Single-Camera Knee Adduction Moment Estimation for Individuals With Knee Osteoarthritis via a Novel Spatio-Temporal Graph Transformer Network.

Accurately estimating knee adduction moment (KAM) is essential for understanding knee biomechanics and may aid in mitigating medial knee osteoarthritis (OA) progression. However, traditional motion capture and force plate systems are costly and restricted to laboratories. Multi-wearable sensor and multi-camera setups have been explored but remain complex and lack validation across diverse gait modifications in individuals with knee OA. This study proposes a novel Spatio-Temporal Graph Transformer Network (STGTN) for estimating KAM using a single-camera setup. Fourteen individuals with medial compartment knee OA performed gait modifications, including variations in walking speed, foot progression angle, step width, trunk sway, and dual-task walking. The proposed model achieved a KAM root mean square error of 0.48% BW $\cdot $ BH and a peak KAM mean absolute error of 0.43% BW $\cdot $ BH, both within clinically meaningful error thresholds (0.5-2.1% BW $\cdot $ BH). Additionally, the model demonstrated sensitivity to gait modifications, identifying significant reductions in peak KAM during slow walking, toe-in gait, wide step width, and increased trunk sway conditions ( ${p} \lt 0.05$ ). These findings suggest that single-camera-based KAM estimation could be a feasible approach for real-world gait assessment, with potential applications in clinical and rehabilitation settings for monitoring knee joint loading and exploring feedback-driven KAM reduction.