Growth differentiation factor-15 is a biomarker for all-cause mortality but less evident for cardiovascular outcomes: a prospective study: Abbreviated title: GDF-15, all-cause mortality and CVD.

BACKGROUND
Previous studies have proposed growth differentiation factor-15 (GDF-15) as a predictor of adverse cardiovascular outcomes and mortality. The present study aimed to determine if such associations remain after accounting for death as a competing risk, and if GDF-15 provides superior prediction performance than other biomarkers.


METHODS
Plasma GDF-15 levels and cardiovascular risk factors were measured in individuals without cardiovascular diseases (n =4,143, aged 57.4 ± 5.96 years, 38.6 % men) from Malmö Diet and Cancer-Cardiovascular Cohort and were followed up for more than 20 years. Incidence of coronary events, ischemic stroke, cardiovascular mortality and all-cause mortality was studied in relation to GDF-15 using Cox proportional hazards regression, with adjustment for potential confounders. Confounding from death as competing risk was carefully checked using the Fine and Gray subdistribution hazard model. Predictive capabilities were further evaluated using C-statistics, continuous net reclassification improvement, and integrated discrimination improvement.


RESULTS
During follow-up, 424 coronary events, 327 ischemic stroke, 368 cardiovascular deaths, and 1,308 all-cause deaths occurred. After controlling for death from other causes as competing events, only all-cause mortality remained significantly related to GDF-15. The addition of GDF-15 significantly improved prediction for all-cause mortality in addition to the traditional risk factors, high-sensitive C-reactive protein and N-terminal prohormone of brain natriuretic peptide (NT-proBNP). Only NT-proBNP improved prediction for CVD mortality.


CONCLUSIONS
GDF-15 is a robust biomarker for all-cause mortality but less reliable for coronary event, ischemic stroke or cardiovascular mortality. Competing risk from death is an important consideration when interpreting the results.


Background
Growth differentiation factor-15 (GDF-15), also known as macrophage inhibitory cytokine-1 (MIC-1), is a member of TGF-β cytokine superfamily [1].Under healthy conditions, GDF-15 is expressed at low levels in all organs and at high level in the placenta [2].However, its expression in many tissues is highly activated by stress and injury and is linked to in ammatory and apoptotic pathways essential for development, differentiation and tissue repair [1].In cultured rat and human cardiomyocytes, GDF-15 expression and secretion were strongly induced by ischemia/reperfusion and were closely associated with smaller infarct sizes and less apoptosis in the infarct border zone [3].In contrast, in patients with cardiovascular diseases (CV diseases, CVDs), high plasma GDF-15 has consistently been associated with adverse prognosis [4].High GDF-15 levels have also been found to predict subsequent cardiovascular event, adding incremental value to traditional cardiovascular risk factors and C-reactive protein (CRP) [5].Pareek et al. demonstrated that among multiple candidate biomarkers, GDF-15 showed superior capabilities in predicting a composite of adverse CV outcomes, independent of N-terminal prohormone of brain natriuretic peptide (NT-proBNP) [6].Due to a modest number of endpoints, only a composite endpoint was examined, despite the potentially different pathogenic mechanisms for individual CV outcomes [6].More recently, Ho et al. used a proteomic platform to investigate 85 protein biomarkers of CVD risk in relation to incident atherosclerotic CVD, heart failure, CVD death and all-cause mortality in 3523 participants with a median follow-up of 14.3 years [7].As a result, GDF15 was the only biomarker that was associated with all outcomes in multi-marker models.However, again in this study, a composite endpoint of atherosclerotic CV events was used and detailed information regarding the GDF15-CVD association was not reported [7].In addition, previous studies examining the associations between GDF-15 and CV outcomes were based on traditional Cox proportional hazards regression.It is worth noting that the competing risk from death could substantially alter the risk estimates, particularly in studies of older adults [8,9].
In the present study, we aim to explore whether plasma GDF15 is associated with incidence of coronary events, ischemic stroke, CVD mortality or all-cause mortality in a population-based cohort study, and whether the associations are independent of markers of in ammation (CRP) and cardiac dysfunction (NT-proBNP) as well as accounting for the competing risk of death from other causes.

Participants Malmö Diet and Cancer-Cardiovascular Cohort (MDC-CV)
The Malmö Diet and Cancer Study (MDC study) is a prospective population-based cohort study from the city of Malmö in southern Sweden, with 230,000 inhabitants at the time of recruitment [10].A total of 28,449 individuals (men = 11,246, born 1923-1945 and women = 17,203, born 1923-1950) underwent a baseline examination between March 1991 and September 1996.Samples of peripheral venous blood, blood pressure measurements and anthropometric measures were taken on the rst visit.Participants also lled out self-administered questionnaires [10].Between October 1991 and February 1994, a group of 6,103 individuals was randomly selected to be recruited in MDC-CV, which was designed to study the epidemiology of carotid artery disease [11].Out of these, 5,540 participants attended the second visit when fasting plasma samples were collected.Subjects were excluded from the present study due to incomplete data on confounders (n = 415), insu cient plasma samples stored for protein biomarker measurement (n = 380) or failing the internal quality control for the biomarker analysis (n = 118).
Individuals with a history of CVDs (n = 109) at baseline were also excluded.After exclusions, at total of 4,518 individuals remained for the analysis (1,773 men and 2,745 women, aged 57.4 ± 5.96 years) (Fig. 1).
All participants provided written informed consent.The study was approved by the Regional Ethical Review Board in Lund, Sweden (LU 51/90) and was carried out in accordance with the Helsinki Declaration.

Measurements And De nitions
Information on the current use of blood pressure-lowering medications, smoking habits, and alcohol consumption was gathered through self-administered questionnaires.Subjects were grouped into current smokers, former smokers and never-smokers.High alcohol consumption was de ned as > 40 g alcohol per day for men and > 30 g per day for women.Waist circumference was measured midway between the lowest rib margin and iliac crest in centimetres.Blood pressure was measured using mercury column sphygmomanometer after 10 minutes of rest in the supine position.Fasting blood glucose and lipids were measured from fresh blood samples using standard procedures at the Department of Clinical Chemistry, University Hospital Malmö.Low density lipoprotein (LDL) concentrations were determined according to Friedewald's formula.The biomarkers were measured from fasting plasma samples that had been frozen at − 80 °C immediately after collection.High sensitive CRP (hsCRP) was analysed using the Tina-quant CRP latex assay (Roche Diagnostics) on an ADVIA® 1650 Chemistry System (Bayer Healthcare).GDF-15 and NT-proBNP levels were measured using Proseek® Multiplex CVD I 96 × 96 reagent kit by the SciLifeLab analysis service (Uppsala, Sweden), based on a Proximity Extension Assay technology [12].The assay procedure is described in detail elsewhere [13].Brie y, the assay involved three core steps: incubation, extension and detection.Raw Proseek data were adjusted using a preprocessing normalization procedure and set relative to a xed background level, which generated Normalized Protein Expression (NPX, log 2 scale) values.A high NPX value corresponds to a high protein concentration.The NPX values of GDF-15 by Proseek assay correlated closely with values obtained using an electrochemiluminescence immunoassay (Roche Diagnostics) [13].

Outcomes
All participants were followed from baseline examination until the rst end point (coronary events, ischemic stroke, cardiovascular mortality or all-cause mortality), emigration from Sweden, death or December 31st, 2016, whichever came rst.Separate follow-up times were calculated for each endpoint.Ascertainment of cases was carried out with the help of validated national and local registers [14,15].A coronary event was de ned as a fatal or non-fatal myocardial infarction according to International Classi cation of Diseases 9th revision (ICD-9: Code 410) or 10th revision (ICD-10: Code I21), or death attributed to ischemic heart disease (ICD-9: Codes 411-414; ICD-10: Codes I22-I25).Ischemic stroke was de ned according to ICD-9 code 434 or ICD-10 code I63.Cardiovascular mortality was de ned as an ICD-10 diagnosis in chapter I as underlying cause of death.

Statistical Analysis
HsCRP and fasting glucose were logarithmically transformed due to skewed distributions.Subjects were classi ed into quartiles with sex-speci c cut points; i.e. equal number of men and women in each quartile based on the concentrations of GDF-15.Baseline characteristics are presented as mean ± standard deviation (SD) or median (interquartile range) for continuous variables and percentages for categorical variables.Cross-sectional relations of GDF-15 quartiles to CV risk factors were assessed using linear regression for continuous variables and logistic regression for dichotomous variables.P-values from trend tests across quartiles were used.
The associations between GDF-15 quartiles and outcomes were examined using Cox proportional hazards regression models with time-on-study as the time-scale and the lowest quartile as the reference category.HR with 95% con dence intervals (CI) were estimated in Model 1 adjusted for age and sex, and Model 2 adjusted for CV risk factors.HsCRP and NT-proBNP was additionally added into the model one after another (Model 3 and Model 4) to determine whether the observed associations were dependent of existing systematic in ammation or CV dysfunction.A Cox regression restricted cubic spline model [13] was used to check the assumption of linearity of the associations between GDF-15 and outcomes with knots placed at equal intervals of GDF-15 concentration (20%, 40%, 60% and 80%).HR per 1 SD increase of GDF-15 (1 SD = 0.56 NPX, log2 scale) was calculated for linear associations.For non-linear associations, HR was obtained by tting piecewise linear functions (post hoc).The competing risks of death was accounted for in a sensitivity analysis by the Fine and Gray proportional subdistribution hazards models method [16].The cumulative incidence of outcomes at different GDF-15 levels was plotted with mortality treated as a censored observation or as a competing risk, respectively, while all other covariates were xed at their average values [16].The interplay between GDF-15 levels and risk factors was examined by introducing an interaction term into the multivariate model one by one.A two tailed p-value < 0.05 was considered signi cant.Statistical Analysis System version 9.4 for Windows (SAS Institute, Cary, NC, USA) was used for the statistical analysis.

Baseline Characteristics
The study population included 4,518 individuals aged 57.4 ± 5.96 years, of whom 39.2% were male.
Baseline characteristics across GDF-15 quartiles are presented in Table 1.Age, waist circumference, current smoking, glucose, systolic blood pressure (SBP), BP-lowering medication, LDL, hsCRP and NT-proBNP showed positive associations with GDF-15 levels at baseline, while former smoking and never smoking showed the opposite pattern.No statistical signi cance was observed between GDF-15 and alcohol consumption.

Follow-up Analysis
A total of 473 individuals (273 men and 200 women) suffered a coronary event after a mean follow-up of 20.9 ± 5.55 years, 336 individuals (181 men and 185 women) developed ischemic stroke after a mean follow-up of 20.9 ± 5.53 years.In addition, during a 21.5 ± 5.01 years follow-up period, a total of 1,445 death (687 men and 758 women) occurred, of which 405 (212 men and 193 women) were attributed to CVD.The incidences per 1000 person years for coronary events, ischemic stroke, CVD mortality and allcause mortality were 5.01, 3.88, 4.17 and 14.90, respectively.
The crude and adjusted associations between GDF-15 quartiles and outcomes can be seen in Table 2.In multivariate model (Model 2), GDF-15 was signi cantly associated with all the outcomes except for ischemic stroke (p for trend = 0.07).Adding hsCRP into the model slightly attenuated the associations (Model 3).However, after additionally adjusting for NT-proBNP (Model 4), only the associations of GDF-15 with coronary events and all-cause mortality remained signi cant.The HR (95% CI; p for trend) for coronary events and all-cause mortality in the 4th vs. 1st quartile of GDF-15 was 1.39 (1.04-1.87;p = 0.01) and 1.60 (1.34-1.91;p < 0.0001), respectively.The restricted cubic spline analysis con rmed no association between GDF-15 and ischemic stroke (p for effect = 0.20) while a signi cant and linear association (p for effect, p for non-linearity) was detected for GDF-15 with coronary events (0.049, 0.59), CVD mortality (0.04, 0.41) and all-cause mortality (< 0.0001, 0.49).The adjusted HRs (95% CI; p) for coronary events, CVD mortality and mortality per SD increase in GDF-15 were 1.15 (1.03-1.28;p = 0.01), 1.16 (1.03-1.30;p = 0.01) and 1.29 (1.21-1.37;p < 0.0001), respectively (Table 2).Table 3 presents results generated by subdistribution hazards models.In multivariate model without NT-proBNP, considering death from other causes as competing events completely eliminated the association between GDF-15 and coronary events (HR per SD increase, 1.08; 95% CI, 0.98-1.20;p = 0.11), and substantially moderated the association between GDF-15 and CVD mortality (HR per SD increase, 1.12; 95% CI, 1.00-1.26;p = 0.04).This association was no longer signi cant after additionally adjusting for NT-proBNP (HR per SD increase, 1.06; 95% CI, 0.94-1.19;p = 0.36).Correspondingly, as visualized in Additional le 1, traditional Cox analysis compared with the competing risk approach exaggerated the effect of GDF-15 on the cumulative incidence of coronary events, ischemic stroke and CVD mortality.In contrast, the cumulative incidence of all-cause mortality clearly increased with elevated GDF-15 levels.Signi cant interaction between GDF-15 and hsCRP was found on the risk of all-cause mortality (p for interaction < 0.01).Subset analyses after stratifying participants by median hsCRP (1.3 mg/L) showed that the association of GDF-15 with all-cause mortality was stronger in participants with higher baseline hsCRP concentration as compared to those with lower baseline hsCRP concentration.The adjusted HR per SD increase in GDF-15 was 1.37 (95% CI, 1.27-1.48;p < 0.0001) for those with CRP above median and 1.19 (95% CI, 1.08-1.30;p < 0.001) for those with lower CRP.

Discussion
This large, population-based cohort study showed a robust association between plasma GDF-15 levels and all-cause mortality.GDF-15 was also observed to be associated with coronary events, ischemic stroke and CVD mortality, but at least in part, the associations were likely to be confounded preexisting CV dysfunction or by the competing risk of death.
In line with previous studies [13,[17][18][19], the present study observed positive correlations between GDF-15 and traditional CV risk factors, including age, waist circumference, smoking, glucose, SBP, LDL and hsCRP.Since GDF-15 is an anti-in ammatory cytokine by nature [20] and all of these CV risk factors are related to systemic in ammation, the correlations could at least be partly explained by a compensatory mechanism of GDF-15 in response to in ammation.Of note, GDF-15 correlated strongly with NT-proBNP [17,19], a commonly used marker of left ventricular dysfunction in clinical practice.Under pressure or volume overload, NT-proBNP is produced on cardiomyocytes to cope with LV hypertrophy and brosis due to myocardial stretch [21].Similar to NT-proBNP, the production of GDF-15 by cardiomyocytes is stimulated in response to CV injury [3,22,23].Thus, individuals with simultaneously elevated GDF-15 and NT-proBNP in this study represented a group with suboptimal CV health even in the absence of overt cardiac pathology.
In longitudinal analyses, GDF-15 was found to be associated with incidence of coronary events, ischemic stroke, CVD mortality and all-cause mortality before correction for the confounding from NT-proBNP.This observation was in agreement with previous studies, supporting an increased risk of a wide range of CVDs in community-dwelling individuals with elevated GDF-15 levels [6,7,17,19].Our data, along with evidence from another Swedish cohort [24], suggested that the association between GDF-15 and ischemic stroke was confounded by NT-proBNP.Whereas additional adjustment for NT-proBNP greatly affected all associations with cardiovascular outcomes, its impact on the association between GDF-15 and all-cause mortality was relatively small.Previous cohort studies have consistently shown that GDF-15 and NT-proBNP are prognostic markers for adverse CV outcomes and mortality among asymptomatic subjects [6,7,19].When comparisons were made between these two biomarkers, Pareek et al. [6] demonstrated that, as compared to NT-proBNP, GDF-15 showed stronger association with composite outcomes including all-cause mortality, but weaker association with incident heart failure or CVD mortality (sample size 1,324, aged 66 years, followed up for 8.6 years).Ho et al. [7] showed that GDF-15 is a better predictor of atherosclerotic CVD and all-cause mortality but a weaker predictor of CVD death than NT-proBNP (sample size 3,523, aged 62 years, followed up for 14.3 years).Similarly, Daniels et al. [19] reported that GDF-15 was a stronger predictor of all-cause mortality while NT-proBNP was a stronger predictor of CVD mortality (sample size 1,391, aged 70 years, followed up for 11 years).In addition, GDF-15 but not NT-proBNP was independently associated with non-CV death and cancer death [19].Therefore, it seems that whereas GDF-15 might not be a better marker than NT-proBNP in a cardiovascular context, it is a powerful marker for all-cause mortality.
Noteworthy, although almost all relevant studies were conducted on an elderly population and most of them were with a long-term follow-up, hardly any study has carefully evaluated the death as a competing risk.Emphasis has been made that competing risk of death is an important consideration in studies evaluating the long-term association among older adults, ignoring of which might lead to substantially overestimated associations [8,9].Kaplan-Meier estimates and traditional Cox proportional hazards regression models assume that subjects are censored from causes unrelated to the primary outcome.Therefore, those who died from other causes are still considered to be at risk of the primary outcomes at the point of censoring and are thought to have the same distribution of time-to-event as incident cases.However, such subjects cannot possibly develop the primary outcomes, thus ignoring that the competing risk of mortality could exaggerate the HRs for the primary outcomes, particularly for subjects with high mortality, e.g.older adults with long-term follow-up [8,9].Our study added important information on the competing risk of death and indicates that the associations of GDF-15 with coronary events and CVD mortality may be markedly overestimated if this competing risk is ignored, especially given that high GDF-15 contributed largely to mortality.As a result, only all-cause mortality remained signi cantly related to GDF-15 in this study.The association between GDF-15 and all-cause mortality was strengthened in subjects with higher baseline hsCRP, which may be due to an enhanced secretion and function of GDF-15 in pro-in ammatory conditions.
Of note, accumulating evidence has indicated that GDF-15 has a role beyond in ammation [1,17].It has multiple roles in, for example, oxidative stress, tumorigenesis, strain and ischemia, and its elevation is associated with a wide range of diseases involving different systems [1,17].Therefore, it is speculated that GDF-15 concentration re ects overall health status and thus correlates with all-cause mortality.
Elevated GDF-15 levels mirror cellular stress and senescence, which might also explain the association between GDF-15 and life span.As a downstream target of p53, GDF15 gene is expressed via a p53dependent pathway in cellular response to a broad range of stress [25,26].During this process, p53 mediate cellular senescence and apoptosis by functioning as a transcription factor [26], and thus plays a role in regulating longevity [26].At structural levels, p53 directly in uences mitochondrial respiration and impairs clearance of defective mitochondria [27,28], leading to mitochondrial dysfunction, a well-known driver of biologic ageing [29].Consistently, emerging studies have proposed GDF-15 as a speci c biomarker of mitochondrial disorder [30,31].In further support of GDF-15 as a marker of ageing, GDF-15 has recently been identi ed as a soluble senescence-associated secretory protein and a key of cellular senescence [32].Nevertheless, it remains largely obscure whether and exactly how GDF-15 itself promotes ageing.
The study is based on a large number of subjects from the general population, with a long-term follow-up and su cient number of cases.Less than 1% was loss to follow-up [10].CV endpoints were retrieved from validated national and local registers [14,15].However, there are some limitations that might have affected our study.High-risk individuals are less likely to participate in population-based studies.Thus, the proportion of individuals with high risk could be underestimated.Information on biomarkers was available for a sub-cohort, MDC-CV, of the whole MDC cohort.Nevertheless, the sub-cohort was randomly selected and should be regarded as representative of the whole MDC cohort in terms of age, sex, biological, lifestyle and socio-economic characteristics [33].Furthermore, the risk factors used for adjustments in the study are well-known CV risk factors but still we cannot exclude the possibility of residual confounding.Information about incident coronary events and ischemic stroke was retrieved from local and hospital registers.The validity of this information has been con rmed in many previous studies [14,15].

Conclusions
Our results suggest that high GDF-15 levels are associated with increased incidence of all-cause mortality while its association with coronary events, ischemic stroke and CVD mortality might be dependent of preexisting cardiovascular dysfunction or confounded by death.

Declarations
Ethics approval and consent to All participants provided written informed consent.The study was approved by the Regional Ethical Review Board in Lund, Sweden (LU 51/90) and was carried out in accordance with the Helsinki Declaration.edfrom all individual participants included in the study.

Table 1
Quartiles of GDF-15 and risk factors of participants in MDC-CV (n = 4,518).
systolic blood pressure; BP, blood pressure; hsCRP, high sensitive C-reactive protein; LDL, low density lipoprotein; NT-proBNP, N-terminal prohormone of brain natriuretic peptide.aGlucose and hsCRP are presented as median (interquartile range in brackets) due to skewed distributions.All other continuous values are presented as means ± SD, unless otherwise stated.b p value for log-transformed value of Glucose and hsCRP.

Table 2
CVD outcomes and mortality in relation to quartiles and per SD change of GDF-15.Adjusted for age, sex, waist circumference, smoking, alcohol consumption, glucose level, systolic blood pressure, BP-lowering drug, LDL, high sensitive C-reactive protein level, and NT-proBNP.
a Adjusted for age and sex.bAdjusted for age, sex, waist circumference, smoking, alcohol consumption, glucose level, systolic blood pressure, BP-lowering drug, and LDL.c Adjusted for age, sex, waist circumference, smoking, alcohol consumption, glucose level, systolic blood pressure, BP-lowering drug, LDL, and high sensitive C-reactive protein level.d a Adjusted for age, sex, waist circumference, smoking, alcohol consumption, glucose level, systolic blood pressure, BP-lowering drug, LDL, and high sensitive C-reactive protein level.b Estimated based on traditional Cox proportional hazards