Accurate identification of older adults at high risk of falling is essential to prevent injuries and implement effective interventions. This study evaluated the performance of several machine learning models in predicting fall risk using both accelerometric data from wearable sensors and non-accelerometric data including demographic, functional, and clinical variables. A dataset comprising 146 older women was used to train and compare seven algorithms: random forest, XGBoost, AdaBoost, LightGBM, Bayesian ridge regression, decision trees, and support vector regression. Predictive accuracy was assessed using mean squared error, mean absolute error, and the coefficient of determination. Models trained with combined accelerometric and non-accelerometric data consistently outperformed those using a single data source. XGBoost achieved the lowest MSE, while Bayesian ridge regression reached the highest R2, highlighting its superior explanatory capability. In contrast, support vector regression exhibited poor predictive performance. Non-accelerometric variables, especially age and comorbidities, emerged as major predictors, whereas accelerometric data alone yielded limited accuracy. The integration of multiple data types significantly improved model robustness and clinical applicability. These findings underscore the value of multi-source data fusion in enhancing fall risk prediction among older adults and support the implementation of hybrid machine learning models in clinical settings.