Serum markers of biological ageing provide long-term prediction of life expectancy—a longitudinal analysis in middle-aged and older German adults

Abstract

Background

lifestyle behaviours and chronic co-morbidities are leading risk factors for premature mortality and collectively predict wide variability in individual life expectancy (LE). We investigated whether a pre-selected panel of five serum markers of biological ageing could improve predicting the long-term mortality risk and LE in middle-aged and older women and men.

Methods

we conducted a case-cohort study (n = 5,789 among which there were 2,571 deaths) within the European Prospective Investigation into Cancer-Heidelberg cohort, a population cohort of middle-aged and older individuals, followed over a median duration of 18 years. Gompertz models were used to compute multi-adjusted associations of growth differentiation factor-15, N-terminal pro-brain natriuretic peptide, glycated haemoglobin A1c, C-reactive protein and cystatin-C with mortality risk. Areas under estimated Gompertz survival curves were used to estimate the LE of individuals using a model with lifestyle-related risk factors only (smoking history, body mass index, waist circumference, alcohol, physical inactivity, diabetes and hypertension), or with lifestyle factors plus the ageing-related markers.

Results

a model including only lifestyle-related factors predicted a LE difference of 16.8 [95% confidence interval: 15.9; 19.1] years in men and 9.87 [9.20; 13.1] years in women aged ≥60 years by comparing individuals in the highest versus the lowest quintiles of estimated mortality risk. Including the ageing-related biomarkers in the model increased these differences up to 22.7 [22.3; 26.9] years in men and 14.00 [12.9; 18.2] years in women.

Conclusions

serum markers of ageing are potentially strong predictors for long-term mortality risk in a general population sample of older and middle-aged individuals and may help to identify individuals at higher risk of premature death, who could benefit from interventions to prevent further ageing-related health declines.

life expectancy, ageing biomarkers, lifestyle factors, biological ageing, prevention older people

Key Points

• Combined unhealthy lifestyle behaviours are associated with wide variation in life expectancy (LE).

• We hypothesised that five ageing-related markers (growth differentiation factor-15, N-terminal pro-brain natriuretic peptide, glycated haemoglobin A1c, C-reactive protein and cystatin-C) could further help predict the LE.

• Using only lifestyle factors, our models predicted a LE difference of 16.8 years in men and 9.87 years in women.

• Adding the biomarkers provided an additional discrimination in LE difference of up to 5.9 years in men and 4.13 years in women.

• These markers may improve the long-term LE prediction and identify individuals at an increased risk of premature mortality.

Background

Life expectancy (LE) at birth has been significantly increasing in the past decades in economically developed countries [1]. However, large parts of the population still experience premature death resulting from avoidable lifestyle factors such as smoking, alcohol consumption, obesity and physical inactivity [2], or due to socio-economic factors affecting psychological health [3] or access to health care [4]. Thus, there is still room for major improvements in human lifespan at a population level.

In studies in Europe, the USA, Canada, China or Japan, combined unhealthy lifestyle behaviours (heavy smoking and alcohol drinking, an unhealthy body weight, physical inactivity and unbalanced diets) were associated with a loss in LE up to 18.9 years compared to combined healthy behaviours [5–15] and might contribute to almost 6 in every 10 premature deaths [16, 17]. Biological ageing, however, covers a wide array of alterations on cellular, physiologic and functional levels [18] and depends not only on modifiable risk factors but also on psychosocial and economic determinants and genetic determinants of individual host response. Thus, improving the prevention of premature biological ageing and related risks of functional decline and mortality may require individualised recommendations and treatments [19].

While hallmarks of biological ageing are complex and diverse [20], in part, these may be reflected in circulating biomarkers [19, 21, 22] that can be used for health monitoring. In a comprehensive literature review to prepare for a large-scale ‘geroscience’-guidance clinical trial [19], a multi-disciplinary expert group identified about 10 blood-based biomarkers, selected from 258 initial candidates, on the basis of measurement reliability and feasibility, ability to predict all-cause mortality, clinical and functional outcomes and potential responsiveness to lifestyle or medical interventions. Five major biomarkers on this shortlist were: growth differentiation factor-15 (GDF-15), N-terminal pro-brain natriuretic peptide (NT-proBNP), glycated haemoglobin A1c (HbA1C), C-reactive protein (CRP) and cystatin-C. These same biomarkers were also among the selected candidates of the ‘Deep Knowledge Group’ consortium, part of the Aging Analytics Agency [23]. Beyond their well-known roles in miscellaneous functional declines (i.e. cardiovascular health for NT-proBNP [24], diabetes for HbA1C [19] and renal function for cystatin-C [25]), these five biomarkers are also involved in biological ageing. GDF-15 is induced by oxidative stress and inflammation and is a marker of mitochondrial dysfunction [26–28]. NT-proBNP released from cardiomyocytes undergoing wall stress; or ischemia is strongly correlated with chronological age [19, 29] and might be stimulated by several pro-inflammatory cytokines, including tumour necrosis factor-α and some interleukins [30]. HbA1C is, in addition to its role in diabetes diagnosis, a marker of metabolically unhealthy ageing [31], whereas CRP is a marker of systemic chronic inflammation, which is also linked to biological ageing [32]. Epidemiologic studies have shown associations of these biomarkers with increased relative risks of all-cause and cardiovascular mortality [32–38]; so far, however, no studies have been performed in general population cohorts to estimate the variations in overall LE which may be predicted by this targeted five-marker panel either alone or in combination with lifestyle factors.

In the German European Prospective Investigation into Cancer (EPIC)-Heidelberg cohort, we previously estimated a difference in LE of 17.0 years for 40-year-old men and 13.9 years for 40-year-old women [17] by comparing healthiest with unhealthiest combined lifestyle patterns. We here present findings of a further study in the (EPIC)-Heidelberg cohort with a longer follow-up to explore whether, beyond lifestyle-related risk factors, the panel of five ageing-associated biomarkers (GDF-15, NT-proBNP, HbA1C, CRP and cystatin-C) provides further meaningful discrimination in life expectancies of middle-aged and older individuals.

Methods

Study setting, sampling strategy and laboratory measurements

The current study used a nested case-cohort design, which is embedded within the EPIC-Heidelberg Study—a population-based cohort study that was initiated to investigate associations between diet, metabolic factors and lifestyle with the risks of cancer and other chronic diseases [39]. The EPIC-Heidelberg cohort, comprising a total of 25,540 women and men aged 35–65 at recruitment, with available data from questionnaires, clinical and biological measurements, is detailed in Supplementary Appendix S1 available in Age and Ageing online. The sampling strategy leading to the case-cohort sample is described in Supplementary Appendix S2 and Supplementary Figure S1 available in Age and Ageing online. The laboratory measurement method of the five ageing biomarkers is described in Supplementary Appendix S3 available in Age and Ageing online.

Mortality ascertainment

In EPIC–Heidelberg, mortality outcomes were ascertained from death certificates that were collected from mortality registries. Within the entire EPIC-Heidelberg cohort, 2,571 deceased until end of follow-up (December 2014) of whom 1,134 (44.1%) died of cancer, 622 (24.2%) died of cardiovascular events and the remaining died of miscellaneous conditions that were also coded according to the ICD-10.

Statistical analyses

Descriptive statistics

Descriptive statistics for subject characteristics in the sub-cohort are reported as medians and interquartile ranges, or as means and standard deviations for continuous variables and as frequencies for categorical variables. Age-adjusted partial Spearman coefficients were used to examine the sex-specific correlations of GDF-15, NT-proBNP, HbA1C, CRP and cystatin-C with body mass index (BMI), waist circumference, alcohol consumption and number of pack-years. Multi-adjusted regression analyses were used to estimate the percentage of the variance (adjusted R²) in each biomarker potentially explained by the individual and combined lifestyle factors. Median follow-up time estimates for the cases of mortality and for participants in the sub-cohort and their confidence intervals (CIs) were calculated using the reverse Kaplan Meier method [40].

Associations with all-cause mortality and LE

Gompertz proportional hazards models with chronological age as time-scale were used to model time-to-death. This parametric proportional hazard survival model allows an extrapolation of the survival curve and thus provides estimates for LE, while allowing hazard ratio (HR) calculation. All models were adjusted for age (time-scale), and further adjustment was made for specific risk variables to be examined (see below) and sex, as adjustment variables. The assumption of proportional hazards was assessed using log–log survival plots. We found that sex interacted significantly with smoking, BMI and biomarkers. Therefore, different models were constructed for men and women separately. In addition, in order to satisfy the proportional hazards assumption, we used a step-wise function of age at follow-up with jumps at 50 and 60 years (i.e. allowing the HRs for the biomarkers to take on different values for individuals when followed before the age 50, between 50 and 60 and after the age of 60 years).

General adequacy of the Gompertz model was verified by investigating log–log survival plots for linearity. The area under the Gompertz survival curves (area under the curve) was used to estimate the LE of individuals with specific covariate values corresponding to combinations of optimally healthy versus less healthy lifestyle factors, as well as biomarker values, as further described below.

Modelling of mortality risk in relation to lifestyle-related factors

For the estimation of lifestyle-related risks, factors included in the Gompertz models were smoking status, smoking pack-years (continuous), BMI (kg/m², continuous), waist circumference (cm, continuous), self-reported baseline diabetes (yes/no), self-reported hypertension (yes/no), alcohol intake (g/day, continuous) and physical activity.

Modelling of mortality risk in relation to biomarkers

To investigate the effects of each biomarker on mortality, we constructed models simultaneously adjusted for each biomarker in addition to the lifestyle factors above. Prior to fitting Gompertz models, biomarker values were log-transformed and scaled to account for skewness and to reduce the impact of outliers. For each marker, we used additional quadratic polynomial terms to test for deviations from (log)linear dose–response.

To determine which biomarkers were associated with the largest loss in LE, HRs were estimated. To avoid giving unrealistic HR estimates (due to comparisons between individuals with unhealthy lifestyle behaviours having extremely high values of biomarkers and individuals with relatively low biomarkers’ values, adopting healthier lifestyle behaviours), we present HRs corresponding to a likely difference of the biomarker between individuals having similar lifestyle behaviours. To achieve this, we estimated conditional distributions for each biomarker, conditional on median or mode levels of the other lifestyle-related variables. The HR for each biomarker then corresponds to the ratio between the first and fifth quintiles of this conditional distribution, where the conditional distribution represents the range of values that are likely for an individual with lifestyle-related variables at their median or mode levels. The conditional distribution corresponds to a reduction in variance from the original, unconditional distribution, ensuring a likely difference between the first and the fifth quintiles. Using an unconditional distribution would have led to over-exaggerated HRs.

Next, we examined the overall LE difference between the highest and lowest quintiles of predicted risk. To find the difference in LE associated with lifestyle, a Gompertz model was constructed with only the lifestyle-related factors described above. This model was then used to estimate the LE of each individual in our cohort. The difference between the first quintile and the fifth quintile of this LE distribution was then considered as the difference in LE related to lifestyle factors. To determine the additional discriminatory power provided by the biomarkers, the difference in LE due to both lifestyle and the biomarkers was then obtained analogously using a model including lifestyle-related factors and all biomarkers simultaneously.

All analyses were performed with R (R core team 2020) using the flexsurv [41] and MICE [42] packages. Associations with P values < 0.05 were considered to be as statistically significant.

Results

This case-cohort analysis included a total of 1,706 men (of whom 323 in the sub-cohort) and 865 women (of whom 163 in the sub-cohort) who had died prior to December 2014 and a representative sub-cohort of 1,947 women and 1,845 men randomly sampled from the EPIC-Heidelberg cohort (Table 1). By the study design, women in the sub-cohort were slightly younger at baseline recruitment (range: 35–66 years, average: 51 years) than men (range: 40–65, average: 54). For those who had deceased, the median duration of prospective follow-up was 17.55 years [95% CI: 17.16, 17.95] for men and 22.76 years [22.26, 22.35] for women compared to 18.08 [17.96, 18.15] and 17.94 [17.86, 18.09], respectively, for men and women in the sub-cohort. The average age at death was 68.3 [67.5, 69.1] and 69.3 [67.9, 70.7] for men and women, respectively.

Table 1

Baseline characteristics of the study population, EPIC-Heidelberg sub-cohort (n = 3,794)

	Sub-cohort
	Women	Men	Overall
	(N = 1,947)	(N = 1,847)	(N = 3,794)
Age at baseline
Mean (min, max)	51.5 (35.2, 66.0)	54.1 (40.3, 65.4)	52.7 (35.2, 66.0)
Age at death  Mean (min, max) Smoking status	69.3 (40.4, 84.3)	68.3 (43.7, 83.5)	68.6 (40.4, 84.3)
Never	1,026 (52.7%)	573 (31.0%)	1,599 (42.1%)
Short time quitters	151 (7.8%)	202 (10.9%)	353 (9.3%)
Long-time quitters	375 (19.3%)	663 (35.9%)	1,038 (27.4%)
Current, heavy smokers	221 (11.4%)	305 (16.5%)	526 (13.9%)
Current, light smokers	174 (8.9%)	104 (5.6%)	278 (7.3%)
BMI (Kg/m2)
Mean (SD)	25.5 (4.7)	26.9 (3.6)	26.2 (4.3)
Waist circumference
Mean (SD)	81.8 (11.9)	96.3 (10.3)	88.9 (13.3)
Self-reported diabetes
No	1,904 (97.8%)	1,738 (94.1%)	3,642 (96.0%)
Yes	43 (2.2%)	109 (5.9%)	152 (4.0%)
Physical activity level
Active	193 (9.9%)	114 (6.2%)	307 (8.1%)
Moderately active	1,013 (52.0%)	557 (30.2%)	1,570 (41.4%)
Inactive	244 (12.5%)	594 (32.2%)	838 (22.1%)
Moderately inactive	497 (25.5%)	582 (31.5%)	1,079 (28.4%)
Alcohol consumption (g/day)
Mean (SD)	7.59 (15.9)	28.4 (31.8)	17.7 (27.0)
Level of education
None/primary school	571 (29.3%)	585 (31.7%)	1,156 (30.5%)
Technical/professional	803 (41.2%)	500 (27.1%)	1,303 (34.3%)
Secondary school	136 (7.0%)	94 (5.1%)	230 (6.1%)
Longer (incl. university)	437 (22.4%)	668 (36.2%)	1,105 (29.1%)
CRP (mg/l)
Median (Q1, Q3)	1.52 (0.64, 3.60)	1.56 (0.73, 3.53)	1.54 (0.68, 3.57)
GDF-15 (pg/ml)
Median (Q1, Q3)	576 (451, 737)	629 (503, 830)	604 (475, 785)
HbA1C (mmol/mol)*
Median (Q1, Q3)	34 (32, 37)	35 (33, 38)	35 (32, 38)
Cystatin-C (ng/ml)
Median (Q1, Q3)	418 (294, 603)	458 (324, 647)	440 (308, 623)
NT-proBNP (pg/ml)
Median (Q1, Q3)	180 (77, 326)	103 (50, 213)	137 (60, 276)

^*HbA1C in % can be obtained from mmol/mol as follows: HbA1C (%) = (0.0915 × (HbA1C mmol/mol)) + 2.15%.

In the sub-cohort, both women and men on average were slightly overweight with average BMI of 25.5 and 26.9 kg/m² and mean waist circumferences of 82 and 96 cm, respectively. Women reported to be more physically active than men. Half of the women in the sub-cohort had never smoked; while only 31% of men never smoked. Regarding educational level, about 22% among women and 36% among men had a university degree, whereas 29 and 32% reported primary school as highest level of formal education.

Among men and women in the sub-cohort, adjusting for age, CRP showed moderately strong correlations with BMI and waist circumference, with Spearman coefficients between 0.33 and 0.4 (Figure 1). Among men, moderate correlations with the number of pack-years were observed with GDF-15 (correlation coefficient = 0.30) and CRP (correlation coefficient = 0.22); while among women, only GDF-15 showed a mildly strong correlation (correlation coefficient = 0.22) with the number of pack-years. The mutual correlations between the five biomarkers were weak to moderate: the strongest correlations in men were observed for GDF-15 with CRP (correlation coefficient = 0.31) and HbA1C with CRP (correlation coefficient = 0.24). Among women, the strongest correlation was observed between GDF-15 and CRP (correlation coefficient = 0.21). Moderate proportions of the variance of the biomarkers were associated with age (R² = 0.016–0.196), or with modifiable lifestyle factors, where R² ranged between 0.035 (for NT-proBNP) and 0.24 (for CRP) in women and between 0.025 (for NT-proBNP) and 0.198 (for CRP) in men (Supplementary Appendix S4 available in Age and Ageing online).

Figure 1

Spearman Partial Correlation Coefficients after adjustment for age at baseline in the sub-cohort, EPIC-Heidelberg, n = 3,794; values for women are above the main diagonal, values for men are below.

The associations between each of the five biomarkers with all-cause mortality risk are presented in Table 2. Adjusting for age and for all lifestyle-related risk factors, GDF-15, NT-proBNP, HbA1C and CRP, but not cystatin-C, were found to be associated with mortality risk in men above the age of 60. Participants with biomarker levels ≥the 80th percentile (highest quintile) of the conditional marker distribution, adjusted for the above risk factors, had an increased relative mortality risk ranging between 1.13 [1.05, 1.35] (for HbA1C) and 1.49 [1.33, 1.66] (for NT-proBNP) compared with those having biomarker levels ≤the 20th percentile (lowest quintile) of the conditional distribution. For women, within the follow-up time frame of >60 years of age, associations were found to be significant only for NT-proBNP, GDF-15 and HbA1C, with increased relative risks ranging between 1.21 [1.10; 1.50] (for HbA1C) and 1.52 [1.28, 1.79] (for GDF-15). Results for follow-up times between ages 50 and 60 for both sexes were similar to those of participants followed-up from ages >60 but were slightly larger in most cases (Table 2).

Table 2

HRs for men and women, comparing the highest to the lowest quintile of each biomarker’s conditional distribution, EPIC-Heidelberg case-cohort (n = 5,879).

Age at follow-up between 50 and 60
	Men	Women
CRP	1.28 (1.14; 1.47)	1.15 (0.93; 1.42)
GDF-15	1.38 (1.17; 1.90)	1.98 (1.29; 2.89)
NT-proBNP	1.47 (1.36; 1.66)	1.30 (1.07; 1.56)
HbA1C	1.12 (1.04; 1.34)	1.20 (1.07; 1.54)
Cystatin-C	0.93 (0.79; 1.09)	0.96 (0.78; 1.15)
Age at follow-up above 60
	Men	Women
CRP	1.27 (1.13; 1.45)	1.16 (0.94; 1.41)
GDF-15	1.21 (1.05; 1.40)	1.52 (1.28; 1.79)
NT-proBNP	1.49 (1.33; 1.66)	1.25 (1.05; 1.49)
HbA1C	1.13 (1.05; 1.35)	1.21 (1.10; 1.50)
Cystatin-C	1.02 (0.87; 1.21)	1.05 (0.85; 1.27)

HRs were obtained from Gompertz proportional hazard models, and we compared the highest to the lowest quintile of each biomarker’s conditional distribution (conditional of lifestyle factors). Models are adjusted for age (timescale), level of education (none/primary school, technical/professional education, secondary school, longer (including university), smoking status (never, long time quitters (>10 years), short time quitters (≤10 years), current light([≤ 10 cigarettes/day) and current heavy smokers (>10 cigarettes/day), number of pack-years, lifetime alcohol consumption, physical activity level (inactive, moderately inactive, moderately active and active), baseline self-reported diabetes and hypertension, BMI, waist circumference and an interaction term between BMI and waist circumference. To account for log-linearity, quadratic terms for CRP, cystatin-C, NT-proBNP and GDF-15 were added in the model. An interaction term with timescale was also introduced to avoid violation of the PH assumption. HRs were corrected to match case-cohort design using the inverse-probability weighting technique of Kalbfleisch and Lawless.

Gompertz models showed an average predicted life-expectancy of 82.2 years for men and 88.9 years for women in the EPIC-Heidelberg cohort. Among men, within the follow-up times from above age 60, the difference in LE for a model based on lifestyle-related risk factors-only was 16.8 [15.9, 19.1] years when comparing men in the highest versus lowest quintiles of the model risk scores. When including the five biomarkers as additional risk predictors, this difference increased to 22.7 [22.3, 26.9] years (Table 3; Figure 2). In women followed up above the age of 60, the difference between those in the highest versus lowest quintiles of model risk is less pronounced but still significant at 9.87 [9.20, 13.1] years for the model based on lifestyle-related risk factors-only and 14.00 [12.9, 18.2] years for the model based on lifestyle factors plus biomarkers. These trends were similar in participants followed from ages 50–60 (Table 3; Figure 2).

Figure 2

Gompertz survival curves comparing individuals in the highest quintile to the lowest quintile of predicted risk, separately for lifestyle–related factors only, and the combination of lifestyle-related factors in addition to the biomarkers.

Table 3

Difference in LE and HR for men and women associated with lifestyle-related variables alone and with ageing biomarkers in addition to lifestyle factor, EPIC-Heidelberg case-cohort (n = 5,879).

Age at follow-up between 50 and 60
	Men		Women
	Difference in LE*	HR*	Difference in LE*	HR*
Lifestyle factors	19.9 (18.4; 24.8)	7.34 (5.75; 9.76)	11.2 (10.5; 14.5)	3.46 (2.78; 5.57)
Lifestyle factors + ageing markers	24.4 (23.5; 30.1)	9.03 (7.39; 13.37)	16.6 (15.0; 22.7)	5.33 (3.96; 9.25)
Age at follow-up above 60
	Men		Women
Lifestyle factors	16.8 (15.9; 19.1)	5.39 (4.53; 7.11)	9.87 (9.20; 13.1)	2.98 (2.48; 4.02)
Lifestyle factors + ageing markers	22.7 (22.3; 26.9)	7.74 (6.65; 10.3)	14.00 (12.9; 18.2)	4.08 (3.53; 5.92)

^*HRs and LE were obtained from Gompertz proportional hazard models and the area under the estimated Gompertz curve, and we compared the highest to the lowest quintile of predicted risk with a model including lifestyle factors-only and with a model additionally including GDF-15, NT-proBNP, CRP, HbA1C and cystatin-C. Models are adjusted for age (timescale), level of education (none/primary school, technical/professional education, secondary school, longer (including university), smoking status (never, long-time quitters (>10 years), short time quitters (≤10 years), current light (≤ 10 cigarettes/day) and current heavy smokers (>10 cigarettes/day)), number of pack-years, lifetime alcohol consumption, physical activity level (inactive, moderately inactive, moderately active and active), baseline self-reported diabetes and hypertension, BMI, waist circumference and an interaction term between BMI and waist circumference. To account for log-linearity, quadratic terms for CRP, cystatin-C, NT-proBNP and GDF-15 were added in the model. An interaction term with timescale was also introduced to avoid violation of the PH assumption. HRs were corrected to match case-cohort design using the inverse-probability weighting technique of Kalbfleisch and Lawless.

Discussion

In this long-term prospective analysis, we confirmed risk associations for all-cause mortality with four out of five ageing-related biomarkers (NT-proBNP, GDF-15, HbA1C, CRP and cystatin-C), which were pre-selected on the basis of their measurement reliability and relevance as physiologic indicators of biological ageing [19, 23]. Furthermore, we found that, compared to models based on lifestyle-related risk factors alone, models integrating the combined biomarker panel predicted a 5.9-year (22.7 versus 16.8 years) greater difference in LE among men, and a 4.13-year (14.00 versus 9.87 years) additional difference in LE among women, for individuals in the highest versus lowest quintiles of predicted mortality risk.

Regarding relative hazards for all-cause mortality, our findings for NT-proBNP, GDF-15, HbA1C and CRP (among men) are in line with previously published meta-analyses [32–36] reporting increased relative risks for all-cause mortality risk among men or women with higher circulating marker levels. Comparing between the highest and lowest quintiles of conditional marker distributions, conditional on age and lifestyle-related risk factors, we found HRs ranging from 1.13 [1.05, 1.35] for HbA1C to 1.49 [1.33, 1.66] for NT-proBNP among men followed over the age of 60 and from 1.21 [1.10, 1.50] for HbA1C to 1.52 [1.28, 1.79] for GDF-15 among women followed over the age of 60. While these relative hazard estimates may appear modest in comparison to estimates from earlier studies, especially for GDF-15 and NT-proBNP (HR > 2 in the meta-analyses), we note that our estimates were for quintiles of the estimated residual variation in biomarker levels which were adjusted for age and lifestyle factors so as to completely eliminate the role of lifestyle factors (especially smoking and adiposity) in the increased risk of all-cause mortality. This ensured having a realistic distribution of the biomarkers within specific categories or values of the lifestyle factors (e.g. avoiding attributing particularly low levels of GDF-15 to heavy smokers, GDF-15 being strongly correlated with smoking).

With regard to absolute LE, our modelling results show that lifestyle-related risk factors alone (i.e. BMI, smoking, alcohol, diabetes, hypertension and physical activity) were associated with a difference of 16.8 years of LE in men and 9.87 years of LE in women when comparing individuals in the highest versus lowest quintiles of predicted mortality risks. These results are in line with those from published studies in several countries [5–8, 10–15, 43], which also showed that unhealthy lifestyle behaviours might be linked with a loss in LE up to 18.9 years compared to healthy behaviours. Furthermore, our current results remain consistent with our previously published findings in the EPIC-Heidelberg cohort based on a shorter follow-up duration and a lower number of overall-mortality cases [17]. Regarding the sex-heterogeneity in the associations between biomarkers and LE, it can be linked not only to the healthier lifestyle behaviours in women compared with men (Table 1) but also to the fact that women have, on average, a longer LE than men, which might explain why the increase in life-expectancy associated with the biomarkers in women is smaller than in men. Compared with relative mortality risks, LE estimates provide an easier-to-understand metric for the general public and a relevant indicator for clinical practice, given the fact that death is an inevitable outcome, whereas occurrences of premature deaths as individualised outcomes are difficult to define.

While biological ageing may depend on genetic factors, increasing evidence suggests that it can still be delayed or targeted through individually modifiable lifestyle changes [44, 45], environmental changes [46] or pharmacological interventions [47]. After adjustment for chronological age, we observed mild-to-moderate correlations between GDF-15, CRP and HbA1C with smoking and adiposity, implying a potential role of smoking and adiposity in increasing the levels of these ageing-related biomarkers. Thus, our findings suggest that adopting a healthy lifestyle—e.g. by refraining from smoking and achieving or maintaining a healthy body weight—might impact LE both directly and indirectly by lowering levels of ageing biomarkers, the latter providing additional discrimination in LE and all-cause mortality risk even when using the conditional distribution.

The present study comes with several strengths. To our knowledge, this was the first study to investigate the loss of LE associated with lifestyle factors combined with a panel of ageing biomarkers in a cohort of initially healthy participants from the general population, having no cardiovascular conditions or cancer at baseline. The present analyses were well-powered, thanks to a case-cohort design [48] and an adapted sampling strategy for ageing-related diseases. Our estimates of relative mortality hazards for the five biomarkers were comprehensively adjusted for potential confounding by age, smoking, adiposity and other lifestyle-related factors, referring to conditional biomarker distributions to avoid excessive relative risk estimates due to unrealistic combinations of biomarker levels with age and lifestyle factors. Differences in LE were estimated using the LE distribution of the actual participants of the cohort and not for hypothetical and potentially unrealistic combinations of risk factors.

The relatively long follow-up duration (median around 18 years) enabled us to investigate long-term associations with mortality risk. However, a dilution effect with regard to baseline risk determinants could not be entirely ruled out, especially that associations above the age of 60 seemed to be slightly attenuated. The proportional hazards assumption was met after the use of a stepwise function of age, therefore assuming a change in the effects of lifestyle factors and biomarkers before and after the age of 60. The second assumption we made was the Gompertz distribution of the baseline hazard function for all-cause mortality, which describes an exponential increase of mortality rate with age [49]. This distribution has been extensively applied to investigate human mortality and lifespan [50]. In addition, the predicted survival curve by the Gompertz PH models show a good consistency with the Kaplan–Meier curve (Figure 2), suggesting that the Gompertz distribution is adapted for our data. Lastly, the validity of our findings also relies on the accuracy of risk factor assessments, assuming the absence of information bias linked to measurement errors. However, some lifestyle risk factors, such as alcohol consumption, physical activity or blood pressure, might be subject to non-differential substantial measurement errors, which might have caused an underestimation of real associations; consequently, we cannot entirely rule out that the additional discrimination provided by the biomarkers partly reflects inaccuracies in lifestyle factors’ measurements.

Conclusion

Our findings suggest that, beyond losses in LE predicted by unfavourable lifestyles, a selected panel of ageing-associated biomarkers, notably NT-proBNP, GDF-15, HbA1C and CRP, may lead to a major further discrimination of LE difference in a general population cohort of apparently healthy middle-aged and older adults and therefore help identify individuals at increased premature mortality risk, who should be prioritised to benefit from pharmacological or non-medical interventions. These results should be interpreted cautiously as they need to be replicated in other populations and settings: these findings do not allow, with the current scientific evidence, to provide a validated universal algorithm for LE prediction using a simple biomarker-kit, and the usefulness of a systematic measurement of these markers in healthy ageing monitoring is yet to be proved.

Declaration of Conflicts of Interest

None.

Declaration of Sources of Funding

The present project was supported by the Helmholtz Association’s Initiative on Aging and Metabolic Programming (AMPro ZT-0026). B.S. was awarded with a ‘Prix Jeunes Chercheurs’ by the Bettencourt-Schueller Foundation.

Acknowledgements

The authors are grateful for the continuous participation of our cohort participants—without their commitment, this work would not have been possible. EPIC–Heidelberg was launched in the 1990s. Unlike in new studies run today, public access to data from the EPIC population was not part of the study protocol at that time. Thus, the data protection statement and informed consent of the EPIC participants do not cover the provision of data in public repositories. Nevertheless, we are open to providing our dataset upon request for (i) statistical validation by reviewers and (ii) pooling projects under clearly defined and secure conditions and based on valid data transfer agreements.

Author notes

Joint first authorship, with equal contribution.