sections by immunofluorescence Immunohistochemistry of the ROS marker 8-hydroxyguanosine (8-OHG) was determined in whole lung sections as described previously [10]. Data analysis All data are expressed as mean with standard error of the mean. Data were analysed using one-way or two-way ANOVA with post hoc analyses or the unpaired t-test (as appropriate and indicated in the figure legends) using Prism version 5 (GraphPad, La Jolla, CA, USA). A p-value <0.05 was considered statistically significant. Results Aromatase expression is upregulated in VAT of male obese mice Peri-renal adipose tissue is an example of white VAT, and allows direct comparison between males and females. Lean female mice express greater levels of aromatase in VAT than males (figure 1a). Aromatase expression was elevated in VAT from male ob/ob but not female ob/ob mice compared with their lean controls (figure 1b and c). Increased aromatase expression was also confirmed in VAT from male HFD mice, but not females (supplementary figure S2).

press greater levels of aromatase in VAT than males (figure 1a). Aromatase expression was elevated in VAT from male ob/ob but not female ob/ob mice compared with their lean controls (figure 1b and c). Increased aromatase expression was also confirmed in VAT from male HFD mice, but not females (supplementary figure S2). FIGURE 1 Characterisation of changes in aromatase expression in visceral adipose tissue (VAT) from lean versus obese mice. GAPDH: glyceraldehyde phosphate dehydrogenase. Representative immunoblot and quantification of aromatase protein expression in VAT from a) lean female and male mice, b) lean and ob/ob male mice, and c) lean and ob/ob female mice (n=3–4 per group). Data are presented as mean±sem. **: p<0.01; ***: p<0.001, determined by two-tailed unpaired t-test. Inhibition of aromatase attenuates parameters of pulmonary hypertension in both male and female obese mice An increase in RVSP was observed in male ob/ob mice under normoxic conditions, whereas RVH remained unchanged (figure 2a and b). This increase in RVSP was reversed by ANA (figure 2a). An increase in pulmonary vascular remodelling was also observed in normoxic male ob/ob mice and this was unaffected by ANA treatment (figure 2c and d). ANA had no effect on hypoxia-induced changes in RVSP, RVH and pulmonary vascular remodelling in lean male mice, but attenuated hypoxia-induced increases in RVSP and vascular remodelling in male ob/ob mice (figure 2a–d).

lso observed in normoxic male ob/ob mice and this was unaffected by ANA treatment (figure 2c and d). ANA had no effect on hypoxia-induced changes in RVSP, RVH and pulmonary vascular remodelling in lean male mice, but attenuated hypoxia-induced increases in RVSP and vascular remodelling in male ob/ob mice (figure 2a–d). FIGURE 2 Inhibition of aromatase attenuates parameters of pulmonary hypertension in ob/ob mice. RVSP: right ventricular systolic pressure; ANA: anastrozole; RVH: right ventricular hypertrophy; RV/(LV+S): right ventricle/(left ventricle+septum) ratio. a–d) Male: effects of ANA 3 mg·kg−1·day−1 for 14 days on a) RVSP (n=5–10 per group), b) RVH (n=5–10 per group) (as determined by RV/(LV+S) ratio) and c) percentage of remodelled pulmonary arteries (n=4 per group) in normoxic and hypoxic male ob/ob mice. d) Representative images of pulmonary arteries from normoxic and hypoxic male ob/ob mice treated with or without ANA 3 mg·kg−1·day−1. Scale bar: 50 µm. e–h) Female: effects of ANA 3 mg·kg−1·day−1 for 14 days on e) RVSP (n=5–10 per group), f) RVH (n=5–10 per group) (as determined by RV/(LV+S) ratio) and g) percentage of remodelled pulmonary arteries (n=4 per group) in normoxic and hypoxic female ob/ob mice. h) Representative images of pulmonary arteries from normoxic and hypoxic female ob/ob mice treated with or without ANA 3 mg·kg−1·day−1. Scale bar: 50 µm. Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni post-test.

in normoxic and hypoxic female ob/ob mice. h) Representative images of pulmonary arteries from normoxic and hypoxic female ob/ob mice treated with or without ANA 3 mg·kg−1·day−1. Scale bar: 50 µm. Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni post-test. Similar findings were observed in normoxic female ob/ob mice compared with males. However, in contrast to males, ANA decreased RVH and pulmonary vascular remodelling in female lean mice, and was effective in attenuating RVSP, RVH and pulmonary vascular remodelling in female ob/ob mice (figure 2e–h). No significant changes in mean systemic arterial pressure, left ventricular end-diastolic pressure (LVEDP) or left ventricle plus septum (LV+S) weight were observed between groups in males or females (supplementary figure S3). To further assess right ventricular function, gene expression analysis of markers of heart failure and fibrosis was carried out in vehicle- and ANA-treated ob/ob mice. Fibronectin, connective tissue growth factor, atrial natriuretic peptide and brain natriuretic peptide expression levels were assessed. Changes in right ventricular remodelling were also determined by Picrosirius Red staining. No significant changes in right ventricular remodelling or gene expression were observed (supplementary figure S4).

e tissue growth factor, atrial natriuretic peptide and brain natriuretic peptide expression levels were assessed. Changes in right ventricular remodelling were also determined by Picrosirius Red staining. No significant changes in right ventricular remodelling or gene expression were observed (supplementary figure S4). Following 20 weeks on a HFD, male mice displayed no significant increases in RVSP or pulmonary vascular remodelling, although an increase in RVH was observed (supplementary figure S5a–d). Under hypoxic conditions, male HFD mice developed significantly more pulmonary vascular remodelling than those on a normal diet (supplementary figure S5d). Administration of ANA attenuated hypoxia-induced increases in RVSP, RVH and pulmonary vascular remodelling in male HFD mice (supplementary figure S5a–d). Similarly, female mice showed no significant changes in RVSP or pulmonary vascular remodelling following a HFD; however, RVH was significantly increased. ANA reduced RVSP in females (supplementary figure S5e–h). No differences in mean systemic arterial pressure, LV+S weight or LVEDP (supplementary figure S6) were observed between groups in males or females. Both ob/ob and HFD mice displayed increased body weight compared with their lean controls. ANA treatment decreased body weight in hypoxic male ob/ob mice and female normoxic ob/ob mice only (supplementary figure S7).

Following 20 weeks on a HFD, male mice displayed no significant increases in RVSP or pulmonary vascular remodelling, although an increase in RVH was observed (supplementary figure S5a–d). Under hypoxic conditions, male HFD mice developed significantly more pulmonary vascular remodelling than those on a normal diet (supplementary figure S5d). Administration of ANA attenuated hypoxia-induced increases in RVSP, RVH and pulmonary vascular remodelling in male HFD mice (supplementary figure S5a–d). Similarly, female mice showed no significant changes in RVSP or pulmonary vascular remodelling following a HFD; however, RVH was significantly increased. ANA reduced RVSP in females (supplementary figure S5e–h). No differences in mean systemic arterial pressure, LV+S weight or LVEDP (supplementary figure S6) were observed between groups in males or females. Both ob/ob and HFD mice displayed increased body weight compared with their lean controls. ANA treatment decreased body weight in hypoxic male ob/ob mice and female normoxic ob/ob mice only (supplementary figure S7). Effects of obesity on circulating E2 and testosterone levels A decrease in plasma E2 was detected in normoxic ob/ob male mice compared with their lean controls (figure 3a). Hypoxia alone, or in the presence of ANA, had no effect on circulating E2 levels in either lean or ob/ob mice (figure 3a).

Both ob/ob and HFD mice displayed increased body weight compared with their lean controls. ANA treatment decreased body weight in hypoxic male ob/ob mice and female normoxic ob/ob mice only (supplementary figure S7). Effects of obesity on circulating E2 and testosterone levels A decrease in plasma E2 was detected in normoxic ob/ob male mice compared with their lean controls (figure 3a). Hypoxia alone, or in the presence of ANA, had no effect on circulating E2 levels in either lean or ob/ob mice (figure 3a). FIGURE 3 Effects of obesity on circulating oestradiol (E2) and visceral adipose tissue (VAT) expression of cytochrome P450 1B1 (CYP1B1). ANA: anastrozole; GAPDH: glyceraldehyde phosphate dehydrogenase. a, b) Circulating plasma E2 levels in a) normoxic and hypoxic lean and ob/ob male mice (n=5–10 per group) and b) normoxic and hypoxic lean and ob/ob female mice (n=4–7 per group) treated with or without ANA 3 mg·kg−1·day−1. c, d) Representative immunoblot and quantification of CYP1B1 protein expression in VAT from c) male and d) female lean and ob/ob mice (n=3–4 per group). Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni post-test.

reated with or without ANA 3 mg·kg−1·day−1. c, d) Representative immunoblot and quantification of CYP1B1 protein expression in VAT from c) male and d) female lean and ob/ob mice (n=3–4 per group). Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni post-test. In ob/ob females, no changes in plasma E2 were observed between the groups (figure 3b). Similarly, plasma E2 was decreased in HFD males compared with their lean controls and E2 levels were unaffected by hypoxia alone or ANA treatment (supplementary figure S8a). Neither a HFD nor hypoxia had any effect on plasma E2 levels in females. However, in female HFD hypoxic mice, ANA treatment reduced the levels of circulating E2 (supplementary figure S8b). Uterus weights in lean mice and their E2 levels suggest they are in oestrus, and the E2 concentrations observed are comparable with those previously detected in female mice (∼0.3–10 pg·mL−1) [15, 16]. Additionally, ob/ob females are infertile and have significantly lower uterine weights (supplementary figure S9a). Uterine weight was unaffected by a HFD, but reduced by ANA (supplementary figure S9b). Testosterone levels were also assessed in male ob/ob and HFD mice. No significant differences in testosterone were observed between ob/ob and HFD study groups (supplementary figure S9c and d). We were unable to detect DHEA-S in plasma samples from the mice studied (data not shown).

Uterus weights in lean mice and their E2 levels suggest they are in oestrus, and the E2 concentrations observed are comparable with those previously detected in female mice (∼0.3–10 pg·mL−1) [15, 16]. Additionally, ob/ob females are infertile and have significantly lower uterine weights (supplementary figure S9a). Uterine weight was unaffected by a HFD, but reduced by ANA (supplementary figure S9b). Testosterone levels were also assessed in male ob/ob and HFD mice. No significant differences in testosterone were observed between ob/ob and HFD study groups (supplementary figure S9c and d). We were unable to detect DHEA-S in plasma samples from the mice studied (data not shown). CYP1B1 and 16αOHE1 are increased in obese male mice CYP1B1 expression was elevated in VAT from male and female ob/ob mice compared with their lean controls (figure 3c and d). In keeping with this observation, urinary 16αOHE1 levels were elevated in male normoxic ob/ob mice (figure 4a). In the male study groups, hypoxia resulted in an increase in urinary 16αOHE1 in lean animals that was unaffected by ANA treatment. However, ANA did increase urinary 16αOHE1 in hypoxic ob/ob males (figure 4a). No changes in 16αOHE1 were observed in female ob/ob animals under normoxic conditions. In the female hypoxic study groups, hypoxia resulted in an increase in 16αOHE1 and this was attenuated by ANA treatment in lean mice, whilst ANA augmented 16αOHE1 in ob/ob mice (figure 4b).

16αOHE1 in hypoxic ob/ob males (figure 4a). No changes in 16αOHE1 were observed in female ob/ob animals under normoxic conditions. In the female hypoxic study groups, hypoxia resulted in an increase in 16αOHE1 and this was attenuated by ANA treatment in lean mice, whilst ANA augmented 16αOHE1 in ob/ob mice (figure 4b). FIGURE 4 16α-hydroxyestrone (16αOHE1) levels and the effects of cytochrome P450 1B1 (CYP1B1) inhibition in obese mice. ANA: anastrozole; RVSP: right ventricular systolic pressure. a, b) Urinary 16αOHE1 levels in a) male and b) female ob/ob mice treated with and without ANA 3 mg·kg−1·day−1 for 14 days in normoxic and hypoxic conditions (n=4–8 per group). c, d) The effects of CYP1B1 inhibition with TMS 3 mg·kg−1·day−1 for 14 days on c) RVSP (n=5–7 per group) and d) percentage of remodelled pulmonary arteries (n=5 per group) in male normoxic lean and ob/ob mice. Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni post-test.

ects of CYP1B1 inhibition with TMS 3 mg·kg−1·day−1 for 14 days on c) RVSP (n=5–7 per group) and d) percentage of remodelled pulmonary arteries (n=5 per group) in male normoxic lean and ob/ob mice. Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni post-test. CYP1B1 expression was increased in VAT from HFD males; however, no comparable increase was apparent in HFD female mice (supplementary figure S10a and b). Similar to normoxic ob/ob males, a HFD resulted in an increase in urinary 16αOHE1 levels. In the case of the HFD study, only lean male animals displayed an increase in 16αOHE1 levels in hypoxia. This hypoxia-induced increase in 16αOHE1 was attenuated in male HFD animals and ANA treatment had no effect on this (supplementary figure S10c). No effect of hypoxia or ANA treatment on 16αOHE1 was observed in the female arm of the HFD study (supplementary figure S10d). 16αOHE1 is often reported as the ratio of 2OHE1 to 16αOHE1. However, the levels of 2OHE1 are often very low in mouse plasma and we are not able to consistently detect its presence by the ELISA method used. Therefore, results are expressed as 16αOHE1 concentrations only.

CYP1B1 expression was increased in VAT from HFD males; however, no comparable increase was apparent in HFD female mice (supplementary figure S10a and b). Similar to normoxic ob/ob males, a HFD resulted in an increase in urinary 16αOHE1 levels. In the case of the HFD study, only lean male animals displayed an increase in 16αOHE1 levels in hypoxia. This hypoxia-induced increase in 16αOHE1 was attenuated in male HFD animals and ANA treatment had no effect on this (supplementary figure S10c). No effect of hypoxia or ANA treatment on 16αOHE1 was observed in the female arm of the HFD study (supplementary figure S10d). 16αOHE1 is often reported as the ratio of 2OHE1 to 16αOHE1. However, the levels of 2OHE1 are often very low in mouse plasma and we are not able to consistently detect its presence by the ELISA method used. Therefore, results are expressed as 16αOHE1 concentrations only. Pharmacological inhibition of CYP1B1 attenuates the pulmonary hypertension phenotype in male ob/ob mice Given the significant increase in CYP1B1 and its product 16αOHE1 in obese male mice we assessed the effect of TMS, a selective CYP1B1 inhibitor, on the pulmonary hypertension phenotype observed in male ob/ob mice. TMS treatment significantly attenuated the elevated RVSP and pulmonary vascular remodelling observed in male ob/ob mice (figure 4c and d). Large amounts of intra-thoracic fat were observed in the ob/ob mice studied, but this was not so apparent in lean animals. Intra-thoracic fat from male ob/ob mice was found to contain 0.94±0.19 ng·mL−1 16αOHE1 (n=5).

Pharmacological inhibition of CYP1B1 attenuates the pulmonary hypertension phenotype in male ob/ob mice Given the significant increase in CYP1B1 and its product 16αOHE1 in obese male mice we assessed the effect of TMS, a selective CYP1B1 inhibitor, on the pulmonary hypertension phenotype observed in male ob/ob mice. TMS treatment significantly attenuated the elevated RVSP and pulmonary vascular remodelling observed in male ob/ob mice (figure 4c and d). Large amounts of intra-thoracic fat were observed in the ob/ob mice studied, but this was not so apparent in lean animals. Intra-thoracic fat from male ob/ob mice was found to contain 0.94±0.19 ng·mL−1 16αOHE1 (n=5). Effects of VAT-CM on mPASMC proliferation and oxidative stress To investigate the role of CYP1B1 in male ob/ob mice further, cell culture media was incubated with VAT harvested from male ob/ob mice. Analysis of the VAT-CM revealed it contained significantly lower levels of E2 compared with control media (figure 5a). ANA pre-treatment had no effect on this, but TMS resulted in a significant increase in E2 levels (figure 5a). Conversely, VAT-CM contained significantly higher levels of 16αOHE1 compared with control media, and this was attenuated by both ANA and TMS (figure 5b). The stimulation of mPASMCs with VAT-CM increased proliferating cell nuclear antigen (PCNA) expression (figure 5c) and resulted in cell proliferation. The proliferative effects of VAT-CM prepared in the presence of ANA or TMS were significantly reduced (figure 5d).

nd this was attenuated by both ANA and TMS (figure 5b). The stimulation of mPASMCs with VAT-CM increased proliferating cell nuclear antigen (PCNA) expression (figure 5c) and resulted in cell proliferation. The proliferative effects of VAT-CM prepared in the presence of ANA or TMS were significantly reduced (figure 5d). FIGURE 5 Effects of visceral adipose tissue conditioned media (VAT-CM) on mouse pulmonary artery smooth muscle cell (mPASMC) proliferation. E2: oestradiol; ANA: anastrozole; TMS: 2,2′,4,6′-tetramethoxystilbene; 16αOHE1: 16α-hydroxyestrone; PCNA: proliferating cell nuclear antigen. a) E2 levels and b) 16αOHE1 levels in control media and VAT-CM prepared in the absence or presence of ANA or TMS (n=4 per group). c) Representative immunoblot and quantification of PCNA expression in mPASMCs following 24 h treatment with VAT-CM (n=3 per group). d) The effects of 24 h VAT-CM prepared in the presence or absence of ANA or TMS on mPASMC proliferation (data normalised to control group, n=3 per group). Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by two-tailed unpaired t-test or one-way ANOVA with Bonferroni post-test, as appropriate.

he effects of 24 h VAT-CM prepared in the presence or absence of ANA or TMS on mPASMC proliferation (data normalised to control group, n=3 per group). Data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by two-tailed unpaired t-test or one-way ANOVA with Bonferroni post-test, as appropriate. VAT-CM induced proliferation to a similar extent as 16αOHE1 and this was attenuated by the ROS scavenger, 4-hydroxy-TEMPO (figure 6a). Furthermore, VAT-CM induced H2O2 production and irreversible oxidative modification of PTPs, a marker of oxidative stress (figure 6b and c). mPASMCs isolated from male ob/ob mice displayed increased levels of PCNA and were more proliferative than mPASMCs from lean mice (figure 7a and b).

r, 4-hydroxy-TEMPO (figure 6a). Furthermore, VAT-CM induced H2O2 production and irreversible oxidative modification of PTPs, a marker of oxidative stress (figure 6b and c). mPASMCs isolated from male ob/ob mice displayed increased levels of PCNA and were more proliferative than mPASMCs from lean mice (figure 7a and b). FIGURE 6 Effects of visceral adipose tissue conditioned media (VAT-CM) stimulation on markers of oxidative stress. TEMPOL: 4-hydroxy-TEMPO; 16αOHE1: 16α-hydroxyestrone; H2O2: hydrogen peroxide; RFU: relative fluorescence unit; ANA: anastrozole; TMS: 2,2′,4,6′-tetramethoxystilbene; Oxy-PTP: oxidised protein tyrosine phosphatase; mPASMC: mouse pulmonary artery smooth muscle cell. a) The effects of 24 h VAT-CM and 16αOHE1 (1 nM) stimulation in the presence or absence of TEMPOL (10 µM) on mPASMC proliferation (data normalised to vehicle group, n=5 per group). b, c) The effect of 24 h stimulation with VAT-CM prepared in the presence of absence of ANA or TMS on b) H2O2 production (expressed as RFUs) and c) Oxy-PTPs in mPASMCs (Oxy-PTP blot shown is representative of n=3 experiments as quantified in the corresponding histogram with data normalised to vehicle group). All data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni or Dunnett post hoc analyses.

FUs) and c) Oxy-PTPs in mPASMCs (Oxy-PTP blot shown is representative of n=3 experiments as quantified in the corresponding histogram with data normalised to vehicle group). All data are presented as mean±sem. *: p<0.05; **: p<0.01; ***: p<0.001, determined by one-way ANOVA with Bonferroni or Dunnett post hoc analyses. FIGURE 7 Mouse pulmonary artery smooth muscle cells (mPASMCs) isolated from male ob/ob mice are more proliferative than mPASMCs from their lean littermates. PCNA: proliferating cell nuclear antigen; GAPDH: glyceraldehyde phosphate dehydrogenase. a) Representative immunoblot and quantification (n=4) of PCNA expression, and b) proliferation (expressed as live cell count in cells·mL−1) of lean versus ob/ob male mPASMCs after 48 h in normal growth conditions. Data are presented as mean±sem. *: p<0.05; ***: p<0.001, determined by two-tailed unpaired t-test. Effects of ANA on obesity-induced oxidative damage in mouse lung Immunofluorescence of the ROS marker 8-OHG was determined in whole lung sections of male ob/ob mice. An increase in 8-OHG immunofluorescence was observed in normoxic male ob/ob mice and this was attenuated by ANA (figure 8a). Hypoxia-induced increases in 8-OHG staining were unaffected by ANA in lean mice. In hypoxic male ob/ob mice, 8-OHG staining was increased compared with lean hypoxic males and this was attenuated by ANA treatment (figure 8a). Similar results were observed in normoxic ob/ob females, but no differences in hypoxic female mice with or without ANA were determined (figure 8b).

ffected by ANA in lean mice. In hypoxic male ob/ob mice, 8-OHG staining was increased compared with lean hypoxic males and this was attenuated by ANA treatment (figure 8a). Similar results were observed in normoxic ob/ob females, but no differences in hypoxic female mice with or without ANA were determined (figure 8b). FIGURE 8 Effects of aromatase inhibition on obesity-induced oxidative damage in mouse lung. ANA: anastrozole; 8-OHG: 8-hydroxyguanosine. Representative images of reactive oxygen species marker 8-OHG and corresponding quantification in whole lung sections from a) male and b) female lean and ob/ob mice treated with or without ANA. Green: 8-OHG; blue: 4′,6-diamidino-2-phenylindole. Scale bar: 50 µm. *: p<0.05; **: p<0.01, determined by one-way ANOVA with Bonferroni post-test of normoxic and hypoxic groups independently.

corresponding quantification in whole lung sections from a) male and b) female lean and ob/ob mice treated with or without ANA. Green: 8-OHG; blue: 4′,6-diamidino-2-phenylindole. Scale bar: 50 µm. *: p<0.05; **: p<0.01, determined by one-way ANOVA with Bonferroni post-test of normoxic and hypoxic groups independently. Changes in antioxidant enzymes were also assessed in the lungs from ob/ob mice and no significant differences in superoxide dismutase 1 (SOD1) or catalase were observed between the groups in male animals (supplementary figure S11a and b). NADPH oxidase 4, which generates intracellular superoxide, was significantly increased in the lung tissue of male hypoxic ob/ob mice compared with normoxic and this was attenuated by ANA (supplementary figure S11c). In female lung, ANA treatment resulted in increases in SOD1 in normoxic conditions and catalase expression in hypoxic conditions (supplementary figure S11d and e). Antioxidant enzymes were also assessed in VAT; SOD1 was found to be decreased in ob/ob mice and this was unchanged by ANA (supplementary figure S12a). Catalase and glutathione peroxidase expression levels were unchanged in ob/ob VAT and unaffected by ANA (supplementary figure S12b and c)

pplementary figure S11d and e). Antioxidant enzymes were also assessed in VAT; SOD1 was found to be decreased in ob/ob mice and this was unchanged by ANA (supplementary figure S12a). Catalase and glutathione peroxidase expression levels were unchanged in ob/ob VAT and unaffected by ANA (supplementary figure S12b and c) Discussion Adipose tissue is metabolically active, expressing high levels of aromatase within its stromal fraction, and is a known source of oestrogen production [4]. Given that 32% of PAH patients may be obese, as reported by the REVEAL registry [1], the current pre-clinical study was designed to assess the potential effect of obesity on changes in oestrogen metabolism and the pathogenesis of experimental pulmonary hypertension in males and females.

rce of oestrogen production [4]. Given that 32% of PAH patients may be obese, as reported by the REVEAL registry [1], the current pre-clinical study was designed to assess the potential effect of obesity on changes in oestrogen metabolism and the pathogenesis of experimental pulmonary hypertension in males and females. Our results provide novel mechanistic insight into epidemiological observation linking pulmonary hypertension and obesity. As we studied many variables in vivo, the results are complex. In summary, however, the main findings were that, in normoxic female ob/ob mice, there is E2-dependent increased RVSP. Aromatase expression in VAT is 4–5-fold higher in lean females than males and does not increase further with obesity. Plasma levels of E2 are some 3-fold higher in females than males and unaffected by obesity. Although VAT CYP1B1 is elevated this did not result in any depletion of E2 plasma levels or changes in urinary 16αOHE1. This suggests that CYP1B1 activity may not drive the obesity-induced changes in RVSP in females. This is summarised in figure 9. In hypoxia there are also E2-dependent increases in RVSP and vascular remodelling.

T CYP1B1 is elevated this did not result in any depletion of E2 plasma levels or changes in urinary 16αOHE1. This suggests that CYP1B1 activity may not drive the obesity-induced changes in RVSP in females. This is summarised in figure 9. In hypoxia there are also E2-dependent increases in RVSP and vascular remodelling. FIGURE 9 Summary of results in wild-type (WT) lean and ob/ob mice. PAH: pulmonary arterial hypertension; CYP1B1: cytochrome P450 1B1; VAT: visceral adipose tissue; RVSP: right ventricular systolic pressure; RVH: right ventricular hypertrophy; PVR: pulmonary vascular remodelling; ANA: anastrozole; E2: oestradiol; 16αOHE1: 16α-hydroxyestrone; ROS: reactive oxygen species; PASMC: pulmonary arterial smooth muscle cell. Where inhibition by ANA is indicated this suggests dependency on endogenous oestrogen (E2). In normoxic female ob/ob mice there is E2-dependent increased RVSP. Aromatase expression in VAT is increased in WT lean females compared with males and does not increase further with obesity. Plasma levels of E2 are higher in lean WT females than males and unaffected by obesity. In hypoxia there are E2-dependent increases in RVSP and PVR. There is also E2-dependent ROS-driven oxidative damage in vivo. ob/ob male mice demonstrate E2-dependent elevations in RVSP and this is accompanied by increased aromatase expression in adipose tissue. There is, however, a decrease in circulating E2. CYP1B1 expression is increased in VAT and consistent with this there is an increase in urinary 16αOHE1. In hypoxia there are also E2-dependent increases in RVSP and PVR. VAT from male ob/ob mice converts E2 to 16αOHE1 via CYP1B1. This 16αOHE1 is produced in sufficient amounts to cause CYP1B1-dependent proliferation of PASMCs. This proliferation was greatest in ob/ob mice and was ROS dependent. In vivo, ROS lung production in male ob/ob mice was dependent on endogenous E2, as was hypoxia-induced ROS.

ob/ob mice converts E2 to 16αOHE1 via CYP1B1. This 16αOHE1 is produced in sufficient amounts to cause CYP1B1-dependent proliferation of PASMCs. This proliferation was greatest in ob/ob mice and was ROS dependent. In vivo, ROS lung production in male ob/ob mice was dependent on endogenous E2, as was hypoxia-induced ROS. ob/ob male mice demonstrate E2-dependent elevations in RVSP and this is accompanied by increased aromatase expression in adipose tissue. There is, however, a decrease in circulating E2. As there is increased CYP1B1 expression in VAT this suggests that there may be increased metabolism of E2 and consistent with this there is an increase in urinary 16αOHE1. Inhibition of CYP1B1 with TMS also reversed the obesity-induced increase in RVSP in males. This is summarised in figure 9. In hypoxia there are also E2-dependent increases in RVSP and vascular remodelling. As obesity in male mice may be uniquely driven by CYP1B1-induced E2 metabolism we focused on the ob/ob males to investigate this further. In summary, we demonstrated that VAT from male ob/ob mice converts E2 to 16αOHE1 via CYP1B1. This 16αOHE1 is produced in sufficient amounts to cause CYP1B1-dependent proliferation of PASMCs. This proliferation was greatest in ob/ob mice and was ROS dependent. In vivo, ROS lung production in male ob/ob mice was dependent on endogenous E2, as was hypoxia-induced ROS. This is summarised in figure 9.

1 via CYP1B1. This 16αOHE1 is produced in sufficient amounts to cause CYP1B1-dependent proliferation of PASMCs. This proliferation was greatest in ob/ob mice and was ROS dependent. In vivo, ROS lung production in male ob/ob mice was dependent on endogenous E2, as was hypoxia-induced ROS. This is summarised in figure 9. Both male and female ob/ob mice develop pulmonary hypertension spontaneously, an effect that can be attenuated by aromatase inhibition, suggesting a role for endogenous oestrogen in both male and female obese mice. This is consistent with the observation that in humans, increased body mass index is associated with an increase in pulmonary arterial systolic pressure in both males and females [17]. We have previously demonstrated that ANA treatment is only therapeutic in female hypoxic rodents and not males [18]. This is due to the unique phenotype of female PASMCs, whereby endogenous oestrogen produced by aromatase in these cells predisposes female PASMCs to proliferation and the development of pulmonary hypertension [18, 19]. Thus, the ability of ANA to attenuate pulmonary hypertension in obese hypoxic male mice but not lean as observed in the current study suggests that endogenous oestrogens are involved in the development of pulmonary hypertension in obese males. This may be due to obesity-mediated changes in oestrogen metabolism. In males, ∼60% of circulating oestradiol is derived from direct testicular secretion or from conversion of testicular androgens and the remainder derived from peripheral conversion of adrenal androgens [20]. It has been suggested that aromatase is less suppressed in the testis compared with adipose and muscle tissue by third-generation aromatase inhibitors such as ANA [21]. Therefore, the increased peripheral production of oestrogen and it metabolites in obese mice may account for the therapeutic effect of aromatase inhibition observed in obese but not lean mice in this study.

in the testis compared with adipose and muscle tissue by third-generation aromatase inhibitors such as ANA [21]. Therefore, the increased peripheral production of oestrogen and it metabolites in obese mice may account for the therapeutic effect of aromatase inhibition observed in obese but not lean mice in this study. Previous studies using ob/ob mice have yielded conflicting results regarding the development of pulmonary hypertension. The ob/ob genotype has been reported to attenuate hypoxia-induced pulmonary hypertension by inhibiting proliferation of PASMCs [22], while another study suggests it recapitulates many of the histological features of pulmonary hypertension [23]. We therefore chose to further investigate the effects of obesity on pulmonary hypertension using a diet-induced model of obesity where mice were maintained on a HFD. As previously reported, a HFD on its own did not induce pulmonary hypertension [24]. Therefore, we studied the effect of a HFD on the development of pulmonary hypertension in hypoxic conditions. In this instance, obese males and females developed significantly more pulmonary vascular remodelling than lean males. ANA attenuated parameters of pulmonary hypertension in both males and females, but appeared to be more beneficial in males, attenuating all pulmonary hypertension indices. In females, ANA only decreased RVSP, and had no effect on RVH and pulmonary vascular remodelling. We also confirmed that in male HFD obese mice there was a decrease in plasma E2 with an increase in urinary 16αOHE1, and that VAT expressed increased aromatase and CYP1B1 (supplementary figure S13). These findings support the hypothesis of sexual dimorphism in the mechanisms underlying pulmonary hypertension that may be accentuated in obesity due to the pronounced changes in oestrogen metabolism observed in males.

in urinary 16αOHE1, and that VAT expressed increased aromatase and CYP1B1 (supplementary figure S13). These findings support the hypothesis of sexual dimorphism in the mechanisms underlying pulmonary hypertension that may be accentuated in obesity due to the pronounced changes in oestrogen metabolism observed in males. The authors acknowledge that the phenotype observed in the mouse models used in this study is not as severe as that seen in other hypoxic models. However, it is comparable to other transgenic mouse models that have not been exposed to hypoxia, such as SERT+ mice [6], Smad-1+/− mice [19] and bone morphogenetic protein receptor type 2 (BMPR2) R899X mice [25]. The pulmonary hypertension phenotype observed is not due to left-sided heart failure as no changes in LV+S weight, LVEDP or systemic arterial pressure were detected between lean and ob/ob mice. Indeed, it has been previously documented that ob/ob mice maintain normal systemic blood pressure despite displaying severe obesity [26]. Additionally, no significant differences in markers of right ventricular remodelling were observed between sexes or ANA treatment groups. Furthermore, the ability of ANA to attenuate pulmonary hypertension in obese male mice but not lean suggests that inhibition of aromatase may influence obesity-induced elevations in pulmonary arterial pressures and thereby modify the mechanisms underlying pulmonary hypertension development. ANA treatment resulted in a reduction in body weight in normoxic ob/ob females and hypoxic ob/ob males that may also have contributed to the decrease in RVSP in these groups given the association of body mass index with pulmonary arterial systolic pressure.

the mechanisms underlying pulmonary hypertension development. ANA treatment resulted in a reduction in body weight in normoxic ob/ob females and hypoxic ob/ob males that may also have contributed to the decrease in RVSP in these groups given the association of body mass index with pulmonary arterial systolic pressure. The disparity between increased levels of aromatase in VAT and the decrease in circulating E2 levels suggests oestrogen may be metabolised in VAT rather than excreted. Therefore, we assessed the expression of the oestrogen-metabolising enzyme CYP1B1 in VAT from the obese models studied. CYP1B1 has previously been reported to be highly expressed in VAT and its expression increases during adipogenesis [27]. Here, we demonstrate that CYP1B1 is upregulated in VAT from obese ob/ob and HFD male mice, and that this correlates with an increase in urinary 16αOHE1 levels. Low DHEA-S has been identified as a risk factor for PAH in male patients; however, the circulating levels of DHEA-S in the mice studied were below the limits of detection of the assay used. Others have also reported low or undetectable DHEA-S in mice [15]. Changes in testosterone were also assessed and no significant difference observed between the treatment groups of mice studied. This suggests changes specific to oestrogen and its metabolism by CYP1B1, rather than their upstream mediators, occur in the obese models studied. The pathogenic role of CYP1B1 and 16αOHE1 in obese mice was confirmed by the beneficial effect of the CYP1B1 inhibitor TMS observed in this study. Changes in oestrogen metabolism have previously been observed in PAH. CYP1B1 is highly upregulated within pulmonary arterial lesions of PAH patients and pharmacological inhibition of CYP1B1 can attenuate the development of experimental pulmonary hypertension. These studies have also demonstrated increased pulmonary arterial expression of CYP1B1 in lungs from hypoxic mice (males and females) and the Sugen/hypoxic mouse model (males and females) [8, 28]. Oestrogen metabolism is a strong predictor of penetrance in heritable PAH. Polymorphisms in CYP1B1 that cause preferential metabolism of oestrogen to 16αOHE1 result in the development of PAH in females, whereas females who preferentially metabolise oestrogen into 2OHE1 or 4OHE1 do not [29, 30]. Oestrogen metabolism can drive PAH penetrance in males, but not to the same degree as in females [31].

ymorphisms in CYP1B1 that cause preferential metabolism of oestrogen to 16αOHE1 result in the development of PAH in females, whereas females who preferentially metabolise oestrogen into 2OHE1 or 4OHE1 do not [29, 30]. Oestrogen metabolism can drive PAH penetrance in males, but not to the same degree as in females [31]. The increased CYP1B1 expression in VAT and the subsequent increase in urinary 16αOHE1 levels observed in male ob/ob and HFD mice may directly contribute to the pulmonary hypertension phenotype observed in these animals, given the beneficial effects of CYP1B1 inhibition on pulmonary hypertension observed in ob/ob mice. Intra-thoracic fat present in male ob/ob mice contains 16αOHE1. As this VAT is in direct contact with both heart and lung tissue it may have significant effects on these tissues and contribute to the development of pulmonary hypertension. Further contributing to this hypothesis is the finding that VAT-CM has significantly lower levels of E2 and higher levels of 16αOHE1 than control media, an effect that can be attenuated by CYP1B1 inhibition. This suggests VAT can metabolise E2 via CYP1B1, resulting in the secretion of 16αOHE1. ANA was effective at reducing indicators of pulmonary hypertension in the hypoxic male ob/ob and hypoxic HFD male mice, whereas it was not effective in the male lean hypoxic mice. This suggests that in hypoxia, endogenous oestrogen plays an increased role in the development of pulmonary hypertension in male obese mice. The relationships between the development of pulmonary hypertension, CYP1B1 and effectiveness of ANA are, however, less clear in hypoxic male mice. Hypoxia itself increased urinary 16αOHE1 in the lean males, but in the ob/ob male mice this was not increased any further. A HFD in male mice actually reduced the elevated urinary 16αOHE1 seen in hypoxia. This suggests that hypoxia itself affects oestrogen metabolism and that endogenous oestrogen is exerting pathogenic effects independent of CYP1B1 activity in hypoxic obese males. Hypoxia and oestrogen are known to reduce BMPR2 signalling [32, 33]. Indeed, BMPR2 signalling is reduced in male and female hypoxic mouse lung [18]. In male hypoxic mice, hypoxia elevates 16αOHE1, which has been shown to synergise with BMPR2 deficiency and uncover a pulmonary hypertension phenotype in BMPR2-deficient transgenic mice. This pulmonary hypertension phenotype in BMPR2-deficient mice is reversed by ANA [34].

ed in male and female hypoxic mouse lung [18]. In male hypoxic mice, hypoxia elevates 16αOHE1, which has been shown to synergise with BMPR2 deficiency and uncover a pulmonary hypertension phenotype in BMPR2-deficient transgenic mice. This pulmonary hypertension phenotype in BMPR2-deficient mice is reversed by ANA [34]. We know that oestrogen synthesis occurs via aromatase in the hypoxic mouse lung [18]. Hence, endogenous oestrogen and increased 16αOHE1 may be synergising with reduced BMPR2 in obese hypoxic mice to contribute to the pulmonary hypertension phenotype. Paradoxically, ANA increased urinary 16αOHE1 in both male and female mice while maintaining levels of plasma oestrogen. This suggests that the effectiveness of ANA at reversing pulmonary hypertension in these hypoxic mice is closely related to the effects of hypoxia on lung oestrogen synthesis via aromatase and the direct effects of this endogenous oestrogen on lung pathology and BMPR2 signalling. These effects may be more influential than the effects of dysregulated oestrogen metabolism.

ing pulmonary hypertension in these hypoxic mice is closely related to the effects of hypoxia on lung oestrogen synthesis via aromatase and the direct effects of this endogenous oestrogen on lung pathology and BMPR2 signalling. These effects may be more influential than the effects of dysregulated oestrogen metabolism. The results do suggest, however, a close relationship between adipose tissue CYP1B1 expression, the development of pulmonary hypertension and effectiveness of ANA in normoxic ob/ob male mice. Indeed, incubation of PASMCs with VAT-CM leads to ROS generation, oxidative damage and activation of cell survival pathways, resulting in their proliferation in a redox-sensitive manner similar to that observed with 16αOHE1 alone. These effects are attenuated by both aromatase and CYP1B1 inhibition. In keeping with this finding, an increase in 8-OHG staining, a marker of ROS production, was observed in whole lung tissue from ob/ob mice. ANA treatment significantly attenuated this phenomenon in male and female mice in normoxia, but was only effective in male hypoxic mice, again highlighting the sexual dimorphism in mechanisms underpinning pulmonary hypertension development. No significant changes in the antioxidant enzymes SOD1 or catalase were observed in lung tissue of the mice studied; however, NADPH oxidase 4 was upregulated in male ob/ob mice in hypoxia and this was attenuated by ANA. These findings suggest an increase in ROS production rather than a decrease in the levels of antioxidants is occurring in obesity and that ANA has antioxidant properties.

e observed in lung tissue of the mice studied; however, NADPH oxidase 4 was upregulated in male ob/ob mice in hypoxia and this was attenuated by ANA. These findings suggest an increase in ROS production rather than a decrease in the levels of antioxidants is occurring in obesity and that ANA has antioxidant properties. We have previously comprehensively demonstrated mechanisms of 16αOHE1-induced proliferation and redox signalling in PASMCs, and so have not addressed this in the current study. 16αOHE1 increases NADPH oxidase 1 expression and ROS production, leading to irreversible oxidation of PTPs, decreased activity nuclear factor erythroid-related factor 2 (Nrf2) and increased cell proliferation in human PASMCs [10]. The signalling pathways stimulated by 16αOHE1 are unique to the pulmonary circulation as 16αOHE1 failed to induce ROS production or proliferation in vascular smooth muscle cells from the systemic circulation [10]. Although the effects of 16αOHE1 on ROS production appear to be specific to the pulmonary circulation, increases in ROS production have been observed globally in obesity [35]. Indeed, levels of the antioxidant enzyme SOD1 are reduced in VAT from male ob/ob mice. ROS production is also known to be increased in hypoxia in a variety of tissues [36], and this may well have systemic effects in the obese models studies and drive other changes in oestrogen responsiveness and metabolism that account for some of the differences in 16αOHE1 observed in hypoxic animals. The effects of hypoxia on oestrogen metabolism have been most widely studied in cancer cells, where it is known to alter responsiveness to oestrogen and its metabolism [37]. The beneficial effects of ANA observed may therefore be due to a general reduction in ROS as well as a reduction in 16αOHE1-mediated ROS production.

e effects of hypoxia on oestrogen metabolism have been most widely studied in cancer cells, where it is known to alter responsiveness to oestrogen and its metabolism [37]. The beneficial effects of ANA observed may therefore be due to a general reduction in ROS as well as a reduction in 16αOHE1-mediated ROS production. 16αOHE1 is also known to promote insulin resistance and other metabolic disorders, and the ob/ob mice used in this study are indeed glucose intolerant. Therefore, it is possible insulin resistance plays a role in the phenotype observed, but this was not specifically assessed in the study undertaken. It is of note that ANA has previously been shown to reduce metabolic defects in mouse models of PAH [34]. Metabolic defects have been increasingly associated with PAH clinically. The mechanisms underlying this remain unclear, although ROS have been implicated [38]. Metformin is a well-established drug for use in type 2 diabetes mellitus and has also been proposed as treatment for PAH. We have previously shown metformin can inhibit aromatase expression and activity to decrease oestrogen and its metabolism in experimental pulmonary hypertension [7].

h ROS have been implicated [38]. Metformin is a well-established drug for use in type 2 diabetes mellitus and has also been proposed as treatment for PAH. We have previously shown metformin can inhibit aromatase expression and activity to decrease oestrogen and its metabolism in experimental pulmonary hypertension [7]. Taken together, these findings suggest that metabolic changes in males during obesity result in the increased synthesis and excretion of 16αOHE1 from VAT that can then act in an endocrine fashion on PASMCs, resulting in ROS generation, oxidative damage and increased proliferation, thus contributing to pulmonary vascular remodelling and the pulmonary hypertension phenotype observed in the animal models studied (supplementary figure S13). The ability of intra-thoracic fat to produce 16αOHE1 lends further weight to this hypothesis given its direct contact with heart and lung tissue. ANA, a third-generation, nonsteroidal, highly selective aromatase inhibitor that is widely used clinically and has been shown to be safe in a small clinical trial in PAH patients [13], had therapeutic effects in both male and female obese mice. The female predisposition to develop PAH is well established [1], mediated partially by the pathogenic effects of oestrogen in the pulmonary circulation [18, 19]. We have demonstrated that endogenous oestrogens can play a role in pulmonary hypertension in males in the presence of modifying factors such as obesity. Obesity may particularly predispose males to the development of PAH due to obesity-mediated VAT dysfunction resulting in altered oestrogen production and metabolism.

. We have demonstrated that endogenous oestrogens can play a role in pulmonary hypertension in males in the presence of modifying factors such as obesity. Obesity may particularly predispose males to the development of PAH due to obesity-mediated VAT dysfunction resulting in altered oestrogen production and metabolism. Supplementary material 10.1183/13993003.01524-2018.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-01524-2018.Supplement This article has supplementary material available from erj.ersjournals.com Support statement: This research was funded by a grant from the British Heart Foundation (RG/16/2/32153). Funding information for this article has been deposited with the Crossref Funder Registry. Conflict of interest: K.M. Mair has nothing to disclose. Conflict of interest: K.Y. Harvey has nothing to disclose. Conflict of interest: A.D. Henry has nothing to disclose. Conflict of interest: D.Z. Hillyard has nothing to disclose. Conflict of interest: M. Nilsen has nothing to disclose. Conflict of interest: M.R. MacLean has nothing to disclose.

Introduction Lung function has an estimated heritability of between 30% and 70% [1]. The variance in phenotype remains incompletely explained by genetic variation, but the impact of environmental exposure on respiratory health and lung function over the life course is well recognised. In particular, pro-inflammatory and oxidative inhalants such as cigarette and environmental tobacco smoke, air pollution, and occupational exposures are important contributors to the increased risk of respiratory symptoms, accelerated lung function decline in adults and poor lung growth in children. DNA methylation (DNAme) has been associated with a wide variety of traits and chronic diseases. A large body of evidence including results from epigenome-wide association studies (EWASs) shows differentially methylated CpG (5′-cytosine-phosphate-guanine-3′ dinucleotide) sites throughout the genome in response to environmental exposures, in particular cigarette smoking [2–4]. In contrast, reports of DNAme associated with respiratory diseases and lung function show inconsistent findings [5, 6]. Most recently, however, independent reports pointed to the consistent association of DNAme in the AHRR gene, cg05575921, with lung function in adults [4, 6, 7].

in particular cigarette smoking [2–4]. In contrast, reports of DNAme associated with respiratory diseases and lung function show inconsistent findings [5, 6]. Most recently, however, independent reports pointed to the consistent association of DNAme in the AHRR gene, cg05575921, with lung function in adults [4, 6, 7]. The current study aimed at agnostically identifying lung function-specific DNAme signals. We undertook a covariate-adjusted EWAS using questionnaire data, spirometry and peripheral blood samples collected in the same participants (discovery cohorts: ECRHS (European Community Respiratory Health Study), NFBC1966 (Northern Finland Birth Cohort 1966) and SAPALDIA (Swiss Study on Air Pollution Heart and Lung Disease in Adults); cohort description in supplementary material) at two time-points 6–15 years apart. EWAS analyses were performed on lung function parameters of forced expiratory volume in 1 s (FEV1), forced vital capacity (FVC) and their ratio (FEV1/FVC). The analyses focused on cross-sectional associations at different time-points and on identifying DNAme markers predicting change in lung function. We tested discovery-identified CpGs (p<5×10−7 for at least one lung function parameter) for replication in adult samples from five adult cohorts (LBC1936 (Lothian Birth Cohort 1936, adult inception birth cohort), KORA (Cooperative Health Research in the Augsburg Region Study), LifeLines (LifeLines cohort study), NSPHS (North Sweden Population Health Study) and FTC (Finnish Twin Cohort study)) and in childhood samples from two birth cohorts (ALSPAC (Avon Longitudinal Study of Parents and Children) and IOWBC (Isle of Wight Birth Cohort)).

perative Health Research in the Augsburg Region Study), LifeLines (LifeLines cohort study), NSPHS (North Sweden Population Health Study) and FTC (Finnish Twin Cohort study)) and in childhood samples from two birth cohorts (ALSPAC (Avon Longitudinal Study of Parents and Children) and IOWBC (Isle of Wight Birth Cohort)). Methods Study design and participants The discovery sample (n=2043) comprised three population-based cohort studies, part of the Aging Lungs in European Cohorts (ALEC) project. ECRHS (n=470) and SAPALDIA (n=962) are adult cohorts designed to investigate respiratory health. NFBC1966 (n=611) is a birth cohort with follow-up to adult age. The replication sample consisted of five adult cohorts (KORA (n=628), LifeLines (n=1622), NSPHS (n=535), LBC1936 (n=449) and FTC (n=93)) and two childhood birth cohorts (ALSPAC (n=258) and IOWBC (n=162)). Replication data from two time-points were available only for KORA and LBC1936 (adult) and ALSPAC and IOWBC (childhood). For cohort details and contribution to analysis, refer to the supplementary material and supplementary figure S1. All cohorts comply with the Declaration of Helsinki, and ethical approval was obtained from the respective national and regional ethical review committees.

Methods Study design and participants The discovery sample (n=2043) comprised three population-based cohort studies, part of the Aging Lungs in European Cohorts (ALEC) project. ECRHS (n=470) and SAPALDIA (n=962) are adult cohorts designed to investigate respiratory health. NFBC1966 (n=611) is a birth cohort with follow-up to adult age. The replication sample consisted of five adult cohorts (KORA (n=628), LifeLines (n=1622), NSPHS (n=535), LBC1936 (n=449) and FTC (n=93)) and two childhood birth cohorts (ALSPAC (n=258) and IOWBC (n=162)). Replication data from two time-points were available only for KORA and LBC1936 (adult) and ALSPAC and IOWBC (childhood). For cohort details and contribution to analysis, refer to the supplementary material and supplementary figure S1. All cohorts comply with the Declaration of Helsinki, and ethical approval was obtained from the respective national and regional ethical review committees. Procedures In the discovery cohorts, DNAme measurements using Infinium technology (Illumina, San Diego, CA, USA) were obtained from peripheral blood samples collected at two consecutive follow-up surveys several years apart. The Infinium 450K BeadChip was used for samples of 984 SAPALDIA participants from both time-points and of 732 NFBC1966 participants collected at time-point 1. The Infinium EPIC BeadChip was used for samples of 509 ECRHS participants from both time-points and of 716 NFBC1966 participants collected at time-point 2. For cohort-specific EWAS analyses, we used all autosomal markers available for each time-point and cohort-specific EWAS marker results were meta-analysed without restriction to markers common to both arrays. DNAme data used for replication were restricted to discovery-identified (sentinel) CpGs and analysed on various arrays.

ort-specific EWAS analyses, we used all autosomal markers available for each time-point and cohort-specific EWAS marker results were meta-analysed without restriction to markers common to both arrays. DNAme data used for replication were restricted to discovery-identified (sentinel) CpGs and analysed on various arrays. Epidemiological data, including covariate information at the subject level, were collected by interview-assisted questionnaires and objective measures. Pre-bronchodilation spirometric data were obtained by performing American Thoracic Society/European Respiratory Society-compliant spirometry (supplementary material). Statistical analyses Epigenome-wide methylation data were analysed in R version 3.4.3 (R Foundation for Statistical Computing, Vienna, Austria). Differential blood cell count was estimated using a reference dataset and the R package minfi [8, 9]. DNAme used as predictors in the statistical models for the adult cohorts were obtained by deriving residuals from linear regression of the normalised absolute DNAme (β-values) on the Illumina control probe-derived 30 first principal components to correct for correlation structures within the data, including technical bias. Thus, effect sizes reported here of the association are not comparable to effect sizes reported elsewhere using normalised β-values as predictor. In the childhood data, batch effect was corrected at the analysis level by regressing the DNAme values against the technical covariates.

n the data, including technical bias. Thus, effect sizes reported here of the association are not comparable to effect sizes reported elsewhere using normalised β-values as predictor. In the childhood data, batch effect was corrected at the analysis level by regressing the DNAme values against the technical covariates. Epigenome-wide covariate-adjusted linear regression was performed to assess the association of single CpG markers with FEV1 (L), FVC (L), their ratio (FEV1/FVC) and their change during follow-up. This multilevel EWAS design tested different models in all participants and never-smoking participants (figure 1). First, cross-sectional EWASs were examined separately at time-point 1 (EWAS1) and time-point 2 (EWAS2) to assess the consistency of the association over follow-up time. Second, the association of DNAme at the first time-point (DNAme1) with change in lung function during follow-up was assessed (prediction EWAS (EWASpredict)). Covariate-adjusted mixed linear regressions with a random intercept on the subject were undertaken using data from both time-points (repeat cross-sectional analysis (EWASrepeat)).

iation of DNAme at the first time-point (DNAme1) with change in lung function during follow-up was assessed (prediction EWAS (EWASpredict)). Covariate-adjusted mixed linear regressions with a random intercept on the subject were undertaken using data from both time-points (repeat cross-sectional analysis (EWASrepeat)). FIGURE 1 Flow of the multilevel discovery design of the epigenome-wide association study (EWAS) on lung function parameters: forced expiratory volume in 1 s (FEV1), forced vital capacity (FVC) and FEV1/FVC. DNAme1: DNA methylation at time-point 1; DNAme2: DNA methylation at time-point 2. #: base model (Mbase) EWAS was covariate adjusted for age, age squared, height, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. ¶: smoking model EWAS (Msmok) additionally adjusted for smoking covariates: history of smoking intensity as pack-years smoked up to the time-point of data collection for regressions and for smoking status (current smoker, ex-smoker and never-smoker). EWAS longitudinally predicting the change in lung function (EWASpredict) was additionally adjusted for lung function at time-point 1.

covariates: history of smoking intensity as pack-years smoked up to the time-point of data collection for regressions and for smoking status (current smoker, ex-smoker and never-smoker). EWAS longitudinally predicting the change in lung function (EWASpredict) was additionally adjusted for lung function at time-point 1. All associations were adjusted for a set of a priori selected covariates known to influence respiratory outcomes from previous research conducted by SAPALDIA and ECRHS. The covariate-adjusted model (Mbase) included age, age squared, height, squared deviation from the mean of height, sex and interaction terms of sex with four covariates (age, age squared, height and squared deviation of height), education, body mass index, spirometer type, study centre as well as estimated cell composition (CD8 cells, CD4 cells, natural killer cells, B-cells, monocytes, eosinophils and neutrophils). Analyses in all participants were run without (Mbase) and with additional smoking adjustment including smoking status and pack-years (Msmok). Mbase covariate adjustment was applied in never-smokers. Prediction associations of DNAme1 were additionally adjusted for lung function at time-point 1. The same covariate adjustment was applied in adult replication analyses, whereas childhood covariates did not include squared terms.

ing status and pack-years (Msmok). Mbase covariate adjustment was applied in never-smokers. Prediction associations of DNAme1 were additionally adjusted for lung function at time-point 1. The same covariate adjustment was applied in adult replication analyses, whereas childhood covariates did not include squared terms. Cohort-specific EWAS results were summarised by inverse-variance-weighted meta-analyses using METAL [10]. Meta-analysis results were not controlled for genomic inflation after confirming its negligible influence. Epigenome-wide significance level was set to p<1×10−7 (Bonferroni correction, 450 000 tests). The selection criteria for replication of sentinel CpGs was less stringent (p<5×10−7). Successful replication was defined as a p-value below the outcome-specific Bonferroni correction threshold.

rming its negligible influence. Epigenome-wide significance level was set to p<1×10−7 (Bonferroni correction, 450 000 tests). The selection criteria for replication of sentinel CpGs was less stringent (p<5×10−7). Successful replication was defined as a p-value below the outcome-specific Bonferroni correction threshold. Replicated CpGs were characterised by enrichment, pathway and functional analyses, and additional post hoc analyses were performed (details in supplementary material). 1) A two-sample Mendelian randomisation analysis based on publicly available data was applied to investigate the causality of replicated CpG associations. 2) A replication of a recently published mediation analysis [4] evidencing 10 smoking-related CpGs mediating the effect of smoking on lung function was undertaken in one discovery cohort (SAPALDIA). 3) To assess the combined effects of smoking-related CpGs on lung function in three discovery cohorts, we built two different DNAme smoking indices based on CpGs: a) predicting lung function effects of smoking [4] and b) located in genome-wide association study (GWAS)-identified lung function genes [2]. These smoking indices were tested for association with lung function in covariate-adjusted linear regression analyses, in all participants and in subgroups stratified by smoking status.

icting lung function effects of smoking [4] and b) located in genome-wide association study (GWAS)-identified lung function genes [2]. These smoking indices were tested for association with lung function in covariate-adjusted linear regression analyses, in all participants and in subgroups stratified by smoking status. Data availability statement Statistical codes and full discovery/replication EWAS effect estimates (meta-analysed and cohort-specific) are made publically available with no end date on the public repository DRYAD (http://datadryad.org/) at the time of publication. Access restrictions apply to the individual methylome data underlying the analysis. Contact details for data requests to the contributing cohorts can be found in the supplementary material. Results Differences in the cohorts' age structure and smoking habits are shown in tables 1 and 2. Mean age was highest for LBC1936 (69.9 years) and youngest for FTC (30.4 years). Self-report of current smoking status was lowest in LBC1936 (5.8%) and highest in LifeLines (43.5% due to oversampling of current smokers for the DNAme-typed subset). TABLE 1 Characteristics of discovery cohorts

Results Differences in the cohorts' age structure and smoking habits are shown in tables 1 and 2. Mean age was highest for LBC1936 (69.9 years) and youngest for FTC (30.4 years). Self-report of current smoking status was lowest in LBC1936 (5.8%) and highest in LifeLines (43.5% due to oversampling of current smokers for the DNAme-typed subset). TABLE 1 Characteristics of discovery cohorts SAPALDIA 2 time-point 1 SAPALDIA 3 time-point 2 ECRHS II time-point 1 ECRHS III time-point 2 NFBC1966 (age 31 years) time-point 1 NFBC1966 (age 46 years) time-point 2 Subjects n 962 962 470 470 611 611 Female 53.5 53.5 56 56 55.3 55.3 Age years 50.5±11.3 58.8±11.3 43.6±6.8 54.5±6.8 31.0±0.3 46.3±0.4 Height cm 169.4±9.2 168.7±9.4 170.0±9.2 169.2±9.3 171±8.8 171±8.9 Weight kg 74.2±14.7 75.5±15.4 72.6±14.6 76.2±15.5 71.3±13.6 78.7±16.3 Body mass index kg·m−2 25.8±4.4 26.5±4.6 25.0±4.0 26.5±4.4 24.2±3.7 26.7±4.8 Smoking status Never-smoker# 41.7 41.1 43.2 41.7 54.5 54.5 Ex-smoker 30.0 37.0 31.1 40.4 21.3 30.2 Current smoker 28.3 21.9 25.7 17.9 24.1 15.3 Pack-years 20.4±20.2 22.6±22.1 16.6±16.9 20.0±21.3 7.7±5.9 11.0±9.6 Education¶ Low 5.4 5.4 11.5 11.5 0.7 0.7 Intermediate 65.7 65.7 29.2 29.2 55.9 55.9 High 28.9 28.9 59.3 59.3 43.3 43.3 FVC L+ 4.4±1.0 4.1±1.1 4.3±1.0 3.9±1.0 4.8±1.0 4.5±0.9 FEV1 L+ 3.3±0.8 3.0±0.8 3.4±0.7 3.0±0.8 4.0±0.8 3.5±0.7 FEV1/FVC+ 0.75±0.07 0.73±0.08 0.78±0.06 0.75±0.06 0.83±0.06 0.77±0.06 Airflow obstruction FEV1/FVC <0.7+ 20.4 29.5 8.9 16.2 1.8 10.8 FEV1/FVC <LLN+,§ 12.9 14.1 8.7 10.4 3.3 9.5 Doctor-diagnosed asthma 13.8 16.5 14.3 16.8 10.7 15.8 Respiratory medication (% missing values) 22.2 (0.8) 23.7 (0.3) 13.4 14.2 NA NA Data are presented as % or mean±sd, unless otherwise stated; percentages may not total 100% due to rounding. FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity; LLN: lower limit of normal; NA: not assessed. #: self-reported lifetime nonsmoking. ¶: the categorical variable “education” is defined differently in cohorts (in SAPALDIA low corresponds to primary education; intermediate to secondary, middle or vocational school and high to technical college or university; in ECRHS and NFBC1966 information of age reached at end of studies is used to define low as ≤16 years, intermediate as 17–19 years and high as ≥20 years). +: values derived from pre-bronchodilation spirometry (lung function values corrected for spirometer device change in SAPALDIA at time-point 2). §: LLN values estimated using Global Lung Initiative 2012 reference equations [11].

ies is used to define low as ≤16 years, intermediate as 17–19 years and high as ≥20 years). +: values derived from pre-bronchodilation spirometry (lung function values corrected for spirometer device change in SAPALDIA at time-point 2). §: LLN values estimated using Global Lung Initiative 2012 reference equations [11]. TABLE 2 Characteristics of adult replication cohorts

ies is used to define low as ≤16 years, intermediate as 17–19 years and high as ≥20 years). +: values derived from pre-bronchodilation spirometry (lung function values corrected for spirometer device change in SAPALDIA at time-point 2). §: LLN values estimated using Global Lung Initiative 2012 reference equations [11]. TABLE 2 Characteristics of adult replication cohorts KORA time-point 1 KORA time-point 2 LBC1936 time-point 1 LBC1936 time-point 2 LifeLines time-point 1 NSPHS time-point 1 FTC time-point 1 Subjects n 628 628 449 449 1622 535 93 Female 53.2 53.2 46.8 46.8 42.8 53.1 47.3 Age years 53.6±4.5 60.1±4.5 69.6±0.9 76.3±0.7 46.7±10.8 55.1±16.0 30.4±3.8 Height cm 169.5±9.3 168.7±9.4 167.2±8.8 166.1±8.8 176.9±9.1 163.8±9.8 173.0±10.5 Weight kg 79.0±16.7 79.9±17.3 77.2±14.6 76.5±14.8 82.1±14.7 74.0±15.2 82.0±18.8 Body mass index kg·m−2 27.4±4.7 28.0±5.1 27.5±4.3 27.7±4.6 26.2±3.9 27.5±4.7 27.3±5.4 Smoking status Never-smoker# 38.2 38.2 52.3 52.3 56.6 83.2 53.8 Ex-smoker 43.8 45.5 40.8 41.9 0ƒ NA## 26.9 Current smoker 18.0 16.2 6.9 5.8 43.5 16.5 19.4 Pack-years 12.8±19.3 13.5±20.2 13.9±24.0 14.1±24.6 21.0±11.7 8.1±21.6 NA Education¶ Low 47.6 47.6 49.7 49.7 23.1 NA 1.1 Intermediate 26.4 26.4 32.3 32.3 40.8 NA 38.6 High 26.0 26.0 18.0 18.0 35.4 NA 60.2 FVC L+ 4.3±1.0 3.9±1.0 3.2±0.9 2.8±0.9 4.7±1.1 3.4±1.1 4.8±1.1 FEV1 L+ 3.3±0.8 3.0±0.7 2.5±0.7 2.1±0.7 3.5±0.9 2.8±0.9 3.9±0.9 FEV1/FVC+ 0.78±0.06 0.75±0.07 0.79±0.09 0.76±0.12 0.73±0.09 0.83±0.09 0.81±0.07 Airflow obstruction FEV1/FVC <0.7+ 8.1 20.1 15.4 26.3 38.4 8.8 5.0 FEV1/FVC <LLN+,§ 5.0 9.6 7.6 14.9 27.5 4.3 11.3 Doctor-diagnosed asthma 7.2 8.6 4.5 7.1 9.9 14.2 0 Respiratory medication 3.3 4.9 6.7 11.8 8.0 7.7 0 Data are presented as % or mean±sd, unless otherwise stated; percentages may not total 100% due to rounding. FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity; LLN: lower limit of normal; NA: not assessed. #: self-reported lifetime nonsmoking. ¶: the categorical variable “education” is defined differently in different cohorts. +: values derived from pre-bronchodilation spirometry. §: LLN values estimated using Global Lung Initiative 2012 reference equations [11]. ƒ: LifeLines: nonrandom selection of samples for DNA methylation typing (current smokers versus never-smokers). ##: NSPHS: information obtained on current smoking status (yes/no).

. +: values derived from pre-bronchodilation spirometry. §: LLN values estimated using Global Lung Initiative 2012 reference equations [11]. ƒ: LifeLines: nonrandom selection of samples for DNA methylation typing (current smokers versus never-smokers). ##: NSPHS: information obtained on current smoking status (yes/no). Across all discovery EWAS meta-analyses, we identified 111 CpG markers for replication (p<5×10−7: 74 for FEV1, 16 for FVC and 47 for FEV1/FVC) (supplementary tables S1 and S2). We present here the results for FEV1/FVC (for FEV1 and FVC, refer to the supplementary material).

. +: values derived from pre-bronchodilation spirometry. §: LLN values estimated using Global Lung Initiative 2012 reference equations [11]. ƒ: LifeLines: nonrandom selection of samples for DNA methylation typing (current smokers versus never-smokers). ##: NSPHS: information obtained on current smoking status (yes/no). Across all discovery EWAS meta-analyses, we identified 111 CpG markers for replication (p<5×10−7: 74 for FEV1, 16 for FVC and 47 for FEV1/FVC) (supplementary tables S1 and S2). We present here the results for FEV1/FVC (for FEV1 and FVC, refer to the supplementary material). Cross-sectional associations without smoking adjustment In the study-specific and meta-analysed discovery EWAS, the number of lung function-associated DNAme increased from the first to second cross-sectional time-point in the same participants, despite age adjustment (figure 2a and b). We therefore meta-analysed cross-sectional discovery and replication results from the older participants' age time-point available. We observed 29 cross-sectional CpG associations with FEV1/FVC. 27 of them replicated formally (Bonferroni correction, p<0.0011; 47 tests on FEV1/FVC) (table 3 and supplementary table S3). All replicated CpG lung function associations were exclusively DNAme previously associated with smoking [2]. Successful replication was observed for cg05575921 (AHRR), showing the strongest signal for FEV1 and FEV1/FVC (FEV1/FVC: p-value combining discovery and replication cohorts (pcombined)=7.22×10−50) among all identified lung function DNAme markers. Methylation at this CpG, previously shown to be hypomethylated with increased smoking, showed positive cross-sectional lung function association. The top 10 CpGs associated with FEV1/FVC (table 3) were located in six loci: cg03636183 (F2RL3), cg21566642, cg01940273 and cg03329539 (vicinity of ALPPL2), cg05575921 and cg21161138 (AHRR), cg23771366 and cg11660018 (PRSS23), cg21611682 (LRP5), and cg15342087 (IER3). The same CpGs, along with cg19572487 (RARA), were also among the top 11 markers cross-sectionally associated with FEV1. Formal replication of cross-sectional associations with FEV1 was observed for 44 CpGs and with FVC for three CpGs (supplementary tables S4 and S5). Similar results were found for repeat cross-sectional analyses (EWASrepeat) (supplementary table S6 and supplementary figure S3). FIGURE 2 a, b) Effect of ageing on the associations between DNA methylation (DNAme) and lung function: quantile–quantile plots of the cross-sectional covariate-adjusted discovery epigenome-wide association study (EWAS) (Mbase#) on forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) at a) time-point 1 and b) time-point 2, all participants. Increase in numbers of signals with ageing.

Ame) and lung function: quantile–quantile plots of the cross-sectional covariate-adjusted discovery epigenome-wide association study (EWAS) (Mbase#) on forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) at a) time-point 1 and b) time-point 2, all participants. Increase in numbers of signals with ageing. For FEV1/FVC, we identified 21 CpGs at time-point 2 compared with three CpGs at time-point 1 to be statistically significant. Meta-analyses were performed without genomic control (for time-point 1 inflation factor λ=1.15 and for time-point 2 inflation factor λ=1.14). For analogous figure for cross-sectional associations with FEV1 and FVC, see supplementary figure S2. c, d) Effect of smoking adjustment on the associations between DNAme and lung function: quantile–quantile plots of c) the repeat cross-sectional covariate-adjusted discovery EWAS (Mbase#; inflation factor λ=1.13) and d) additionally smoking adjusted (Msmok¶; inflation factor λ=1.05), all participants. Decrease in numbers of signals after smoking adjustment. #: base model (Mbase) EWAS was covariate adjusted for age, age squared, height, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. ¶: smoking-adjusted model (Msmok): covariates applied for Mbase and additionally smoking status and pack-years smoked.

, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. ¶: smoking-adjusted model (Msmok): covariates applied for Mbase and additionally smoking status and pack-years smoked. TABLE 3 Combined epigenome-wide association study (EWAS) meta-analyses of cross-sectional associations# of CpG markers with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in all participants: base model covariate-adjusted EWAS (Mbase¶) CpG Chr.

, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. ¶: smoking-adjusted model (Msmok): covariates applied for Mbase and additionally smoking status and pack-years smoked. TABLE 3 Combined epigenome-wide association study (EWAS) meta-analyses of cross-sectional associations# of CpG markers with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in all participants: base model covariate-adjusted EWAS (Mbase¶) CpG Chr. Position (hg19) Locus β#±se p-value meta-analysis Direction of effects+ p-value between-study heterogeneity Replicated p<0.0011§ Previously reported smoking CpGƒ Previously reported smoking pFDR-valueƒ Previously reported smoking association direction of effectƒ cg05575921 5 373 378 AHRR 0.124±0.008 7.22×10−50 +/+/+/+/+/+/+ 0.023 Yes Yes## 6.10×10−22 (−) cg03636183 19 17 000 585 F2RL3 0.201±0.015 4.50×10−43 +/+/+/+/+/+/+ 0.008 Yes Yes## 5.70×10−17 (−) cg21566642 2 233 284 661 ALPPL2 0.151±0.011 5.02×10−43 +/+/+/+/+/+/+ 0.043 Yes Yes## 4.50×10−21 (−) cg01940273 2 233 284 934 ALPPL2 0.206±0.015 4.09×10−41 +/+/+/+/+/+/+ 0.031 Yes Yes## 9.80×10−30 (−) cg03329539 2 233 283 329 ALPPL2 0.257±0.023 5.58×10−30 +/+/+/+/+/+/+ 0.628 Yes Yes 9.70×10−16 (−) cg21161138 5 399 360 AHRR 0.243±0.021 9.72×10−30 +/+/+/+/+/+/+ 0.152 Yes Yes## 7.90×10−13 (−) cg23771366 11 86 510 998 PRSS23 0.233±0.022 5.38×10−27 +/+/+/+/+/+/+ 0.286 Yes Yes 1.90×10−14 (−) cg11660018 11 86 510 915 PRSS23 0.238±0.023 3.40×10−26 +/+/+/+/+/+/+ 0.318 Yes Yes 4.40×10−21 (−) cg21611682 11 68 138 269 LRP5 0.309±0.030 1.26×10−25 +/+/+/+/+/+/+ 0.049 Yes Yes 4.20×10−15 (−) cg15342087 6 30 720 209 IER3 0.359±0.036 5.44×10−24 +/+/+/+/+/+/+ 0.169 Yes Yes 3.90×10−14 (−) cg26703534 5 377 358 AHRR 0.266±0.026 7.34×10−24 +/+/+/+/+/+/+ 0.101 Yes Yes 7.20×10−18 (−) cg25648203 5 395 444 AHRR 0.250±0.026 9.84×10−22 +/+/+/+/+/+/+ 0.194 Yes Yes 2.70×10−11 (−) cg19572487 17 38 476 024 RARA 0.196±0.021 8.87×10−21 +/+/+/+/+/+/+ 0.018 Yes Yes 1.60×10−16 (−) cg00310412 15 74 724 918 SEMA7A 0.261±0.028 4.01×10−20 +/+/+/+/+/+/+ 0.275 Yes Yes 1.20×10−13 (−) cg24859433 6 30 720 203 IER3 0.303±0.034 2.05×10−19 +/+/+/+/+/+/+ 0.067 Yes Yes## 2.20×10−9 (−) cg09935388 1 92 947 588 GFI1 0.105±0.012 7.05×10−19 +/+/+/+/+/+/+ 0.034 Yes Yes## 7.00×10−14 (−) cg14753356 6 30 720 108 IER3 0.189±0.021 9.08×10−19 +/+/+/+/+/+/+ 0.405 Yes Yes 2.30×10−14 (−) cg04885881 1 11 123 118 SRM/EXOSC10 0.168±0.020 5.66×10−18 +/+/+/+/+/+/+ 0.670 Yes Yes 2.70×10−11 (−) cg25949550 7 145 814 306 CNTNAP2 0.335±0.039 6.04×10−18 +/+/+/+/+/+/+ 0.013 Yes Yes 9.30×10−21 (−) cg19859270 3 98 251 294 GPR15 0.467±0.055 2.80×10−17 +/+/+/+/+/+/+ 0.029 Yes Yes 6.30×10−17 (−) cg

s Yes 2.30×10−14 (−) cg04885881 1 11 123 118 SRM/EXOSC10 0.168±0.020 5.66×10−18 +/+/+/+/+/+/+ 0.670 Yes Yes 2.70×10−11 (−) cg25949550 7 145 814 306 CNTNAP2 0.335±0.039 6.04×10−18 +/+/+/+/+/+/+ 0.013 Yes Yes 9.30×10−21 (−) cg19859270 3 98 251 294 GPR15 0.467±0.055 2.80×10−17 +/+/+/+/+/+/+ 0.029 Yes Yes 6.30×10−17 (−) cg 03450842 10 80 834 947 ZMIZ1 0.265±0.031 2.92×10−17 +/+/+/+/+/+/+ 0.003 Yes Yes 2.40×10−11 (−) cg03707168 19 49 379 127 PPP1R15A 0.206±0.025 1.27×10−16 +/+/+/+/+/+/+ 0.668 Yes Yes 3.50×10−7 (−) cg17087741 2 233 283 010 ALPPL2 0.161±0.020 4.48×10−16 +/+/+/+/+/+/− <0.001 Yes Yes 6.10×10−7 (−) cg21140898 1 51 442 318 CDKN2C 0.120±0.017 4.46×10−13 +/+/+/+/+/+/+ 0.103 Yes Yes 3.70×10−8 (−) cg01899089 5 369 969 AHRR 0.172±0.027 1.47×10−10 +/+/+/+/+/+/+ 0.005 Yes Yes 1.80×10−12 (−) cg08763102 4 3 079 751 HTT 0.225±0.039 1.20×10−8 +/+/+/+/+/+/− 0.001 Yes Yes 3.80×10−15 (−) cg21282907 6 74 289 980 SLC17A5 0.176±0.031 1.28×10−8 +/+/+/+/+/+/− 0.003 No Yes 1.28×10−2 (−) cg20853880 2 10 184 444 KLF11 0.077±0.014 6.05×10−8 +/+/+/+/+/+/+ 0.052 No Yes 3.70×10−7 (−) cg16391678 16 30 485 597 ITGAL 0.164±0.031 1.15×10−7 +/+/+/+/+/+/− 0.003 Yes Yes 3.00×10−11 (−) Chr.: chromosome; hg19: human genome build 19; β: coefficient of association; FDR: false discovery rate.

0.003 No Yes 1.28×10−2 (−) cg20853880 2 10 184 444 KLF11 0.077±0.014 6.05×10−8 +/+/+/+/+/+/+ 0.052 No Yes 3.70×10−7 (−) cg16391678 16 30 485 597 ITGAL 0.164±0.031 1.15×10−7 +/+/+/+/+/+/− 0.003 Yes Yes 3.00×10−11 (−) Chr.: chromosome; hg19: human genome build 19; β: coefficient of association; FDR: false discovery rate. Meta-analyses of cross-sectional associations obtained using data from the oldest time-point available: time-point 2 of ECRHS, NFBC1966, SAPALDIA and LBC1936; time-point 1 of KORA, LifeLines and NSPHS. For complete results for FEV1/FVC associations, see supplementary table S3. See supplementary tables S4 and S5 for analogous results for FEV1 and FVC, respectively. #: presentation of CpG markers showing meta-analysis p<5×10−7 in the combined meta-analysis. Note that DNA methylation predictors used were technical bias-adjusted, normalised residuals and thus effect sizes of the association (β) are not directly comparable to effect sizes reported elsewhere using normalised % methylation as predictor. ¶: base model (Mbase) epigenome-wide association study was covariate adjusted for age, age squared, height, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre as well as cell composition. +: order of cohorts: ECRHS, NFBC1966, SAPALDIA, KORA, LBC1936, LifeLines and NSPHS (FTC was excluded from this meta-analysis, given the smaller sample size and lower mean age (30.4 years) compared with the other adult cohorts (ECRHS (54.5 years), NFBC1966 (46.3 years), SAPALDIA (58.8 years) and LBC1936 (76.3 years), and the single available time-point for KORA (60.1 years), LifeLines (46.7 years) and NSPHS (55.1 years)). §: replication was defined for association if replication p<0.0011 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2]. ##: smoking CpG previously reported to mediate the effect of smoking on lung function [4].

1 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2]. ##: smoking CpG previously reported to mediate the effect of smoking on lung function [4]. Cross-sectional smoking-adjusted associations The smoking-adjusted EWAS (Msmok) resulted in fewer genome-wide significant results (figure 2c and d). Yet, despite adjustment for self-report of smoking history, the top five CpGs were known smoking-related CpGs. DNAme at cg05575921 (AHRR) remained the top cross-sectional association signal for FEV1/FVC (pcombined=2.21×10−11) (supplementary table S7). Predictive associations without smoking adjustment The prediction EWAS results (table 4 and figure 3) revealed that DNAme at time-point 1 (DNAme1) at six of nine sentinel CpGs (p<5×10−7) associated with change in FEV1/FVC was replicated (cg05575921 and cg21161138 (AHRR), cg21566642, cg01940273 and cg03329539 (vicinity of ALPPL2), and cg03636183 (F2RL3)). These six replicated CpGs were smoking-related markers. They were also associated with cross-sectional FEV1/FVC and four of them also with predicting change in FEV1 (AHRR (cg05575921), ALPPL2 (cg05951221 and cg01940273) and F2RL3 (cg03636183)) (supplementary table S8). TABLE 4 Combined meta-analyses of the prediction associations# of CpG markers on annual change in forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in all participants: base model adjustment (Mbase¶)

Predictive associations without smoking adjustment The prediction EWAS results (table 4 and figure 3) revealed that DNAme at time-point 1 (DNAme1) at six of nine sentinel CpGs (p<5×10−7) associated with change in FEV1/FVC was replicated (cg05575921 and cg21161138 (AHRR), cg21566642, cg01940273 and cg03329539 (vicinity of ALPPL2), and cg03636183 (F2RL3)). These six replicated CpGs were smoking-related markers. They were also associated with cross-sectional FEV1/FVC and four of them also with predicting change in FEV1 (AHRR (cg05575921), ALPPL2 (cg05951221 and cg01940273) and F2RL3 (cg03636183)) (supplementary table S8). TABLE 4 Combined meta-analyses of the prediction associations# of CpG markers on annual change in forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in all participants: base model adjustment (Mbase¶) CpG Chr. Position (hg19) Locus Combined meta-analysis (ECRHS/NFBC1966/SAPALDIA/KORA/LBC1936) Previously reported smoking CpGƒ Previously reported smoking pFDR-valueƒ Previously reported smoking association direction of effectƒ β#±se p-value meta-analysis Direction of effects+ p-value between-study heterogeneity Replicated p<0.0011§ cg05575921 5 373 378 AHRR 0.006±0.001 2.77×10−13 +/+/+/+/+ 0.005 Yes Yes## 6.10×10−22 (−) cg21566642 2 233 284 661 ALPPL2 0.006±0.001 3.17×10−11 +/+/+/+/+ 0.235 Yes Yes## 4.50×10−21 (−) cg01940273 2 233 284 934 ALPPL2 0.009±0.001 4.93×10−11 +/+/+/+/+ 0.023 Yes Yes## 9.80×10−30 (−) cg21161138 5 399 360 AHRR 0.011±0.002 5.81×10−9 +/+/+/+/+ 0.103 Yes Yes## 7.90×10−13 (−) cg03636183 19 17 000 585 F2RL3 0.008±0.001 6.22×10−9 +/+/+/+/+ 0.001 Yes Yes## 5.70×10−17 (−) cg01377124 2 237 172 609 ASB18 −0.018±0.003 7.38×10−8 −/−/+/−/+ 0.005 No No NA NA cg03329539 2 233 283 329 ALPPL2 0.011±0.002 7.66×10−8 +/+/+/+/+ 0.015 Yes Yes 9.70×10−16 (−) cg07222133 5 179 499 488 RNF130 −0.009±0.002 2.45×10−7 ?/−/+/−/+ <0.001 No No NA NA cg14366110 9 133 779 382 FIBCD1 0.014±0.003 9.62×10−7 +/+/+/−/− 0.206 No No NA NA Chr.: chromosome; hg19: human genome build 19; β: coefficient of association; FDR: false discovery rate; NA: not assessed. For complete results for FEV1/FVC and analogous results for FEV1 and FVC, see supplementary table S8. #: predictive associations of DNA methylation at first time-point (DNAme1) with annual change in lung function during follow-up, defined as (lung function at second time-point−lung function at first time-point)/time of follow-up (years). Presentation of CpG markers showing meta-analysis p-value<5×10−7 at discovery or combined meta-analyses level. CpGs shown sorted by statistical significance of combined meta-analysis results. Note that DNAme predictors used were technical bias-adjusted, normalised residuals and thus effect sizes of the association (β) are not directly comparable to effect sizes reported elsewhere using normalised % methylation as predictor.

el. CpGs shown sorted by statistical significance of combined meta-analysis results. Note that DNAme predictors used were technical bias-adjusted, normalised residuals and thus effect sizes of the association (β) are not directly comparable to effect sizes reported elsewhere using normalised % methylation as predictor. ¶: base model (Mbase) epigenome-wide association study was covariate adjusted for age, age squared, height, FEV1/FVC at time-point 1, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre as well as cell composition. +: order of cohorts: ECRHS, NFBC1966, SAPALDIA, KORA and LBC1936. §: replication was defined for association if replication p<0.0011 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2]. ##: smoking CpG previously reported to mediate the effect of smoking on lung function [4].

1 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2]. ##: smoking CpG previously reported to mediate the effect of smoking on lung function [4]. FIGURE 3 a) Manhattan and b) quantile–quantile plots of the covariate-adjusted prediction# epigenome-wide association study (EWAS) (Mbase¶) on forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC), all participants. Meta-analysis of the prediction association was performed without genomic control (inflation factor λ=0.95). For analogous figure for associations with change in FEV1 and FVC, see supplementary figure S4. #: predictive associations of DNA methylation at first time-point with change in lung function during follow-up. ¶: base model (Mbase) EWAS was covariate adjusted for age, age squared, height, FEV1/FVC at time-point 1, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition.

as covariate adjusted for age, age squared, height, FEV1/FVC at time-point 1, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. Associations in never-smokers The agnostic discovery EWAS (Mbase) in never-smokers, similar to the entire sample, showed more statistically significant associations at time-point 2 (older age). Eight CpGs were cross-sectionally associated with FEV1/FVC in never-smokers (p<5×10−7), but none replicated (table 5 and supplementary figure S5). The CpG cg14366110 (FIBCD1) showed predictive association of DNAme1 with change in FEV1/FVC (pdiscovery=4.2×10−9, pcombined=3.6×10−9) in never-smokers, but it did not replicate in KORA and LBC1936 (preplication=0.439; replication cohorts with lung function at two time-points). The direction of effect, however, was consistent (table 6; see supplementary table S9 for cross-sectional associations and supplementary table S10 for prediction associations) in discovery and replication cohorts. TABLE 5 Combined meta-analyses# of cross-sectional associations on forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in never-smokers only: base model adjustment (Mbase¶)

Associations in never-smokers The agnostic discovery EWAS (Mbase) in never-smokers, similar to the entire sample, showed more statistically significant associations at time-point 2 (older age). Eight CpGs were cross-sectionally associated with FEV1/FVC in never-smokers (p<5×10−7), but none replicated (table 5 and supplementary figure S5). The CpG cg14366110 (FIBCD1) showed predictive association of DNAme1 with change in FEV1/FVC (pdiscovery=4.2×10−9, pcombined=3.6×10−9) in never-smokers, but it did not replicate in KORA and LBC1936 (preplication=0.439; replication cohorts with lung function at two time-points). The direction of effect, however, was consistent (table 6; see supplementary table S9 for cross-sectional associations and supplementary table S10 for prediction associations) in discovery and replication cohorts. TABLE 5 Combined meta-analyses# of cross-sectional associations on forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in never-smokers only: base model adjustment (Mbase¶) CpG Chr. Position (hg19) Locus Combined meta-analysis (ECRHS/NFBC1966/SAPALDIA/KORA/LBC1936/Lifelines/NSPHS) Previously reported smoking CpGƒ Previously reported smoking pFDR-valueƒ Previously reported smoking association direction of effectƒ β#±se p-value meta-analysis Direction of effects+ p-value between-study heterogeneity Replicated p<0.0011§ cg09884077 15 23 086 698 NIPA1 −0.308±0.084 0.0003 −/−/−/+/−/−/− 0.001 No No NA NA cg25758394 1 3 623 859 TP73 0.213±0.083 0.0107 ?/?/+/−/−/+/− <0.001 No No NA NA cg18664508 3 169 487 465 ARPM1 −0.308±0.072 2.02×10−5 +/−/−/−/−/−/− <0.001 No No NA NA cg19268386 15 23 086 595 NIPA1 −0.263±0.140 0.0615 ?/?/−/−/−/−/− <0.001 No No NA NA cg15981995 3 169 487 311 ARPM1 −0.231±0.073 0.0016 ?/?/−/−/−/−/+ <0.001 No No NA NA cg05785298 1 204 654 622 LRRN2 −0.423±0.111 1.41×10−4 −/+/−/+/−/+/− 0.001 No No NA NA cg20278790 20 57 583 474 CTSZ 0.319±0.070 5.01×10−6 −/+/+/+/−/−/− <0.001 No No NA NA cg13562246 8 33 368 277 C8orf41 0.349±0.074 2.67×10−6 +/+/+/+/+/−/+ 0.206 No No NA NA Chr.: chromosome; hg19: human genome build 19; β: coefficient of association; FDR: false discovery rate; NA: not assessed. For complete results for FEV1/FVC and for FEV1 and FVC in never-smokers, see supplementary table S8. #: presentation of CpG markers showing meta-analysis p<5×10−7 at discovery level for cross-sectional association at time-point 2, using data from time-point 2 of ECRHS, NFBC1966, SAPALDIA and LBC1936 and from time-point 1 of KORA, LifeLines and NSPHS. Note that DNA methylation predictors used were technical bias-adjusted, normalised residuals and thus effect sizes of the association (β) are not directly comparable to effect sizes reported elsewhere using normalised % methylation as predictor. ¶: base model (Mbase) epigenome-wide association study was covariate adjusted for age, age squared, height, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre as well as cell composition. +: order of cohorts: ECRHS, NFBC1966, SAPALDIA, KORA, LBC1936, LifeLines and NSPHS.

ation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre as well as cell composition. +: order of cohorts: ECRHS, NFBC1966, SAPALDIA, KORA, LBC1936, LifeLines and NSPHS. §: replication was defined for association if replication p<0.0011 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2]. TABLE 6 Combined meta-analyses of the prediction associations# of CpG markers on annual change in forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in never-smokers only: base model adjustment (Mbase¶)

§: replication was defined for association if replication p<0.0011 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2]. TABLE 6 Combined meta-analyses of the prediction associations# of CpG markers on annual change in forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) in never-smokers only: base model adjustment (Mbase¶) CpG Chr. Position (hg19) Locus Combined meta-analysis (ECRHS/NFBC1966/SAPALDIA/KORA/LBC1936) Previously reported smoking CpGƒ Previously reported smoking pFDR-valueƒ Previously reported smoking association direction of effectƒ β#±se p-value meta-analysis Direction of effects+ p-value between-study heterogeneity Replicated p<0.0011§ cg14366110 9 133 779 382 FIBCD1 0.017±0.003 3.60×10−9 +/+/−/+/+ 0.315 No No NA NA cg11216682 2 131 113 867 PTPN18 −0.017±0.003 1.10×10−7 +/−/+/−/− 0.282 No No NA NA Chr.: chromosome; hg19: human genome build 19; β: coefficient of association; FDR: false discovery rate; NA: not assessed. For complete results for FEV1/FVC and for analogous results for FEV1 and FVC, see supplementary table S9. #: predictive associations of DNA methylation at first time-point (DNAme1) with annual change in lung function during follow-up, defined as (lung function at second time-point−lung function at first time-point)/time of follow-up (years). Presentation of CpG markers showing meta-analysis p<5×10−7 at discovery or replication level. Note that DNAme predictors used were technical bias-adjusted, normalised residuals and thus effect sizes of the association (β) are not directly comparable to effect sizes reported elsewhere using normalised % methylation as predictor. ¶: base model (Mbase) epigenome-wide association study was covariate adjusted for age, age squared, height, FEV1/FVC at time-point 1, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre as well as cell composition. +: order of cohorts: ECRHS, NFBC1966, SAPALDIA, KORA and LBC1936. §: replication was defined for association if replication p<0.0011 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2].

PALDIA, KORA and LBC1936. §: replication was defined for association if replication p<0.0011 (multiple testing correction, 47 tests for FEV1/FVC). ƒ: smoking CpGs defined on the reported FDR-corrected p<0.05 for association reported with smoking status and reported direction of effects for association with smoking [2]. Characterisation of replicated CpGs None of the not-smoking-related discovery-identified sentinel CpGs (n=25) were confirmed by replication. In contrast, 78% of the sentinel CpGs (n=86) had previously been identified as smoking related, and 57 of these (mapping to 43 loci) formally replicated across all models and lung function outcomes tested (supplementary table S11). They were used jointly for functional annotation and pathway analyses (supplementary tables S12–S16). Briefly, these 57 lung function-associated CpGs displayed enrichment for transcription factors, such as RELA (false discovery rate-adjusted p-value (pFDR)=0.002) and EP300 (pFDR=0.004), and suggestive enrichment (pFDR<0.1) for the chromatin state model of flanking active transcription start sites, of transcription at gene 5′ and 3′, and of enhancers. No significant pathways were revealed using Ingenuity Pathway Analysis database or Gene Ontology term enrichment. Transcriptional misregulation in cancer, pathways in cancer and regulation of actin cytoskeleton were identified (pFDR<0.05) using KEGG (Kyoto Encyclopedia of Genes and Genomes) pathways enrichment.

ancers. No significant pathways were revealed using Ingenuity Pathway Analysis database or Gene Ontology term enrichment. Transcriptional misregulation in cancer, pathways in cancer and regulation of actin cytoskeleton were identified (pFDR<0.05) using KEGG (Kyoto Encyclopedia of Genes and Genomes) pathways enrichment. Using the weighted Kolmogorov–Smirnov test on the entire EWAS discovery results, we noted statistically significant enrichment for smoking-related CpGs among the lung function-associated CpGs. This enrichment was also present in the smoking-adjusted EWAS and even in the EWAS restricted to never-smokers (supplementary table S17). Association of adult lung function CpG markers with childhood lung function Using the same scheme of analysis as for the adult replication cohorts, none of the sentinel CpGs showed associations with FEV1, FVC and FEV1/FVC in the childhood replication cohorts (ALSPAC and IOWBC) (supplementary table S18). The strongest associations observed in children (p<0.01) were for five CpGs not known to be smoking-related DNAme markers and one smoking-related CpG (cg00310412 (SEMA7A)).

he sentinel CpGs showed associations with FEV1, FVC and FEV1/FVC in the childhood replication cohorts (ALSPAC and IOWBC) (supplementary table S18). The strongest associations observed in children (p<0.01) were for five CpGs not known to be smoking-related DNAme markers and one smoking-related CpG (cg00310412 (SEMA7A)). Comparison with published DNAme–lung function association reports Our agnostic results were compared with previously reported lung function-specific [4, 6, 7, 12] or chronic obstructive pulmonary disease (COPD)-specific [13, 14] DNAme. We retrieved all CpGs reported being associated with lung function (n=376) for a look-up in the cross-sectional FEV1, FVC and FEV1/FVC associations at time-point 2. Only 12 out of 376 CpGs showed evidence for association (Bonferroni correction for 376 tests: p<1.3×10−4) (supplementary table S19). Notably, the most recently reported CpG markers [4, 6], having also been related to smoking, showed consistent associations with lung function, e.g. cg05575921 and cg21161138 (AHRR), cg05951221 (near ALPPL2), and cg06126421 (IER3). They were among our top replicated lung function association signals.

entary table S19). Notably, the most recently reported CpG markers [4, 6], having also been related to smoking, showed consistent associations with lung function, e.g. cg05575921 and cg21161138 (AHRR), cg05951221 (near ALPPL2), and cg06126421 (IER3). They were among our top replicated lung function association signals. Two-sample Mendelian randomisation investigation To assess the causality of replicated DNAme–lung function association, we conducted a post hoc Mendelian randomisation look-up using publicly available databases [15, 16]. Genetic instruments were identified for 12 replicated CpGs. A two-sample Mendelian randomisation on cross-sectional lung function could be completed for seven CpGs (supplementary table S20). Results support causal effects for cg23771366 and cg11660018 (PRSS23), cg21990700 (C1RL), and cg00073460 (ZC3H12D) on FEV1, and for cg00073460 (ZC3H12D) and cg24086068 (SHROOM3) on FVC. Integration of DNAme into a smoking index A recent smoking EWAS followed-up by a mediation analysis identified 10 CpGs as mediators of the smoking–lung function association [4]. Eight of these mediating CpGs were among our replicated lung function-associated CpGs (supplementary table S21). In a post hoc mediation analysis in SAPALDIA, we showed statistically significant average causal mediation on lung function for nine of these mediating CpGs (FEV1/FVC: table 7; FEV1 and FVC: supplementary table S22).

these mediating CpGs were among our replicated lung function-associated CpGs (supplementary table S21). In a post hoc mediation analysis in SAPALDIA, we showed statistically significant average causal mediation on lung function for nine of these mediating CpGs (FEV1/FVC: table 7; FEV1 and FVC: supplementary table S22). TABLE 7 Mediation# analysis on the role of previously reported CpGs in the smoking association with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC): the SAPALDIA cohort

these mediating CpGs were among our replicated lung function-associated CpGs (supplementary table S21). In a post hoc mediation analysis in SAPALDIA, we showed statistically significant average causal mediation on lung function for nine of these mediating CpGs (FEV1/FVC: table 7; FEV1 and FVC: supplementary table S22). TABLE 7 Mediation# analysis on the role of previously reported CpGs in the smoking association with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC): the SAPALDIA cohort CpG¶ Locus ACME ADE Total effect Proportion Estimate (95% CI) p-value Estimate (95% CI) p-value Estimate (95% CI) p-value Estimate (95% CI) p-value cg01940273 ALPPL2 −0.0079 (−0.0119– −0.0041) <0.0001 −0.0026 (−0.0129–0.0077) 0.604 −0.0106 (−0.0203– −0.0014) 0.026 0.7313 (0.2616–3.4325) 0.026 cg03636183 F2RL3 −0.0080 (−0.0122– −0.0040) <0.0001 −0.0029 (−0.0126–0.0062) 0.556 −0.0108 (−0.0197– −0.0021) 0.018 0.7312 (0.2819–2.9097) 0.018 cg05575921 AHRR −0.0102 (−0.0147– −0.0055) <0.0001 −0.0008 (−0.0109–0.0086) 0.870 −0.0110 (−0.0202– −0.0020) 0.012 0.9213 (0.3818–4.0453) 0.012 cg05951221 ALPPL2 −0.0075 (−0.0122– −0.0030) 0.002 −0.0033 (−0.0131–0.0062) 0.520 −0.0109 (−0.0197– −0.0022) 0.020 0.6836 (0.1942–2.7656) 0.022 cg06126421 IER3 −0.0054 (−0.0093– −0.0017) <0.0001 −0.0049 (−0.0148–0.0049) 0.328 −0.0103 (−0.0194– −0.0012) 0.030 0.5233 (0.1050–2.5558) 0.030 cg09935388 GFI1 −0.0033 (−0.0058– −0.0010) 0.002 −0.0073 (−0.0168–0.0022) 0.122 −0.0105 (−0.0198– −0.0013) 0.034 0.3009 (0.0568–1.4190) 0.036 cg21161138 AHRR −0.0056 (−0.0089– −0.0025) <0.0001 −0.0052 (−0.0146–0.0043) 0.282 −0.0108 (−0.0194– −0.0020) 0.020 0.5127 (0.1647–2.0961) 0.020 cg21566642 ALPPL2 −0.0098 (−0.0145– −0.0057) <0.0001 −0.0014 (−0.0116–0.0089) 0.796 −0.0112 (−0.0209– −0.0011) 0.024 0.8663 (0.3453–4.6567) 0.024 cg22994830 PRKAR1B −0.0002 (−0.0009–0.0003) 0.542 −0.0103 (−0.0201– −0.0010) 0.028 −0.0105 (−0.0202– −0.0013) 0.024 0.0103 (−0.0470–0.1595) 0.550 cg24859433 IER3 −0.0024 (−0.0053–0.0002) 0.068 −0.0082 (−0.0179–0.0013) 0.112 −0.0107 (−0.0201– −0.0014) 0.022 0.2186 (−0.0438–1.2776) 0.090 ACME: average causal mediation effect; ADE: average direct effect. For analogous results for FEV1 and FVC, see supplementary table S22. #: performed using the R package mediation [17]. ¶: previously reported candidate CpG for mediation of effect of smoking on lung function [4].

0014) 0.022 0.2186 (−0.0438–1.2776) 0.090 ACME: average causal mediation effect; ADE: average direct effect. For analogous results for FEV1 and FVC, see supplementary table S22. #: performed using the R package mediation [17]. ¶: previously reported candidate CpG for mediation of effect of smoking on lung function [4]. To assess the combined effect of these smoking exposure-mediating CpGs on lung function, we constructed a mediation smoking index (Mediation-SI). Its association with lung function by smoking status was tested in covariate-adjusted regression models in the discovery cohorts and following EWAS models (SAPALDIA, ECRHS and NFBC1966). Meta-analysed results of Mediation-SI showed strong association with cross-sectional FEV1/FVC in all participants and ever-smokers (table 8 and figure 4; FEV1 and FVC: supplementary table S23). Mediation-SI association in all participants was more pronounced for cross-sectional (β±se −1.2±0.13; p=2.65×10−20) than for prediction association (β±se −0.03±0.01; p=0.0072). We noted comparable associations of Mediation-SI and of pack-years with lung function (figure 5). Both were inversely associated with level of FEV1/FVC. Adding Mediation-SI or self-reported smoking history (smoking status and pack-years) to the different Mbase-adjusted statistical models showed a comparable increase in total adjusted R2. The highest total adjusted R2 was obtained when including both DNAme score and self-reported smoking history. Covariate-adjusted mean Mediation-SI values decreased from never- to ex- to current smokers and from more distant to more recent smoking exposure, with increase in pack-years in current smokers and with fewer years since quitting in ex-smokers (figure 6).

en including both DNAme score and self-reported smoking history. Covariate-adjusted mean Mediation-SI values decreased from never- to ex- to current smokers and from more distant to more recent smoking exposure, with increase in pack-years in current smokers and with fewer years since quitting in ex-smokers (figure 6). TABLE 8 Meta-analyses# of the discovery cohort-specific association of mediation smoking index (Mediation-SI) with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) (%), cross-sectionally at time-point 2 and longitudinally predicting the annual change during follow-up, in all study participants, ever and never-smokers: base model adjustment (Mbase¶)

pecific association of mediation smoking index (Mediation-SI) with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) (%), cross-sectionally at time-point 2 and longitudinally predicting the annual change during follow-up, in all study participants, ever and never-smokers: base model adjustment (Mbase¶) Cross-sectional meta-analysis at time-point 2# Prediction on change in lung function¶ β±se p-value+ Direction of effects§ p-value between-study heterogeneity β±se p-value+ Direction of effects§ p-value between-study heterogeneity All −0.012±0.0013 1.05×10−20 −/−/− 0.44 −0.0005±0.0001 8.66×10−9 −/−/− 0.006 Ever-smokers −0.014±0.0016 3.28×10−18 −/−/− 0.30 −0.0004±0.0001 4.94×10−4 −/−/− 0.13 Never-smokers −0.0033±0.0041 0.423 −/−/+ 0.62 −0.0007±0.0002 1.73×10−4 +/−/+ 0.003 β: coefficient of association. For analogous results of associations of Mediation-SI with FEV1 and FVC, see supplementary table S23. #: cohort-specific association results for Mediation-SI were meta-analysed. The 10 CpGs contributing Mediation-SI values are shown in supplementary table S21. Note that DNA methylation predictors used were technical bias-adjusted, normalised residuals and thus effect sizes of the association (β) are not directly comparable to effect sizes reported elsewhere using normalised % methylation as predictor. ¶: base model (Mbase) covariate adjustment: age, age squared, height, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre as well as cell composition. Prediction models were additionally adjusted for FEV1/FVC at time-point 1. +: p-value of meta-analysis: p<0.008 was considered statistically significant, Bonferroni correction for six tests per lung function outcome. §: order of cohorts: ECRHS, NFBC1966 and SAPALDIA.

eter type, study centre as well as cell composition. Prediction models were additionally adjusted for FEV1/FVC at time-point 1. +: p-value of meta-analysis: p<0.008 was considered statistically significant, Bonferroni correction for six tests per lung function outcome. §: order of cohorts: ECRHS, NFBC1966 and SAPALDIA. FIGURE 4 Forest plots of cohort-specific results and meta-analyses of the association of the mediation smoking index (Mediation-SI) with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) and change in FEV1/FVC in a, b) ever-smokers and c, d) never-smokers in the discovery cohorts: a, c) time-point 2 and b, d) prediction. Associations run applying base model adjustment (Mbase#). #: base model (Mbase) epigenome-wide association study was covariate adjusted for age, age squared, height, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. Prediction models were additionally adjusted for FEV1/FVC at time-point 1.

t, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. Prediction models were additionally adjusted for FEV1/FVC at time-point 1. FIGURE 5 Distribution and association# of a) mediation smoking index (Mediation-SI)¶ and b) self-reported smoking history (pack-years) with forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) with 95% confidence intervals (shaded). Box plots of a) Mediation-SI (median (range) 0.3 (−1.7–5.2)) and b) pack-years (median (range) 2.0 (0–145.9)) in all participants of SAPALDIA are shown at the top of each panel. Red dotted lines indicate box plot interquartile range (IQR) borders. Whiskers indicate 1.5 IQR of the lower and upper quartile; outliers are indicated. For analogous figures for associations of Mediation-SI with FEV1 and FVC, see supplementary figures S6 and S7, respectively. #: associations were adjusted for the base model (Mbase): age, age squared, height, squared deviation from the mean of height, sex and interaction terms of age, age squared, height and squared deviation of height with sex, education (low, medium and high), body mass index, spirometer type, study centre, and cell composition. ¶: Mediation-SI can be constructed for all participants irrespective of their smoking status. The Mbase-adjusted model explained 17.5% of the variance in the outcome. The Mbase-adjusted model additionally adjusted for the Mediation-SI explained 19.6% of the FEV1/FVC variance (total adjusted R2=0.196) of which 2.8% of the variance was specifically explained by the Mediation-SI variable. This was comparable to the variance explained by the Mbase-adjusted model additionally adjusted for pack-years and smoking status corresponding to the Msmok model (R2=0.198, and with 1.6% of the variance specifically explained by the pack-years variable). Model including both smoking adjustments (Msmok and additionally Mediation-SI) explained 20.1% of the FEV1/FVC variance.

se-adjusted model additionally adjusted for pack-years and smoking status corresponding to the Msmok model (R2=0.198, and with 1.6% of the variance specifically explained by the pack-years variable). Model including both smoking adjustments (Msmok and additionally Mediation-SI) explained 20.1% of the FEV1/FVC variance. FIGURE 6 Distribution of adjusted mediation smoking index (Mediation-SI) in SAPALDIA at time-point 2. a) Smoking status: adjusted for age, sex and education. Never-smokers (n=395), ex-smokers (n=356) and current smokers (n=211). b) Years since quitting: adjusted for age, sex, education, pack-years and cigarettes per day. Ex-smokers (n=356). c) Pack-years: adjusted for age, sex, education and cigarettes per day. Current smokers (n=211). d) Cigarettes per day: adjusted for age, sex, education and pack-years. Current smokers (n=211). Data are presented as median with interquartile range (IQR) (boxes) and 1.5 IQR of the lower and upper quartile (whiskers); outliers are indicated. The assessment of a second DNAme smoking score (Lung-Function-Genes-SI), based on smoking-related CpGs located in 18 GWAS-identified lung function candidate genes (supplementary table S24), showed less prominent associations with lung function (strongest association observed in ever-smokers for FEV1: β±se −0.196±0.053; p=0.0002) (supplementary table S25).

ing score (Lung-Function-Genes-SI), based on smoking-related CpGs located in 18 GWAS-identified lung function candidate genes (supplementary table S24), showed less prominent associations with lung function (strongest association observed in ever-smokers for FEV1: β±se −0.196±0.053; p=0.0002) (supplementary table S25). Discussion The understanding of how environmental exposure and disease are related to site-specific DNAme status is growing [18, 19]. Our agnostic EWAS on lung function contributes to this body of evidence. Lung function-associated DNAme markers were strongly enriched for smoking-associated loci. More than 50 known smoking CpGs were consistently, and in several cases causally, associated with lung function and its change in adults. The current agnostic approach converges with recent results of DNAme–lung function studies [4, 6, 7] that were a priori focusing on smoking-related loci, and included pyrosequencing in blood [7] and lung tissue [4] of some of our strongest association signals, including AHRR hypomethylation at cg05575921 and cg21161138, cg05951221 and cg21566642 (ALPPL2), and cg06126421 (IER3). A methylation index integrating 10 DNAme that reportedly mediate the effect of smoking on lung function [4] was associated with lung function level and its change in adults.

ngest association signals, including AHRR hypomethylation at cg05575921 and cg21161138, cg05951221 and cg21566642 (ALPPL2), and cg06126421 (IER3). A methylation index integrating 10 DNAme that reportedly mediate the effect of smoking on lung function [4] was associated with lung function level and its change in adults. Smoking is an important risk factor for poor lung function and accelerated decline. Several EWASs identified a large number of differentially methylated CpG markers to be associated with smoking [2–4]. In particular, the hypomethylation of cg05575921, a CpG located in the third intron of the aryl hydrocarbon receptor repressor (AHRR) gene, investigated for lung function and respiratory symptoms [4], stands out as a robust indicator of smoking status and smoking history [20]. Given the consistency of the associations observed for cg05575921 and the smoking index containing it in this study, the latter may have potential as a biomarker of clinical utility in predicting smoking-related morbidity and mortality [20, 21]. The positive direction of effects observed in identified DNAme–lung function association is in accordance with the reported hypomethylation of smoking-related DNAme sites. The identified lung function-associated CpGs in this study have been previously reported to be associated with smoking-related molecular phenotypes [22], with increased risk of noncommunicable disease, including cancer [20, 23], and with epigenetically defined accelerated ageing [24].

on of smoking-related DNAme sites. The identified lung function-associated CpGs in this study have been previously reported to be associated with smoking-related molecular phenotypes [22], with increased risk of noncommunicable disease, including cancer [20, 23], and with epigenetically defined accelerated ageing [24]. Whether most smoking-related DNAme markers are only markers of exposure or indirectly associated with lung function [7] or whether some inform on causal disease pathways cannot be answered conclusively by the current study. First, DNAme may just be a more precise measure of smoking exposure than self-reporting, as AHRR DNAme was previously shown to correlate with genetic smoking dependency [20]. Second, DNAme identified by previous smoking EWASs [2, 4] may not exclusively have picked up methylation effects of smoking, but methylation related to phenotypes also affected by smoking. In this case, the observed DNAme–lung function associations may result from comorbidity between lung function and other smoking-related phenotypes. However, some of the results are consistent with a causal disease pathway. First, Mendelian randomisation results support causal effects from some DNAme. Unfortunately, no genetic instrument was available for the top ranked AHRR signal. Second, our report confirms nine CpGs, including cg05575921 (AHRR), previously shown to mediate the effect of smoking of lung function [4]. The observation that many smoking DNAme–lung function associations withstood smoking adjustment is consistent with the mediating role of DNAme between smoking behaviour (more distant predictor) and lung function. Third, smoking was also observed to influence methylation in lung tissue at several lung function CpGs, including at cg05575921 in AHRR, and these methylation levels correlated with AHRR gene expression [25] and expression of other genes [4]. Hypotheses for a mediating and causal role of smoking-related DNAme include altered AHRR DNAme inducing altered phase 2 enzyme activity and toxicant metabolism, and altered inflammatory pathways in the lung [7]. Other inhalants impacting on the same pathways could in part explain the observed enrichment for smoking DNAme among never-smokers. Methylation of AHRR cg05575921 was previously associated with lung function and chronic bronchitis in never-smokers [7]. Maternal smoking, passive smoking and environmental exposures other than cigarette smoking (e.g.

the same pathways could in part explain the observed enrichment for smoking DNAme among never-smokers. Methylation of AHRR cg05575921 was previously associated with lung function and chronic bronchitis in never-smokers [7]. Maternal smoking, passive smoking and environmental exposures other than cigarette smoking (e.g. air pollution) are known to modify DNAme patterns across the genome [26–32]. Maternal smoking during pregnancy has been shown to alter the offspring's DNA markers in a number of genes known to contain smoking-related CpGs [27, 28] and some of these epigenetic patterns, including in AHRR, persist to adulthood [29].

the same pathways could in part explain the observed enrichment for smoking DNAme among never-smokers. Methylation of AHRR cg05575921 was previously associated with lung function and chronic bronchitis in never-smokers [7]. Maternal smoking, passive smoking and environmental exposures other than cigarette smoking (e.g. air pollution) are known to modify DNAme patterns across the genome [26–32]. Maternal smoking during pregnancy has been shown to alter the offspring's DNA markers in a number of genes known to contain smoking-related CpGs [27, 28] and some of these epigenetic patterns, including in AHRR, persist to adulthood [29]. From our findings in two well-characterised childhood birth cohorts, there was no evidence for shared common epigenetic mechanisms underlying lung function in adults and children. The comparison was driven by results from the lung function EWAS in adults, given sample size limitations in the available birth cohorts. Lung function in childhood versus adulthood is expected to be influenced in part by different biological processes. The nonreplication of the mostly smoking-related lung function DNAme signals might reflect the nonsmoking status of the children and adolescents. Our findings in SAPALDIA point to a dose–response effect of smoking history and intensity on the smoking index. Effects of maternal exposure in utero, passive smoking or other inhalants on smoking DNAme are likely smaller than the effects of active smoking [30]. Our EWAS findings generally showed an age-related increase in number and strength of DNAme–lung function associations in adults, despite covariate adjustment for age, as also observed by others [6]. This result is consistent with the observed dose–response effect of smoking and possibly other inhalants on DNAme. However, the inherent interdependency of lung function decline, cumulative smoking exposure and DNAme with ageing prohibits attributing associations to single factors.

age, as also observed by others [6]. This result is consistent with the observed dose–response effect of smoking and possibly other inhalants on DNAme. However, the inherent interdependency of lung function decline, cumulative smoking exposure and DNAme with ageing prohibits attributing associations to single factors. A systematic review of peripheral DNAme associated with lung function in population-based cohorts pointed to the lack of consistent evidence [5]. Epigenome-wide DNAme profiling studies of lung tissue suggested DNAme in genes such as NOS1AP, TNFAIP2 and CHRM1 to be associated with COPD [13, 14]. An EWAS meta-analysis, adjusted for smoking status and pack-years, identified differential DNAme related to COPD and lung function in Koreans. Five loci (CTU2, USP36, ZNF516, KLK10 and CPT1B) were associated with at least two respiratory traits [12]. Evidence of associations in the current EWAS was only observed for 12 out of 376 CpGs associated with lung function phenotypes in these previous studies. This inconsistency may be due to differences in population ancestry, disease status, exposure status, tissue-specific methylation or covariate adjustment. Furthermore, limited sample size and false discovery findings could contribute to nonreplication, as could the absence of post-bronchodilation lung function in the current EWAS. However, our results confirm the associations of two recently published population-based reports [4, 6] investigating smoking, DNAme and lung function. Both reports and our results reveal the same smoking CpGs as prominent signals.

ication, as could the absence of post-bronchodilation lung function in the current EWAS. However, our results confirm the associations of two recently published population-based reports [4, 6] investigating smoking, DNAme and lung function. Both reports and our results reveal the same smoking CpGs as prominent signals. The strength of this EWAS investigation is the robust and extensive study design with availability of repeat measures of DNAme and spirometry data in the same cohort participants, as well as its population-based design. The utilisation of a multilevel analysis scheme, including cross-sectional and longitudinal EWAS analyses at two time-points in the same participants, and EWAS with and without smoking adjustment in all participants and in never-smokers, allowed for a better understanding of lung function DNAme being affected by ageing and smoking. The lung function-associated smoking index derived is building on robust evidence that DNAme in blood is correlated with DNAme and gene expression in lung tissue [4, 23, 33], and that it is a valid biomarker for capturing the effect of smoking on DNAme in the lung [7, 20].

on DNAme being affected by ageing and smoking. The lung function-associated smoking index derived is building on robust evidence that DNAme in blood is correlated with DNAme and gene expression in lung tissue [4, 23, 33], and that it is a valid biomarker for capturing the effect of smoking on DNAme in the lung [7, 20]. There are several limitations to this study. Limitations in sample size may explain the inability to find association signals in never-smokers and therefore signals common to lung function in childhood and adulthood. The estimation of decline in lung function from only two spirometry time-points is likely to misclassify decline. Additionally, not all replication cohorts had data available for more than one time-point. Pre-bronchodilation lung function is less robust than post-bronchodilator values and may increase variability of the findings. The meta-analysed EWAS results of the cross-sectional analyses showed evidence of inflation (inflation factor λ>1.1) indicating insufficient genomic control; however, adjusting for genomic inflation did not alter our main results. The relevance of the smoking index derived from CpGs in or close to lung function GWAS genes can be questioned given evidence on the complex trans-regulation of gene expression [34].

inflation factor λ>1.1) indicating insufficient genomic control; however, adjusting for genomic inflation did not alter our main results. The relevance of the smoking index derived from CpGs in or close to lung function GWAS genes can be questioned given evidence on the complex trans-regulation of gene expression [34]. In conclusion, our agnostic investigation shows that DNAme at CpGs related to smoking behaviour are the predominant signals associated cross-sectionally and prospectively with lung function in adults. The findings stimulate further research into the involvement of smoking-related CpGs in lung function-relevant mechanisms and potentially their role as exposure markers beyond active smoking. From our EWAS results it has become clear that larger samples are required to confidently identify CpGs involved in lung function and its age-related decline in persons who never smoked. Supplementary material 10.1183/13993003.00457-2019.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-00457-2019.Supplement Spreadsheet containing additional tables S3, S4, S5, S9 and S19. ERJ-00457-2019.Additional_Tables Acknowledgements For a full list of acknowledgements for each cohort, please refer to the supplementary material. We gratefully acknowledge the contribution of co-author John M. Starr, who died prior to the publication of this manuscript. This article has supplementary material available from erj.ersjournals.com

Spreadsheet containing additional tables S3, S4, S5, S9 and S19. ERJ-00457-2019.Additional_Tables Acknowledgements For a full list of acknowledgements for each cohort, please refer to the supplementary material. We gratefully acknowledge the contribution of co-author John M. Starr, who died prior to the publication of this manuscript. This article has supplementary material available from erj.ersjournals.com Data available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.pr10c20 Conflict of interest: M. Imboden has nothing to disclose. Conflict of interest: M. Wielscher has nothing to disclose. Conflict of interest: F.I. Rezwan has nothing to disclose. Conflict of interest: A.F.S. Amaral has nothing to disclose. Conflict of interest: E. Schaffner has nothing to disclose. Conflict of interest: A. Jeong has nothing to disclose. Conflict of interest: A. Beckmeyer-Borowko has nothing to disclose. Conflict of interest: S.E. Harris reports grants from Medical Research Council, Biotechnology and Biological Sciences Research Council, Age UK and The Wellcome Trust, during the conduct of the study. Conflict of interest: J.M. Starr has nothing to disclose. Conflict of interest: I.J. Deary reports grants from Age UK and Medical Research Council, during the conduct of the study. Conflict of interest: C. Flexeder has nothing to disclose. Conflict of interest: M. Waldenberger has nothing to disclose. Conflict of interest: A. Peters has nothing to disclose. Conflict of interest: H. Schulz reports grants from German Federal Ministry of Education and Research (BMBF), during the conduct of the study.

Conflict of interest: I.J. Deary reports grants from Age UK and Medical Research Council, during the conduct of the study. Conflict of interest: C. Flexeder has nothing to disclose. Conflict of interest: M. Waldenberger has nothing to disclose. Conflict of interest: A. Peters has nothing to disclose. Conflict of interest: H. Schulz reports grants from German Federal Ministry of Education and Research (BMBF), during the conduct of the study. Conflict of interest: S. Chen has nothing to disclose. Conflict of interest: S.K. Sunny has nothing to disclose. Conflict of interest: W.J.J. Karmaus has nothing to disclose. Conflict of interest: Y. Jiang has nothing to disclose. Conflict of interest: G. Erhart has nothing to disclose. Conflict of interest: F. Kronenberg has nothing to disclose. Conflict of interest: R. Arathimos has nothing to disclose. Conflict of interest: G.C. Sharp has nothing to disclose. Conflict of interest: A.J. Henderson reports grants from Medical Research Council and Wellcome Trust, during the conduct of the study. Conflict of interest: Y. Fu has nothing to disclose. Conflict of interest: P. Piirilä has nothing to disclose. Conflict of interest: K.H. Pietiläinen has nothing to disclose. Conflict of interest: M. Ollikainen has nothing to disclose. Conflict of interest: A. Johansson has nothing to disclose. Conflict of interest: U. Gyllensten has nothing to disclose. Conflict of interest: M de Vries has nothing to disclose. Conflict of interest: D.A. van der Plaat has nothing to disclose. Conflict of interest: K. de Jong has nothing to disclose. Conflict of interest: H.M. Boezen has nothing to disclose.

Conflict of interest: A. Johansson has nothing to disclose. Conflict of interest: U. Gyllensten has nothing to disclose. Conflict of interest: M de Vries has nothing to disclose. Conflict of interest: D.A. van der Plaat has nothing to disclose. Conflict of interest: K. de Jong has nothing to disclose. Conflict of interest: H.M. Boezen has nothing to disclose. Conflict of interest: I.P. Hall reports grants from GSK and Boehringer Ingelheim, outside the submitted work. Conflict of interest: M.D. Tobin reports grants from Pfizer and GSK, outside the submitted work. Conflict of interest: M-R. Jarvelin has nothing to disclose. Conflict of interest: J.W. Holloway reports grants from European Union and National Institutes of Health, during the conduct of the study. Conflict of interest: D. Jarvis reports grants from European Union, Medical Research Council and Asthma UK, during the conduct of the study. Conflict of interest: N.M. Probst-Hensch has nothing to disclose. Support statement: This work has been conducted within the Aging Lungs in European Cohorts (ALEC) project, funded from the European Union's Horizon 2020 research and innovation programme under grant agreement 633212. The funding agency had no role in the design, data collection and analysis of the data. Cohort-specific funding details are provided in the supplementary material. Funding information for this article has been deposited with the Crossref Funder Registry.

To the Editor: The advent of antifibrotic agents [1, 2] as standard of care in idiopathic pulmonary fibrosis (IPF) requires that new non-inferiority IPF drug trials will need to identify smaller declines of forced vital capacity (FVC). Marginal annualised FVC declines (between 5.00 and 9.99%) are particularly challenging to interpret as they might reflect measurement variation or genuine clinical deterioration [3]. Following on from previous baseline-only computed tomography (CT) analyses [4], the current study examined whether changes in computer features (CALIPER) across serial CT examinations could be considered as a trial co-endpoint, particularly with regard to adjudicating marginal FVC declines, and therefore improve the sensitivity of IPF drug trials.

seline-only computed tomography (CT) analyses [4], the current study examined whether changes in computer features (CALIPER) across serial CT examinations could be considered as a trial co-endpoint, particularly with regard to adjudicating marginal FVC declines, and therefore improve the sensitivity of IPF drug trials. Previous baseline IPF analyses identified that variable initiation time, dosages, durations and types of antifibrotic medication in study participants had a profound confounding effect on mortality relationships [4]. Consequently, analyses in the current manuscript were restricted to IPF patients not receiving anti-fibrotic therapy (discovery cohort: n=71 Royal Brompton Hospital patients presenting from January 2007 to December 2014); validation cohort: n=24 St Antonius Hospital, Nieuwegein patients presenting from January 2005 to June 2014 and n=23 Mayo Clinic Rochester patients presenting from January 2009 to June 2015). All patients had two non-contrast volumetric CT scans between 5 and 30 months apart (mean CT interval: discovery cohort 1.1 years; validation cohort 1.2 years) as part of their clinical care. Baseline diffusion capacity of the lung for carbon monoxide (DLCO) and FVC (baseline and longitudinal) were collected if performed within 3 months of the respective CTs. No patients were lost to follow-up.

apart (mean CT interval: discovery cohort 1.1 years; validation cohort 1.2 years) as part of their clinical care. Baseline diffusion capacity of the lung for carbon monoxide (DLCO) and FVC (baseline and longitudinal) were collected if performed within 3 months of the respective CTs. No patients were lost to follow-up. Annualised FVC change was measured using a linear mixed effects model on all eligible timepoints to derive the best linear unbiased predictor (BLUP) as previously described using the lmer function from the R package lme4 [5]. A naïve estimate of FVC change was also examined using FVC measurements at the first and second CT timepoint. For the naïve estimate we computed annual relative change by dividing the absolute annual change by the baseline FVC value (relative). Dichotomised relative FVC declines (≥5% or ≥10%) were derived based on the naïve and BLUP estimates. Of the 27 CALIPER features examined [4], nine were measured on a whole lung level: total lung volume, normal parenchyma, vessel-related structures (CAL VRS), emphysema, honeycombing, reticular pattern and ground-glass opacity. Fibrosis extent summed reticular pattern and honeycombing. Interstitial lung disease extent additionally summed ground-glass opacification. 18 CAL VRS subdivisions were evaluated, separated according to lung zonal location: upper (UZ), middle (MZ) and lower zones (LZ), and cross-sectional area of structures in each zone: <5 mm−2, 5–10 mm−2, 10–15 mm−2, 15–20 mm−2, >20 mm−2. Volumes for all CALIPER features were converted into a percentage using CALIPER-derived total lung volume measurements [6, 7]. Absolute change in the derived 27 CT variables was annualised by dividing by the time interval between the two measurements (in years). Cox proportional hazards models examined CALIPER and FVC change variables in separate discovery and validation cohorts. Time was measured from the second CT. An event was either death (n=90) or transplantation (n=8). Each predictor variable was tested alone while correcting for patient age (at the second CT) and gender. Model fit was evaluated using the concordance index, which assesses how well the ordering of subjects for the actual time of the event agrees with the predicted time of the event. That is, for all subject pairs it checks whether the subject who had the event first was also the subject predicted to have the event first.

Model fit was evaluated using the concordance index, which assesses how well the ordering of subjects for the actual time of the event agrees with the predicted time of the event. That is, for all subject pairs it checks whether the subject who had the event first was also the subject predicted to have the event first. A C-index of 0.5 indicates random performance, where in 50% of cases the subject with the earlier event was predicted to be the subject with the later event. Approval for this study of clinically indicated CT and pulmonary function data was obtained from Liverpool Research Ethics Committee (reference: 14/NW/0028) and the Institutional Ethics Committee of the Royal Brompton Hospital, Mayo Clinic Rochester and St Antonius Hospital, Nieuwegein. Informed patient consent was not required. Our study findings identified absolute CAL VRS and UZ VRS increases as the strongest survival predictors in discovery and validation cohorts (figure 1a). Both variables were at least equivalent to FVC change when evaluated using C-indices. Significant but weak correlations (r=−0.42, p=1.8×10−6) were identified between FVC change and absolute VRS change (Pearson's correlation).

ncreases as the strongest survival predictors in discovery and validation cohorts (figure 1a). Both variables were at least equivalent to FVC change when evaluated using C-indices. Significant but weak correlations (r=−0.42, p=1.8×10−6) were identified between FVC change and absolute VRS change (Pearson's correlation). FIGURE 1 a) Scatterplot of -log10 p-values for various computer-derived (CALIPER) variables (blue points) and forced vital capacity (FVC) decline (green points) in patients not exposed to antifibrotic medication in the discovery cohort (x-axis, n=71) and validation cohort (y-axis, n=47). Horizontal and vertical dotted lines represent the Li and Ji corrected cut-off for statistical significance. FVC decline was calculated using two methods: naïve estimate from two timepoints aligned with the two computed tomography (CT) timepoints (simple) and using best linear unbiased predictions. FVC change was expressed as a continuous variable (FVC change), and at ≥5% decline and ≥10% decline thresholds. The FVC value at the timepoint of the second CT scan (red dot) was used to benchmark expressions of FVC decline. The pulmonary vessel-related structure score (CAL VRS) was subdivided according to zonal location (UZ VRS: upper zone; MZ VRS: middle zone; LZ VRS: lower zone) and structure cross-sectional area in each zone (<5 mm−2, 5–10 mm−2, 10–15 mm−2, 15–20 mm−2, >20 mm−2). b) C-indices (a measure of goodness of fit of a model) for models examining thresholds of change in CAL VRS examined against a 10% FVC decline threshold. The horizontal dotted black line indicates the C-index for a 10% FVC decline threshold model examining the relevant FVC threshold alone. The blue line demonstrates the C-indices for models when a CAL VRS threshold alone was examined. The red line demonstrates the C-indices for models where a binary variable indicated a “joint endpoint”, i.e. either the CALIPER or FVC threshold had been reached. c) Additional patients that would reach an endpoint (y-axis), if CAL VRS (red) or upper-zone vessel related structure (UZ VRS, blue) thresholds of change (x-axis) were examined in addition to FVC decline thresholds. The FVC decline thresholds examined included a ≥5% FVC decline threshold (solid line) and a ≥10% FVC decline threshold (dotted line). d) C-indices (y-axis) for models containing varying thresholds (x-axis) of CAL VRS (red) or UZ VRS (blue) in patients with an FVC between 5% and 10%. The horizontal dashed black line indicates the C-index 0.5, i.e.

xamined included a ≥5% FVC decline threshold (solid line) and a ≥10% FVC decline threshold (dotted line). d) C-indices (y-axis) for models containing varying thresholds (x-axis) of CAL VRS (red) or UZ VRS (blue) in patients with an FVC between 5% and 10%. The horizontal dashed black line indicates the C-index 0.5, i.e. random performance. Both study populations were then combined and Cox proportional hazards models examined thresholds of CAL VRS and UZ VRS change measured against relative FVC decline thresholds of ≥5% and ≥10% adjusted for patient age and gender. The predictive performance of FVC-based indicator variables was compared to VRS indicator variables, either used alone or when combined with an FVC-based indicator variable as a “joint endpoint”. The joint endpoint reflected whether the FVC decline or the VRS increase was achieved with estimates based on 500 bootstrap replicates (n=118). We estimated the number of additional patients that would reach either a ≥5% or ≥10% predicted FVC threshold or a preselected CAL VRS/UZ VRS change threshold in a drug trial setting. Further, we computed the Kaplan–Meier estimator for different subgroups of patients (n=118) depending on whether they reached the FVC or VRS threshold or both (using the SPSS Kaplan–Meier function [8]). Finally, we examined mortality prediction in patients with a BLUP estimated relative decline in FVC of >5% but <10% not receiving antifibrotics (n=41).

ier estimator for different subgroups of patients (n=118) depending on whether they reached the FVC or VRS threshold or both (using the SPSS Kaplan–Meier function [8]). Finally, we examined mortality prediction in patients with a BLUP estimated relative decline in FVC of >5% but <10% not receiving antifibrotics (n=41). In the combined study population, multivariate Cox mortality models demonstrated that a CAL VRS increase of ≥0.30% independently predicted mortality when evaluated against a ≥10% FVC decline threshold. When CAL VRS increased by ≥0.50%, a ≥10% FVC decline threshold no longer significantly contributed to mortality prediction. At CAL VRS (figure 1b) and UZ VRS increases of ≥0.40%, no difference in model C-index was seen when compared to a ≥10% FVC decline threshold. The C-index was unchanged when using either a solitary CALIPER endpoint (CAL VRS or UZ VRS increase), or a combined endpoint of an increase in a CALIPER variable and an FVC ≥10% decline threshold. Results were maintained when CALIPER variable change thresholds were compared to a ≥5% FVC decline threshold.

ne threshold. The C-index was unchanged when using either a solitary CALIPER endpoint (CAL VRS or UZ VRS increase), or a combined endpoint of an increase in a CALIPER variable and an FVC ≥10% decline threshold. Results were maintained when CALIPER variable change thresholds were compared to a ≥5% FVC decline threshold. 79/118 (67%) patients reached a CAL VRS of ≥0.40% change whilst 54/118 (46%) reached a ≥10% FVC decline threshold (p=0.0003). 89/118 (75%) patients reached either the CAL VRS threshold of ≥0.40% change or ≥10% FVC decline threshold (figure 1c). Use of a CAL VRS increase threshold of ≥0.40% change identified 35/118 (30%) more patients reaching an endpoint than the ≥10% FVC decline threshold alone. Similarly, at least 30% more patients reached an endpoint when an UZ VRS threshold was used alongside a ≥10% FVC decline threshold (figure 1c). When CAL VRS and UZ VRS elevation thresholds were examined against a ≥5% FVC decline threshold, additional patients reaching an endpoint were again identified. When all patients with an FVC decline more than 5% and less than 10% were subanalysed, CAL VRS thresholds ≥0.40% change demonstrated C-indices that were at least equivalent to a ≥10% FVC decline threshold (figure 1d).

d against a ≥5% FVC decline threshold, additional patients reaching an endpoint were again identified. When all patients with an FVC decline more than 5% and less than 10% were subanalysed, CAL VRS thresholds ≥0.40% change demonstrated C-indices that were at least equivalent to a ≥10% FVC decline threshold (figure 1d). Our findings demonstrate that in independent discovery and validation populations, an absolute increase in a computer-derived variable, the vessel-related structures (CAL VRS), strongly predicts mortality in IPF patients not exposed to antifibrotic medication. Patients exhibiting a CAL VRS increase ≥0.40% were different to those experiencing an FVC decline ≥10%. Accordingly, if a composite endpoint of CAL VRS ≥0.40% increase and/or ≥10% FVC decline were used in a drug trial setting, 30% more patients would reach the composite endpoint than a solitary endpoint of ≥10% FVC decline. Our findings also suggest the utility of a CAL VRS threshold ≥0.40% increase as an arbitration tool for marginal FVC declines (between 5.0 and 9.9%).

40% increase and/or ≥10% FVC decline were used in a drug trial setting, 30% more patients would reach the composite endpoint than a solitary endpoint of ≥10% FVC decline. Our findings also suggest the utility of a CAL VRS threshold ≥0.40% increase as an arbitration tool for marginal FVC declines (between 5.0 and 9.9%). The weak correlations between FVC change and VRS change indicate that both variables represent important yet distinct surrogate measures of mortality and argues for their integration as co-endpoints rather than selecting one over another. A ≥0.40% increase in VRS across a cohort appeared to be the most accurate measure of change in VRS, when considering both its prognostic effect when judged against FVC decline and its sensitivity as an endpoint. In an individual, whilst the most accurate threshold for VRS change may also be a ≥0.40% threshold, further work is necessary to establish optimal thresholds for use in clinical practice, as just having knowledge of the range of change of a variable does not of course provide any statement of the clinical significance of that change. For example, it was noticeable that more extreme VRS cut-offs, e.g. 0.75%, made even more of a difference in model fit and C-index than a ≥0·40% threshold. But we cannot know how often such a magnitude of VRS change would be seen in a clinical trial population. A logical next analytic step would therefore be to evaluate VRS change in a well-controlled drug trial population receiving antifibrotics at a standardised dosing regimen.

l fit and C-index than a ≥0·40% threshold. But we cannot know how often such a magnitude of VRS change would be seen in a clinical trial population. A logical next analytic step would therefore be to evaluate VRS change in a well-controlled drug trial population receiving antifibrotics at a standardised dosing regimen. The validity of VRS change was considered according to the OMERACT filter criteria for IPF clinical trial domains [9]. Regarding truth and discrimination criteria, VRS change was considered to be more discriminatory than FVC change at predicting outcome, with potential for use as a continuous variable (with no loss of signal strength), or as a binary threshold alongside an FVC decline threshold to improve endpoint sensitivity. The variable therefore satisfies construct, content and criterion validity and demonstrates sensitivity to change. The specific impact on VRS change of differing inspiratory effort, acquisition or reconstruction parameters has not been systematically investigated, and further study is indicated. However, our analysis of this measure in a heterogeneous dataset from multiple institutions suggests this is robust. CALIPER outputs are eminently interpretable and feasible to perform but real-world utility of VRS for clinical trials relies on availability of repeated CTs and the computer algorithm, and is therefore limited when compared to FVC measurements.

a heterogeneous dataset from multiple institutions suggests this is robust. CALIPER outputs are eminently interpretable and feasible to perform but real-world utility of VRS for clinical trials relies on availability of repeated CTs and the computer algorithm, and is therefore limited when compared to FVC measurements. There were limitations to the current study. Though there were similar average CT intervals between the two study cohorts and change in CT variables were reported as annualised change, the CTs time intervals were not standardised in this retrospective analysis. This lack of standardisation reflects real world clinical practice but may have biased our findings in patients with shorter or longer CT follow up intervals. Whilst the ideal study would have rigorous protocol-led control of serial CT and functional measurements and antifibrotic use, no such study yet exists and were it to begin today, outcome data may only be available several years hence. Accordingly, we believe our analyses capture a realistic contemporary cross-section of IPF data points.

udy would have rigorous protocol-led control of serial CT and functional measurements and antifibrotic use, no such study yet exists and were it to begin today, outcome data may only be available several years hence. Accordingly, we believe our analyses capture a realistic contemporary cross-section of IPF data points. In conclusion, we have demonstrated for the first time that change in a computer-derived variable, vessel-related structures, which has no visual correlate is a powerful surrogate for mortality in IPF. VRS change correlates weakly with FVC change and identifies different poor-outcome patients than a ≥10% FVC decline threshold. Use of a VRS threshold of ≥0.40% change alongside a ≥10% FVC decline threshold can identify 30% more patients that reach an endpoint and argues for the consideration of VRS change as an IPF drug trial co-endpoint to adjudicate indeterminate FVC declines of 5.0–9.9%. Shareable PDF 10.1183/13993003.02341-2018.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-02341-2018.Shareable

In conclusion, we have demonstrated for the first time that change in a computer-derived variable, vessel-related structures, which has no visual correlate is a powerful surrogate for mortality in IPF. VRS change correlates weakly with FVC change and identifies different poor-outcome patients than a ≥10% FVC decline threshold. Use of a VRS threshold of ≥0.40% change alongside a ≥10% FVC decline threshold can identify 30% more patients that reach an endpoint and argues for the consideration of VRS change as an IPF drug trial co-endpoint to adjudicate indeterminate FVC declines of 5.0–9.9%. Shareable PDF 10.1183/13993003.02341-2018.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-02341-2018.Shareable Author contributions: J. Jacob, A. Altmann, F.T. van Beek, C.H.M. van Moorsel, M. Veltkamp, H.W. van Es, R. Clay, T.M. Jacob, T. Moua, E.P. Judge, A. de Lauretis, A. Devaraj, T. Peikert, F. Maldonado, M. Kokosi, T.M. Maher, E. Renzoni and A.U. Wells were involved in either the acquisition, or analysis or interpretation of data for the study. J. Jacob and A.U. Wells were also involved in the conception and design of the study. B.J. Bartholmai, R. Karwoski and S. Rajagopalan invented and developed CALIPER. They were involved in processing the raw CT scans and in generation of figures but were not involved with the analysis or interpretation of the data in the study. All authors revised the work for important intellectual content and gave final approval for the version to be published. J. Jacob agrees to be accountable for the all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

All authors revised the work for important intellectual content and gave final approval for the version to be published. J. Jacob agrees to be accountable for the all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. Conflict of interest: J. Jacob reports personal fees for advisory board work from Boehringer Ingelheim, outside the submitted work. Conflict of interest: B.J. Bartholmai reports grants (paid to Mayo Clinic) from Royal Brompton Hospital, personal fees from Promedior and Boehringer Ingelheim, during the conduct of the study; royalties (paid to Mayo Clinic) from Imbio, LLC, outside the submitted work; and has a patent Systems and Methods for Analysing in vivo Tissue Volumes using Medical Imaging Data licensed to Imbio LLC. Conflict of interest: C.H.M. van Moorsel has nothing to disclose. Conflict of interest: S. Rajagopalan reports grants (paid to Mayo Clinic) from Royal Brompton Hospital, during the conduct of the study; royalties (paid to Mayo Clinic) from Imbio, LLC, outside the submitted work; and has a patent Systems and Methods for Analysing in vivo Tissue Volumes using Medical Imaging Data licensed to Imbio LLC. Conflict of interest: A. Devaraj reports personal fees from Boehringer Ingelheim and Roche, outside the submitted work. Conflict of interest: H.W. van Es has nothing to disclose. Conflict of interest: T. Moua has nothing to disclose. Conflict of interest: F.T. van Beek has nothing to disclose. Conflict of interest: R. Clay has nothing to disclose.

Conflict of interest: A. Devaraj reports personal fees from Boehringer Ingelheim and Roche, outside the submitted work. Conflict of interest: H.W. van Es has nothing to disclose. Conflict of interest: T. Moua has nothing to disclose. Conflict of interest: F.T. van Beek has nothing to disclose. Conflict of interest: R. Clay has nothing to disclose. Conflict of interest: M. Veltkamp has nothing to disclose. Conflict of interest: M. Kokosi has nothing to disclose. Conflict of interest: A. de Lauretis has nothing to disclose. Conflict of interest: E.P. Judge has nothing to disclose. Conflict of interest: T.M. Jacob has nothing to disclose. Conflict of interest: T. Peikert has nothing to disclose. Conflict of interest: R. Karwoski reports grants (paid to Mayo Clinic) from Royal Brompton Hospital, during the conduct of the study; royalties (paid to Mayo Clinic) from Imbio, LLC, outside the submitted work; and has a patent Systems and Methods for Analysing in vivo Tissue Volumes using Medical Imaging Data licensed to Imbio LLC. Conflict of interest: F. Maldonado has nothing to disclose. Conflict of interest: E. Renzoni reports personal fees for lectures from Roche and Takeda, personal fees for lectures and advisory board meetings from Boehringher, outside the submitted work.

Conflict of interest: R. Karwoski reports grants (paid to Mayo Clinic) from Royal Brompton Hospital, during the conduct of the study; royalties (paid to Mayo Clinic) from Imbio, LLC, outside the submitted work; and has a patent Systems and Methods for Analysing in vivo Tissue Volumes using Medical Imaging Data licensed to Imbio LLC. Conflict of interest: F. Maldonado has nothing to disclose. Conflict of interest: E. Renzoni reports personal fees for lectures from Roche and Takeda, personal fees for lectures and advisory board meetings from Boehringher, outside the submitted work. Conflict of interest: T.M. Maher is an investigator in an ongoing Phase 2b study for Gilead; reports grants and personal fees for advisory board work from GSK, grants from Novartis, personal fees from Boehringer Ingelheim InterMune, Lanthio, Sanofi Aventis, AstraZeneca, Roche, Bayer, Biogen Idec, Cipla, Prometic and Apellis, grants, personal fees and research fees (paid to institution) from UCB, outside the submitted work. Conflict of interest: A. Altmann has nothing to disclose. Conflict of interest: A.U. Wells reports personal fees for lectures and advisory board work from Intermune, Boehringer Ingelheim, Roche and Bayer, personal fees for advisory board work from Gilead and MSD, personal fees for lectures from Chiesi, outside the submitted work.

Conflict of interest: A. Altmann has nothing to disclose. Conflict of interest: A.U. Wells reports personal fees for lectures and advisory board work from Intermune, Boehringer Ingelheim, Roche and Bayer, personal fees for advisory board work from Gilead and MSD, personal fees for lectures from Chiesi, outside the submitted work. Support statement: A. Altmann holds an MRC eMedLab Medical Bioinformatics Career Development Fellowship. This work was supported by the Medical Research Council (grant number MR/L016311/1), the National Institute of Health Research Respiratory Disease Biomedical Research Unit at the Royal Brompton and Harefield NHS Foundation Trust and Imperial College London. J. Jacob was supported by Wellcome Trust Clinical Research Career Development Fellowship 209553/Z/17/Z

Introduction Infection of the pleural space causes serious morbidity and is often life threatening [1]. Despite advances in management, 30-day mortality remains high, reported at between 9% and 10.5% in a recent Danish cohort [2]. This is especially true among older patients, in whom 30-day mortality has been reported at 20.2% in patients aged >80 years [2–4]. Pleural infection is common, with >30 000 diagnoses in the years 2000–2011 in the largest and most recent population-based cohort in Taiwan [5]. In recent years, incidence rates have been trending upwards [2, 3, 6], and coupled with advancing therapeutic techniques, the management of pleural infection represents a growing resource strain, with reported median length of hospital stay in a Canadian study averaging nearly 22 days [6]. The use of intrapleural fibrinolytics [7] as well as the improved safety profile for endoscopic thoracic surgery increased the average cost of hospitalisation in a Taiwan-based study to reach USD 4400 per admission in 2008, an increase of >60% over the preceding 12 years of the study [4].

averaging nearly 22 days [6]. The use of intrapleural fibrinolytics [7] as well as the improved safety profile for endoscopic thoracic surgery increased the average cost of hospitalisation in a Taiwan-based study to reach USD 4400 per admission in 2008, an increase of >60% over the preceding 12 years of the study [4]. The underlying drivers of the rise of pleural infection cases are not fully established. Possible mechanisms include the rise of multimorbidity in ageing populations, as well as immunosuppressive states such as HIV, predisposing individuals to the condition. This is supported by data from large population-based cohorts demonstrating that incidence is skewed towards older persons and is rising more quickly in this group [2, 6]. Furthermore, rates of comorbidity in pleural infection have been reported as being as high as 74% [2, 5], and patients with increased pre-existing comorbidity have higher mortality rates (20.6% if Charlton comorbidity score (CCS) >2 points, 6% if CCS 0 points) [2].

rsons and is rising more quickly in this group [2, 6]. Furthermore, rates of comorbidity in pleural infection have been reported as being as high as 74% [2, 5], and patients with increased pre-existing comorbidity have higher mortality rates (20.6% if Charlton comorbidity score (CCS) >2 points, 6% if CCS 0 points) [2]. This tripartite trend of increasing incidence of pleural infection, accelerated cases among older persons and higher mortality among older and comorbid persons are consistently reported in large population-based cohorts from Canada, Taiwan, Denmark and USA [2–6]. However, these data do not necessarily represent worldwide patterns and to date no study has comprehensively reviewed the published data available. To address this, we performed a systematic review of the literature reporting the clinical characteristics and outcomes of patients with pleural infection, with a comparison between high-income and lower-income economies, of which reports from the latter are sparse. More research in low-economic settings will be essential going forward to understand regional trends and inform local resource provision. Methods This review was performed according to Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and the protocol registered on the PROSPERO international prospective register of systematic reviews (CRD42017076418) [8]. Search strategy Ovid MEDLINE and Embase were searched between 2000 and 2017 using the keywords “empyema”, “pleural infection” and “pleuritis”. The full search strategy is reported in detail elsewhere [8, 9].

Methods This review was performed according to Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and the protocol registered on the PROSPERO international prospective register of systematic reviews (CRD42017076418) [8]. Search strategy Ovid MEDLINE and Embase were searched between 2000 and 2017 using the keywords “empyema”, “pleural infection” and “pleuritis”. The full search strategy is reported in detail elsewhere [8, 9]. Data extraction All records were screened independently by two authors (TC and MH). The following inclusion criteria were used. 1) Population: adults (age >18 years) with bacterial pleural infection/empyema acquired in any setting (community, secondary or tertiary hospital care); 2) intervention: any intervention including conservative management with antibiotics and chest tube, intrapleural medication or any form of surgical procedure; 3) comparator: no comparator assessed; and 4) outcomes: in-hospital mortality, length of stay, escalation to surgical intervention in mixed cohort studies and any recorded comorbidity on admission. Randomised and non-randomised controlled trials as well as observational or cross-sectional studies were included. Records with <20 participants were excluded due to the case selective nature of these reports.

The following inclusion criteria were used. 1) Population: adults (age >18 years) with bacterial pleural infection/empyema acquired in any setting (community, secondary or tertiary hospital care); 2) intervention: any intervention including conservative management with antibiotics and chest tube, intrapleural medication or any form of surgical procedure; 3) comparator: no comparator assessed; and 4) outcomes: in-hospital mortality, length of stay, escalation to surgical intervention in mixed cohort studies and any recorded comorbidity on admission. Randomised and non-randomised controlled trials as well as observational or cross-sectional studies were included. Records with <20 participants were excluded due to the case selective nature of these reports. Reports where over half of participants were aged <18 years or with tuberculous, fungal or post-pneumonectomy pleural infection were excluded, as the aetiology and outcome in these groups are not comparable to bacterial pleural infection in adults. Non-English language studies were included when suitable translation was available.

e over half of participants were aged <18 years or with tuberculous, fungal or post-pneumonectomy pleural infection were excluded, as the aetiology and outcome in these groups are not comparable to bacterial pleural infection in adults. Non-English language studies were included when suitable translation was available. Data was extracted where available into a Microsoft Excel proforma. Countries of studies included in this review were classified by income as per the World Bank definition for the 2019 fiscal year [10] with gross national income (GNI) per capita in the year 2017 [11] as follows: low-income economies (GNI USD 995 per capita or less); lower-middle-income economies (GNI USD 996–3895 per capita); upper-middle-income economies (GNI USD 3896–12 055 per capita); high-income economies (GNI USD 12 055 or more per capita). Collectively in this article low, lower-middle and upper-middle economies are referred to as lower-income economies. Subgroup analysis Comorbidity data are reported as number and percentage of total participants in each study. Studies that specifically recruited patients with empyema and a specific disease exclusively (e.g. HIV) were excluded from the analysis of that particular comorbidity. Data were presented as prevalence of comorbidity both by affected organ/system and by specific disease.

percentage of total participants in each study. Studies that specifically recruited patients with empyema and a specific disease exclusively (e.g. HIV) were excluded from the analysis of that particular comorbidity. Data were presented as prevalence of comorbidity both by affected organ/system and by specific disease. Outcome analysis Data on 30-day/in-hospital mortality, length of hospital stay, need for surgical intervention and intrapleural fibrinolytic therapy were collected. Studies where an entire cohort was comprised of a single intervention were excluded from analysis of that particular outcome. Statistical analysis was performed in Prism (version 8.0; GraphPad, San Diego, CA, USA). Median values were transformed where possible to means for the age variable using the formula: mean=((lower limit+(2×median)+upper limit))/4. A t-test for parametric data or Mann–Whitney U-test for non-parametric data was performed for statistical comparison between two groups. Results Cohort characteristics Of the 20 980 publications returned from the initial search, 211 studies met the inclusion criteria. 134 articles (totalling 227 898 patients) [2, 4–7, 12–140] reported comorbidity and/or outcome data (characteristics summarised in supplementary table S1). The remaining papers were excluded due to lack of relevant data (n=48), duplicate datasets (n=12), special populations predefined for exclusion in the protocol (n=6), case-series of <20 participants (n=10) or the original article was unobtainable (n=1), as shown in figure 1.

stics summarised in supplementary table S1). The remaining papers were excluded due to lack of relevant data (n=48), duplicate datasets (n=12), special populations predefined for exclusion in the protocol (n=6), case-series of <20 participants (n=10) or the original article was unobtainable (n=1), as shown in figure 1. FIGURE 1 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart showing the identification, screening, eligibility and inclusion process. The majority of studies (n=104, 78%) were retrospective observational cohorts, while 17 (13%) studies were prospective observational studies, 11 (8%) were randomised controlled trials and two (2%) were diagnostic accuracy studies. The published data were skewed towards reports from high-income economies (n=100, 74%; totalling 224 476 patients) with over a quarter from low, low-middle and upper-middle economies combined (n=34, 26%; totalling 3422 patients, 1.5%). The most common region of reporting was East Asia (n=33, 24%) followed by North America (n=27, 20%). The mean age of individuals in all studies was 52.8 years (95% CI 51.0–54.7 years). Patients in cohorts from high-income economies were significantly older than patients from lower-income economies (56.5 years versus 42.5 years, p<0.0001).

The majority of studies (n=104, 78%) were retrospective observational cohorts, while 17 (13%) studies were prospective observational studies, 11 (8%) were randomised controlled trials and two (2%) were diagnostic accuracy studies. The published data were skewed towards reports from high-income economies (n=100, 74%; totalling 224 476 patients) with over a quarter from low, low-middle and upper-middle economies combined (n=34, 26%; totalling 3422 patients, 1.5%). The most common region of reporting was East Asia (n=33, 24%) followed by North America (n=27, 20%). The mean age of individuals in all studies was 52.8 years (95% CI 51.0–54.7 years). Patients in cohorts from high-income economies were significantly older than patients from lower-income economies (56.5 years versus 42.5 years, p<0.0001). Patients with pleural infection have a high prevalence of pre-existing comorbidity 85 articles reported comorbidity data (totalling 50 756 patients, summarised in supplementary table S2). The majority of published data were from countries with high-income economies (n=68, 79%; totalling 48 703 patients, 96%). Most reports were single-centre/multicentre retrospective or prospective observational cohorts (n=64, 74% and n=10, 12%, respectively). 38 studies were cohorts including empyema treated exclusively by surgery or fibrinolysis (n=27, 31% and n=11, 13%, respectively).

economies (n=68, 79%; totalling 48 703 patients, 96%). Most reports were single-centre/multicentre retrospective or prospective observational cohorts (n=64, 74% and n=10, 12%, respectively). 38 studies were cohorts including empyema treated exclusively by surgery or fibrinolysis (n=27, 31% and n=11, 13%, respectively). 28 studies reported the presence of overall comorbidity levels within their dataset. The percentage prevalence of pre-existing comorbidity in patients with empyema was high (median 72%, interquartile range (IQR) 58–83%; figure 2a). FIGURE 2 Pre-existing prevalence of comorbidity in studies of patients with pleural infection. Percentage prevalence of comorbidities in each study were extracted and data from high-income and lower-income economies were compared. a) Percentage prevalence of comorbidity, smoking, alcohol excess and disease by organ system affected; b) percentage prevalence of comorbidity by specific diseases. Data are presented as median (interquartile range). Mann–Whitney test was used to compare median prevalence of diabetes mellitus in high-income and lower-income economies. COPD: chronic obstructive pulmonary disease.

ess and disease by organ system affected; b) percentage prevalence of comorbidity by specific diseases. Data are presented as median (interquartile range). Mann–Whitney test was used to compare median prevalence of diabetes mellitus in high-income and lower-income economies. COPD: chronic obstructive pulmonary disease. Figure 2a shows the median percentage prevalence of reported comorbidity grouped by organ system affected. The percentage prevalence of smokers in patients with empyema had a median (IQR) 41% (30–51%; data reported in 21 studies). The median (IQR) percentage prevalence of alcohol excess in patients with empyema was 15% (8–25%; 30 studies). The median (IQR) percentage prevalence of respiratory comorbidity in patients with empyema was 20% (16–32%; 17 studies). This was similar to the percentage prevalence of cardiac disease (19%, 15–27%; 21 studies) and higher than the percentage prevalence of malignancy (12%, 8–23%; 32 studies) and liver disease (5%, 3–11%; 33 studies).

) percentage prevalence of respiratory comorbidity in patients with empyema was 20% (16–32%; 17 studies). This was similar to the percentage prevalence of cardiac disease (19%, 15–27%; 21 studies) and higher than the percentage prevalence of malignancy (12%, 8–23%; 32 studies) and liver disease (5%, 3–11%; 33 studies). Figure 2b shows the median percentage prevalence of different comorbidities by specific disease. The median (IQR) percentage prevalence of hypertension in patients with empyema was 23% (17–38%; 21 studies). This was higher than the percentage prevalence of diabetes (17%, 11–27%; 66 studies), stroke (13%, 5–20%; 21 studies), ischaemic heart disease (11%, 5–16%; 12 studies), chronic obstructive pulmonary disease (11%, 6–20%; 40 studies) and chronic kidney disease (7%, 5–13%; 33 studies). The reported presence of immunosuppressive states was relatively low. The median (IQR) percentage prevalence of HIV was 4% (1–9%; 16 studies), steroid use 4% (2–16%; six studies) and recent chemotherapy in 4% (1–15%; three studies). Where possible, we compared comorbidity prevalence between high-income and lower-income economies. There were no significant differences between studies reporting from high-income economies compared to low-income economies in prevalence of overall pre-existing comorbidity (median 73% versus 58%, p=0.623) or diabetes mellitus (median 20% versus 14%, p=0.0835). Comparisons for other specific comorbidities were not attempted due to the paucity of data reported from lower-income economies.

gh-income economies compared to low-income economies in prevalence of overall pre-existing comorbidity (median 73% versus 58%, p=0.623) or diabetes mellitus (median 20% versus 14%, p=0.0835). Comparisons for other specific comorbidities were not attempted due to the paucity of data reported from lower-income economies. Patients with pleural infection have long hospital stays Data on outcome of pleural infection was reported in 125 papers (totalling 192 298 patients). Studies reported long inpatient hospital stays (median 19 days, IQR 13–27; reported in 79 studies, totalling 180 931 patients) and median (IQR) mortality in hospital or within 30 days was 4% (1–11%, from 105 studies totalling 179 031 patients). Prevalence of patients requiring either fibrinolytic treatment (median 31%, IQR 17–52%; 38 studies, 30 071 patients) or surgery (median 20%, IQR 1–32%; 65 studies, 37 330 patients) were also reported. Figure 3 shows the differences in outcome parameters according to the income category of the country of study. There was no significant difference between studies reporting from high-income compared to lower-income economies in mean length of stay (18.7 days versus 19.7 days; figure 3a), percentage prevalence of patients receiving surgery (median 19.5% versus 20.0%; figure 3b), fibrinolytic treatment (median 41% versus 24%, p=0.1) or 30-day/in-hospital mortality (median 5% versus 4%).

ing from high-income compared to lower-income economies in mean length of stay (18.7 days versus 19.7 days; figure 3a), percentage prevalence of patients receiving surgery (median 19.5% versus 20.0%; figure 3b), fibrinolytic treatment (median 41% versus 24%, p=0.1) or 30-day/in-hospital mortality (median 5% versus 4%). FIGURE 3 Prevalence of outcomes in studies of patients with pleural infection. Percentage prevalence of outcomes in each study were extracted and data from high-income and lower-income economies were compared. a) Mean length of hospital stay; b) percentage prevalence of mortality, patients requiring fibrinolysis and patients requiring surgical treatment. Data are presented as median (interquartile range). Discussion This is the first systematic review describing the comorbidities and outcomes of studies reporting on patients with pleural infection since the turn of the 21st century. We found that the percentage prevalence of pre-existing comorbidity was high (median 72%) with a wide range of chronic conditions affecting the major organ systems. This is consistent with large population-based studies, which have reported comorbidity prevalence of up to 74% [5] and supports the hypothesis that the rise in the incidence of pleural infection in recent years might be associated with an increasingly ageing, multi-morbid population.

affecting the major organ systems. This is consistent with large population-based studies, which have reported comorbidity prevalence of up to 74% [5] and supports the hypothesis that the rise in the incidence of pleural infection in recent years might be associated with an increasingly ageing, multi-morbid population. Chronic respiratory and cardiovascular conditions had the highest percentage prevalence and where specific conditions were reported, hypertension, diabetes mellitus, stroke and ischaemic heart disease all had median prevalence rates between 11% and 23%. As the majority of studies were from high-income settings where these conditions are endemic, this finding may seem unsurprising, and indeed, known risk factors for these diseases including smoking and alcohol excess were also reported (median 42% and 15%, respectively). When comparing studies from high-income and lower-income economies, there were no significant differences in the prevalence of pre-existing comorbidities. In high-income economies there was a higher prevalence of diabetes mellitus, but this was not significant (median 20% versus 14%, p=0.08). The mean age of patients in studies from high-income economies was also significantly higher than from lower-income economies (56.5 versus 42.5 years, p<0.001), which probably reflects the longer life expectancy in high-income economies and the increasing prevalence of diabetes mellitus with age [141].

us 14%, p=0.08). The mean age of patients in studies from high-income economies was also significantly higher than from lower-income economies (56.5 versus 42.5 years, p<0.001), which probably reflects the longer life expectancy in high-income economies and the increasing prevalence of diabetes mellitus with age [141]. Evidence from large prospective population-based cohort studies will be required to investigate whether there is a causative link between the rise in chronic noncommunicable conditions such as diabetes mellitus and the increased incidence in pleural infection and whether there are true differences between high-income and lower-income economies. Immunosuppressive states acquired through diseases such as HIV or iatrogenically induced by steroids, immunomodulatory and chemotherapeutic agents can be associated with pleural infection [48, 69]. We found a relatively low prevalence of these conditions among studies of patients with pleural infection; however, this is probably an underestimation as only a small number of studies collected these data. Studies from lower-income economies reported higher levels of HIV compared to high-income economies (median 14% versus 4%), in keeping with current HIV trends, but this was not statistically significant [142]. Future studies should focus on the routine collection of these data to better understand the risk, pathogenesis and outcomes of pleural infection in these specific groups where the microbiological milieu and immune response are likely to differ from those in immunocompetent persons.

not statistically significant [142]. Future studies should focus on the routine collection of these data to better understand the risk, pathogenesis and outcomes of pleural infection in these specific groups where the microbiological milieu and immune response are likely to differ from those in immunocompetent persons. When analysing outcomes of pleural infection, we confirmed that patients have long hospital stays (median 19 days), comparable to previously reported length-of-stay data [143, 144] This supports the premise that pleural infection is an important use of healthcare resources.

not statistically significant [142]. Future studies should focus on the routine collection of these data to better understand the risk, pathogenesis and outcomes of pleural infection in these specific groups where the microbiological milieu and immune response are likely to differ from those in immunocompetent persons. When analysing outcomes of pleural infection, we confirmed that patients have long hospital stays (median 19 days), comparable to previously reported length-of-stay data [143, 144] This supports the premise that pleural infection is an important use of healthcare resources. We found a median in-hospital/30-day mortality of 4%, which is lower than has been previously reported. The British Thoracic Society guidelines quote mortality at 20% [145], based on data from a large prospective UK cohort of patients reported in 1996 with pleural infection with an 18% mortality rate at 6 months [143] and the MIST-1 (Multicentre Intrapleural Sepsis Trial) cohort in which 12% had died by 1 year [146]. A recent outcome study of pleural infection of >600 patients from Australia found 1-year mortality of 32.4% in patients with community-acquired pleural infection [147]. We recorded 30-day or in-hospital mortality rather than 3, 6 or 12 months, as this was most commonly reported in the studies we analysed. Our finding of a median 4% in-hospital mortality is consistent with a prospective single centre study of pleural infection reported in 1999, where in-hospital deaths were 4.7%, rising to 14% mortality within 400 days of chest tube insertion [144]. However, in the largest, most recently published population-based cohort study of pleural infection cases in Denmark, overall unadjusted 30-day mortality was reported at between 9% and 10.5% [2]

reported in 1999, where in-hospital deaths were 4.7%, rising to 14% mortality within 400 days of chest tube insertion [144]. However, in the largest, most recently published population-based cohort study of pleural infection cases in Denmark, overall unadjusted 30-day mortality was reported at between 9% and 10.5% [2] One explanation for this is that our dataset comprises observational studies with low numbers of participants and significant risk of bias. There was a weakly positive correlation between the percentage mortality reported and number of participants (r=0.2, p=0.05) (supplementary figure S1a). Studies with <300 participants had a lower median mortality than studies with >300 participants (4% versus 9%, p=0.072) (supplementary figure S1b) and reported a wider range of mortality estimates. There was no association between numbers of participants and age or year of publication, but as expected there is an association between mortality and age in line with previous data (r=0.35, p=0.0007; data not shown). We were not able to investigate whether there was an association between comorbidity and outcome as the data were not reported in such a way in the original studies.

e or year of publication, but as expected there is an association between mortality and age in line with previous data (r=0.35, p=0.0007; data not shown). We were not able to investigate whether there was an association between comorbidity and outcome as the data were not reported in such a way in the original studies. Studies with more participants are less susceptible to inclusion bias, as data are obtained from databases using International Classification of Diseases (ICD)-10 codes assigned at hospital discharge rather than locally curated cohorts. These larger studies therefore likely provide more reliable mortality estimates than our combined unadjusted estimate, and this is a weakness of our approach. Overall our analysis supports the seriousness of pleural infection, but probably underestimates the true mortality prevalence.

ther than locally curated cohorts. These larger studies therefore likely provide more reliable mortality estimates than our combined unadjusted estimate, and this is a weakness of our approach. Overall our analysis supports the seriousness of pleural infection, but probably underestimates the true mortality prevalence. Despite a significant difference in mean age of patients in studies from high- and lower-income economies, the length of hospital stay, percentage of patients requiring surgery and 30-day/in-hospital mortality were similar between groups. Studies from higher-income economies reported a trend towards increased use of fibrinolytics (median 41% versus 24%), which offers an option for symptomatic drainage of loculated pleural effusions in patients who are not suitable for surgery. This is consistent with a trend towards the increasing use of fibrinolytics over time in one high-income-based cohort [4]. However, prospective randomised trials have failed to show a mortality benefit from fibrinolytic treatment, and further studies are needed to explore this area of practice [7, 86].

ble for surgery. This is consistent with a trend towards the increasing use of fibrinolytics over time in one high-income-based cohort [4]. However, prospective randomised trials have failed to show a mortality benefit from fibrinolytic treatment, and further studies are needed to explore this area of practice [7, 86]. This study describes the results of 134 articles reporting data from >200 000 patients, and thus is the most comprehensive work to date examining the comorbidities and outcomes of patients with pleural infection worldwide. However, the findings may not be generalisable to all settings or fully representative of real-world trends, as the majority of studies are relatively small retrospective observational cohorts from secondary care institutions in high-income settings. Mixed cohorts were included in the analysis to maximise inclusion, and therefore some of the results will reflect patient populations with a proportion of tuberculosis, post-surgical and childhood empyema, which differ in aetiology and outcome to adult bacterial pleural infection. In conclusion, this study confirms that pleural infection remains an important disease. Patients have a high prevalence of pre-existing comorbidity and are older in high-income economies. Importantly this study highlights the paucity of data on pleural infection from lower-income economies and calls for large prospective registries at the population level in these settings to better understand regional trends in pleural infection and to enable optimal resource provision.

older in high-income economies. Importantly this study highlights the paucity of data on pleural infection from lower-income economies and calls for large prospective registries at the population level in these settings to better understand regional trends in pleural infection and to enable optimal resource provision. Supplementary material 10.1183/13993003.00541-2019.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-00541-2019.SUPPLEMENT Shareable PDF 10.1183/13993003.00541-2019.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-00541-2019.Shareable This article has supplementary material available from erj.ersjournals.com This systematic review is registered at www.crd.york.ac.uk/prospero with registration number CRD42017076418. Data available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.5d3s2vg Author contributions: M. Hassan, J.P. Corcoran and N.M. Rahman conceptualised and planned the study. E. Harriss conducted the literature review. T.N. Cargill and M. Hassan screened abstracts for inclusion. T.N. Cargill extracted and analysed the data and prepared the manuscript. D.J. McCracken and N.M. Rahman critically revised the first draft. All authors reviewed the final manuscript prior to submission.

the study. E. Harriss conducted the literature review. T.N. Cargill and M. Hassan screened abstracts for inclusion. T.N. Cargill extracted and analysed the data and prepared the manuscript. D.J. McCracken and N.M. Rahman critically revised the first draft. All authors reviewed the final manuscript prior to submission. Support statement: T.N. Cargill is funded as an Academic Clinical Fellow by the National Institute of Health Research and by a Wellcome Trust PhD Training Fellowship for Clinicians (grant number 211042/Z/18/Z). M. Hassan is a recipient of the European Respiratory Society Long Term Research Fellowship (ERS 2016 7333). Conflict of interest: T.N. Cargill has nothing to disclose. Conflict of interest: M. Hassan has nothing to disclose. Conflict of interest: J.P. Corcoran has nothing to disclose. Conflict of interest: E. Harriss has nothing to disclose. Conflict of interest: R. Asciak has nothing to disclose. Conflict of interest: R.M. Mercer has nothing to disclose. Conflict of interest: D.J. McCracken has nothing to disclose. Conflict of interest: E.O. Bedawi has nothing to disclose. Conflict of interest: N.M. Rahman has nothing to disclose.

To the Editor: Sarcoidosis is characterised by an accumulation of CD4+ T-cells in the lungs and an increased bronchoalveolar lavage fluid (BALF) CD4/CD8 ratio (>3.5) [1]. In sarcoidosis, an expansion of BALF CD4+ T-cells expressing the T-cell receptor Vα2.3 has been associated with good prognosis and with specific HLA-alleles, i.e. HLA-DRB1*0301 and HLA-DRB3*0101 (which is often carried together with HLA-DRB1*13). HLA-DRB1*03 and HLA-DRB3*0101 molecules show similarities in the region important for antigen presentation and both may therefore be capable of presenting identical antigens to the lung T-cells [2]. Furthermore, an expansion defined as >10.5% CD4+ Vα2.3+ BALF T-cells is commonly seen in patients with Löfgren's syndrome [3], which is characterised by an acute onset with bilateral ankle arthritis and/or erythema nodosum, bilateral hilar lymphadenopathy and, in some cases, with parenchymal infiltrates and usually fever [4]. We have previously shown that very high expansions of CD4+ Vα2.3+ T-cells are associated with Löfgren's syndrome and a disease duration <2 years [3]. However, not all patients with an expansion of CD4+ Vα2.3+ T-cells have Löfgren's syndrome and resolving disease. In this much enlarged study on a HLA-typed sarcoidosis cohort, we aimed at investigating the clinical characteristics of patients with an expansion of CD4+ Vα2.3+ T-cells in BALF and to analyse if the degree of expansion may predict the prognosis of sarcoidosis.

3+ T-cells have Löfgren's syndrome and resolving disease. In this much enlarged study on a HLA-typed sarcoidosis cohort, we aimed at investigating the clinical characteristics of patients with an expansion of CD4+ Vα2.3+ T-cells in BALF and to analyse if the degree of expansion may predict the prognosis of sarcoidosis. In a registry of sarcoidosis patients (n=661, including 252 with Löfgren's syndrome), all investigated with bronchoscopy and BAL for diagnostic purposes and HLA-typed and followed for at least 2 years, 248 subjects were identified with BALF CD4+ Vα2.3+ T-cells expansions. An expansion was defined as three times the median percentage of Vα2.3+ CD4+ T cells in peripheral blood of healthy subjects, as previously described (3×3.5%) [3]. Disease activity was evaluated 2 years after disease onset, considering presence of symptoms (e.g. cough, fatigue, dyspnoea, fever), serum-ACE activity, spirometry values and chest radiographic findings. Patients without any pathological findings were regarded to have a resolving disease.

iously described (3×3.5%) [3]. Disease activity was evaluated 2 years after disease onset, considering presence of symptoms (e.g. cough, fatigue, dyspnoea, fever), serum-ACE activity, spirometry values and chest radiographic findings. Patients without any pathological findings were regarded to have a resolving disease. We focused on patients with Vα2.3+ CD4+ T-cell expansions, out of which 73% had classical Löfgren's syndrome (table 1). They were all judged to have active disease at the time for bronchoscopy. The percentage of Vα2.3+ CD4+ T-cells in BALF is known to be normalised when the patients recover [5]. All patients were without immunosuppressive treatment at the time for bronchoscopy. After 2 years follow-up very few patients with Löfgren's syndrome, but some more with non-Löfgren's syndrome, had been treated with immunosuppressants. The sarcoidosis diagnosis was made through typical clinical and radiographic manifestations, findings at bronchoscopy with BAL including an elevated CD4/CD8-ratio (>3.5) and/or positive biopsies, in accordance with the criteria of the World Association of Sarcoidosis and other Granulomatous Disorders [6]. Chest radiographs were evaluated as previously described [7]. Written informed consent was obtained from all subjects, and approval was granted from the regional ethical review board. TABLE 1 Clinical characteristics of patients with Vα2.3+ T-cells >10.5% in bronchoalveolar lavage fluid (BALF)

We focused on patients with Vα2.3+ CD4+ T-cell expansions, out of which 73% had classical Löfgren's syndrome (table 1). They were all judged to have active disease at the time for bronchoscopy. The percentage of Vα2.3+ CD4+ T-cells in BALF is known to be normalised when the patients recover [5]. All patients were without immunosuppressive treatment at the time for bronchoscopy. After 2 years follow-up very few patients with Löfgren's syndrome, but some more with non-Löfgren's syndrome, had been treated with immunosuppressants. The sarcoidosis diagnosis was made through typical clinical and radiographic manifestations, findings at bronchoscopy with BAL including an elevated CD4/CD8-ratio (>3.5) and/or positive biopsies, in accordance with the criteria of the World Association of Sarcoidosis and other Granulomatous Disorders [6]. Chest radiographs were evaluated as previously described [7]. Written informed consent was obtained from all subjects, and approval was granted from the regional ethical review board. TABLE 1 Clinical characteristics of patients with Vα2.3+ T-cells >10.5% in bronchoalveolar lavage fluid (BALF) LS Non-LS Subjects 180 68 Male/female 111/69 41/27 Age years# 37 (21–62) 45 (26–72) Radiographic stage 0/I/II/III/IV# 0/123/57/0/0 4/16/36/8/4 Resolving/non-resolving# 166/14 25/43 CD4/CD8 ratio* 9.8 (0.9–56.8) 7.1 (1.2–24.0) % Vα2.3 BALF cells# 28.4 (11.0–50.0) 17.8 (11.4–44.3) HLA-DRB1*03+/− 155/25 36/32 Vα2.3 BALF cells % HLA-DRB1*03+# 29.9 20.5 HLA-DRB1*03−** 19.8 14.8 Patients recovered % HLA-DRB1*03+# 94 39 HLA-DRB1*03−** 80 34 Data are presented as n or mean (range), unless otherwise stated. LS: Löfgreńs syndrome. *: p<0.05, **: p<0.001 and #: p<0.0001, comparing differences between patients with LS and non-LS and for radiological stage differences between stage I and II.

9.8 14.8 Patients recovered % HLA-DRB1*03+# 94 39 HLA-DRB1*03−** 80 34 Data are presented as n or mean (range), unless otherwise stated. LS: Löfgreńs syndrome. *: p<0.05, **: p<0.001 and #: p<0.0001, comparing differences between patients with LS and non-LS and for radiological stage differences between stage I and II. Bronchoscopy with BAL was carried out as described before [8]. Surface markers expressed on T-cells were analysed using flow cytometry and all patients were HLA-typed as previously described [9, 10]. Statistical analyses were performed with Graph Pad Prism 6 (GraphPad Software Inc., San Diego, CA, USA). When comparing several groups such as differences between HLA-DRB1* alleles, p<0.003 (p<0.05 divided by 13) was regarded as significant after Bonferroni correction for the number of alleles (n=13), and otherwise p<0.05 was regarded as significant. High percentages of CD4+ Vα2.3+ T-cells (i.e. Vα2.3+ CD4+ T-cells >10.5% in BALF) associated with a resolving disease, as 77% (191 out of 248) of these patients resolved within 2 years compared with 28% (114 out of 413) of patients with normal levels (p<0.0001). The proportion of patients who recovered increased gradually with the increasing proportion of CD4+ Vα2.3+ T-cells in BALF, for example in patients with 0–5% of CD4+ Vα2.3+ T-cells 25% had resolving disease; in the range 11–15%, 44% resolved and when there were 21–25% Vα2.3+ T-cells, 82% resolved. If >30%, 95% resolved.

he proportion of patients who recovered increased gradually with the increasing proportion of CD4+ Vα2.3+ T-cells in BALF, for example in patients with 0–5% of CD4+ Vα2.3+ T-cells 25% had resolving disease; in the range 11–15%, 44% resolved and when there were 21–25% Vα2.3+ T-cells, 82% resolved. If >30%, 95% resolved. Patients with Löfgren's syndrome had higher proportion of CD4+ Vα2.3+ T-cells in BALF compared to non-Löfgren's syndrome patients and were also younger at disease onset (table 1). Furthermore, patients with Löfgren's syndrome who carried the HLA-DRB1*03 allele had a higher median CD4+ Vα2.3+ T-cell proportion in BALF compared to HLA-DRB1*03− with Löfgren's syndrome (p<0.0001). Among the HLA-DRB1*03− patients, HLA-DRB1*13 was carried by 88% of the patients with Löfgren's syndrome and by 63% with non-Löfgren's syndrome. In this study, we chose to focus on patients with an expansion of CD4+ Vα2.3+ T-cells in BALF. The highest proportion of CD4+ Vα2.3+ T-cells in the present study was seen in Löfgren's syndrome patients who were HLA-DRB1*03+. The non-LS group was characterised by a less pronounced expansion of Vα2.3+ T-cells and disease onset at a higher age. That older patients have less favourable outcome has been shown in another cohort [11].

ortion of CD4+ Vα2.3+ T-cells in the present study was seen in Löfgren's syndrome patients who were HLA-DRB1*03+. The non-LS group was characterised by a less pronounced expansion of Vα2.3+ T-cells and disease onset at a higher age. That older patients have less favourable outcome has been shown in another cohort [11]. Our hypothesis is that patients with expansion of CD4+ Vα2.3+ T-cells (i.e. Vα2.3+ CD4+ T-cells >10.5% in BALF) may have a more effective eradication of a presumed disease-promoting antigen. An influx of CD4+ Vα2.3+ T-cells to the lungs may then explain the concomitant pronounced CD4/CD8 ratio. We have in a previously study showed that the BALF CD4+ Vα2.3+ T-cells express significantly reduced levels of FOXP3 versus CD4+ Vα2.3− T-cells [12], suggesting the CD4+ Vα2.3+ T-cells function as effector cells rather than regulatory cells, in line with a hypothetically more efficient elimination of a hypothetical sarcoidosis-antigen by such T-cells. The clinical presentation, i.e. Löfgren's syndrome or non-Löfgren's syndrome, may reflect an altered immune and inflammatory reaction influenced by different exposures or genetic differences, which also include other inflammatory genes (e.g. tumour necrosis factor gene variants linked to HLA-DRB1*03 variants). A hypothetical antigen might itself also have properties that may influence the inflammatory reaction, e.g. by inducing auto-immune reactions due to similarities of the inciting antigen and some self-structures or by preferentially stimulating a T helper (Th) 1-, Th 2- or a Th 17-dominant response.

A-DRB1*03 variants). A hypothetical antigen might itself also have properties that may influence the inflammatory reaction, e.g. by inducing auto-immune reactions due to similarities of the inciting antigen and some self-structures or by preferentially stimulating a T helper (Th) 1-, Th 2- or a Th 17-dominant response. In conclusion, the findings in this study indicate that the more pronounced the expansion of CD4+ Vα2.3+ T-cells in the BAL fluid is, the better the prognosis. The usefulness of Va2.3+ T-cells as a prognostic marker is described here for a Scandinavian cohort. Whether they may be of clinical interest in other populations needs to be analysed in future studies. Shareable PDF 10.1183/13993003.01450-2019.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-01450-2019.Shareable Author contributions: P. Darlington designed the study, characterised patients, summarised data and drafted the manuscript. S. Kullberg characterised patients, interpreted data and helped writing the manuscript. A. Eklund and J. Grunewald co-designed the study and characterised patients, interpreted data and helped writing the manuscript. All authors read and approved the final manuscript. Conflict of interest: P. Darlington has nothing to disclose. Conflict of interest: S. Kullberg has nothing to disclose. Conflict of interest: A. Eklund has nothing to disclose. Conflict of interest: J. Grunewald has nothing to disclose.

Author contributions: P. Darlington designed the study, characterised patients, summarised data and drafted the manuscript. S. Kullberg characterised patients, interpreted data and helped writing the manuscript. A. Eklund and J. Grunewald co-designed the study and characterised patients, interpreted data and helped writing the manuscript. All authors read and approved the final manuscript. Conflict of interest: P. Darlington has nothing to disclose. Conflict of interest: S. Kullberg has nothing to disclose. Conflict of interest: A. Eklund has nothing to disclose. Conflict of interest: J. Grunewald has nothing to disclose. Support statement: This work was supported by the Swedish Heart-Lung Foundation, the King Gustaf V's and Queen Victoria's Freemasons' Foundation, the Swedish Research Council and Karolinska Institutet. Support was also given through the regional agreement on medical training and clinical research (ALF) between Stockholm County Council and Karolinska Institutet. None of the funding sources had any influence on the production of this manuscript. Funding information for this article has been deposited with the Crossref Funder Registry.

To the Editor: Diagnosing tuberculosis (TB) in people living with HIV (PLHIV) remains challenging in part, because of its diversity of clinical manifestations, including high rates of extra-pulmonary and disseminated disease [1]. In particular, disseminated TB, involving multiple organ systems, is associated with high mortality but often presents non-specifically, which may hinder prompt diagnosis [2, 3]. Xpert MTB/RIF (Xpert; Cepheid, Sunnyvale, CA, USA), is currently recommended by the World Health Organization (WHO) as the first line assay for evaluating a subset of extra-pulmonary TB disease (EPTB) manifestations [4]. To detect specific forms of EPTB, such as pleural TB, TB meningitis or TB lymphadenitis, Xpert may require an invasive sample to be collected, which often limits its use for EPTB detection to hospitals where appropriate equipment is available and invasive sampling can be safely performed. Furthermore, even when concomitant pulmonary disease is present, it can be very difficult to obtain sputum in the sickest HIV patients to submit for Xpert testing [5, 6]. Therefore, an urgent priority for improving TB detection among PLHIV remains the development of rapid, point-of-care (POC) assays that use an easily obtainable clinical specimen, such as urine, and that have good diagnostic accuracy for both pulmonary and EPTB, including disseminated disease [7].

Xpert testing [5, 6]. Therefore, an urgent priority for improving TB detection among PLHIV remains the development of rapid, point-of-care (POC) assays that use an easily obtainable clinical specimen, such as urine, and that have good diagnostic accuracy for both pulmonary and EPTB, including disseminated disease [7]. The commercially available Alere Determine TB LAM (AlereLAM; Abbott, Chicago, IL, USA) assay is a rapid, inexpensive, urinary POC TB test [8]. While its use is associated with a mortality benefit in severely ill and immunocompromised PLHIV [9, 10], it has only moderate sensitivity that is limited to patients with low CD4 counts, which has led to limited programmatic uptake [11]. We have previously reported on the Fujifilm SILVAMP TB LAM (FujiLAM; Fujifilm, Tokyo, Japan) POC assay that, similar to AlereLAM, detects the presence of lipoarabinomannan (LAM) in urine [12]. It offers on average 30% improved sensitivity for detecting TB (independent of whether it is PTB or EPTB) compared to AlereLAM across subgroups stratified by CD4 strata, while maintaining high specificity. Here we report the sensitivity of FujiLAM in comparison to AlereLAM specifically for detecting EPTB in the same patient cohorts.

verage 30% improved sensitivity for detecting TB (independent of whether it is PTB or EPTB) compared to AlereLAM across subgroups stratified by CD4 strata, while maintaining high specificity. Here we report the sensitivity of FujiLAM in comparison to AlereLAM specifically for detecting EPTB in the same patient cohorts. This post hoc analysis utilised data from two previously published, prospective cohort studies of adults (>18 years) living with HIV who were admitted to South African district hospitals on the outskirts of Cape Town [13, 14]. Cohort A enrolled patients without a current TB diagnosis regardless of presenting signs or symptoms, and independent of CD4 count [13]. Cohort B enrolled patients with a CD4 count <350 cells·μL−1 in whom TB was considered the most likely diagnosis on admission [14]. A third previously published cohort was not included in the present analysis as it excluded patients with exclusively extra-pulmonary TB disease [12]. Informed consent was obtained from patients who had capacity or regained capacity and all study-related activities were approved by the Human Research Ethics Committee of the University of Cape Town.

ohort was not included in the present analysis as it excluded patients with exclusively extra-pulmonary TB disease [12]. Informed consent was obtained from patients who had capacity or regained capacity and all study-related activities were approved by the Human Research Ethics Committee of the University of Cape Town. Patients were systematically evaluated for the presence of TB. Whenever possible, patients provided two sputum samples, a blood sample and a urine sample for mycobacteriology; those unable to produce sputum or urine samples were not excluded from study enrolment. Sputum specimens were tested using smear fluorescence microscopy, mycobacteria growth indicator tube (MGIT) liquid culture (Becton Dickinson, Franklin Lakes, NJ, USA), and Xpert MTB/RIF Version G4. Blood specimens were tested using BACTEC Myco/F Lytic culture (Becton Dickinson, Franklin Lakes, NJ, USA). Sediments from urine specimens were tested using Xpert after centrifugation of 30–40 mL. The routine clinical team obtained additional specimens (sputum and non-sputum) as clinically indicated. FujiLAM and AlereLAM were performed on biobanked urine samples according to manufacturers' instructions and read by two investigators blinded to patient status and all other test results [12]. Microbiologically confirmed TB was defined by the detection of M. tuberculosis on any clinical specimen using either culture or Xpert. All patients with microbiologically-confirmed TB were classified into one of three mutually-exclusive groups: pulmonary TB (PTB) (TB detected in sputum only), EPTB (TB detected in extra-pulmonary specimen(s) only), or PTB+EPTB (TB detected in both sputum and at least one extra-pulmonary specimen). The sensitivity (and corresponding 95% confidence intervals) of FujiLAM and AlereLAM was calculated for each form of TB as well as for individual forms of EPTB.

ed in sputum only), EPTB (TB detected in extra-pulmonary specimen(s) only), or PTB+EPTB (TB detected in both sputum and at least one extra-pulmonary specimen). The sensitivity (and corresponding 95% confidence intervals) of FujiLAM and AlereLAM was calculated for each form of TB as well as for individual forms of EPTB. Of 1079 eligible patients, 111 had a TB status that could not be classified, 90 did not have urine samples and six had missing urine results; therefore, 872 patients (420 from cohort A and 659 from cohort B) had complete results and were included in this analysis. The median age was 36 (IQR 30–43) years, 54% were female, the median CD4 count was 84 (IQR 32–188) cells·μL−1, and 45% had previously been treated for TB. Among 872 patients, 553 (138 from cohort A and 415 from cohort B) had microbiologically confirmed TB (prevalence 56%) on at least one specimen, 88 (37 from cohort A and 51 from cohort B) had possible TB and 231 (189 from cohort A and 42 from cohort B) had no evidence of TB. Of those with confirmed TB, 126 (23%) out of 553 had PTB, 156 (28%) out of 553 had EPTB, and 271 (49%) out of 553 had both PTB+EPTB. The urine LAM assays performed best in those with PTB+EPTB, with FujiLAM detecting 91% (95% CI 87–94; 246 out of 271) of cases compared with 61% (95% CI 55–67; 165 out of 271) using AlereLAM (figure 1a). In patients with PTB or EPTB only, FujiLAM detected 60% (95% CI 51–69; 76 out of 126) and 67% (95% CI 59–75; 105 out of 156) of cases, respectively, which was compared with 19% (95% CI 12–27; 24 out of 126) and 41% (95% CI 33–49; 64 out of 156), respectively for AlereLAM (figure 1a).

of 271) using AlereLAM (figure 1a). In patients with PTB or EPTB only, FujiLAM detected 60% (95% CI 51–69; 76 out of 126) and 67% (95% CI 59–75; 105 out of 156) of cases, respectively, which was compared with 19% (95% CI 12–27; 24 out of 126) and 41% (95% CI 33–49; 64 out of 156), respectively for AlereLAM (figure 1a). FIGURE 1 The diagnostic sensitivity of FujiLAM and AlereLAM by (a) type of tuberculosis disease (pulmonary, extra pulmonary or both; n=553), (b) site of disease involvement in patients with confirmed extra-pulmonary tuberculosis (EPTB). Bars represent 95% confidence intervals. The numbers in parenthesis denote 95% confidence intervals. Of note, the same patient may have multiple sites of confirmed disease (e.g., pulmonary, blood, urine, etc). Sputum, blood and urine were obtained from all patients whenever possible and additional specimens were obtained at this discretion of routine medical team, however the ability to produce sputum was not a requirement for study entry. This analysis was limited to among those with both FujiLAM and AlereLAM results available. PTB: pulmonary tuberculosis.

ined from all patients whenever possible and additional specimens were obtained at this discretion of routine medical team, however the ability to produce sputum was not a requirement for study entry. This analysis was limited to among those with both FujiLAM and AlereLAM results available. PTB: pulmonary tuberculosis. The sensitivity for FujiLAM across different extra-pulmonary forms of TB disease ranged from 47 to 94% as shown in figure 1b. Notably, FujiLAM detected TB in 94% (95% CI 90–97) of patients with TB mycobacteraemia and 88% (95% CI 84–92) of those with TB confirmed by urine Xpert or culture. It also demonstrated moderate sensitivity in patients with microbiologically-confirmed pleural TB (68%; 95% CI 55–80) and with TB meningitis (47%; 95% CI 24–71). AlereLAM's sensitivity ranged from 16 to 70% and performed best in those with TB mycobacteraemia (70%; 95% CI 64–76) and TB confirmed by urine Xpert or culture (61%; 95% CI 55–67).

oderate sensitivity in patients with microbiologically-confirmed pleural TB (68%; 95% CI 55–80) and with TB meningitis (47%; 95% CI 24–71). AlereLAM's sensitivity ranged from 16 to 70% and performed best in those with TB mycobacteraemia (70%; 95% CI 64–76) and TB confirmed by urine Xpert or culture (61%; 95% CI 55–67). Overall, FujiLAM showed substantially higher sensitivity over the commercially available AlereLAM, for detecting both pulmonary and extra-pulmonary TB in HIV inpatients. This suggests that FujiLAM may have clinical utility as a first-line test for the rapid detection of TB in HIV patients, independent of disease location. Given that a large proportion of patients with HIV-associated TB have EPTB and a diagnosis may only be possible using a non-sputum sample that may be challenging to obtain, an up-front FujiLAM test could substantially reduce the time to diagnosis. FujiLAM was able to detect TB in 67% (105 out of 156) of patients who could not produce a sputum sample or did not have evidence of pulmonary disease; such patients comprised 28% of the study cohort.

tum sample that may be challenging to obtain, an up-front FujiLAM test could substantially reduce the time to diagnosis. FujiLAM was able to detect TB in 67% (105 out of 156) of patients who could not produce a sputum sample or did not have evidence of pulmonary disease; such patients comprised 28% of the study cohort. FujiLAM performed best in those with TB mycobacteraemia as well as those with concomitantly positive sputum and non-respiratory cultures, detecting >90% of cases. Mycobacterium tuberculosis bacteraemia is one of the most common blood stream infections among PLHIV in sub-Saharan Africa [3] and such patients have an extremely high mortality risk. FujiLAM's excellent performance in those with mycobacteraemia suggests a mechanistic association between disease dissemination and urinary LAM. This finding is supported by our recent study that showed a good association between detection of LAM in urine and serum of TB patients, independent of HIV status [15]. However, even for patients with forms of disease such as pleural TB and TB meningitis that may be compartmentalised, FujiLAM had moderate sensitivity, which could add substantial benefit in these cases. Taken together, these findings suggest that LAM antigenuria is likely indicative of glomerular filtration of circulating LAM (or LAM fragments) in addition to renal TB [16]. Further research, that aims to detect LAM with ultra-sensitive platforms, as well as characterisation of LAM structure in urine, is needed to better understand the mechanisms by which LAM enters the bloodstream and urine. This may help to further refine urine-based diagnostics and catalyse the development of blood-based assays. As the overall load of mycobacteria is expected to be higher in HIV-positive patients, our findings should not be generalised to HIV-negative individuals with EPTB.

by which LAM enters the bloodstream and urine. This may help to further refine urine-based diagnostics and catalyse the development of blood-based assays. As the overall load of mycobacteria is expected to be higher in HIV-positive patients, our findings should not be generalised to HIV-negative individuals with EPTB. Patients were not systematically evaluated for the presence of EPTB beyond mycobacterial blood cultures and urine Xpert; additional systematic sampling (e.g. pleural fluid or cerebrospinal fluid) could not be justified as it would require invasive sampling where this was not clinically indicated. Coupled with challenges in universally obtaining clinically indicated samples (e.g. sputum), some misclassification of TB category (PTB, EPTB or PTB+EPTB) is likely present. Furthermore, because the overall cohort was severely immunocompromised (increasing the likelihood of disease dissemination) and because those with lower, site-specific mycobacterial burdens (CSF or pleural fluid) may not have been diagnosed by Xpert or culture, the true sensitivity of urinary FujiLAM (and AlereLAM) for localised extra-pulmonary disease is possibly an overestimate and may not be generalisable to all patients with these forms of disease. However, the additional specimens collected by the routine clinical team mirrored common practice in settings with a high-burden HIV-associated TB and followed clinical symptomatology. Finally, we did not evaluate the specificity of FujiLAM for specific disease forms given a lack of systematic EPTB sampling. We have previously reported the cohort-specific specificity as well as the estimated specificity using a Bayesian bivariate random-effects model in three cohorts: using a composite reference standard, the specificity of FujiLAM was 95.7% (95% CI 92.0–98.0%) compared with 98.2% (95% CI 95.7–99.6%) for AlereLAM [12].

ng. We have previously reported the cohort-specific specificity as well as the estimated specificity using a Bayesian bivariate random-effects model in three cohorts: using a composite reference standard, the specificity of FujiLAM was 95.7% (95% CI 92.0–98.0%) compared with 98.2% (95% CI 95.7–99.6%) for AlereLAM [12]. In conclusion, our results suggest that the POC FujiLAM has good sensitivity for detecting both pulmonary and extra-pulmonary forms of TB in patients with advanced HIV, in which conventional diagnostics may be slow, require infrastructure and equipment or rely on samples that are difficult to obtain. While appropriate sampling should still be undertaken to allow for drug susceptibility testing, FujiLAM may be an appropriate first microbiological TB investigation for all hospitalised PLHIV, allowing for more rapid initiation of anti-TB therapy. Shareable PDF 10.1183/13993003.01259-2019.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-01259-2019.Shareable Acknowledgments The investigators are grateful to the clinical and administrative staff of the Western Cape Department of Health as well as the patients who contributed to this data. The authors also wish to thank the late Stephen D. Lawn who designed and led the Cohort1 study.

Shareable PDF 10.1183/13993003.01259-2019.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-01259-2019.Shareable Acknowledgments The investigators are grateful to the clinical and administrative staff of the Western Cape Department of Health as well as the patients who contributed to this data. The authors also wish to thank the late Stephen D. Lawn who designed and led the Cohort1 study. Author contributions: T. Broger, B. Sossen, A.D. Kerkhoff, C. Schutz, E.I. Reipold, M.P. Nicol, G. Meintjes and C.M. Denkinger designed the parent study and it was overseen by T. Broger, B. Sossen, A. Trollip, M.P. Nicol, G. Meintjes and C.M. Denkinger. B. Sossen, A.D. Kerkhoff, C. Schutz, R. Burton, A. Ward and G. Meintjes coordinated the individual study sites. A.D. Kerkhoff, T. Broger and C.M. Denkinger designed the statistical analysis and A.D. Kerkhoff undertook the statistical analyses. A.D. Kerkhoff, T. Broger and B. Sossen developed the first manuscript draft. All authors contributed to interpretation of data and editing of the article and approved the final version of the manuscript before submission. Conflict of interest: A.D. Kerkhoff has nothing to disclose. Conflict of interest: B. Sossen has nothing to disclose. Conflict of interest: C. Schutz has nothing to disclose. Conflict of interest: E.I. Reipold has nothing to disclose. Conflict of interest: A. Trollip has nothing to disclose. Conflict of interest: E. Moreau reports grants from GHIT and KfW, during the conduct of the study.

Conflict of interest: B. Sossen has nothing to disclose. Conflict of interest: C. Schutz has nothing to disclose. Conflict of interest: E.I. Reipold has nothing to disclose. Conflict of interest: A. Trollip has nothing to disclose. Conflict of interest: E. Moreau reports grants from GHIT and KfW, during the conduct of the study. Conflict of interest: S.G. Schumacher is an employee of FIND. FIND is a non-for-profit foundation, whose mission is to find diagnostic solutions to overcome diseases of poverty in LMICs. It works closely with the private and public sectors and receives funding from some of its industry partners. It has organisational firewalls to protect it against any undue influences in its work or the publication of its findings. All industry partnerships are subject to review by an independent Scientific Advisory Committee or another independent review body, based on due diligence, TTPs and public sector requirements. FIND catalyses product development, leads evaluations, takes positions, and accelerates access to tools identified as serving its mission. It provides indirect support to industry (e.g. access to open specimen banks, a clinical trial platform, technical support, expertise, laboratory capacity strengthening in LMICs) to facilitate the development and use of products in these areas. FIND also supports the evaluation of prioritised assays and the early stages of implementation of WHO-approved (guidance and PQ) assays using donor grants. In order to carry out test validations and evaluations, has product evaluation agreements with several private sector companies for the diseases FIND works in which strictly define its independence and neutrality vis a vis the companies whose products get evaluated, and describes roles and responsibilities.

nor grants. In order to carry out test validations and evaluations, has product evaluation agreements with several private sector companies for the diseases FIND works in which strictly define its independence and neutrality vis a vis the companies whose products get evaluated, and describes roles and responsibilities. Conflict of interest: R. Burton has nothing to disclose. Conflict of interest: A. Ward has nothing to disclose. Conflict of interest: M.P. Nicol reports grants from FIND, during the conduct of the study. Conflict of interest: G. Meintjes has nothing to disclose. Conflict of interest: C.M. Denkinger is a former employee of FIND. Conflict of interest: T. Broger reports grants from GHIT, personal fees from FIND, during the conduct of the study; and has a patent pending in the field of lipoarabinomannan detection for TB diagnosis.

Conflict of interest: G. Meintjes has nothing to disclose. Conflict of interest: C.M. Denkinger is a former employee of FIND. Conflict of interest: T. Broger reports grants from GHIT, personal fees from FIND, during the conduct of the study; and has a patent pending in the field of lipoarabinomannan detection for TB diagnosis. Support statement: This work was funded by Global Health Innovative Technology (GHIT) Fund grant number G2015-201, UK Department for International Development (DFID) grant number 300341-102, Dutch Ministry of Foreign Affairs grant number PDP15CH14, Bill and Melinda Gates Foundation grant number OPP1105925, Australian Department of Foreign Affairs and Trade (DFAT) grant number 70957 and the German Federal Ministry of Education and Research (BMBF) through KfW. The Cohort1 study was funded by the Wellcome Trust (088590 and 085251). B. Sossen received salary support from the Wellcome Trust (grant number 088316). G. Meintjes was supported by Wellcome Trust (098316 and 203135/Z/16/Z), the South African Research Chairs Initiative of the Department of Science and Technology and National Research Foundation (NRF) of South Africa (grant number 64787), NRF incentive funding (UID: 85858) and the South African Medical Research Council through its TB and HIV Collaborating Centres Programme, with funds received from the National Department of Health (RFA#SAMRC-RFA-CC:TB/HIV/AIDS-01-2014). C. Schultz received funding from the South African Medical Research Council through the National Health Scholarship Programme. A.D. Kerkhoff was supported by the National Institute of Allergy and Infectious Diseases (grant number T32 AI060530). The funders had no role in the study design, data collection, data analysis, data interpretation, or writing of this report. The opinions, findings and conclusions expressed in this manuscript reflect those of the authors alone.

Introduction Pulmonary arterial hypertension (PAH) is a rare disease with an estimated prevalence of 7–26 cases per million in the developed world [1]. It is characterised by increased pulmonary vascular resistance due to vasoconstriction and structural remodelling of pulmonary arterioles, leading to right ventricular hypertrophy and end-stage right heart failure [2]. Despite increased awareness and new therapeutic options, annual mortality remains ∼10% [1]. Approximately 70–80% of heritable PAH and 10–20% of idiopathic PAH patients are known to harbour mutations in the bone morphogenetic protein type II receptor (BMPR2) gene [3]. A recent large study of >1000 PAH patients confirmed the prevalence of causal mutations in BMPR2, as well as in five other established genes (TBX4, ACVRL1, ENG, SMAD9 and KCNK3), and identified PAH-associated mutations in four new genes (ATP13A3, SOX17, AQP1 and GDF2), altogether accounting for 23.5% of the cases studied [4]. The rare mutations in all these genes are inherited in an autosomal dominant manner and exhibit reduced penetrance, indicating that other genetic, epigenetic and/or environmental factors influence the development of PAH.

ew genes (ATP13A3, SOX17, AQP1 and GDF2), altogether accounting for 23.5% of the cases studied [4]. The rare mutations in all these genes are inherited in an autosomal dominant manner and exhibit reduced penetrance, indicating that other genetic, epigenetic and/or environmental factors influence the development of PAH. We and others have demonstrated that one factor strongly correlated with survival in PAH is red cell distribution width (RDW) [5, 6]. A recent hypothesis-free phenome-wide analysis indicated PAH, among several disease descriptors, as the most strongly associated with RDW (OR 2.0, 95% CI 1.75–2.4 per % increase in RDW) in a hospital population [7]. RDW (a measure of red blood cell (RBC) size variability in an individual) is part of the full blood count in standard hospital practice and readily available as a biomarker. RDW correlates with iron status biomarkers and high values can signal early stage iron deficiency or iron deficiency anaemia [8]. RDW increases with decreasing iron as available body iron stores fail to meet the iron demand of RBC synthesis, resulting in RBCs of varied size.

actice and readily available as a biomarker. RDW correlates with iron status biomarkers and high values can signal early stage iron deficiency or iron deficiency anaemia [8]. RDW increases with decreasing iron as available body iron stores fail to meet the iron demand of RBC synthesis, resulting in RBCs of varied size. Iron deficiency is commonly observed in PAH patients and is under investigation as a therapeutic target [9–14]. An observed correlation between two traits does not necessarily imply that interventions on one trait will change the other and there are numerous examples where false-positive associations have led to unsuccessful randomised controlled trials [15, 16]. Clinical trials are expensive and time-consuming, and recruitment can be challenging, especially in rare diseases. With the growing availability of genetic data in large disease-focused and population-based studies, testing causal relationships between traits of interest has become possible by harnessing the naturally occurring genetic variation in the population. The collection of methods used to test causal relationships using genetic variants is called Mendelian randomisation (MR) [17, 18]. MR has been used successfully to help prioritise intervention and drug targets and to identify causal factors for several diseases [19–23]. In general, candidate drugs with genetic evidence for effectiveness are more successful in drug trials compared to those without such genetic support [24].

omisation (MR) [17, 18]. MR has been used successfully to help prioritise intervention and drug targets and to identify causal factors for several diseases [19–23]. In general, candidate drugs with genetic evidence for effectiveness are more successful in drug trials compared to those without such genetic support [24]. It remains to be investigated whether elevated RDW is largely a consequence of PAH or plays a causal role in the condition. The common genetic variation determining RDW levels has been defined in very large population studies [25], with more power than equivalent studies of iron status. This makes RDW a strong candidate for MR analysis and our primary aim was to dissect the epidemiological association between RDW and PAH by assessing causality using MR. We refined the estimate of association between RDW and PAH using biomarker data in 642 PAH cases and >15 000 controls with common diseases from the Imperial College Pulmonary Hypertension Biobank, the UK PAH cohort and the Vanderbilt University Medical Centre (VUMC) and applied MR to investigate whether being genetically predisposed to higher levels of RDW increases the odds of developing PAH. Methods Definition of PAH PAH was defined by internationally agreed criteria [26], specifically mean pulmonary artery pressure >25 mmHg, pulmonary vascular resistance >3 Woods units and pulmonary capillary wedge pressure <15 mmHg. Patients with concurrent diseases known to cause pulmonary hypertension were excluded, such that all were considered to have idiopathic, heritable or anorexigen-induced PAH.

pecifically mean pulmonary artery pressure >25 mmHg, pulmonary vascular resistance >3 Woods units and pulmonary capillary wedge pressure <15 mmHg. Patients with concurrent diseases known to cause pulmonary hypertension were excluded, such that all were considered to have idiopathic, heritable or anorexigen-induced PAH. Data: genetic and phenotype data in contributing studies For our analyses we used both individual-level and summary-level data. Individual-level data including clinical laboratory RDW were available for 524 PAH patients from the Imperial College Pulmonary Hypertension Biobank and the multicentre UK PAH cohort, a study that facilitates collaboration and the sharing of data and samples between the specialist pulmonary hypertension centres in the UK (www.ipahcohort.com). In addition, the hospital population-based VUMC study provided longitudinal clinical laboratory RDW measurements, detailed clinical diagnoses and genome-wide genotype array data (genotyping platform: Illumina MEGAex) for an additional 118 PAH patients and 15 889 common disease controls (supplementary table S1). VUMC hosts a collection of electronic medical records linked to genetic data derived from blood collected during routine clinical assessment in outpatient clinics where all patients are shown the consent form during check-in [27] (https://victr.vanderbilt.edu/pub/biovu/) The exclusion criteria (supplementary figure S1) and the imputation of genotype array data are described in the supplementary methods.

rom blood collected during routine clinical assessment in outpatient clinics where all patients are shown the consent form during check-in [27] (https://victr.vanderbilt.edu/pub/biovu/) The exclusion criteria (supplementary figure S1) and the imputation of genotype array data are described in the supplementary methods. Summary-level data for both RDW and PAH susceptibility were used in the MR analyses to maximise the sample size and therefore the statistical power. Genetic instruments serving as a proxy for RDW were obtained from the largest-to-date (>170 000 individuals) genome-wide association study (GWAS) on haematological traits (here referred to as the “RDW GWAS”) in a population-based study by Astle et al. [25].

maximise the sample size and therefore the statistical power. Genetic instruments serving as a proxy for RDW were obtained from the largest-to-date (>170 000 individuals) genome-wide association study (GWAS) on haematological traits (here referred to as the “RDW GWAS”) in a population-based study by Astle et al. [25]. A large proportion of the idiopathic and familial PAH cases from the UK PAH cohort were whole-genome sequenced as part of the UK National Institute for Health Research BioResource (NIHRBR) [4, 28] study. For genetic effects on PAH susceptibility we used a recent study published by our group and others investigating, in the largest-to-date GWAS, the effects of common genetic variation on PAH risk (“PAH GWAS”) involving 11 744 individuals, of which 2085 were PAH cases. The results of the PAH GWAS were combined through the meta-analysis of four independent studies, one of which is the NIHRBR, with 847 PAH cases and 5048 rare disease controls. The other major contributing study, the US PAH Biobank (PAHB), used a control population with mixed common diseases recruited from outpatient clinics (694 PAH cases and 1560 controls), while the two smaller studies used population-based controls, with 269 and 275 PAH cases and 1068 and 1983 controls in the Pulmonary Hypertension Allele-Associated Risk and British Heart Foundation Pulmonary Arterial Hypertension studies, respectively [4].

ited from outpatient clinics (694 PAH cases and 1560 controls), while the two smaller studies used population-based controls, with 269 and 275 PAH cases and 1068 and 1983 controls in the Pulmonary Hypertension Allele-Associated Risk and British Heart Foundation Pulmonary Arterial Hypertension studies, respectively [4]. Statistical analyses To confirm and refine the estimate for the association between RDW and PAH, we combined PAH patients from the Imperial College Pulmonary Hypertension Biobank, the UK PAH Cohort and the VUMC study and compared them to common disease controls from the VUMC recruited in outpatient clinics (supplementary methods and supplementary table S1).

ine the estimate for the association between RDW and PAH, we combined PAH patients from the Imperial College Pulmonary Hypertension Biobank, the UK PAH Cohort and the VUMC study and compared them to common disease controls from the VUMC recruited in outpatient clinics (supplementary methods and supplementary table S1). To test for causality between RDW and PAH, we applied a two-sample MR strategy that requires effect estimates for the genetic instrument on the risk factor (here RDW) and the outcome (here PAH) from two non-overlapping datasets (supplementary methods and supplementary figure S2). The genetic instrument for RDW comprised genetic variants associated with RDW levels in the RDW GWAS at a study-specific genome-wide level of significance (p<8.31×10−9). In the PAH GWAS, 179 variants (inclusive of 12 proxy variants with a minimum r2 of 0.8) out of the 212 independent (r2<0.01) RDW quantitative trait loci (QTL) were available after excluding 13 palindromic variants (A/T or C/G) with intermediate allele frequencies (minor allele frequency >45%) to ensure that the effects of the variants for the two traits can be harmonised to the same allele. To obtain the causal estimate, we applied the inverse variance weighted (IVW) [29] and weighted median [30] methods. We assessed heterogeneity between the causal estimates from each QTL using Cochran's Q-test (supplementary methods).

t the effects of the variants for the two traits can be harmonised to the same allele. To obtain the causal estimate, we applied the inverse variance weighted (IVW) [29] and weighted median [30] methods. We assessed heterogeneity between the causal estimates from each QTL using Cochran's Q-test (supplementary methods). In the primary MR analysis, we included all available genome-wide significant RDW QTL (n=179). The secondary analysis explicitly tested the role of iron deficiency in the RDW–PAH association using five RDW QTL mapped to genes involved in iron homeostasis (HFE, TMPRSS6, TFRC and TFR2) from the full set of genome-wide significant RDW QTL (supplementary figure S2). All of these five RDW QTL concomitantly increase serum iron, ferritin and transferrin saturation and decrease transferrin, reflecting systemic iron status (supplementary table S2) and were first reported by an independent GWAS (the Genetics of Iron Status) as genome-wide significant signals for at least two of the aforementioned iron status biomarkers [31]. These five RDW QTL are among the signals which explain the highest proportion of variance in the RDW GWAS, highlighting the importance of iron availability in RDW levels.

independent GWAS (the Genetics of Iron Status) as genome-wide significant signals for at least two of the aforementioned iron status biomarkers [31]. These five RDW QTL are among the signals which explain the highest proportion of variance in the RDW GWAS, highlighting the importance of iron availability in RDW levels. Furthermore, we validated the RDW genetic instrument as a proxy for RDW levels in our common disease controls from VUMC. To do so, we regressed the first available RDW measurement on the RDW genetic risk score (GRS) constructed for each individual (supplementary methods) and obtained the proportion of variance explained (R2). The R2 for the same RDW GRS in the RDW GWAS study populations (UK Biobank (UKB) and INTERVAL) was calculated from the summary statistics of the RDW–QTL associations (supplementary methods). Since genetic variants typically explain a small proportion of the variability in the associated trait, MR studies often require large sample sizes to detect a non-zero causal effect. Our power to detect a causal association in the current MR analyses was calculated [32] using the R2 values estimated in VUMC and those estimated in the RDW GWAS populations (supplementary methods). Results Defining the association of RDW and PAH Within this observational analysis, each standard unit (1.4%) increase in RDW was associated with 90% higher odds of prevalent PAH after adjusting for the effects of age and sex (OR 1.90, 95% CI 1.80–2.01). There were no marked between-cohort differences in RDW levels (figure 1 and supplementary table S1).

RDW and PAH Within this observational analysis, each standard unit (1.4%) increase in RDW was associated with 90% higher odds of prevalent PAH after adjusting for the effects of age and sex (OR 1.90, 95% CI 1.80–2.01). There were no marked between-cohort differences in RDW levels (figure 1 and supplementary table S1). FIGURE 1 Boxplot of red cell distribution width (RDW) levels in the merged cohort of pulmonary arterial hypertension (PAH) cases (n=642) and the disease control cohort (n=15 889). The bottom and the top lines of the box indicate the 25th and 75th percentiles, while the centre line indicates the median value. The whiskers extend to 1.5 times the interquartile range from both ends of the box with individual points being more extreme observations. VUMC: Vanderbilt University Medical Center.

5 889). The bottom and the top lines of the box indicate the 25th and 75th percentiles, while the centre line indicates the median value. The whiskers extend to 1.5 times the interquartile range from both ends of the box with individual points being more extreme observations. VUMC: Vanderbilt University Medical Center. Genetic risk score using RDW QTL We estimated that the 179 RDW QTL would explain >12% of the variability (R2 12.7%, 95% CI 12.32%–12.99%) in RDW levels in the RDW GWAS population (UK Biobank and INTERVAL) in which they were discovered. The RDW GRS constructed for the VUMC controls explained 2.6% (95% CI 2.17–3.19%) of the variability in the first available RDW measurement (supplementary table S3). The set of five variants specific to iron status explain an estimated 1.7% (95% CI 1.62–1.87%) of the RDW variability in the UKB and INTERVAL populations while they explain 0.7% (95% CI 0.43–0.92%) of the total variability in RDW in the VUMC controls (supplementary table S3). The RDW GWAS study populations had a lower mean RDW level than our common disease controls (UKB and INTERVAL 13.45; VUMC 13.60) and lower standard deviation (UKB and INTERVAL 0.82; VUMC 1.40), which could explain in part differences between the R2 estimates in these studies.

the VUMC controls (supplementary table S3). The RDW GWAS study populations had a lower mean RDW level than our common disease controls (UKB and INTERVAL 13.45; VUMC 13.60) and lower standard deviation (UKB and INTERVAL 0.82; VUMC 1.40), which could explain in part differences between the R2 estimates in these studies. Statistical power to detect causal effect With a genetic instrument that explains a relatively high proportion (R2 12%) of the total variation in RDW levels (figure 2, red curve) we have 80% power to detect a causal effect as small as 1.25 (odds ratio). If the true variance explained lies closer to that estimated in the VUMC controls (R2 2.6%; figure 2, green curve), this changes to 1.52. When the variance explained by the genetic instrument is small (R2 1.7%; figure 2, blue curve), we are limited, with our sample size, to an odds ratio of ≥1.7. However, if the effect of RDW calculated in our observational analysis (OR 1.80–2.01) was causal in nature, either of the two MR analyses, based on R2 estimates from the RDW GWAS, would have >80% power to detect an effect of that magnitude. Using the estimates based on the VUMC data, the analysis using all 179 RDW QTL (figure 2, green curve) would have sufficient (>80%) power in our sample, while the iron-specific model (figure 2, purple curve) would be underpowered.

estimates from the RDW GWAS, would have >80% power to detect an effect of that magnitude. Using the estimates based on the VUMC data, the analysis using all 179 RDW QTL (figure 2, green curve) would have sufficient (>80%) power in our sample, while the iron-specific model (figure 2, purple curve) would be underpowered. FIGURE 2 Power (%) to detect a causal association (y-axis) given the size of the true underlying causal effect of one standard unit increase in red cell distribution width (RDW) on pulmonary arterial hypertension (PAH) risk (x-axis). n=11 744 (2085 cases). The effect estimate obtained from the observational study is indicated with the vertical black line at OR 1.90 while the red dotted line marks the desired power of 80%. Red curve: Mendelian randomisation (MR) using all overlapping genome-wide significant variants from the RDW genome-wide association study (GWAS), given the true R2 12% as per estimated in the UK Biobank (UKB) and INTERVAL cohorts; green curve: MR using all genome-wide significant quantitative trait loci (QTL) from the RDW GWAS, given the true R2 2.6% as per estimated in our Vanderbilt University Medical Center (VUMC) control cohort; blue curve: MR using five genome-wide significant variants from the RDW GWAS reflecting systemic iron status, given the true R2 1.7% as per estimated in the UKB and INTERVAL cohorts; purple curve: MR using five genome-wide significant QTL from the RDW GWAS reflecting systemic iron status, given the true R2 0.7% as per estimated in our VUMC control cohort.

significant variants from the RDW GWAS reflecting systemic iron status, given the true R2 1.7% as per estimated in the UKB and INTERVAL cohorts; purple curve: MR using five genome-wide significant QTL from the RDW GWAS reflecting systemic iron status, given the true R2 0.7% as per estimated in our VUMC control cohort. RDW–PAH causal relationship We tested for a causal effect of RDW on development of PAH in our primary MR analysis using 179 RDW QTL and found no significant relationship (IVW ORcausal 1.07, 95% CI 0.93–1.23; Q p-value 0.57; figure 3). A secondary MR analysis based on five RDW QTL provided no evidence for a causal effect of iron status on PAH (IVW ORcausal 1.09, 95% CI 0.77–1.54; Q p-value 0.91; figure 3). The weighted median method, which is more robust to violations of MR instrument assumptions, yielded similar estimates for both the primary (ORcausal 1.11, 95% CI 0.89–1.38) and the secondary (ORcausal 1.04, 95% CI 0.68–1.59) MR analyses.

on status on PAH (IVW ORcausal 1.09, 95% CI 0.77–1.54; Q p-value 0.91; figure 3). The weighted median method, which is more robust to violations of MR instrument assumptions, yielded similar estimates for both the primary (ORcausal 1.11, 95% CI 0.89–1.38) and the secondary (ORcausal 1.04, 95% CI 0.68–1.59) MR analyses. FIGURE 3 Scatterplot of variant (red cell distribution width (RDW)) associations (x-axis) plotted against variant (pulmonary arterial hypertension (PAH)) associations (y-axis) where each dot represents a single RDW quantitative trait locus (QTL). The effect estimates and their standard errors (grey bars) are given in standard units for RDW and in odds ratios for PAH. The solid black line denotes an odds ratio of 1 (no effect), while the dashed blue line is the overall causal effect from the inverse variance weighted regression using all 179 RDW QTL. The five iron-specific RDW QTL used as instruments in the secondary Mendelian randomisation analysis are labelled with their corresponding gene names and the red dotted line denotes the corresponding causal effect.

ue line is the overall causal effect from the inverse variance weighted regression using all 179 RDW QTL. The five iron-specific RDW QTL used as instruments in the secondary Mendelian randomisation analysis are labelled with their corresponding gene names and the red dotted line denotes the corresponding causal effect. If the odds ratio estimated in the primary MR analysis was indicative of the magnitude of a real causal effect, the number of PAH cases needed to detect such a causal effect with a genetic instrument that explains 10% of the variance in RDW would be ∼20 600 (with the same 1:4.6 ratio of cases:controls as in this study) to achieve 80% power at a false-positive rate of 5% (p=0.05). We tested for heterogeneity between causal effects estimated in the four studies contributing to the PAH GWAS separately (supplementary figures S1 and S2) to assess if differences in the nature of their control populations yielded heterogenous effect estimates for the instrumental variants. The two heterogeneity tests on the IVW estimates (main MR: Q=0.83, df=3, p=0.84; secondary MR: Q=0.55, df=3, p=0.91) did not detect considerable variability between the causal effects in the four contributing studies based on a random-effects model (supplementary figures S3 and S4).

mates for the instrumental variants. The two heterogeneity tests on the IVW estimates (main MR: Q=0.83, df=3, p=0.84; secondary MR: Q=0.55, df=3, p=0.91) did not detect considerable variability between the causal effects in the four contributing studies based on a random-effects model (supplementary figures S3 and S4). Discussion We applied a two-sample MR approach to test whether the epidemiological relationship between elevated RDW levels, which are associated with iron deficiency, and PAH is causal in nature. We estimated the effect of RDW on PAH in a large sample of cases and common disease controls. By using genetic variants as instruments for RDW, we found no evidence for a causal effect of RDW on PAH of the magnitude suggested by observational studies.

ch are associated with iron deficiency, and PAH is causal in nature. We estimated the effect of RDW on PAH in a large sample of cases and common disease controls. By using genetic variants as instruments for RDW, we found no evidence for a causal effect of RDW on PAH of the magnitude suggested by observational studies. Previous work has shown that iron deficiency is common in PAH and associated with a poor prognosis, reduced exercise capacity and worsening haemodynamics [9, 11, 13]. A physiological link has been described in healthy volunteers where iron infusion attenuated the rise in pulmonary artery pressure induced by acute hypoxia [33, 34] and in rats where chronic iron deficiency results in pulmonary hypertension [35]. This relationship could be driven by the role of iron in destabilising the hypoxia-inducible factor, thereby deficiency can mimic the hypoxic state [36]. Our study confirmed the association of raised RDW with PAH using controls from a hospital population and cases from multiple centres (with an effect size 1.90 for one standard unit RDW, 1.4%), supporting a recent hypothesis-free phenome-wide analysis which indicated, among several disease descriptors, PAH as the most strongly associated with RDW with a similar effect size (OR 2.0, 95% CI 1.75–2.4 per % increase in RDW) [7]. Our MR analysis was adequately powered to detect a causal role for RDW with an effect of this magnitude. The fact that we did not detect a causal effect at this level suggests that at least some of the observed association is secondary to the disease.

lar effect size (OR 2.0, 95% CI 1.75–2.4 per % increase in RDW) [7]. Our MR analysis was adequately powered to detect a causal role for RDW with an effect of this magnitude. The fact that we did not detect a causal effect at this level suggests that at least some of the observed association is secondary to the disease. The results of this study may appear to be at odds with previous clinical studies of the efficacy of iron supplementation in PAH patients, which have focused on functional capacity rather than disease pathology. An open-label study of 20 patients with idiopathic PAH with iron deficiency reported improved iron status, 6-min walk distance and quality of life (QoL) 2 months after a single infusion of 1000 mg ferric carboxymaltose [14]. Another open-label study in 15 iron-deficient idiopathic PAH patients reported improvement of iron status, QoL and exercise endurance capacity on cardiopulmonary exercise testing after receiving 1000 mg of intravenous iron [10]. Neither of the clinical studies were placebo controlled, although Viethen et al. [14] compared their intervention group to a group of matched iron-replete patients who did not receive iron infusion. It remains possible that iron supplementation in PAH could have benefits through mechanisms distinct from those driving the cardiovascular pathology, for example on muscle function [37].

Viethen et al. [14] compared their intervention group to a group of matched iron-replete patients who did not receive iron infusion. It remains possible that iron supplementation in PAH could have benefits through mechanisms distinct from those driving the cardiovascular pathology, for example on muscle function [37]. MR studies using data from large consortia support a causal effect of iron status in other diseases. The genetic instruments (two variants in HFE and one variant in TMPRSS6) used in these studies were also used in our secondary MR analysis. Gill and co-workers found iron to have a protective effect against coronary artery disease (IVW ORcausal 0.94 per sd change in serum iron) [38], but increased the risk of cardioembolic stroke (IVW ORcausal 1.16 per sd iron) [39]. The authors suggest the opposing effects of iron status on coronary artery disease and stroke might be due to different underlying mechanisms. Pichler et al. [19] have reported that iron protects against the risk of developing Parkinson's disease (IVW ORcausal 0.88 per sd iron). Iron deficiency is a common risk factor and these causal effect estimates in common diseases are modest; this study was not powered to detect an effect if this is also true of PAH.

s. Pichler et al. [19] have reported that iron protects against the risk of developing Parkinson's disease (IVW ORcausal 0.88 per sd iron). Iron deficiency is a common risk factor and these causal effect estimates in common diseases are modest; this study was not powered to detect an effect if this is also true of PAH. In the light of this MR analysis, alternative explanations for the association of RDW and PAH have to be considered; namely, that PAH causes raised RDW (reverse causation, for example reduced oxygen delivery and/or haemolysis related to PAH may stimulate reticulocytosis, which would increase RDW) or that PAH and elevated RDW are caused by an independent common mechanism. One such mechanism is chronic inflammation, which is a common feature of PAH [40] and leads to intracellular sequestration of iron. Other mechanisms which may modify RDW, such as folate or vitamin B12 could be studied for their association with PAH. To directly test whether PAH is causal for raised RDW levels a stronger genetic instrument for PAH is required than the currently known common variation identified by PAH GWAS.

ar sequestration of iron. Other mechanisms which may modify RDW, such as folate or vitamin B12 could be studied for their association with PAH. To directly test whether PAH is causal for raised RDW levels a stronger genetic instrument for PAH is required than the currently known common variation identified by PAH GWAS. An important strength of our study lies in the sample size available with phenotype and genetic data achieved through extensive collaboration. Given the rarity of PAH, data had to be pooled from several centres to allow the investigation of common genetic variation in PAH and to test causal relationships. Our study has some limitations. The control population for our observational study was not specifically selected to represent a population at risk of developing PAH. An example of a population at risk of PAH are relatives of patients with pathogenic BMPR2 mutations (and other rare pathogenic variants). Given the reduced penetrance of familial/heritable PAH (∼42% in females and ∼14% in males carrying known mutations) [41], following-up the families of affected individuals, especially relatives harbouring mutations associated with PAH, would be an invaluable source to identify environmental triggers of PAH.

variants). Given the reduced penetrance of familial/heritable PAH (∼42% in females and ∼14% in males carrying known mutations) [41], following-up the families of affected individuals, especially relatives harbouring mutations associated with PAH, would be an invaluable source to identify environmental triggers of PAH. Despite our efforts to exclude controls with conditions that probably affect RDW and to use the first available RDW measurement, we expect genetic effects of RDW levels to differ between individuals with common diseases and a cohort of healthy volunteers. The R2 of a genetic variant describes the variance explained in the phenotype in a given population at a given time. Therefore, there can be no single population parameter that applies to multiple populations or the same population at multiple time points. Although the R2 estimated in the RDW GWAS is likely to be upwardly biased (since it contains the discovery as well as the replication samples), it might reflect better the extent to which genetic variation influences long-term RDW levels in disease-free populations than the R2 estimate from the VUMC disease controls, who may be expected to have more variable RDW levels due to comorbidities. This highlights a potential challenge in estimating power for MR studies. We now estimate that >20 000 PAH patients would be required to detect any likely causal effect of iron deficiency.

ons than the R2 estimate from the VUMC disease controls, who may be expected to have more variable RDW levels due to comorbidities. This highlights a potential challenge in estimating power for MR studies. We now estimate that >20 000 PAH patients would be required to detect any likely causal effect of iron deficiency. It is important to note that in our MR analyses, causal effects were estimated using the results of the PAH GWAS [4]. Four independent studies contributed to the overall results of the PAH GWAS comparing allele frequencies of their PAH cases to control cohorts selected according to different criteria. The NIHRBR study used a control population with a mixture of rare diseases, while the other major contributing study from the United States used mixed common disease controls. The two smaller contributing studies used population-based controls; a preferable design for estimating the effects of surrogate genetic variants for common risk factors. Selection bias can affect the estimates of the instrumental genetic variants on disease susceptibility. This is especially true if the control cohort is enriched for conditions also affected by the risk factor of interest.

eferable design for estimating the effects of surrogate genetic variants for common risk factors. Selection bias can affect the estimates of the instrumental genetic variants on disease susceptibility. This is especially true if the control cohort is enriched for conditions also affected by the risk factor of interest. Conclusions There is strong observational evidence for an association between elevated RDW, a surrogate for iron deficiency, and PAH. However, this MR analysis does not indicate that RDW is causally linked to disease development. Our study was powered to detect a causal effect similar in size to that observed. A more modest causal effect remains possible, but a significantly larger study population would be required to detect it. Extending international collaborations and careful follow-up of populations at risk will allow increasingly sophisticated study designs to investigate causal relationships, shared underlying mechanisms with other conditions and overall genetic susceptibility in PAH. Supplementary material 10.1183/13993003.01486-2019.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-01486-2019.SUPPLEMENT Shareable PDF 10.1183/13993003.01486-2019.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-01486-2019.Shareable

Supplementary material 10.1183/13993003.01486-2019.Supp1Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author. Supplementary material ERJ-01486-2019.SUPPLEMENT Shareable PDF 10.1183/13993003.01486-2019.Shareable1This one-page PDF can be shared freely online. Shareable PDF ERJ-01486-2019.Shareable Acknowledgements We gratefully acknowledge the participation of patients recruited to the UK National Institute of Health Research BioResource – Rare Diseases (NIHR BR-RD) Study, the UK National Cohort of Idiopathic and Heritable PAH, and the National Institutes of Health/National Heart, Lung, and Blood Institute-sponsored National Biological Sample and Data Repository for PAH (aka PAH Biobank). We thank the physicians, research nurses and coordinators in the UK, Europe and at the 38 pulmonary hypertension centres across the USA involved in the PAH Biobank (www.pahbiobank.org). This article has supplementary material available from erj.ersjournals.com Conflict of interest: A. Ulrich has nothing to disclose. Conflict of interest: J. Wharton has nothing to disclose. Conflict of interest: T.E. Thayer has nothing to disclose. Conflict of interest: E.M. Swietlik has nothing to disclose. Conflict of interest: T.R. Assad has nothing to disclose. Conflict of interest: A.A. Desai has nothing to disclose. Conflict of interest: S. Gräf has nothing to disclose. Conflict of interest: L. Harbaum has nothing to disclose.

Conflict of interest: T.E. Thayer has nothing to disclose. Conflict of interest: E.M. Swietlik has nothing to disclose. Conflict of interest: T.R. Assad has nothing to disclose. Conflict of interest: A.A. Desai has nothing to disclose. Conflict of interest: S. Gräf has nothing to disclose. Conflict of interest: L. Harbaum has nothing to disclose. Conflict of interest: M. Humbert reports grants and personal fees from Bayer and GSK, personal fees from Actelion, Merck and from United Therapeutics, outside the submitted work. Conflict of interest: N.W. Morrell reports personal fees from Actelion and Morphogen-IX, outside the submitted work. Conflict of interest: W.C. Nichols has nothing to disclose. Conflict of interest: F. Soubrier has nothing to disclose. Conflict of interest: L. Southgate has nothing to disclose. Conflict of interest: D-A. Trégouët has nothing to disclose. Conflict of interest: R.C. Trembath reports personal fees for advisory board work from Ipsen Pharmaceuticals, personal fees for non-executive board membership from King's College Hospital NHS Foundation Trust, outside the submitted work. Conflict of interest: E.L. Brittain reports personal fees for advisory board work from Bayer, outside the submitted work. Conflict of interest: M.R. Wilkins reports grants from Vifor Pharma, outside the submitted work. Conflict of interest: I. Prokopenko has nothing to disclose. Conflict of interest: C.J. Rhodes reports personal fees from Actelion, outside the submitted work.

Conflict of interest: E.L. Brittain reports personal fees for advisory board work from Bayer, outside the submitted work. Conflict of interest: M.R. Wilkins reports grants from Vifor Pharma, outside the submitted work. Conflict of interest: I. Prokopenko has nothing to disclose. Conflict of interest: C.J. Rhodes reports personal fees from Actelion, outside the submitted work. Support statement: The work cited here is supported by funding from the NIHR BR-RD, the British Heart Foundation (SP/12/12/29836), the BHF Cambridge and Imperial Centres of Cardiovascular Research Excellence (RE/18/4/34215), UK Medical Research Council (MR/K020919/1), the Dinosaur Trust, and BHF Programme grants to R.C. Trembath (RG/08/006/25302), N.W. Morrell (RG/13/4/30107) and M.R. Wilkins (RG/10/16/28575). Funding for the PAH Biobank is provided by NIH/NHLBI HL105333. Vanderbilt University Medical Center's BioVU is supported by numerous sources: institutional funding, private agencies, and federal grants that include the NIH funded Shared Instrumentation Grant S10RR025141; and CTSA grants UL1TR002243, UL1TR000445, and UL1RR024975. Genomic data are also supported by investigator-led projects that include U01HG004798, R01NS032830, RC2GM092618, P50GM115305, U01HG006378, U19HL065962, R01HD074711; and additional funding sources listed at https://victr.vanderbilt.edu/pub/biovu/. E.L. Brittain receives funding from the NIH R01 HL146588, American Heart Association Fellow to Faculty Grant (13FTF16070002) and the Gilead PAH Scholars Award Program. The genotyping of the VESPA samples was supported by RC2GM092618. The authors acknowledge use of BRC Core Facilities provided by the financial support from the Department of Health via the National Institute for Health Research (NIHR) comprehensive Biomedical Research Centre award to Imperial College NHS Trust, Cambridge University Hospitals and Guy's and St Thomas' NHS Foundation Trust in partnership with King's College London and King's College Hospital NHS Foundation Trust and by NIHR funding to the Imperial NIHR Clinical Research Facility. C.J. Rhodes is supported by a British Heart Foundation Intermediate Basic Science Research Fellowship (FS/15/59/31839). L. Southgate is supported by the Wellcome Trust Institutional Strategic Support Fund (204809/Z/16/Z) awarded to St George's, University of London. I. Prokopenko is supported by the Wellcome Trust (WT205915), and the EU H2020 (DYNAhealth, project number 633595). N.W.

e Basic Science Research Fellowship (FS/15/59/31839). L. Southgate is supported by the Wellcome Trust Institutional Strategic Support Fund (204809/Z/16/Z) awarded to St George's, University of London. I. Prokopenko is supported by the Wellcome Trust (WT205915), and the EU H2020 (DYNAhealth, project number 633595). N.W. Morrell is a British Heart Foundation Professor and National Institute of Health Research (NIHR) Senior Investigator. W.C. Nichols is supported by NIH NHLBI HL105333. A.A. Desai receives support from NIH NHLBI R01HL136603. Funding information for this article has been deposited with the Crossref Funder Registry.

Introduction In recent years, the importance of a full growth to maximal lung function in childhood has been reinforced by accumulating evidence that lung function deficits established by school age may track into adult life [1, 2]. Thus, achieving optimal lung function is an important goal in the prevention of chronic respiratory diseases and subsequent mortality, and a major public health objective [3]. However, less is known about factors that might influence lung function trajectories [4, 5]. The association between dietary factors with antioxidant and anti-inflammatory properties and risk of asthma and other chronic respiratory diseases in the general population has been investigated previously [6–8]. Prospective studies examining the association between maternal diet during pregnancy and the occurrence of asthma and other allergic diseases in the offspring have contributed information on the role of dietary exposures early in life [9]. Analyses from the Swedish BAMSE birth cohort also show that a high intake of dietary antioxidants at age 8 years was associated with a reduced risk of subsequent development of IgE sensitisation to inhalant allergens and allergic asthma [10]. A recent prospective study from Japan found a significant inverse association between fruit intake and the onset of respiratory allergic symptoms in schoolchildren [11].

dants at age 8 years was associated with a reduced risk of subsequent development of IgE sensitisation to inhalant allergens and allergic asthma [10]. A recent prospective study from Japan found a significant inverse association between fruit intake and the onset of respiratory allergic symptoms in schoolchildren [11]. Epidemiological studies on the association between dietary antioxidants and lung function show conflicting results [12–14]. Most studies have been cross-sectional, but a prospective study in middle-aged adults from three participating countries of the European Community Respiratory Health Survey indicated that a higher intake of fruits and tomatoes was associated with a slower decline in lung function 10 years later [15]. A case–control study in Puerto Rican children indicated that a diet with frequent consumption of vegetables and grains and low consumption of dairy products and sweets was associated with higher lung function, as measured by forced expiratory volume in 1 s (FEV1) and forced vital capacity (FVC) [16]. Longitudinal studies on childhood diet and subsequent lung function development are still lacking. Thus, it remains unclear if diet at school age influences lung function. The aim of this study was to investigate the association between dietary antioxidant intake at age 8 years and lung function development between 8 and 16 years. In order to estimate the cumulative action of the antioxidants present in foods, total antioxidant capacity (TAC) of the diet was used [10].

ces lung function. The aim of this study was to investigate the association between dietary antioxidant intake at age 8 years and lung function development between 8 and 16 years. In order to estimate the cumulative action of the antioxidants present in foods, total antioxidant capacity (TAC) of the diet was used [10]. Methods Study population and study design The study was conducted within the population-based birth cohort BAMSE (Swedish abbreviation for Children, Allergy, Milieu, Stockholm, Epidemiology), in which 4089 children (born 1994–1996) from predefined areas of Stockholm County, Sweden have been followed repeatedly from infancy [4, 17]. In brief, baseline information was collected through parental questionnaires when the children were aged 2 months on average and follow-up questionnaires eliciting information on symptoms of allergic diseases and selected exposures were answered by the parents when the children were aged 1, 2, 4 and 8 years and by the adolescents themselves at 16 years. At ages 8 and 16 years, participants were invited to clinical examinations, which included anthropometric measurements, lung function testing and blood sampling using standardised methods. Sera were analysed for specific IgE to common inhalant and food allergens. Drop-out rates remained low at all ages, and at 16 years 78% (n=3180) completed the questionnaire and 62% (n=2547) attended the clinical examination. The BAMSE study and respective follow-ups were approved by the regional ethical review board (Karolinska Institutet, Stockholm, Sweden), and written informed consent was obtained from parents at 8 years and study participants at 16 years.

) completed the questionnaire and 62% (n=2547) attended the clinical examination. The BAMSE study and respective follow-ups were approved by the regional ethical review board (Karolinska Institutet, Stockholm, Sweden), and written informed consent was obtained from parents at 8 years and study participants at 16 years. Dietary assessment Diet was assessed at 8 years, using a food frequency questionnaire (FFQ). The FFQ was most often filled out by a parent (57%) or by a parent together with the child (40%) and included questions about 98 foods and beverages commonly consumed in Sweden. Children (n=2614) were asked how often, on average, they had consumed each type of food or beverage during the past 12 months. There were 10 prespecified response categories ranging from “never” to “≥3 times per day”. Calculation of the TAC of the FFQ items has been described previously [10]. Briefly, individual TAC estimates were obtained by combining the information on frequency of consumption of specific food items with information from a database of common foods analysed with the oxygen radical absorbance capacity (ORAC) method [18] on the average ORAC content (μmol Trolox equivalents (TE) per day) of age-specific portion sizes. ORAC values were further energy-adjusted using the residuals method [19]. There were 35 food items (including 20 fruits and vegetables) with available ORAC values, while there was no available information on TAC from dietary supplements.

ORAC content (μmol Trolox equivalents (TE) per day) of age-specific portion sizes. ORAC values were further energy-adjusted using the residuals method [19]. There were 35 food items (including 20 fruits and vegetables) with available ORAC values, while there was no available information on TAC from dietary supplements. Lung function testing Details of lung function testing have been described elsewhere [4]. Briefly, lung function was measured by spirometry at 8 years (n=1832) using a 2200 Pulmonary Function Laboratory (SensorMedics, Anaheim, CA, USA) and by impulse oscillometry (IOS) (n=2452) followed by spirometry at 16 years (n=2056) using a Jaeger MasterScreen-IOS system (Carefusion Technologies, San Diego, CA, USA). The same spirometry test protocol was used at both time points. All participants performed repeated maximal expiratory flow volume (MEFV) measurements. The highest values of FEV1 and FVC were extracted and used for analysis, provided that the subject's effort was accepted as being maximal by the test leader, the MEFV curve passed visual quality inspection and the two highest FEV1 and FVC readings were reproducible according to American Thoracic Society/European Respiratory Society criteria [20]. FEV1/FVC ratios were calculated and expressed as percentages. Standard deviation scores (z-scores) for FEV1, FVC and FEV1/FVC ratio were computed accounting for age, sex, height and ethnicity [21]. Regarding IOS measurements, the mean value of resistance at 5 and 20 Hz, frequency dependence of resistance and the square root of the area of reactance were used for analyses. Measurements of exhaled nitric oxide fraction (FeNO) were performed at 16 years (n=2087) at an expiratory flow of 50 mL·s−1, using an online chemiluminescent analyser (CLD88; Eco Medics AG, Duernten, Switzerland). Details of lung function measurements, as well as asthma and other definitions are described in the supplementary material.

xhaled nitric oxide fraction (FeNO) were performed at 16 years (n=2087) at an expiratory flow of 50 mL·s−1, using an online chemiluminescent analyser (CLD88; Eco Medics AG, Duernten, Switzerland). Details of lung function measurements, as well as asthma and other definitions are described in the supplementary material. Statistical analyses Differences between children who were included and excluded from the study population were analysed by Chi-squared and t-test, for categorical and continuous variables, respectively. The distribution of selected exposure characteristics by tertiles (T1, T2, T3) of TAC (linear relationship not assumed) was compared using Chi-squared test (categorical covariates) and ANOVA (continuous covariates). Multivariate linear regression on the mean was used to analyse associations between dietary TAC in tertiles at age 8 years and lung function parameters at ages 8 and 16 years. Tests for trends were performed by assigning the median value of dietary TAC within each tertile and tested as a continuous variable in the model. Analyses were stratified by sex and potential interactions with sex were tested by the Wald test using an interaction term between TAC and sex in the statistical model. Covariates were identified from previous literature [22] and included maternal age <26 years (yes or no), older siblings (one or more older sibling at birth, yes or no), socioeconomic status (categorised on the basis of parents' occupation as manual and non-manual workers), parental allergic disease (any maternal or paternal history of asthma or hay fever, yes or no), maternal smoking during pregnancy (yes or no) and parental smoking in infancy (yes or no). Additional adjustment for educational level, energy intake, dietary vitamin D and fish intake, supplement use, obesity, physical activity and active smoking at 16 years did not influence the results and was not included in the final models. In order to assess effect modification, stratified analysis by asthma status at 8 years was conducted based on a priori determination. In stratified analysis, the two top tertiles were combined due to small numbers. Sensitivity analysis was conducted using other asthma definitions and symptoms, adjusting for inhaled steroid use, as well as excluding supplement users and children who reported avoidance of fruits or vegetables due to allergic symptoms.

In stratified analysis, the two top tertiles were combined due to small numbers. Sensitivity analysis was conducted using other asthma definitions and symptoms, adjusting for inhaled steroid use, as well as excluding supplement users and children who reported avoidance of fruits or vegetables due to allergic symptoms. Associations between TAC in tertiles at 8 years and spirometry results (main outcome) up to 16 years were further analysed longitudinally by mixed-effects linear regression with a random intercept, an unstructured correlation matrix and restricted maximum likelihood estimation. An interaction term between TAC and the time indicator variable was incorporated into the model to estimate age-specific associations at 8 and 16 years and changes in lung function between 8 and 16 years. For low lung function (binary variable), defined as FEV1 z-score below the 25th percentile (Q1) due to small numbers, logistic regression analysis was used. IOS and FeNO results were analysed on the median using quantile regression, due to non-normally distributed data. Participants who answered the questionnaire with baseline information and follow-up questionnaires at 8 and 16 years, and had a FFQ with a mean energy intake within ±3 log sd, as well as anthropometric and lung function measurements at 8 and/or 16 years were included in the present study. In total, 2307 participants fulfilled these criteria (supplementary figure S1). All analyses were performed using the statistical software Stata (version 13; StataCorp, College Station, TX, USA).

Participants who answered the questionnaire with baseline information and follow-up questionnaires at 8 and 16 years, and had a FFQ with a mean energy intake within ±3 log sd, as well as anthropometric and lung function measurements at 8 and/or 16 years were included in the present study. In total, 2307 participants fulfilled these criteria (supplementary figure S1). All analyses were performed using the statistical software Stata (version 13; StataCorp, College Station, TX, USA). Results Descriptive results on exposure and outcomes The children included in the study population (n=2307) were comparable to the excluded children (n=1782) with regard to distribution of selected characteristics (supplementary table S1). At 8 years, the median TAC intake was 10 067 μmol TE·g−1, which corresponds approximately to two servings of apples per day [10], with males having an 8% lower mean TAC intake compared to females (9963 versus 10 819 μmol TE·g−1, p<0.001). Children with older siblings and children who came from a household with university education level at baseline had a significantly higher TAC intake than those without older siblings and from a household with lower education. Additionally, children with higher TAC intake tended to use less inhaled steroids (table 1). TABLE 1 Distribution of selected characteristics in the study population in relation to the total antioxidant capacity (TAC) of the diet (n=2307)

Results Descriptive results on exposure and outcomes The children included in the study population (n=2307) were comparable to the excluded children (n=1782) with regard to distribution of selected characteristics (supplementary table S1). At 8 years, the median TAC intake was 10 067 μmol TE·g−1, which corresponds approximately to two servings of apples per day [10], with males having an 8% lower mean TAC intake compared to females (9963 versus 10 819 μmol TE·g−1, p<0.001). Children with older siblings and children who came from a household with university education level at baseline had a significantly higher TAC intake than those without older siblings and from a household with lower education. Additionally, children with higher TAC intake tended to use less inhaled steroids (table 1). TABLE 1 Distribution of selected characteristics in the study population in relation to the total antioxidant capacity (TAC) of the diet (n=2307) Tertiles of the TAC of the diet# p-value¶ T1 T2 T3 Subjects 750 779 778 ORAC μmol TE per day 6946 (1768–8615) 10009 (8615–11477) 13530 (11477–33097) Categorical variables Male 433 (57.7) 375 (48.1) 343 (44.1) <0.001 Maternal age <26 years 62 (8.3) 52 (6.7) 45 (5.8) 0.153 Parental allergic disease 251 (33.7) 252 (32.5) 223 (29.1) 0.142 High socioeconomic status 631 (85.4) 658 (85.1) 667 (86.7) 0.625 University education 393 (52.4) 423 (54.4) 457 (58.7) 0.038 Maternal smoking during pregnancy 94 (12.5) 87 (11.2) 90 (11.6) 0.701 Parental smoking during infancy 161 (21.5) 150 (19.4) 160 (20.7) 0.572 Older siblings 316 (42.1) 373 (47.9) 391 (50.3) 0.005 Age 8 years Overweight and obesity 129 (17.2) 157 (20.2) 170 (21.9) 0.070 Physical activity >2 times per week 119 (15.9) 113 (14.5) 112 (14.4) 0.646 Asthma+ 65 (8.8) 54 (7.0) 44 (5.7) 0.064 Inhaled steroid use in the past 12 months 74 (9.9) 58 (7.5) 51 (6.6) 0.046 Inhalant IgE sensitisation 186 (26.3) 210 (28.7) 170 (23.2) 0.058 Food IgE sensitisation 146 (20.6) 156 (21.3) 137 (18.7) 0.455 Allergy to fruits and vegetables§ 88 (12.8) 67 (9.3) 66 (9.3) 0.046 Use of multivitamins 315 (42.5) 316 (41.1) 334 (43.6) 0.622 Fish intake ≥2 times per week 269 (36.0) 293 (37.7) 321 (41.5) 0.079 Age 16 years Overweight and obesity 128 (19.1) 103 (14.5) 111 (15.6) 0.056 Asthma 71 (10.1) 51 (6.9) 62 (8.4) 0.091 Inhaled steroid use in the past 12 months 82 (11.3) 53 (7.1) 52 (6.8) 0.003 Inhalant IgE sensitisation 314 (47.9) 311 (44.7) 285 (40.7) 0.029 Food IgE sensitisation 108 (16.5) 85 (12.2) 87 (12.4) 0.039 Active smoking 76 (10.4) 86 (11.4) 92 (12.1) 0.597 Continuous variables Energy intake kcal 1911.7±467.0 1915.5±451.5 1889.1±464.7 0.474 Data are presented as n, median (range), n (%) or mean±sd, unless otherwise stated. ORAC: oxygen radical absorbance capacity; TE: Trolox equivalents.

8 (16.5) 85 (12.2) 87 (12.4) 0.039 Active smoking 76 (10.4) 86 (11.4) 92 (12.1) 0.597 Continuous variables Energy intake kcal 1911.7±467.0 1915.5±451.5 1889.1±464.7 0.474 Data are presented as n, median (range), n (%) or mean±sd, unless otherwise stated. ORAC: oxygen radical absorbance capacity; TE: Trolox equivalents. #: TAC intake (μmol TE per day) as measured with ORAC assay, energy-adjusted to 1900 kcal per day, presented in tertiles (T1, T2, T3); ¶: p-values were calculated from the Chi-squared test for categorical variables and ANOVA for continuous variables; +: defined based on the parental questionnaire at age 8 years as more than three episodes of wheeze in the past 12 months and/or at least one episode of wheeze in the past 12 months, in combination with inhaled steroids occasionally or regularly; §: allergic symptoms related to fruits and/or vegetables or avoidance of any fruit or vegetable due to allergic symptoms. Distribution of anthropometric and lung function characteristics among children in the 8- and 16-year examination is shown in supplementary tables S2 and S3. Associations between dietary TAC at 8 years and lung function at 8 and 16 years In linear regression analyses, associations between dietary TAC in tertiles at 8 years and spirometry results at 8 and 16 years were not statistically significant, although higher mean FEV1 and FVC were observed for males (supplementary table S4). Tests for trend or interaction with sex were not statistically significant.

linear regression analyses, associations between dietary TAC in tertiles at 8 years and spirometry results at 8 and 16 years were not statistically significant, although higher mean FEV1 and FVC were observed for males (supplementary table S4). Tests for trend or interaction with sex were not statistically significant. Figure 1 presents the results from the mixed-effect model analyses of the longitudinal association between dietary TAC at 8 years and lung function up to 16 years. In analyses of lung function change between 8 and 16 years, there were no associations of TAC and spirometry results for the total study population. Consistent with the linear regression results, higher mean FEV1 and FVC were observed for males, but associations were not significant. FIGURE 1 Associations between total antioxidant capacity (TAC) in tertiles (T1 reference, T2, T3) at 8 years and adjusted spirometry results at 8 and 16 years: a) forced expiratory volume in 1 s (FEV1) z-score; b) forced vital capacity (FVC) z-score; c) FEV1/FVC (%). β-coefficients and 95% confidence intervals were estimated using mixed effect models (n=2115 subjects with 3306 observations), adjusted for maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years. Totals additionally adjusted for sex.

d 95% confidence intervals were estimated using mixed effect models (n=2115 subjects with 3306 observations), adjusted for maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years. Totals additionally adjusted for sex. Associations between dietary TAC at 8 years and lung function at 16 years by asthma status To assess possible effect modification, we stratified our analysis by asthma at 8 years. Asthma prevalence in the study population was 7% (n=163) at 8 years; 106 (65%) out of 163 children with asthma also had IgE sensitisation to inhalant and/or food allergens; 134 (82%) had used inhaled steroids occasionally or regularly and 105 (64%) had used bronchodilators in the past 12 months. Children with asthma had 7% lower mean dietary TAC compared to children without asthma (9708 versus 10 450 μmol TE·g−1, p<0.01). Among children with asthma at 8 years, higher TAC intake (second and third tertiles combined) at 8 years was associated with higher mean FEV1 at 16 years (200.0 mL, 95% CI 38.3–361.6 mL versus −7.3 mL, 95% CI −57.2–42.6 mL among children without asthma, p-value for interaction 0.018) (supplementary table S5). This association remained comparable among children with asthma and IgE sensitisation, as well as after adjustment for inhaled steroid use and exclusion of children who reported avoidance of fruits or vegetables due to allergic symptoms (data not shown), and supplement users (supplementary table S6).

mentary table S5). This association remained comparable among children with asthma and IgE sensitisation, as well as after adjustment for inhaled steroid use and exclusion of children who reported avoidance of fruits or vegetables due to allergic symptoms (data not shown), and supplement users (supplementary table S6). In the longitudinal model, higher TAC intake at 8 years was associated with increased FEV1 at 16 years (0.46 sd, 95% CI 0.11–0.80) among children with asthma (figure 2). Regarding change in lung function between 8 and 16 years, there is some evidence for increased mean FEV1 among children with asthma and higher TAC intake, but results were not statistically significant. In sensitivity analysis using other asthma definitions and symptoms, results were consistent (supplementary table S7). There were no associations among children without asthma, or between TAC intake and FVC. FIGURE 2 Associations between total antioxidant capacity (TAC) (tertiles 2 and 3 combined versus reference tertile 1) at 8 years and adjusted spirometry results at 8 and 16 years stratified by asthma at 8 years: a) forced expiratory volume in 1 s (FEV1) z-score; b) forced vital capacity (FVC) z-score; c) FEV1/FVC (%). β-coefficients and 95% confidence intervals were estimated using mixed effect models (n=1948 subjects without asthma with 3027 observations and n=154 subjects with asthma with 258 observations), adjusted for sex, maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years.

imated using mixed effect models (n=1948 subjects without asthma with 3027 observations and n=154 subjects with asthma with 258 observations), adjusted for sex, maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years. Low lung function at 16 years (defined as Q1 FEV1 z-score) was observed in 36% (46 out of 128) of children with asthma and in 24% (369 out of 1534) of children without asthma. In multivariate logistic regression analysis, higher TAC intake at 8 years was associated with lower odds of low lung function at 16 years among children with asthma (OR 0.28, 95% CI 0.12–0.65), while no association was observed among children without asthma (OR 0.96, 95% CI 0.74–1.25, p-value for interaction between TAC and asthma 0.008) (table 2). TABLE 2 Association between total antioxidant capacity (tertiles (T) 2 and 3 combined versus reference T1) at 8 years and lowest quartile (Q1) of forced expiratory volume in 1 s (FEV1) at 16 years stratified by asthma at 8 years (n=415) No asthma at 8 years Asthma at 8 years p-value for interaction with asthma Subjects n OR (95% CI) Subjects n OR (95% CI) Q1: FEV1 z-score T1 118 Reference 24 Reference 0.008 T2 and T3 251 0.96 (0.74–1.25) 22 0.28 (0.12–0.65) Logistic regression adjusted for sex, height and age at examination, maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years.

core T1 118 Reference 24 Reference 0.008 T2 and T3 251 0.96 (0.74–1.25) 22 0.28 (0.12–0.65) Logistic regression adjusted for sex, height and age at examination, maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years. Finally, higher dietary TAC at 8 years was not associated with any of the measured indices in analyses of lung function using IOS or FeNO (supplementary table S8 and table 3). TABLE 3 Associations between total antioxidant capacity (tertiles 2 and 3 combined versus reference tertile 1) at 8 years and impulse oscillometry (IOS) and exhaled nitric oxide fraction (FeNO) results at 16 years stratified by asthma at 8 years No asthma at 8 years Asthma at 8 years Subjects n β (95% CI) Subjects n β (95% CI) IOS results R5 Pa·L−1·s 1800 −5.1 (−13.8–3.6) 137 4.9 (−31.0–40.9) R20 Pa·L−1·s 1800 −0.2 (−7.7–7.2) 137 −9.5 (−37.7–18.6) R5–20 Pa·L−1·s 1800 −1.5 (−6.3–3.4) 137 −0.7 (−21.9–20.5) AX0.5 (Pa·L−1)0.5 1799 0.2 (−0.3–0.7) 137 0.5 (−1.7–2.7) Additional parameters FeNO ppb 1512 0.6 (−0.4–1.7) 117 −4.6 (−15.5–6.4) IOS and FeNO data were analysed by linear regression on the median, adjusted for sex, height and age at examination, maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years. R5: resistance at 5 Hz: R20: resistance at 20 Hz; R5–20: frequency dependence of resistance; AX0.5: square root of the area of reactance.

nd age at examination, maternal smoking during pregnancy, parental smoking during infancy, parental allergic disease, socioeconomic status, older siblings and maternal age <26 years. R5: resistance at 5 Hz: R20: resistance at 20 Hz; R5–20: frequency dependence of resistance; AX0.5: square root of the area of reactance. Discussion In our study of 2307 children from a population-based birth cohort, higher TAC intake at 8 years was associated with increased FEV1 and decreased odds of low lung function at 16 years among children with asthma. We observed no statistically significant associations between TAC and lung function among children without asthma, or between TAC and other than spirometry measurements.

ohort, higher TAC intake at 8 years was associated with increased FEV1 and decreased odds of low lung function at 16 years among children with asthma. We observed no statistically significant associations between TAC and lung function among children without asthma, or between TAC and other than spirometry measurements. To our knowledge, this is the first prospective study investigating the association between dietary antioxidant intake in early school age and lung function development from school age to adolescence. Fresh fruits and vegetables are dietary sources rich in antioxidants, such as vitamins and minerals, β-carotene, flavonoids, isoflavonoids and polyphenolic compounds [23]. Respiratory airways are highly susceptible to oxidative damage and antioxidants may protect the airways against oxidants from both endogenous (activated inflammatory cells) and exogenous (indoor and outdoor air pollution, smoke exposure) sources [8]. Previous cross-sectional studies in adults [12–14, 24–26] and children [16, 27], and prospective studies in adults [13, 28, 29] have indicated that higher intake of dietary antioxidants may be associated with better lung function. However, responses to antioxidants might be modified by life stage, genetic susceptibility and environmental sources of oxidative stress [7].

4–26] and children [16, 27], and prospective studies in adults [13, 28, 29] have indicated that higher intake of dietary antioxidants may be associated with better lung function. However, responses to antioxidants might be modified by life stage, genetic susceptibility and environmental sources of oxidative stress [7]. In our study, asthma was an effect modifier in the association between TAC and lung function. Oxidative stress plays a major role in the pathophysiology of asthma, due to chronic activation of airway inflammatory cells and a high intake of antioxidants has been reported to be protective against asthma risk and severity [23, 30]. Moreover, changes in gut microbiome modulated by dietary intake have recently been linked to alterations in immune responses and lung disease [31]. Our results are consistent with a previous prospective study showing that fruit and vegetable intake had a beneficial effect on inflammatory response and lung function in asthmatic children [32]. Moreover, a lower mean TAC intake was observed in children with asthma in our study. This is in line with previous studies showing that children with asthma have lower levels of antioxidants in the serum [32, 33]. Thus, additional antioxidants may have greater impact on children with asthma since they have higher demands. In a recent study on diet and allergic symptoms in children, the protective effect observed from higher intake of fruits and vegetables in children aged 6–7 years was less or not observed in children aged 13–14 years [34].

onal antioxidants may have greater impact on children with asthma since they have higher demands. In a recent study on diet and allergic symptoms in children, the protective effect observed from higher intake of fruits and vegetables in children aged 6–7 years was less or not observed in children aged 13–14 years [34]. In our study, we did not observe significant sex differences in the association between TAC and lung function, although males had a lower mean TAC intake compared to females. Sex differences have previously been described among adults, and it was suggested that oxidative stress may be associated with airflow limitation in males, but not in females, due to lower serum antioxidant levels and mediation via hormonal mechanisms [35, 36].

ugh males had a lower mean TAC intake compared to females. Sex differences have previously been described among adults, and it was suggested that oxidative stress may be associated with airflow limitation in males, but not in females, due to lower serum antioxidant levels and mediation via hormonal mechanisms [35, 36]. A major strength of our study is the population-based longitudinal design and the large sample size with limited loss to follow-up. In contrast to most previous studies that have focused on fruits, vegetables and individual antioxidants [11, 25, 27–29], we used TAC, which reflects the sum of dietary antioxidant intake and takes synergistic and antagonistic effects between compounds into account [10]. Nevertheless, associations with specific nutrients may be diluted using the TAC approach. Additionally, TAC was available only at 8 years and potential dietary changes from 8 to 16 years were not taken into account. Of the 98 food items in the FFQ, 35 had available TAC values, including the most important dietary antioxidant sources, such as fruits, vegetables, whole grains, nuts and chocolate [18]. In contrast, dietary supplements were not included in the calculation of TAC. However, we were able to adjust for use of dietary supplements and several other confounding factors. Moreover, we excluded children who used supplements to control for potential misclassification of exposure and children who reported avoidance of fruits or vegetables due to allergic symptoms to control for potential reverse causality [37], but these exclusions did not affect the observed associations. Lung function was measured using standardised protocols at 8 and 16 years, although lung function measurements post-bronchodilator were not available. This repeated assessment is a major strength of our study, which enabled us to study TAC in relation to change in lung function. Additionally, IOS, a method measuring respiratory mechanics in contrast to airway calibre measured by spirometry, has not been described in relation to TAC.