figure b

Introduction

Growth differentiation factor-15 (GDF-15) is a protein of the TGF-β cytokine superfamily [1], which is expressed in several human tissues [2]. The putative effects of GDF-15 have been described in mechanistic studies [3,4,5], which point to its role in oxidative stress, mitochondrial function, energy balance and glucose homeostasis.

Diabetes is an established risk factor for incident heart failure (HF) [6], but the pathways linking diabetes and HF remain poorly understood. Prior epidemiological studies have described a positive association between GDF-15 and diabetes [7, 8], and robust associations of high GDF-15 levels with incident HF [9,10,11,12] and HF prognosis [13,14,15]. GDF-15 is considered a marker of mitochondrial dysfunction [16, 17], reflecting alterations in cellular stress pathways [18]. Diabetes impairs mitochondrial function [19], and adversely affects myocardial energetics even in the absence of overt HF [20, 21]. Laboratory [22] and clinical [20, 23] data suggest that in diabetes-related HF alteration in myocardial metabolism is likely more pronounced than among individuals with HF but without diabetes. These diabetes-related myocardial alterations are directly related to the degree of cellular stress and mitochondrial dysfunction, which can be reflected by GDF-15 levels [16, 17]. This suggests that the association of GDF-15 with incident HF might differ in the presence of diabetes compared with its absence. Despite the accruing evidence on the associations between diabetes and GDF-15, there are limited clinical or population-based data comparing how GDF-15 might improve prediction of HF in people with and without diabetes.

We used data from the community-based Atherosclerosis Risk in Communities (ARIC) study at Visit 3 (1993 to 1995) to examine the associations of GDF-15 and diabetes with incident HF, individually and in combination.

Methods

Study population

The ARIC study recruited 15,792 participants from four US communities [24]. The first study visit took place in 1987–1989; since then, participants have returned for subsequent study visits and received annual telephone calls. The third visit (Visit 3) took place in 1993 to 1995.

Of the 12,887 participants who attended ARIC Visit 3, we excluded individuals with missing GDF-15 measurements (n = 1427), participants who were black people from Minneapolis and Washington County (n = 35) because of their small number, participants with missing diabetes status (n = 32) and participants with prevalent HF (based on Gothenburg criteria and prior hospitalisation related to HF, n = 823), thus leaving 10,570 participants for this analysis.

All participants provided written informed consent and the study protocol was approved by the Institutional Review Board at each study site.

Laboratory measures

GDF-15 was measured in plasma samples collected during ARIC Visit 3 (1993 to 1995) and stored at −70°C prior to analysis using SOMAscan assay (SomaLogic, Boulder, CO, USA) and expressed in relative fluorescence intensity units. For the purposes of analyses, proteins, reported in relative fluorescence units, were log2 transformed because of skewed distributions, and values outside of 5 SDs on the log2 scale were winsorised.

The experimental process for proteomic assessment and data normalisation has been previously described. The relative concentration of plasma proteins or protein complexes was measured using a Slow Off-rate Modified Aptamer (SOMAmer)-based capture array [25]. In brief, this method uses short single strands of DNA with chemically modified nucleotides, called modified aptamers, which act as protein-binding reagents with defined three-dimensional structures and unique nucleotide sequences, which are identifiable and quantifiable using DNA detection technology. The SOMAscan assay has been described in detail previously [26], as have the assay’s performance characteristics [27, 28]. Studies have demonstrated a median intra- and inter-run coefficient of variation of approximately 5% and intra-class correlation coefficients of ~0.9 [25, 29]. The SOMAscan assay has a sensitivity that is comparable to that of immunoassays while extending the lower limit of detection (in the femtomolar range) down to below that offered by conventional immunoassay approaches [30].

Ascertainment of diabetes status

Prevalent diabetes at Visit 3 was defined by a physician-reported diagnosis of diabetes, self-reported use of diabetes medications, a non-fasting blood glucose level ≥11.1 mmol/l (200 mg/dl) or a fasting plasma glucose (FPG) ≥7 mmol/l (126 mg/dl).

Incident outcome assessment

The outcome of interest was incident HF, defined as the first hospitalisation or death related to HF occurring after Visit 3, with follow-up until 31 December 2019. Participants were called on a yearly basis to obtain information regarding hospitalisations, and vital records were examined for all deaths. Hospitalisations and deaths due to incident HF were defined by HF discharge codes (ICD-9 code 428 for hospitalisations early during follow-up and ICD-10 code I50 for later) [31].

Covariates assessment

Information on medical history, medication use, current alcohol use and current smoking was obtained using standardised self-report questionnaires. Physical activity was assessed using the interviewer-administered Baecke questionnaire [32], and categorised as per the American Heart Association guidelines as poor, intermediate and recommended [33]. Systolic and diastolic BP measurements were recorded as the mean of two readings. Hypertension was defined as systolic BP ≥130 mmHg, diastolic BP ≥80 mmHg or use of antihypertension medications. BMI was calculated as weight in kilograms divided by the square of height in metres, and obesity was defined as BMI ≥30 kg/m2. Plasma glucose was measured using the hexokinase method. Serum total cholesterol, triacylglycerol and HDL-cholesterol concentrations were measured by using automated enzymatic assays. LDL-cholesterol was calculated using the Friedewald equation. eGFR was calculated from serum creatinine using the Chronic Kidney Disease Epidemiology Collaboration equation [34]. N-terminal proB-type natriuretic peptide (NT-proBNP) and high-sensitivity cardiac troponin T (hs-cTnT) were also measured using an electrochemiluminescent immunoassay on an automated Cobas e411 analyser (Roche Diagnostics, Mannheim, Germany).

Statistical analysis

We compared the baseline characteristics of participants across GDF-15 quartiles using the ANOVA procedure (for continuous variables) or the χ2 test (for categorical variables).

In cross-sectional analyses, we evaluated the association of diabetes with higher levels of GDF-15 at Visit 3, using multivariable logistic regression. We built a number of sequential models. Model 1 adjusted for age, sex and race/centre. Model 2 included the Model 1 variables as well as current smoking, systolic BP, use of antihypertensive medications, use of cholesterol-lowering medications, total cholesterol, HDL-cholesterol, triacylglycerols, BMI, eGFR and metformin use, as this medication can impact GDF-15 levels [35]. Model 3 adjusted for Model 2 plus NT-proBNP and hs-cTnT.

In prospective analyses, we used Cox proportional hazard regression models to estimate HRs (95% CIs) for the prospective association between GDF-15 at baseline and incident HF by diabetes status, and after adjustment for baseline risk factors, as well as the joint associations of diabetes status and GDF-15 with the incidence of HF. For all the HF incidence models, we initially adjusted for age, sex and race/centre (Model 1). The subsequent adjustments included variables in Model 1 plus education, current smoking, physical activity, systolic BP, use of antihypertensive medications, use of cholesterol-lowering medications, total cholesterol, HDL-cholesterol, triacylglycerols, BMI, eGFR, metformin use, use of diabetes medication other than metformin and diabetes duration (Model 2). We additionally accounted for hs-cTnT and NT-proBNP (Model 3), the use of medication including β-blockers, and angiotensin converting enzyme inhibitors or angiotensin receptor blockers use (Model 4), and for coronary heart disease as a time-varying covariate (Model 5). GDF-15 was modelled in quartiles and as restricted cubic and linear splines to more flexibly evaluate the associations with HF by diabetes status.

We tested for the diabetes and GDF-15 interaction for the incident HF outcome on the multiplicative scale. An interaction between GDF-15 and sex was also investigated. We conducted additional analyses of the prospective associations of cross categories of GDF-15 (in quartiles) and diabetes status (yes vs no) with incident HF; individuals without diabetes and in the lowest GDF-15 quartile served as the reference group.

We assessed the additive predictive value of GDF-15 above and beyond traditional risk factors, including diabetes, by evaluating the changes in C statistic (prediction statistic) associated with the addition of GDF-15 to traditional HF risk factors in the overall sample, as well as, separately, in individuals with and without diabetes.

A p value <0.05 was used to denote statistical significance, including for interaction tests. All analyses were performed using Stata version 15 (StataCorp, USA).

Results

A total of 10,570 individuals were included in our analysis (mean age of 60.0 [SD: 5.7] years, 54% women, 27% black participants, mean GDF-15: 14.4 [SD: 0.50]). Table 1 shows the baseline characteristics of participants by quartile of GDF-15. Individuals in the highest GDF-15 quartile were older and more likely to have hypertension, diabetes or coronary heart disease, as well as elevated hs-cTnT and NT-proBNP, but were less likely to be women, drinkers, smokers or to be obese. The characteristics of participants by diabetes and by HF status are shown in electronic supplementary material (ESM) Tables 1, 2.

Table 1 Baseline characteristics of ARIC study participants at Visit 3 (1993–1995), by quartiles of GDF-15

Diabetes and GDF-15 levels association

Elevated GDF-15 was more common in people with diabetes compared with those without diabetes (32.8% vs 23.6%, p<0.0001). Diabetes status was associated with elevated GDF-15 levels (≥75th percentile), even after adjustment for traditional HF risk factors (ESM Table 3). After adjusting for relevant risk factors, the OR for the association of diabetes with elevated GDF-15 was 1.59 (95% CI 1.38, 1.84) (Model 2, ESM Table 3). After an additional adjustment for NT-proBNP and for hs-cTnT levels the association was attenuated but remained significant (OR 1.56; 95% CI 1.39, 1.78; Model 2, ESM Table 3).

Diabetes, GDF-15 and HF

Over a median of 23 years of follow-up, 2429 incident HF events occurred within the study sample (ESM Table 4). Higher GDF-15 at baseline was associated with an increase in the risk of HF (Model 2, Table 2, ESM Fig. 1), with an HR for the highest GDF-15 quartile (GDF-15 values: 14.7–17.1) vs the lowest quartile (GDF-15 values: 12.4–14.1) of 1.70 (95% CI 1.49, 1.94). There was a statistically significant interaction between GDF-15 and diabetes status on the outcome of incident HF (p for interaction = 0.034). In analyses stratified by diabetes status (Model 2, Table 2), GDF-15 was significantly associated with incident HF among those without diabetes (HR for the highest GDF-15 quartile vs the lowest quartile: 1.64 [95% CI 1.41, 1.91]), but more so among those with diabetes (HR for the highest GDF-15 quartile vs the lowest quartile: 1.72 [95% CI 1.32, 2.23]). Additionally, accounting for cardiac biomarkers (hs-cTnT and NT-proBNP), medication use (β-blocker use, and angiotensin converting enzyme inhibitors or angiotensin receptor blockers) and coronary heart disease as a time-varying covariate did not appreciably affect the magnitude or significance of the effect estimates (ESM Table 5).

Table 2 HRs (95% CIs) for the associations of GDF-15 and incident HF post ARIC Visit 3, stratified by diabetes status

Among individuals with diabetes, we observed roughly J-shaped associations of GDF-15 with HF (Fig. 1b), whereas among those without diabetes, the association of GDF-15 and incident HF was roughly linear (Fig. 1c).

Fig. 1
figure 1

HRs (95% CIs) for the association of GDF-15 with incident HF overall and according to diabetes status. GDF-15 was modelled as a restricted cubic spline (solid line) and as a piece-wise linear spline (dashed line) (knots at percentiles 5, 27.5, 50, 72.5 and 95); y-axes are plotted on a logarithmic scale. (a) Overall sample. (b) Diabetes. (c) No diabetes

In the overall study population, the addition of GDF-15 to a model including traditional risk factors, among which diabetes (Model 2, Table 2), showed that GDF-15 significantly improved risk prediction for HF (C statistic for model without GDF-15: 0.753 vs C statistic for model with GDF-15: 0.758, C statistic improvement [ΔC statistic]: +0.005, p for difference: <0.0001). Among individuals with diabetes, the C statistic for the model without GDF-15 was 0.721 vs C statistic for model with GDF-15: 0.729, and ΔC statistic was +0.008 (p = 0.0001). In those without diabetes, the C statistic for the model without GDF-15 was 0.736 vs C statistic for model with GDF-15: 0.742, and ΔC statistic was +0.006 (p<0.0001).

The examination of the joint association of diabetes and GDF-15 with HF showed that individuals in the top quartile of GDF-15 with diabetes had an HR of 2.46 (95% CI 1.99, 3.03) for incident HF relative to those in the lowest GDF-15 quartile without diabetes (Model 2, Table 3). Additional adjustments for cardiac biomarkers (NT-proBNP and hs-cTnT), the use of medications (β-blockers use, and angiotensin converting enzyme inhibitors or angiotensin receptor blockers) and coronary heart disease as time-varying covariates did not affect the magnitude of the effect estimate and its significance (ESM Table 6).

Table 3 HR (95% CI) for the joint associations of diabetes and GDF-15 with incident HF post ARIC Visit 3

Given the significance of the GDF-15 and sex interaction (p for interaction = 0.0034), we also conducted sex-specific analyses. Among men (ESM Table 7), GDF-15 was significantly associated with incident HF among those without diabetes (HR for the highest GDF-15 quartile vs the lowest quartile: 1.63 [95% CI 1.31, 2.03]), but not among those with diabetes (HR for the highest GDF-15 quartile vs the lowest quartile: 1.32 [95% CI 0.91, 1.92]). In women (ESM Table 7), GDF-15 was significantly associated with incident HF among those without diabetes (HR for the highest GDF-15 quartile vs the lowest quartile: 1.56 [95% CI 1.26, 1.94]), but to a greater extent among those with diabetes (HR for the highest GDF-15 quartile vs the lowest quartile: 2.30 [95% CI 1.58, 3.34]).

The joint association of diabetes and GDF-15 with HF showed that women in the top quartile of GDF-15 with diabetes had an HR of 2.53 (95% CI 1.86, 3.43) for incident HF relative to those in the lowest GDF-15 quartile without diabetes (ESM Table 7). The corresponding estimate in men was 1.69 (95% CI 1.37, 2.09).

Discussion

In a large community-based sample of black and white adults, we found an independent association between GDF-15 and incident HF, which was more pronounced among people with diabetes. There were sex differences in the relation of GDF-15 and HF by diabetes status, with the stronger associations among men without diabetes and women with diabetes. Even after adjustment for traditional HF risk factors, as well as markers of subclinical cardiac disease (hs-cTnT and NT-proBNP), medication use and coronary artery disease, individuals with diabetes and high GDF-15 levels had a greater than threefold higher risk of incident HF than individuals without diabetes and with lower levels of GDF-15. Moreover, GDF-15 added prognostic information to that of diabetes for HF risk prediction. These findings may have clinical implications, as diabetes status and GDF-15 are both independently associated with an increased risk of incident HF. While the observed change in C statistic after the addition of GDF-15 was statistically significant, the magnitude of the change was small and the role of GDF-15 for monitoring HF risk in clinical practice remains unclear.

Previous studies have demonstrated associations between diabetes and HF, between GDF-15 and diabetes, and between GDF-15 and incident HF. Indeed, higher GDF-15 concentrations have been described among individuals with impaired glucose tolerance vs those without glycaemic impairment [36, 37], and have also been prospectively associated with diabetes [7, 8]. Similarly, a number of studies have shown associations of GDF-15 with incident HF [9, 10, 12] and adverse HF prognosis [13,14,15]. However, the existing population-based studies of diabetes and GDF-15 have not examined their combined role in the pathogenesis of HF. The present analysis extends prior research by showing the additional prognostic implications of both dysglycaemia and elevated GDF-15 levels for incident HF risk. Our findings suggest that GDF-15 is an informative biomarker in the setting of diabetes, with the practical implication being that GDF-15 can be used for HF risk stratification among individuals with diabetes, thus allowing a better selection of candidates for effective HF prevention, possibly using novel therapies such as sodium–glucose cotransporter 2 (SGLT2) inhibitors [38, 39]. Indeed, the addition of GDF-15 to diabetes-specific risk prediction tools such as the UK Prospective Diabetes Study (UKPDS) engine [40, 41] could be considered to refine HF risk stratification among individuals with diabetes. There is not agreement upon GDF-15 cut-off for clinical diagnosis or prognosis purposes and specific cut-points merit exploration in future studies.

Mechanistic studies suggest that elevated GDF-15 levels reflect mitochondrial dysfunction, which contributes to the adverse myocardial effects [16, 17]. Mitochondrial dysfunction may be particularly pronounced in the setting of diabetes [19], with potentially more severe consequences on myocardial energetics and function [20, 22, 23], ultimately translating into a higher HF frequency. GDF-15 also has pro-atherogenic effects possibly through LDL oxidisation [42], as well as reflecting myocardial fibrosis [43] and endothelial dysfunction [44]; all of these processes are also more common in the setting of diabetes.

There are limitations to our study. First, the diagnosis of incident HF was based on hospital discharge and death certificate codes, which may have resulted in some misclassification, as the HF cases seen in the outpatient setting (i.e., potentially less severe or chronic stable forms of HF) were not captured. Second, our analysis does not account for the potential impact of all HF- or diabetes-directed therapies during the follow-up period. Third, we only had one measure of GDF-15, an inherently time-varying measure. Fourth, GDF-15 was measured using an aptamer assay and expressed in relative fluorescent units, although the correlation between this assay and GDF-15 measured using a targeted (ELISA) assay is known to be high (Pearson’s correlation >0.8) [25]. Fifth, cardiac imaging data were not available to assess the subtypes of HF (HF with reduced ejection fraction [HFrEF] and HF with preserved ejection fraction [HFpEF]), and the combined effects of diabetes and elevated GDF-15 levels on HF incidences may differ by HF subtype. Sixth, we lacked detailed information on the type of diabetes, and the extent of glycaemic control (as assessed by glycosylated haemoglobin).

The strengths of our study include the community-based design, the large sample of black and white individuals, with long-term follow-up for incident HF events, and the extensive adjustment for potential confounding factors. The high number of HF events provided power to stratify by both the diabetes status and GDF-15 concentrations, in order to fully examine the contributions of both of these variables to incident HF risk.

Conclusion

In this analysis of community-dwelling black and white people, we found an independent association between GDF-15 concentrations and incident HF, which was more pronounced among individuals with diabetes. Individuals with both diabetes and high GDF-15 levels had a markedly increased HF risk. Our results also indicate that GDF-15 can be used to better stratify people with diabetes for HF risk, and thus help select patients who should be aggressively targeted for HF prevention. Our results suggest that GDF-15 may have an eventual role in clinical practice for monitoring cardiovascular risk and guiding preventive strategies.