MSK-NN estimates joint angles from partial muscle signals with physiological accuracy

A team led by Wending Heng and colleagues at the University of Manchester, working with Glen Cooper and Zhenhong Li, has developed a novel neural network called MSK-NN (Musculoskeletal Neural Network) to solve a persistent challenge in biomedical signal processing: estimating joint kinematics from incomplete surface electromyography (sEMG) data. In real-world scenarios, many relevant muscles are inaccessible due to anatomical constraints or sensor limitations. MSK-NN tackles this by jointly estimating multi-degree-of-freedom (DoF) joint angles and the activations of both measured and unmeasured muscles.

The architecture is fully differentiable: a CNN-based module predicts muscle activations, which feed into an embedded musculoskeletal forward dynamics module that computes joint angles. Unlike prior hybrid AI/biomechanical models, MSK-NN requires no ground-truth biomechanical labels like muscle-tendon forces or joint torques. Instead, it uses a composite loss combining kinematics error, a data-driven muscle synergy loss (to enforce coordinated activation patterns), and an anatomy-guided trend loss (to keep estimated activations physiologically realistic). Tested on two-DoF wrist motions—including repetitive cyclic movements and random, unconstrained paths—MSK-NN outperformed standard deep learning baselines (CNN, Bi-LSTM, CNN-LSTM, PET) in both normalized root mean square error and coefficient of determination. Crucially, optimized muscle parameters remained within physiological bounds, and the estimated activation of a deliberately excluded muscle closely matched its recorded sEMG envelope, proving the model's ability to recover hidden physiological signals.