BMC Geriatr. 2026 Feb 18. doi: 10.1186/s12877-026-07161-y. Online ahead of print.
ABSTRACT
BACKGROUND: With the accelerating aging of the global population, muscle health issue occurs commonly as an age-related process in older people. The conventional low muscle mass screening and diagnosis reliant on bulky and costly instruments, remain challenging for regular self-monitoring. If routine physical examination information from primary healthcare settings is integrated and analyzed using appropriate statistical methods, it may be possible to derive robust predictions for low muscle mass screening. By doing so, we seek to explore an interpretable machine learning-based screening manner for low muscle mass among Chinese community-dwelling older adults.
METHODS: We recruited aged ≥ 60 years older adults from the baseline of the elderly nutrition and health cohort. Low muscle mass was assessed by BIA-measured appendicular skeletal muscle mass index (ASMI) using AWGS 2019 consensus cut-offs. Following physical examination in community health settings, individual information about the participants was measured and gathered, including general information, medical history, physical measurements and biochemical indicators. The primary objective of this study was to explore an interpretable machine learning-based screening manner for low muscle mass. For predicting low muscle mass (by classification) or ASMI (by regression), three representative supervised machine learning models were constructed. To make the prediction behavior of the model transparent and ease clinical use, SHAP algorithm and Shiny framework were utilized, respectively.
RESULTS: 569 Chinese community-dwelling older adult were enrolled. Among them, 99 participants (17.4%) were assessed with low muscle mass. Among three models tested, the random forest model exhibited superior overall performance and better generalizability for low muscle mass (AUC = 0.872 in test set), and the elastic net showed the best prediction performance for ASMI (R² = 0.763 in test sets). The identified key predictors of low muscle mass based SHAP algorithm revealed expected patterns, such as the importance of BMI, age, calf circumference, MNA score, but also unexpected variables, such as HDL. The final optimal prediction model was deployed in an interactive and user-friendly decision support application to facilitate the clinical application.
CONCLUSIONS: This study demonstrates that routine physical examination information could be a valuable component to incorporate into targeted assessments to screen low muscle mass among community-dwelling older adults. Building on this foundation, an interpretable machine learning approach was explored, which proves well-suited as a screening manner for low muscle mass to guide further standard assessment. Its suitability stems from superior predictive performance and operational feasibility in resource-constrained community health settings.
PMID:41703458 | DOI:10.1186/s12877-026-07161-y
