Boosting Personalized Musculoskeletal Modeling with Deep Transfer Learning: A Case Study

In recent years, musculoskeletal models have demonstrated significant potential in human-machine interfaces (HMI). However, their computational complexity limits real-time applications. In this paper, a deep transfer learning framework to accelerate personalized musculoskeletal modeling is proposed, enabling rapid extraction of motion features from surface electromyography (sEMG) and precise estimation of joint angles and torques. We exemplify the proposed framework by synchronously estimating the angles and torques of the wrist and metacarpophalangeal (MCP) joints using sEMG. The correlation coefficients between the estimated and ground truth of wrist joint angles and torques both reached 0.97. For MCP joints, they reached 0.95 and 0.92, respectively. Compared to physics-based musculoskeletal models, the proposed method increased the forward inference speed by at least 5 times, demonstrating its effectiveness.