We used a case-cohort study design nested within the EPIC-Heidelberg study. EPIC-Heidelberg is one of the study centres for the European Prospective Investigation into Cancer and Nutrition (EPIC) study, which recruited more than half-million (510,000) participants across 10 European countries who have been followed up for more than 15 years to investigate the associations between diet, metabolic and lifestyle factors with risk of cancer and other chronic conditions. At the EPIC-Heidelberg study center, between 1994 and 1998, 25,540 participants were recruited from the local population in Heidelberg and surrounding municipalities and baseline information about participants’ health, and social and economic status was collected, including anthropometric measurements (weight, height, waist and hip circumferences). A blood sample was also drawn from the participants on the day of recruitment, regardless of fasting status, and kept for a maximum of 24h at 4 to 10 ˚C until centrifugation, and further processing. Blood samples were aliquoted into fractions of plasma, serum, erythrocytes and buffy coat and stored under liquid nitrogen at −196 °C. Informed consent was obtained from all participants at baseline. Participants also provided self-reports about their hypertension and diabetes status. For diabetes status, heamoglobin A1c (HbA1c) levels were also measured from the blood samples collected at recruitment.
Prospective outcome ascertainment
Cases of chronic diseases among EPIC-Heidelberg participants were prospectively ascertained through active follow up of the study participants directly or through their next-of-kin, as well as through linkages to hospitalisation records and cancer and pathology registries. Mortality outcomes were ascertained from death certificates which were collected from mortality registries. For the cardiovascular disease incidence outcome, all verified incident cases of myocardial infarction (MI) (International Classification of Diseases (ICD)-10: I21) and stroke (ICD-10: I60, I61, I63, I64), diagnosed up to the end of December 2014 were included. All cases were validated and coded by a study physician based on medical records and only verified cases are included in this study.
We used the case-cohort design because it allows investigation of several different outcomes while sparing the excessive use of biological samples. The detailed sampling process of the case-cohort has been previously described elsewhere. Briefly, the sub-cohort was selected using a 2-step age-stratified sampling from the EPIC-Heidelberg cohort. The first step of the sampling process consisted of selection of a random 10% of all EPIC-Heidelberg cohort and included all incident cases of chronic diseases diagnosed until December 2009. A second sampling consisted of additional 10% random sampling of the participants who were older than 50 years at baseline and were not part of the initial sample and included all cases of chronic diseases and deaths occurring until December 2014. The 2 samples were then merged to obtain the final sub-cohort, with a total of 3794 randomly selected study participants of whom 3,591 had their adiponectin levels measured. The case-cohort included a total of 1,387 verified cases of cardiovascular disease (MI and stroke), 582 deaths from cardiovascular causes and 2,352 total deaths that occurred until the end of December 2014. There were 253 cases of MI and stroke (18.2% of all cases of MI and stoke), 130 (22.4%) deaths from cardiovascular and 459 (19.5%) deaths from all causes in the subcohort.
Laboratory measurement of Adiponectin and NT-proBNP
Serum and erythrocyte samples from cases and controls were retrieved from long term liquid nitrogen storage (cryo-straws), placed in temporary -80°C storage and sorted into batches on dry ice. Adiponectin and NT-proBNP were measured on the Meso Scale Discovery (Maryland, USA) electrochemiluminescence Quickplex SQ120 platform. All assays were carried out according to the manufacture’s protocols. Briefly, small spot streptavidin plates were coated with a biotin coupled capture antibody for two hours followed by washing. Wells were then coated with assay buffer followed by the addition of standards, QCs and samples with an incubation of one hour with shaking. Wells were washed and detection antibody added for an hour with shaking. Wells were then washed a final time, read buffer added and the plate read within five minutes. HbAc1 was measured using the Variant II Turbo HPLC system (Bio-Rad-Laboratories, Inc., Hercules, California, USA) according to the manufactures protocols. The inter- and intra-batch coefficients of variation (CV) for quality controls were 3.2% and 18.7% for adiponectin, 3.9% and 18.8% for NT-proBNP and 2.0% and 3.9% for HbA1c. All samples were blinded to laboratory personnel and went through only one freeze thaw cycle before measurement.
The characteristics of participants at baseline are described separately for the sub-cohort and for cases of each of the outcomes namely; cardiovascular disease incidence, cardiovascular mortality and all-cause mortality. We conducted analysis of variance to compute marginal means and tests for differences in the mean concentrations of adiponectin by strata of categorical covariates. For continuous covariates, we calculated partial correlations with adiponectin using Pearson’s coefficients. Marginal means and partial correlations were adjusted for sex, BMI (continuous), smoking (never, long time quitters, short time quitters, current light, and current heavy smokers).
Relative hazards for cardiovascular disease risk and mortality were estimated for the associations between log-transformed and standardised adiponectin levels and CVD risk and mortality using Prentice-weighted Cox proportional hazards models. We used inverse sub-cohort sampling probability to account for case-cohort sampling design and the oversampling of older participants. Participants who were 50 years or younger were assigned a probability of 10%. Those who were older than 50 years were assigned a probability of 19% (10% given they were not drawn in the first selection step (a 90% probability): 10% + (10% × 90%)). The standardisation of adiponectin was sex specific and based on the distribution of the adipokine in the subcohort. The hazards ratios were thus estimated per 1-standard deviation (SD) in the log-transformed adiponectin levels. In all models, age was the underlying timescale, and all models were additionally stratified by age at recruitment (5-year category). All models were adjusted for sex, BMI (continuous), and smoking (never, long time quitters, short time quitters, current light, and current heavy smokers). For cardiovascular disease incidence, the hazard ratios (HR) and their 95% confidence intervals (CI) were computed for any first occurrence of incident cardiovascular event (where stroke and MI were considered as mutually competing events). Proportional hazards assumption was tested and was not violated in any of the models according to an extended version of the Schoenfeld residuals test.
First, we examined heterogeneity in the association between adiponectin and three endpoints; cardiovascular disease incidence, cardiovascular mortality and all-cause mortality within sub-groups according to: NT-proBNP, sex, BMI, smoking status, waist circumference, diabetes, and hypertension status. NT-proBNP was binned into tertiles. For diabetes, we had both information on both laboratory measured HbA1c and self-reported diabetes status. HbA1c was categorised using cut-offs; ≥42 mmol/mol as normal, 42 - <48 mmol/mol as prediabetes and ≥48 as diabetes. Non-smokers and quitters were merged together because our earlier analyses showed similar incidence and mortality trends for the two groups. Heterogeneity was examined by including interaction terms between adiponectin and covariates of interest and p-values for interaction/heterogeneity (phet) were based on the Wald test.
For NT-proBNP which showed significant associations with outcomes and significant heterogeneity, we conducted further analyses to examine the shape of the association between adiponectin and incidence and mortality within tertiles of NT-proBNP. For these analyses, adiponectin concentrations were binned into quintiles in a sex-specific manner. Within each tertile of NT-proBNP, hazards ratios were estimated for increasing levels of adiponectin with the lowest quintile (Q1) as the reference. Visualisation of the dose-response association between adiponectin and mortality by strata of NT-proBNP were performed using natural cubic spline plots, specifying 3 knots evenly placed across the range of the data and with median adiponectin levels (19.5 mg/ml) as the reference value. All analyses were performed using SAS v.9.4 (SAS Institute).