Prediction of ground reaction force waveforms using consumer-grade wearable devices and a convolutional neural network

Yang Xiao, Li J, Haijun Dong

Estimating ground reaction forces (GRFs) with consumer-grade wearables could support accessible biomechanical monitoring outside laboratory settings. This study examined whether Apple Watch inertial measurement unit (IMU) signals could predict three-dimensional GRF waveforms during walking, jogging, and running and compared performance across force directions, tasks, and sensor configurations. A total of 227 stance-phase trials from 10 healthy adults were analyzed. Wrist-only, waist-only, and combined wrist–waist IMU signals were used as inputs to an activity-informed residual-learning convolutional neural network, evaluated using leave-one-subject-out cross-validation. Performance was assessed using Pearson’s correlation coefficient (r) and relative root mean square error (relRMSE). Prediction was strongest for vertical GRF ( r = 0.92–0.93; relRMSE = 9.87–10.47%), followed by anteroposterior GRF ( r = 0.87–0.88; relRMSE = 10.92–11.36%), whereas mediolateral GRF was poorly predicted ( r = 0.01–0.04; relRMSE = 22.21–24.64%). Performance for vertical and anteroposterior GRFs was greater during jogging and running than walking, while no overall difference was found among sensor configurations. A single wrist- or waist-worn device may therefore support estimation of vertical and anteroposterior GRF waveforms during periodic locomotion, although mediolateral prediction and external validation remain important challenges.