Enhancing sarcopenia screening in primary care: a machine learning approach using simple physical tests vs. SARC-F in 2,788 community-dwelling older adults.
Key claim: An XGBoost model using brief physical tests (grip strength, 5STS, balance, reaction time) screened community-dwelling older adults for sarcopenia with AUC 0.92 and 86.5% sensitivity—nearly 2.5× that of SARC-F—with 5STS as the strongest predictor and an exploratory 12-s threshold.
Abstract
BACKGROUND: SARC-F, a widely used screening tool for sarcopenia, offers high specificity but poor sensitivity (30-50%), leading to substantial missed diagnoses in community settings. OBJECTIVES: To develop and validate a machine learning model using simple physical function tests to screen for sarcopenia in community-dwelling older adults, and to compare its performance against SARC-F. DESIGN: Cross-sectional study. SETTING AND PARTICIPANTS: Data were collected from 2,788 older adults (≥60 years; 64.8% female) across 45 community health centers in Hangzhou, China (August-September 2025). Confirmed sarcopenia prevalence was 18.2% (AWGS 2019 criteria). METHODS: Predictors included grip strength, five-repetition sit-to-stand (5STS) time, static balance, and reaction time (total <5 min). XGBoost, Random Forest, and Logistic Regression models were developed and evaluated on a temporally independent test set (training = 2,024; test = 764). Model performance was assessed using AUC, sensitivity, and specificity. SHAP analysis provided interpretability. RESULTS: The XGBoost model achieved superior performance (AUC = 0.92; 95% CI: 0.90-0.94), with sensitivity of 86.5% and specificity of 85.1%-nearly 2.5 times the sensitivity of SARC-F (34.8%). 5STS time emerged as the strongest predictor (mean |SHAP| = 0.21). A 12-s 5STS threshold (exploratory, 95% CI: 11-13 s) was identified using Youden’s index and SHAP analysis, warranting prospective validation. Decision curve analysis demonstrated positive net benefit across 10-60% thresholds, with net benefit 0.12 at 20%, equivalent to 12 additional true cases identified per 100 screened individuals without increasing unnecessary referrals. CONCLUSIONS AND IMPLICATIONS: This internally validated machine learning model shows promise for sarcopenia screening using brief, low-cost functional tests. Its superior sensitivity, an exploratory 12-s 5STS threshold, and an estimated 80-90% reduction in per-capita screening costs (based on equipment cost comparison) suggest potential utility in primary care, but external validation is required before widespread deployment.: Community health workers can deploy this tool to enable early identification and timely nutrition and exercise interventions.: Integration into existing community health programs may reduce long-term care burden by delaying functional decline in aging societies.
Why this matters
SARC-F is the dominant primary-care sarcopenia screen but has known low sensitivity; this study shows a short battery of objective physical tests, combined via a tree-based ML model, can raise detection sensitivity to ~87% at AUC 0.92 in a community cohort of 2,788. It also elevates the 5-times sit-to-stand (with an exploratory 12-s cut-point) above grip or gait alone as the single most informative test, tightening the functional-screening front end that gates entry into resistance-training and sarcopenia intervention pathways — pending external validation.