Physiologically Constrained Musculoskeletal Neural Network for Multi-DoF Joint Kinematics Estimation from Partially Observed sEMG

This study examines how to estimate kinematics of multi-degrees of freedom (DoF) joints with partially observable surface electromyography (sEMG). The difficulty lies in the limited measurement to only a subset of task-relevant muscles due to either anatomical limitations or sensor constraints. To tackle this problem, we propose the musculoskeletal neural network (MSK-NN) architecture that aims to estimate multi-DoF joint angles while simultaneously inferring activations for both the measured and unmeasured muscles. The MSK-NN architecture consists of a CNN-based muscle activation estimator and an embedded musculoskeletal forward dynamics module, making it fully differentiable. Unlike other hybrid neural frameworks which Artificial Neural Networks (ANNs) rely on to provide additional biomechanical labels, such

as muscle-tendon forces or joint torques, MSK-NN does not require supervision with respect to those internal biomechanical variables. We pose the first physics-physiology composite loss function, which consists of a joint kinematics loss, a datadriven muscle synergy loss, and an anatomy-guided trend loss. We validate our approach by estimating two-DoF wrist kinematics during rhythmic and random motion. Our approach, MSK-NN, outperformed the baselines (CNN, Bi-LSTM, CNN-LSTM, and PET) with respect to normalized root mean square error (NRMSE) and coefficient of determination (R2), especially in random motion tasks.

Message: Hi FE, I need help with “Physiologically Constrained Musculoskeletal Neural Network for Multi-DoF Joint Kinematics Estimation from Partially Observed sEMG” in “Biotechnology & Biomedical Engineering”