Classification of fall risk across the lifespan using gait derived features from a wearable device

Falls are one of the leading factors of injury and fatality in older adults. Given the importance of early detection of adults at higher risk of falls, we evaluated the ability of machine learning to classify fall risk in adults across the lifespan using wearable sensors embedded in a smartshirt. We evaluated the classification performance of binary and multiclass fall risk classifier models using SciKit Digital Health in adults across the lifespan. Using a k-fold and group k-fold cross-validation strategy, we demonstrate the feasibility of fall risk classification using accelerometer data from 10 second epochs of treadmill walking data from adults across the lifespan. We achieved an 88% accuracy in a binary clasifier of fallers vs. non-fallers, and an 86% accuracy in a multiclass classifier comparing non-fallers, fallers, and recurrent fallers using retrospective fall histories. Comparing group k-fold vs. k-fold cross-validation strategies, we find a 22-27% drop-off in accuracy performance. Furthering the evaluation framework presented in this study would be valuable to the development of more robust and clinically relevant models used in the prediction of fall risk. These models could one day be applied in clinical settings to help better diagnose and monitor fall risk among older adults, improving the care of at-risk individuals and reducing the injury and associated cost of falls.