Aging is a complex, inevitable, and multifactorial process, characterized by functional deterioration, physiological damage, and multiple age-related diseases (1). One key question to address aging-related issues is how to precisely quantify aging, with accumulating evidence supporting the utility of biological age (BA) in predicting age-related outcomes and differentiating individual health status (2–5). To be more specific, one study on 2,844 Chinese Singaporeans developed BA with the Klemera and Doubal method (KDM) and found that BA is better than chronological age (CA) for predicting life span (mortality) and healthspan (frailty) (2). BA has therefore been accepted as a more accurate proxy of aging than CA.

Biological age (BA) is generally referred to as a single latent variable that integrated multiple biomarkers relevant to health (6). Various statistical methods have been used to approximate BA, such as the multiple linear regression method (7), the principal component analysis (8), Hochschild's method (9), and KDM (10). KDM has been suggested as the optimal method for BA estimation (11). Although traditional methods have been demonstrated to perform well in predicting adverse aging outcomes (7–10), they may encounter obstacles when dealing with complex, multidimensional data. Among such multidimensional data, there are complex interactions among the features such as the interaction between vitamin D and albumin on mortality (12), and most of the current models were developed based on regression methods. The researcher needs to manually input the predefined interactions. Missing those complex interactions in the regression model may result in an inaccurate prediction of outcomes. Machine learning (ML) offers tremendous opportunities for researchers to address these issues (1, 13). In the prediction model developed by machine learning methods, the model can automatically identify those interactive relationship from the data and if it is unnecessary to specify interactions. Accordingly, ML methods help to figure out complicated relations among the biomarkers and generate more accurate prediction models. Also, the ML methods can reduce the loss of biological information such as the complex interactions (1). However, the application of ML in the development of aging measures has not been studied thoroughly (2, 14–17). Most of these studies were conducted among adults in Europe and the US (14–16) and ML seems to not provide more accurate aging measures than conventional methods in one study with eight biomarker features in men and 10 features in women (2). The small number of features may be one potential limitation for such methods. More studies are required to validate the application values of ML in other populations and with more features.

China is facing rapid population aging, which brings formidable challenges to policymakers and caregivers. In 2019, the Chinese population accounted for 18% of the world population, with 164.5 million adults aged 65 and over and 26 million adults aged 80 and over (18). Developing aging measures for the Chinese population is of great significance to solve aging-related issues in this large country, such as facilitating the early identification of adults at risk. To date, a few relevant studies have been conducted in the Chinese population (5, 8, 19–23). Most of them used the multiple linear regression method (19, 23) or the principal component analysis method (8, 20, 22). We have previously provided a step forward, i.e., developing a valid physiological biomarker-based aging measure using KDM (hereafter referred to as KDM-BA) (5). As the KDM measurement included limited biomarkers, we considered building the ML-based aging measurement with more features among the Chinese population and evaluating how it behaves relative to the most recent KDM-BA we developed.

Therefore, this study aimed to apply several ML algorithms (e.g., Gradient Boosting Regressor, Random Forest, CatBoost Regressor, and Support Vector Machine) to develop new aging measures (hereafter referred to as ML-BAs). We then examined the associations of the best ML-BA and KDM-BA with physical disability and mortality during the follow-up period. We used data from the China Health and Retirement Longitudinal Study (CHARLS), a nationally representative survey.

The basic characteristics of the study population are presented in Table 1. The mean CA of the study population was 59.1 (SD = 9.2) years. Of the 9,771 middle-aged and older adults, ~44.6% were aged ≥ 60 years, 53.5% were women. The mean CAs of men and women was 59.8 (SD = 9.1) years and 58.5 (SD = 9.2) years, respectively.

In this study, we successfully developed an aging measure using the Gradient Boosting Regression model in a sample of middle-aged and older Chinese adults. We found that this ML-BA was predictive of physical disability and mortality during the follow-up period, and these associations were independent of CA. The results were better than that of KDM-BA, supporting the development of ML-BA. This ML-BA may serve as a proxy of life span in geroscience research and help with the risk stratification in the general Chinese older adults.

To date, several studies have shown that BA calculated using ML has the predictive ability for mortality risk in populations from different countries, such as the US (15, 17), Italy (34), and Singapore (2). Because of differences in genetic and socio-environmental factors, the findings may not be generalizable to various populations in other countries, such as the Chinese population, a rapidly increasing segment worldwide. To the best of our knowledge, no studies have been performed to develop BAs using ML and evaluate the associations of ML-BAs with adverse outcomes in the Chinese population. We filled up this gap in this study. More importantly, we demonstrated that the best ML-BA performed just as well as KDM-BA, which has been regarded as the best biological aging measure (28). The findings support that ML could be used to develop measures of biological aging. Moreover, both ML-BA and KDM-BA could be developed across various populations separately, and they may capture something underlying the aging process.

It should be noted that the strength of the associations of the best ML-BA with physical disability and mortality is slightly stronger than that for KDM-BA. The ML-BA in our study was computed based on 19 biomarkers, while the KDM-BA was computed based on only eight of the 19 biomarkers. The remaining 11 biomarkers included diastolic blood pressure, pulse, white blood cell count, mean corpuscular volume, glucose, high-density lipoprotein, low-density lipoprotein, hemoglobin, cystatin, uric acid, and hematocrit, which have been demonstrated to be associated with aging (35–39). Hence, we assume that the better performance of ML-BA may be due to the more information covered by ML-BA than that by KMD-BA. The ML-BA was developed without prior assumptions and was not dependent on intermediate results from multiple linear regression models (40), allowing ML-BA to be easily verified. In general, ML-BA may therefore provide a useful tool to identify individual risks for adverse outcomes.

The stable associations of ML-BA and KDM-BA with physical disability and mortality risk can be interpreted by looking into the biological biomarkers used to develop the two aging measures. The aging process is subclinical, characterized by various types of biological degradations. So, it is proposed to estimate aging based on cellular and molecular hallmarks (41). In our study, the biomarkers used for ML-BA and KDM-BA computation represent different but important domains of physiological function or systems: immune system (e.g., high-sensitivity C-reactive protein, platelet count, and white blood cell), cardiac-metabolic system (e.g., Total cholesterol, systolic blood pressure, and low-density lipoprotein), and kidney system (e.g., urea, creatinine, cystatin, and uric acid). First, the immune system is a homeostatic system that helps to maintain the function of the organisms, and age-related changes in immune function have been demonstrated to affect longevity (42). Due to infectious diseases, older adults usually have an increased risk of morbidity and mortality (43), emphasizing the importance of maintaining the function of the immune system during the aging process. Second, since the incidence of heart disease increases sharply with age, it has been postulated that aging and cardiovascular disease are interrelated (44) and may share common pathology (45). During the normal aging process, the cardiac-metabolic function is impaired with the increase of age (44), contributing to adverse age-related outcomes. Finally, evidence has suggested that even in the absence of comorbidities, the kidney may experience significant age-related changes in structure and function (46). This implies that the deterioration of kidney function may be one of the important phenotypes of the aging process. The aging measures we developed in the current study integrated various biomarkers of immune function, cardiac-metabolic function, and kidney function; therefore, they could reflect the aging process through multiple physiological systems and work well in predicting physical disability and mortality.

From the perspective of the application, both ML-BA and KDM-BA could be considered since they had satisfactory predictive performance in this study. The choice of methods is largely dependent on sample size, distribution, and data availability. ML is more non-parametric and modeling-based, while KDM is more parametric and theory-based. Non-parametric methods do not require assumptions about sample size and data distribution (47, 48), and thus, are flexible. Therefore, ML-BA is a good choice if the shape of the distribution was not suited for parametric methods. On the contrary, KDM-BA would be efficient in a sound dataset.

In this study, the large sample size of the nationwide prospective cohort study provided us with the opportunity to develop aging measures by ML and explore its associations with adverse health outcomes in middle-aged and older Chinese adults. Nevertheless, limitations in this study should be acknowledged. First, the relatively short follow-up period (i.e., up to 6 years) of the CHARLS has impeded us to explore the long-term effect of aging measures on the outcomes. Longitudinal studies with long-term follow-up are needed to confirm the associations. Second, we did not have data on the exact timing of physical disability incidence and death. Therefore, we cannot evaluate the impact of BAs on survival time and we used the 2-year mortality as an alternative. Third, we did not have data on the incidence of chronic diseases (e.g., diabetes, heart disease, and stroke), impeding us to evaluate the associations of biological aging with chronic diseases. Fourth, it would be useful to validate the predictive performance of ML-BA in another dataset. However, there are only a few large aging cohort studies in China, such as the Chinese Longitudinal Healthy Longevity Survey (CLHLS), the China Health and Nutrition Survey (CHNS), and the CHARLS, which was used in our work. The CLHLS and CHNS do not have all the biomarkers used in our work, we were unable to validate this ML-BA in this regard. Finally, the utility of this ML-BA needs to be further validated as it needs more input information. However, with the further development of medical informatization, more individual-level data will be available and this method will be the workflow for building the BA with more information.

In summary, this study provides a valid ML-based measure of biological aging for middle-aged and older Chinese adults. We further demonstrated that this ML-BA was associated with physical disability incidence and mortality. These associations were comparable with that of KDM-BA, a valid physiological biomarker-based aging measure we have previously developed. The findings support the application of ML in geroscience research and promote further understanding of the aging process. Together with KDM-BA, these aging measures could serve as a proxy of life span and help with the risk stratification in the general Chinese older adults.

The datasets for this study can be found in the CHARLS website at http://charls.pku.edu.cn/index/en.html.

The studies involving human participants were reviewed and approved by the Institutional Review Board at Peking University. The patients/participants provided their written informed consent to participate in this study.

ZL and CW conceived and designed the study and contributed to the critical revision of the manuscript for important intellectual contents. XC, GY, and XJ performed the analysis and wrote the initial draft of the manuscript. LH, XL, ZZ, ZL, and CW helped to interpret the results and edit the manuscript. All authors have read and agreed to the final version of the manuscript.

The research results of this article are sponsored by the Kunshan Municipal Government research funding.

XJ is employed by MindRank AI ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.