Comparison and validation of machine learning models to predict 5-year fall risk among community-dwelling older adults in China.

IMPORTANCE: Falls are a leading cause of injury, disability, and loss of independence among older adults in China. Validated, data-driven approaches to risk stratification are needed, yet no study has systematically compared multiple machine learning models for fall prediction using nationally representative longitudinal data from this rapidly aging population. OBJECTIVE: To develop, compare, and validate five distinct machine learning models for predicting incident falls over 5 years among community-dwelling older adults in China. DESIGN, SETTING, AND PARTICIPANTS: This prospective cohort study used data from the China Health and Retirement Longitudinal Study (CHARLS). A total of 8,072 community-dwelling adults aged ≥ 60 years with baseline assessments in 2015 were followed through 2020. Fall incidence was ascertained via self-report during the 2018 and 2020 follow-up waves. EXPOSURES: Baseline predictors included 39 variables across six domains—demographic, lifestyle, chronic disease, sensory and physical function, anthropometric and biomarker, and psychosocial/environmental factors, after excluding variables with ≥ 50% missingness. MAIN OUTCOMES AND MEASURES: Fall incidence between 2015 and 2020. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUC-PR), Brier score, calibration plots, and decision curve analysis (DCA). The five models compared were: generalized linear model (GLM), gradient boosting machine (GBM), distributed random forest (DRF), deep learning (DL), and a stacked ensemble integrating all four base learners. RESULTS: Over 5 years, 3,074 participants (38.1%) reported at least one fall. The stacked ensemble model achieved the highest discrimination (AUC = 0.865; 95% CI, 0.848–0.882) and excellent calibration (Brier score = 0.105), outperforming GLM (AUC = 0.849), DL (0.860), DRF (0.860), and GBM (0.859). Key predictors included prior fall history (permutation importance ΔAUC = − 0.306), age, glycated hemoglobin, arthritis/rheumatism, white blood cell, total cholesterol and so on. Decision curve analysis indicated that a “treat-all” strategy yielded greater net benefit than any threshold-based intervention across clinically relevant probabilities (1%–50%). CONCLUSIONS AND RELEVANCE: Among five machine learning approaches evaluated, the stacked ensemble model demonstrated superior performance in predicting 5-year fall risk in a nationally representative Chinese cohort. Given the high population-level burden, universal prevention remains the baseline strategy. Our findings indicate that the ensemble model is particularly effective for secondary prevention (i.e., risk stratification for individuals with a prior fall history), where it achieves excellent performance. However, its predictive power is limited for primary prevention in fall-naïve individuals, suggesting that static baseline data alone are insufficient for first-time fall prediction. Ultimately, integrating this model into digital health platforms has the potential to support proactive, risk-stratified fall prevention, directly aligning with the strategic goals of China’s ‘Healthy China 2030’ initiative.