Aging constituted a significant risk factor for age-related chronic diseases, cognitive decline, and mortality. Chronological age (CA), serving as an indicator of individual aging status, was the primary contributor to the Framingham Risk Score (1) and represented the most robust predictor of cardiovascular disease. Nevertheless, owing to variations in epigenetics (2), nutritional intake (3), environmental exposure (4), and access to healthcare (5, 6), individuals with identical CA may have exhibited disparate rates of biological aging. The concept of biological age (BA) was introduced in 1988 (7) to delineate the true pace of aging. Precisely quantifying BA could facilitate the identification of early physiological alterations and elucidated the heterogeneity inherent in the aging process across different individuals.

Recent research had increasingly emphasized the estimation of BA based on clinical and biochemical parameters. Cohort studies had demonstrated that BA was more precise than CA in predicting age-related events. In comparison with omics-based measures of BA, such as DNA methylation age (8, 9), telomere length (10), and the inflammatory aging clock (11), BA derived from clinical indicators was more cost-effective, accessible, and scalable. Furthermore, compared to phenotype-based measures of BA, such as the Healthy Aging Index (12) and Frailty Index (13, 14), clinical-parameter-based BA could better reflect early physiological changes, captured the complexity of the aging process, and exhibited good applicability across both young and elderly populations (15). Common methods for estimating BA include standard linear models, the Klemera and Doubal method (KDM), homeostatic dysregulation (HD), and deep neural networks (DNNs) (16). Among these, DNNs were especially suited for training large datasets comprising tens of thousands or more samples. By explicitly encoding specific biological pathways into neural network architectures or backpropagating information to input features (17), DNNs may have more effectively captured the intricacies of the aging process.

Cohort studies had demonstrated that BA was a reliable predictor of cardiovascular disease, cancer, and all-cause mortality (18–20), but the majority of evidence originated from Western populations, with limited focus on Asian populations. Research exploring the association between BA and the risk of heart failure (HF) remained scarce. Additionally, investigations into heterogeneous aging trajectories had predominantly centred on the relationships between frailty trajectories and mortality in elderly populations (21). To date, no study had examined the relationship between BA trajectories and HF risk across a broad age spectrum. To address this research gap, the present study utilized data from the Kailuan Study, focusing on Chinese adults aged 18 years or older. BA was estimated using a DNNs model. It aimed to address the following core scientific questions: First, it employed a deep neural network (DNN) model to calculate the biological age of the target population, systematically explored the longitudinal association between biological age and the incidence risk of heart failure (HF), and filled the existing research gap regarding the association between these two factors in Asian populations. Second, taking into account the dynamic changes in biological age over time, it further constructed individual-specific biological age trajectories, clarified the association patterns between different aging trajectory types and the subsequent incidence risk of HF, and provided a novel theoretical basis for HF risk stratification and early intervention in the general population across all age groups.

The baseline characteristics of the 76,908 participants according to different aging status were summarized in Supplementary Table S2. The average CA was 51.3 years, and 61,115 participants (79.5%) were male. Compared with the other groups, participants with accelerated aging had the highest proportion of lower education levels and high salt intake.

The demographic and clinical characteristics of 41,489 participants included in the construction of BA percentile trajectories were presented in Table 1. In the 2010–2011 survey, the mean CA was 52.9 years, with 31,798 (76.6%) being male. Compared to individuals in the low-stable trajectory group, participants in the high-stable trajectory group were more likely to have lower educational attainment, current smoking and drinking habits, and lower income levels.

During follow-up from 2006 to 2022 (median follow-up duration: 16 years; interquartile range: 15.61–16.18 years), a total of 2,338 participants developed HF. The incidence of HF was highest among participants exhibiting accelerated aging (2.64 per 1,000 person-years) and lowest among those with decelerated aging (1.54 per 1,000 person-years). Compared to participants with normal aging, those with accelerated aging demonstrated an adjusted HR of 1.30 (95%CI: 1.19–1.43), while those with decelerated aging exhibited an adjusted HR of 0.76 (95% CI: 0.68–0.84) (Table 2).

The association between baseline aging status and HF risk was more pronounced among younger participants (<65 years) compared to those aged 65 years or older (P for interaction <0.01). Additionally, no significant interactions were observed between baseline aging status and other covariates (all P > 0.05) (Supplementary Figure S1).

Compared with the traditional risk model, models that incorporate BA (C-statistic: 0.7490; 95% CI: 0.7399–0.7581) or baseline aging status (C-statistic: 0.7429; 95% CI: 0.7336–0.7521) demonstrated enhanced predictive performance. However, an enhancement in discrimination ability was observed only in the model that incorporated BA (NRI: 0.1869; 95% CI: 0.1460–0.2278) (Supplementary Table S3).

During follow-up from 2010 to 2022 (median follow-up duration: 12.03 years, interquartile range: 11.66–12.32 years), a total of 858 participants developed HF. The HF incidence was highest in the high-stable trajectory group (2.41 per 1,000 person-years) and lowest in the low-stable group (1.34 per 1,000 person-years). Compared with the low-stable group, all other trajectory groups except for the increasing-decreasing group had significantly higher risks of HF. The highest risk was observed in the high-stable group (HR: 1.79, 95% CI: 1.48–2.17), followed by the decreasing-increasing group (HR: 1.45, 95% CI: 1.09–1.93), the high-decreasing group (HR: 1.32, 95% CI: 1.04–1.67), and the low-increasing group (HR: 1.30, 95% CI: 1.03–1.65) (Table 3). Compared with the high-stable group, the high-decreasing trajectory was associated with a lower risk of HF (HR: 0.74, 95% CI: 0.60–0.91), whereas the decreasing-increasing trajectory did not show a significant difference. No significant differences in HF risk were observed between the other trajectory comparisons (decreasing-increasing vs. high-decreasing; increasing-decreasing vs. low-increasing) (Table 4).

Sensitivity analyses were conducted by excluding participants who developed HF within the first year of follow-up and by applying competing risk models to account for the impact of all-cause mortality. The associations between baseline aging status, BA percentile trajectories, and HF risk remained consistent with the primary findings (Supplementary Tables S4–S7).

Our research findings revealed a substantial association between baseline aging status, as defined by BA, and the risk of developing HF. Specifically, individuals exhibiting accelerated aging demonstrated a higher risk of developing HF compared to those with normal aging patterns. Conversely, individuals with decelerated aging exhibited a reduced risk of HF. Furthermore, when compared to participants who consistently maintained the lowest percentile of BA (low-stable trajectory), other trajectory patterns-excluding the increasing-decreasing trajectory-were associated with an elevated risk of HF. Notably, the highest risk of HF was observed within 4 years among individuals maintaining the highest percentile of BA (high-stable trajectory). Additionally, participants whose trajectories transitioned from high to lower BA percentiles (high-decreasing trajectory) exhibited a diminished risk of HF relative to those in the high-stable group.

To our knowledge, this was the first large-scale prospective cohort study to demonstrate a longitudinal association between BA and HF risk. Previous studies had used different indicators and methods to construct phenotype age and other BAs and validated their predictive ability for age-related events in multiple cohorts based on Eastern and Western populations (19, 28–30). Mamoshina et al. trained independent DNNs to construct hematological aging clocks for South Korean, Canadian, and Eastern European populations, thereby predicting the risk of all-cause mortality across different populations (31). A study based on Moli Sani further extended this evidence to other age-related events, demonstrating that BA calculated using 36 circulating biomarkers and trained through DNNs could predict the risk of specific causes of mortality, hospitalization, cardiovascular disease, and cancer (32). However, the application of DNNs-derived BA in predicting HF has not been previously documented in the literature. This study provides novel evidence from a Chinese adult population, revealing heterogeneity in BA among individuals with the same CA and establishing that accelerated aging served as a risk factor, whereas decelerated aging acted as a protective factor against HF.

The association between aging status and the risk of developing HF remained stable across both male and female populations, which was consistent with prior studies on BA and age-related events (33). However, our findings indicated that this association was more pronounced in younger, predominantly middle-aged populations. A cohort study revealed that BA was strongly associated with all-cause mortality across multiple age groups, including young, middle-aged, and elderly individuals (15). Two additional studies demonstrated that the correlation between the frailty index and the risk of all-cause mortality was stronger in younger populations compared to older ones (13, 34). These results aligned with those obtained using continuous BA to estimate HF risk across different age groups. While incorporating CA into the calculation of BA could better reflect an individual's physiological state, it may have introduced bias in risk analysis, even when CA was adjusted for in the model. In our study, normalization methods were employed to eliminate the potential influence of CA, revealing a significant correlation between aging status and the risk of developing HF in the middle-aged population (45–65 years old). Accelerated aging represented a risk factor across all age groups.

Our research further demonstrated that BA could enhance the predictive power of traditional models, which included CA, gender, education level, occupation, physical activity, smoking status, alcohol consumption, salt intake, and income level. Two cohort studies conducted in Asian populations had shown that BA, derived from blood biomarkers, significantly improves mortality prediction (30, 35). Liu et al. calculated BA among adults aged 20 and above in the Third National Health and Nutrition Examination Survey (NHANES IV) and revealed that BA provided additional value for predicting 10-year mortality across all age groups (15). By integrating indicators reflecting the structure and function of the cardiovascular, hepatic, renal, immune, and metabolic systems into the assessment model for health risk, we achieved a more comprehensive understanding of physiological status, thereby enhancing health management and enabling precision prevention.

Our study identified six heterogeneous trajectories of BA percentile patterns and investigated their association with HF. Prior research on the relationship between BA and aging-related events had predominantly focused on single-point BA measurements, with limited studies examining the effects of continuous BA changes over time. The Health ABC study quantified changes in the Healthy Aging Index over a 9-year period and identified four distinct change patterns. Among these, the pattern with the most pronounced increase was associated with the highest mortality risk (36). A large-scale prospective MJ cohort study calculated multidimensional aging measures (MDAges) and delineated three homogeneous accelerated aging trajectories. The finding indicated that moderate or high accelerated aging trajectories were significantly associated with an elevated risk of mortality compared to low accelerated aging trajectories (37). A cohort study focusing on elderly individuals constructed five heterogeneous frailty trajectories based on frailty traits. Specifically, compared with the moderate degrading trajectory, the moderate increasing trajectory, high stable trajectory, and high increasing trajectory were associated with a higher risk of mortality (21). Additionally, another study conducted among older adults identified multiple sets of frailty trajectories across different age groups. This study also demonstrated that the increase in frailty was significantly associated with an elevated risk of all-cause mortality (38). Our study revealed that the risk of HF was lowest in individuals following low-stable aging trajectories, underscoring the critical importance of maintaining consistently low and stable aging status for HF prevention. The risk of developing HF in the low-increasing trajectory group was found to be higher than in the low-stable trajectory group, highlighting the necessity of long-term management and control of aging status in clinical practice. In comparison, participants following high-decreasing trajectories exhibited a lower risk of HF compared to those on high-stable trajectories, whereas no significant differences were observed in the decreasing-increasing trajectory group. This suggested that individuals with initially high aging status may have reduced their HF risk by intervening to improve these levels. However, such improvements required sustained long-term efforts. Notably, the HF risk for participants in the high-decreasing trajectory remained higher than that for participants in the low-stable trajectory, indicating residual risk even after improvement. Furthermore, reducing aging status before significant accumulation of aging-related damage yielded less benefit in terms of HF incidence reduction compared to maintaining a low aging statusthroughout early life. Furthermore, we found that the association between baseline aging status and HF risk appeared more pronounced in younger participants compared to those aged 65 years and older at baseline. This may have been attributed to the fact that physiological functions (such as metabolic regulation, immune response, and organ reserve capacity) in individuals under 65 had not yet undergone irreversible decline, resulting in greater biological plasticity (28). This implied that aging-related characteristics (e.g., biological age percentile, aging status) in this age group had a more direct and sensitive impact on heart failure.

This study has several strengths. First, it was the first to investigate the association between BA and HF in a large-scale prospective cohort of Chinese adults aged 18 years and older. Second, BA was repeatedly measured, allowing us to identify six heterogeneous aging trajectories and to explore the longitudinal association between long-term changes in aging statusand subsequent HF risk. Third, we adjusted for a wide range of confounders—including smoking, alcohol consumption, and physical activity—thereby substantially reducing the potential for unmeasured residual confounding. Lastly, the traditional Cox model assumes “independent risks” and may overestimate the risk of the target event, especially in subgroups with high mortality. In this study, non-heart failure (HF) death was treated as a competing risk event, and a Fine-Gray competing risk model was constructed to calculate a more accurate hazard ratio (HR). Results of the competing risk analysis showed that compared with the primary results, there was no significant change in the associations between baseline aging status, biological age percentile trajectory patterns, and HF. Nonetheless, several limitations should be noted. First, all participants in our cohort are Chinese adults from the Kailuan community, with the majority being coal miners (30% are coal miners), which may limit the generalization of the conclusions to other populations. However, BA has demonstrated predictive utility for aging-related outcomes across multiple racial and ethnic groups, supporting the broader applicability of our findings. Second, the proportion of male participants in the cohort was relatively high. To mitigate sex-related bias, BA and aging status were calculated separately for men and women. Third, we were unable to classify HF into clinical subtypes. This study mainly recruited patients with heart failure with reduced ejection fraction (HFrEF), and the results may not be applicable to other types of heart failure. HF is a complex syndrome with heterogeneous pathophysiology, and future studies should explore whether specific BA patterns are differentially associated with HF subtypes. Lastly, in large cohort studies, perfect uniformity in follow-up intervals is difficult to achieve. As a result, minor misclassification of chronological age strata for a small number of participants may have occurred during BA percentile calculation. However, given the large sample size and the uniform distribution of such cases across age groups, the overall impact of this limitation is likely minimal.

Using a DNNs model, we estimated BA to identify individual biological aging status and long-term aging trajectories. Compared with individuals exhibiting normal aging patterns, those with accelerated aging demonstrated an elevated risk of developing HF, while those with decelerated aging exhibited a reduced risk. Additionally, biological aging trajectories were significantly associated with subsequent HF risk. Specifically, individuals with consistently low aging status exhibited the lowest risk, whereas those transitioning from high to low aging status also experienced a reduction in risk. These findings emphasize the significant predictive value of BA in evaluating the risk of HF and reinforce the critical importance of promoting healthy aging within primary healthcare frameworks. Furthermore, these results establish a solid foundation for future experimental and clinical investigations aimed at elucidating the mechanisms and developing interventions to mitigate the impact of biological aging on HF progression.