Ground Reaction Force Estimation From a Single Wrist-Worn IMU Using Biomechanics-Informed Deep Learning

Long-term gait monitoring in daily environments may support early detection and tracking of sarcopenia-related functional decline. Anterior-posterior propulsive ground reaction force (GRF) reflects push-off function and plantarflexor contribution during walking, but GRF measurement remains largely confined to laboratory force plates. Although wrist-worn devices are widely used, wrist inertial measurement unit (IMU) signals contain both arm-swing kinematics and whole-body dynamics, making GRF estimation from a single distal sensor challenging. This study proposes a Biomechanics-Informed Neural Network (BINN) for estimating walking GRF from a single wrist IMU. A sagittal-plane three-link upper-body kinematics model was used to derive center of mass (CoM)-related acceleration from wrist IMU signals. This information was integrated with the wrist IMU signal through a dual-pathway architecture, in which model-derived and data-driven features were processed separately and combined via Cross-Attention. Leave-one-subject-out cross-validation on 19 healthy young men across three walking speeds showed accurate GRF waveform estimation ($r$ = 0.95 $\pm$ 0.02, NRMSE = 8.2 $\pm$ 1.7%). Under unseen walking-speed conditions, BINN retained high waveform accuracy ($r$ = 0.91 $\pm$ 0.04, NRMSE = 12.2 $\pm$ 2.1%). BINN also outperformed CNN and Transformer baselines under limited training data and distribution shift, while using 60-70% fewer parameters. Ablation analysis showed that architecture-level integration of CoM-related information, rather than simple feature concatenation, was the primary contributor to performance. These findings suggest that biomechanics-informed integration of wrist-derived CoM-related information enables accurate, data-efficient, and lightweight GRF estimation from a single wrist IMU, supporting future wearable systems for long-term gait-function monitoring and sarcopenia-related functional assessment.