Continuous assessment of daily-living gait using self-supervised learning of wrist-worn accelerometer data

Abstract Physical activity and mobility are critical for healthy aging and predict diverse health outcomes. While wrist-worn accelerometers are widely used to monitor physical activity, estimating gait metrics from wrist data remains challenging. We extend ElderNet, a self-supervised deep-learning model previously validated for walking-bout detection, to estimate gait metrics from wrist accelerometry. Validation involved 819 older adults (Rush-Memory-and-Aging-Project) and 85 individuals with gait impairments (Mobilise-D), from six medical centers. In Mobilise-D, ElderNet achieved an absolute error of 8.82 cm/s and an intra-class correlation of 0.87 for gait speed, outperforming state-of-the-art methods (p < 0.001) and models using a lower-back sensor. ElderNet outperformed (percentage error; p < 0.01) competing approaches in estimating cadence and stride length, and better (p < 0.01) classified mobility disability (AUC = 0.80) than conventional gait or physical activity metrics. These results demonstrate the potential of ElderNet a scalable tool for gait assessment using wrist-worn devices in aging and clinical populations.