Machine Learning Models for Fall Risk Assessment Using Wearable-Derived Gait Features

Wearable sensors have been widely used to assess gait parameters in real-world settings, enabling non-invasive and continuous monitoring for fall risk assessment in older adults. In this study, we extracted spatiotemporal gait features and their variability from foot-mounted inertial measurement units (IMUs) using data from the GSTRIDE dataset, which includes 163 older participants. We systematically evaluated the performance of four machine learning models-logistic regression, random forest, support vector machine, and artificial neural network-under three commonly used validation strategies: Hold-out Method, 10-Fold Cross-Validation and Leave-One-Out Cross-Validation. Our results show that the choice of both classifier and validation protocol significantly affects model performance and stability. While the Hold-Out Method yielded higher classification accuracy in some cases (e.g., 94% for artificial neural network using the first 20 strides), it also demonstrated high variability, raising concerns about its reliability in small-sample studies. In contrast, Leave-One-Out Cross-Validation provided more consistent and generalizable outcomes. Additionally, gait features derived from the initial 20 strides outperformed full-sequence features across models, suggesting that early gait dynamics may better reflect fall risk. These findings offer practical insights into model selection, data preprocessing, and validation design for building robust fall risk assessment systems based on wearable-derived gait data.