Growth differentiation factor 15 is positively associated with incidence of diabetes mellitus: the Malmö Diet and Cancer–Cardiovascular Cohort

The Malmö Diet and Cancer (MDC) study is a large prospective cohort study that includes men and women from Malmö, a city in southern Sweden [16]. From 1991 to 1994, 6103 participants were randomly selected from the MDC study population and were invited to undergo additional examinations to study the epidemiology of carotid artery atherosclerosis. These individuals constituted the subcohort of the MDC cohort, the Malmö Diet and Cancer–Cardiovascular Cohort (MDC-CV) [17]. Among them, 291 were excluded on the basis of missing data on waist circumference or smoking. Of the remaining 5812 participants, 5249 attended a second appointment for fasting blood collection. Participants who had missing data on LDL-cholesterol or glucose (n = 21), or whose remaining stored blood samples were insufficient for GDF-15 measurement (n = 380), or who did not pass an internal quality control test for biomarker analysis (n = 118) were excluded, so that 4730 participants with complete data on GDF-15 and covariates remained in the study. Further exclusion of 193 individuals with self-reported diabetes or use of glucose-lowering medication, and 177 with fasting blood glucose ≥6.1 mmol/l (corresponding to a fasting plasma glucose cut-off of 7.0 mmol/l for the diagnosis of diabetes [18]), left 4360 participants for the cohort analysis (1682 men and 2678 women, age = 57.4 ± 5.96 years [mean ± SD]) (Fig. 1). Complete information on insulin and the HOMA index was available for 4336 participants. In sensitivity analyses, a further 93 individuals were excluded because CRP measurements were not available. All participants provided written informed consent. The study was approved by the Regional Ethical Review Board in Lund, Sweden (LU 51/90) (LU 2012/762) and was carried out in accordance with the Declaration of Helsinki.

Anthropometric measurements were made at baseline using standard procedures. Waist circumference was determined midway between the lowest rib margin and iliac crest. Blood pressure was measured after individuals had rested for 10 min in a supine position. Information about current medication, smoking and alcohol consumption was obtained from a self-administered questionnaire. Participants were classified as non-smokers (former or never smokers) or current smokers (regular or occasional smokers). Men with alcohol intake >40 g per day and women with alcohol intake >30 g per day were classified as having a high level of alcohol consumption.

Blood samples were collected from the cubital vein after an overnight fast. Fresh plasma samples were used for measurement of levels of fasting blood glucose (mmol/l), according to standard procedures at the Department of Clinical Chemistry, University Hospital, Malmö. Insulin was measured by radioimmunoassay. The HOMA index was calculated as fasting insulin × fasting glucose/22.5 [17]. LDL-cholesterol concentrations (mmol/l) were estimated using the Friedewald formula [19]. Measurements of GDF-15 and CRP (nmol/l) used fasting plasma samples that had been stored at −80°C immediately after collection. CRP was analysed with the Tina-quant CRP latex assay (Roche Diagnostics, Basel, Switzerland). GDF-15 was measured by the SciLifeLab analysis service (Uppsala, Sweden) using the Proseek Multiplex CVD I ^96 × 96 reagent kit (Olink Proteomics, Uppsala, Sweden) [20]. The high-specificity Proseek assay uses proximity extension technology [21, 22], so that antibody binding of antigens (such as GDF-15) brings oligonucleotide pairs into proximity, enabling PCR amplification of a reporter sequence. For GDF-15, the lower and upper limits of quantification were 15.3 and 62,500 pg/ml, respectively. Intra-assay variation was 9%, which was the mean coefficient of variation for seven individual samples within each of nine separate runs during validation. Inter-assay variation was 11%, which was the mean coefficient of variation for the same seven individual samples between the nine separate runs during validation. Raw Proseek data underwent a pre-processing normalisation procedure. For each data point (corresponding to PCR amplification resulting from binding to a specific protein in one sample well), normalisation for technical variation (such as pipetting differences) was performed by subtraction of the quantification cycle (C_q) value in that well for the extension control (an antibody linked to a pair of oligonucleotides, resulting in antigen-independent PCR amplification). To control for variation between plates, a standard sample containing goat IgG conjugated to each of the 92 oligonucleotide pairs of the assay (the inter-plate control [IPC]) was added to three wells of each plate. Normalisation between runs was performed by subtracting the median IPC C_q from all the extension control-adjusted values on a plate, resulting in normalised protein expression (NPX, log₂ scale) values. The NPX values were finally adjusted to give a background level (from negative controls) of around zero. Levels of plasma GDF-15 measured by the Proseek assay correlated closely with those measured with an electrochemiluminescence immunoassay (Roche Diagnostics, Mannheim, Germany) [23] in a sample of 987 men and women from a different population to the cohort in the current study (r = 0.89 [L. Lind, unpublished data]).

All participants without a previous history of diabetes were followed until incident diabetes, emigration from Sweden, death or the end of follow-up (31 December 2014), whichever came first. Information on new diabetes cases was retrieved from both local and national registers [24], consisting of the Swedish National Diabetes Register (NDR), the regional Diabetes 2000 register of the Scania region, the Malmö HbA_1c register (MHR), the Swedish inpatient register, the Swedish outpatient register and the nationwide Swedish drug prescription register. In the Swedish NDR and the Diabetes 2000 register, new cases of diabetes were diagnosed according to established criteria (fasting plasma glucose concentration ≥ 7.0 mmol/l with two repeated tests on separate occasions). In MHR, individuals were considered to have developed diabetes if they had at least two HbA_1c recordings ≥42 mmol/mol (6.0%) with the Swedish Mono-S standardisation system (corresponding to 53 mmol/mol [7.0%] according to the US National Glycohemoglobin Standardization Program) after the baseline examination. In the Swedish inpatient and outpatient registers, diabetes was diagnosed by a senior physician. In the nationwide prescription register, a filled prescription of insulin or glucose-lowering medications (ATC-code A10) was required for diagnosis of diabetes.

Because the distribution of values was skewed, CRP measurements were logarithmically transformed before analyses. Participants were divided into quartiles according to GDF-15 concentrations, with a stable proportion of men and women across quartiles. To describe baseline characteristics of participants in each quartile, continuous variables are presented as means ± SD (or medians [25–75%] for skewed distributions), and categorical variables are presented as percentages. Differences in characteristics across GDF-15 quartiles were determined by analysis of variance for continuous variables and logistic regression for categorical variables. Multiple linear regression including other known covariates was used for identification of factors independently associated with GDF-15 concentration. The incidence of diabetes across GDF-15 quartiles was plotted with the Kaplan–Meier curve. Cox proportional hazards regression was used for assessment of the associations between baseline GDF-15 quartiles or SD increments of GDF-15 and incident diabetes, with ‘time in study’ as the timescale. In a sensitivity analysis, age was also used as the timescale. The analyses were adjusted for potential confounders including age, sex, waist circumference, LDL-cholesterol, fasting glucose, systolic blood pressure, anti-hypertensive medication, smoking and alcohol consumption. A sensitivity analysis was also adjusted for CRP. In other sensitivity analyses waist circumference was replaced with body mass index or waist:hip ratio, and adjustment was made for physical activity and total energy intake (see [25] for detailed measurement methods).

Possible interactions between GDF-15 levels and risk factors, with respect to incidence of diabetes, were explored by introduction of interaction terms in the multivariate model (one at a time). Restricted cubic spline functions [26, 27] were incorporated into the Cox model to explore the shape of the association between GDF-15 and diabetes in adjusted analyses, with knots placed at 20%, 40%, 60% and 80% of GDF-15 concentration. Similarly, possible time-dependent effects were estimated with knots placed at 5, 10, 15 and 20 years of follow-up. If (post hoc) threshold or time-dependent effects were detected, risk was estimated by fitting piecewise linear models [26] or sub-distribution hazards models with time-dependent effects [28], respectively. A competing-risks model was applied to investigate the association between GDF-15 and incidence of diabetes, with deaths from causes other than diabetes used as competing events [28].