Hybrid Physics-AI Models Estimate Body Center of Mass Dynamics from Wrist Sensors

Summary

Wrist-worn inertial measurement units (IMUs) are convenient for health monitoring but do not directly capture whole-body dynamics. This study proposes a simplified kinematic model that maps wrist IMU measurements to body center of mass (COM) acceleration under assumptions that make the dynamic equations solvable from wrist data alone. It also introduces three human kinematic model-based neural network (HKM-NN) approaches: serial learning and two forms of simultaneous learning, combining biomechanical structure with data-driven modeling. The models were trained and tested on measurements from 10 healthy volunteers performing six gait activities and a sit-to-stand transition, with ground-truth COM measurements available for comparison. The kinematic model produced errors of 6.7% to 12.5% for gait activities and 5.6% for sit-to-stand. HKM-NN methods reduced gait errors to 5.3% to 9.3%, with the best sit-to-stand error reaching 3.9%. Their noise behavior differed: the simultaneous-learning methods generally retained more robustness under Gaussian perturbations, while the kinematic model performed comparatively well under salt-and-pepper noise. The findings support combining biomechanical models with neural networks for wearable sensing when measurements are imperfect.