Abstract
OBJECTIVES: Accurate prediction of fall risk in older adults is essential to prevent injuries and improve quality of life. This study evaluates the predictive performance of various machine learning models using accelerometric data, non-accelerometric data, aiming to improve predictive accuracy and identify key contributing variable. METHODS: We applied random forest, XGBoost, AdaBoost, LightGBM, support vector regression (SVR), decision trees, and Bayesian ridge regression to a dataset of 146 older adults. Models were trained using accelerometric data (movement patterns) and non-accelerometric data (demographic and clinical variables). Performance was evaluated based on mean squared error (MSE) and coefficient of determination ( R2 ), to assess how combining multiple data types influences prediction accuracy. RESULTS: Models trained on combined accelerometric and non-accelerometric data consistently outperformed those based on single data types. Bayesian ridge regression achieved the highest accuracy (MSE = 0.6746, R2 = 0.9941), demonstrating superior performance compared to decision trees (MSE = 0.1907, R2 = 0.8991) and SVR (MSE = 1.5243, R2 = - 2.2532). Non-accelerometric factors, including age and comorbidities, significantly contributed to fall risk prediction. CONCLUSIONS: Integrating accelerometric and non-accelerometric data improves fall risk prediction accuracy in older adults. Bayesian ridge regression trained on combined datasets provides superior predictive power compared to traditional models. These findings highlight the importance of multi-source data fusion for effective fall prevention strategies. Future work should validate these models in larger, more diverse populations to enhance clinical applicability.