Sarcopenia, characterized by the progressive loss of muscle mass, strength, and function, is a prevalent condition among older adults (1). Its prevalence ranges from 10 to 27% in individuals aged over 60 years (2). The impact of sarcopenia on older adults is multifaceted, encompassing various dimensions of health and well-being. Research has demonstrated that sarcopenia is correlated with functional impairment, physical disability, and an elevated susceptibility to adverse health-related outcomes (3). Furthermore, as the largest metabolic organ of the human body, the dysfunction of skeletal muscle is also associated with the increased risk of osteoporosis (4), diabetes metabolism (5), cardiovascular diseases (6), and other chronic diseases, which ultimately leads to the decease of physical activity, and limitations in health-related quality of life (7). Given the detrimental impact of sarcopenia, multiple consensus underscores the significance of early detection, intervention, and management strategies (8, 9).

In recent years, propelled by advancements in machine learning and artificial intelligence, researchers have employed statistical or machine learning models to forecast specific health-related outcomes (10). These models have the capacity to analyze complex healthcare data and provide valuable insights for disease risk, treatment response, or healthcare resource utilization (11). The prediction of sarcopenia has been a subject of extensive research, with various models and tools being developed to identify and predict the risk of sarcopenia in different patient populations. For instance, Shin et al. developed a predictive model for sarcopenia utilizing multiple biomarkers in community-dwelling older adults (12). Similarly, Xu et al. developed a multivariable model based on ultrasound imaging features of the gastrocnemius muscle to identify patients with sarcopenia (13). Several researchers developed and validated of a nomogram for predicting sarcopenia in community-dwelling older adults (14, 15).

While current studies collectively support the feasibility of using physiological, biochemical, and imaging indicators to predict sarcopenia, existing sarcopenia prediction models face critical limitations in community settings. One of the primary limitations is the availability of data. The development of predictive models typically requires a large amount of clinical data. However, compared to hospital settings, there may be difficulties in obtaining older adult patients’ information and test results in community setting. Thus, the reliance on specific modalities such as ultrasound and serum biomarkers for risk stratification may pose limitations in terms of accessibility and standardization (16, 17). Another critical limitation is the transferability of model. Current evidence emphasizes specific population groups, such as cancers patients (18) and chronic patients (19–21), which may hinder the widespread application of community screening of certain prediction models, necessitating the development of cost-effective, universal screening methods. Additionally, the lack of data from China using precise assessments for sarcopenia, such as skeletal muscle area or skeletal muscle mass index, underscores the need for more comprehensive and standardized diagnostic criteria. Thus, this highlights the importance of highly actionable, wide applicability, and well-performance insights in the implementation of prediction models in sarcopenia.

China’s National Basic Public Health Services (NBPHS) program, launched in 2009 by the Chinese government, is a nationwide healthcare initiative designed to enhance population health by delivering essential preventive and primary care services (22). A key component of the program is annual health examinations for adults aged 60 and above, conducted at local community health centers (23). The NBPHS dataset offers several unique advantages. Firstly, unlike hospital-based datasets, NBPHS captures community-dwelling older adults without selection bias (e.g., excluding those with severe comorbidities). Secondly, prior studies relied on niche biomarkers (e.g., serum leptin) or imaging (e.g., CT) that are impractical for community screening. NBPHS data bridge this gap by using accessible, low-cost variables scalable to resource-limited settings. More importantly, the program’s nationwide infrastructure and standardized protocols ensure that any risk models developed from these data can be immediately integrated into existing community health workflows.

In this study, we developed and validated a machine learning model (w-ACT: weight, Age, Calf circumference, Triglycerides) using routinely collected NBPHS variables to predict sarcopenia risk among community-dwelling older adults. Through comparative analysis of eight algorithms, our Random Forest-based model achieved superior performance while maintaining clinical interpretability. The final w-ACT model incorporates four readily accessible predictors—weight, Age, Calf circumference, and Triglycerides—addressing critical gaps in community screening through three key innovations: (1) replacing specialized biomarkers with routine examination metrics, (2) demonstrating cross-regional generalizability through external validation, and (3) providing transparent risk stratification. This approach establishes a practical framework for integrating sarcopenia risk assessment into China’s primary care system while maintaining diagnostic accuracy comparable to resource-intensive methods.

Recent research has sought to develop risk prediction models for sarcopenia in community-dwelling older adult populations, facing challenges in data accessibility, interpretability, and predictive accuracy. To address these issues, we utilized local data from the NBPHS program and applied advanced machine learning techniques to create the w-ACT model. Our findings demonstrate that the w-ACT model outperformed traditional LR and other common ML models in both internal and external validations, highlighting its potential to improve the identification of at-risk older adult populations and inform preventive interventions.

We identified a sarcopenia prevalence of approximately 20.11% among community-dwelling older adults in Chongqing, which is notably higher than rates reported in other regions of China utilizing the AWGS 2019 diagnostic criteria—ranging from 14.62% on the Yunnan-Guizhou Plateau to 13.47% in Tianjin (38) and 10.29 to 13.47% in Shanghai (39, 40). The observed discrepancies may be attributed to regional demographic variations. Chongqing, a typical mountainous city with a severe aging population, has an aging rate of 21.87% according to the results of the seventh national census, ranking fifth nationwide and first in the western region (41). In additional, it may be also contributed by the complex interplay of lifestyle and dietary norms characteristic of Chongqing’s residents (42). A pervasive pattern of high oil and salt intake, common among the city’s inhabitants, may be contributing to this disparity. Moreover, nutritional surveys indicate that over 50.5% of the middle-aged and older adult population are overweight or obese. Future research should explore these factors to develop targeted interventions that preserve muscle health and improve the well-being of the aging population.

The w-ACT tool’s design prioritizes immediate clinical utility through three key characteristics. Firstly, seamless integration with existing workflows. The selected indicators (age, weight, CC, TG) are routinely collected in NBPHS service projects, which promote equal access to public health services. The four risk factors—age, weight, CC, and TG—are readily obtainable across various older adult care settings. This allows direct implementation without requiring additional tests or equipment.

Secondly, these risk factors provide significant interpretability. The model’s clinical adoption potential stems from its deliberate avoidance of computationally significant but clinically obscure features - a common criticism of ML healthcare applications. Our four predictors were selected not only for statistical contribution but for their established clinical meaning and measurability in resource-limited settings Consistent with existing guidelines, age, low weight, and CC are established risk factors for sarcopenia, and their non-invasive measurement methods facilitate assessment in diverse settings (1). Notably, CC serves as an index of free fat muscle mass (FFM) and has been recognized by the WHO as a sensitive indicator of muscle mass in older adults (43). Our study also identified triglycerides as a potential risk factor, aligning with previous research (44–46). Researches has established a bidirectional causal relationship between triglycerides and sarcopenia, particularly regarding muscle quality (47, 48). Although the specific mechanisms remain incompletely understood, theories such as the “cycle metabolism theory” suggest that elevated blood lipid levels may contribute to fat accumulation in skeletal muscle, ultimately leading to mitochondrial dysfunction and insulin resistance (49). Thus, the inclusion of these four risk factors enhances both the credibility of the model and facilitates clinical practitioners’ understanding of its predictive foundation.

Thirdly, Despite the advantages of machine learning algorithms, the comparative performance of ML versus traditional LR remains debated (50). While some studies suggest that ML does not consistently outperform LR (51, 52), aligning with the majority of literature (37, 51, 53), our findings indicate superior predictive performance of the ML, especially RF algorithm, which emerged as the top performer among the models assessed. The observed differences may stem from prior research primarily focusing on singular predictive metrics, without considering a comprehensive range of evaluation metrics. Notably, the RF algorithm is praised for its resistance to overfitting, adaptability to categorical variables, accuracy in error rate estimation, and capability of ranking variables by relative importance. These characteristics have led to its successful application in stoke (54), cancer (55), and postoperative functional (56) prediction. Thus, our findings furnish additional evidence for the application of the RF algorithm in risk prediction and serve as a reference for peers in evaluating model performance and selecting algorithms.

Some limitations should be mentioned. First, the cross-sectional NBPHS design precludes causal inference, though we mitigated this via instrumental variable analysis (57). To definitively address these limitations, we will launch the Chongqing Aging and Sarcopenia Evaluation (CHASE) cohort. This prospective study is specifically designed to document predictor trajectories preceding sarcopenia onset and validate dynamic predictions. Second, although internal-external validation was conducted, demonstrating the same level of reliability, the external validation sample size (n = 182) may limit the assessment of overfitting risks, particularly given the model’s complexity. While our permutation-based feature importance analysis (Figure 4) shows biologically plausible weightings, the possibility of overfitting to specific subpopulations cannot be entirely excluded. Additionally, dealing with imbalanced data is a challenging issue in both deep learning models and traditional models for practical classification problems (58). While oversampling improved model training, its theoretical impact on population prevalence estimates cannot be fully excluded - though our external validation using original unbalanced data suggests robust generalizability. Third, while our model demonstrated consistent performance across internal and external sets, its generalizability to populations with different ethnic compositions or healthcare systems requires further validation, particularly given regional variations in sarcopenia prevalence and risk factors. Normalization assumptions may not perfectly fit all variables, but this approach represents the most widely accepted compromise for mixed-type clinical data (59, 60).

This study developed a ML-based predictive model to identify early-stage high-risk individuals for sarcopenia among older adult residents in the community. Weight, triglycerides, age, and calf circumference were identified as significant factors associated with sarcopenia, and RF exhibited superior predictive performance in eight approaches. The interpretability of this model (termed w-ACT model) was achieved using SHAP values, providing valuable insights into the contributions of these variables to sarcopenia risk.