Physics-informed neural network design towards interpretable and efficient joint moment prediction.

Data-driven approaches are increasingly being adopted in human motion analysis research. Joint moment estimation plays a crucial role in providing valuable insights for robotic design and rehabilitation-related applications. Estimating joint moments using surface electromyography (sEMG) offers the advantages of being non-invasive and cost-effective. However, the end-to-end design of deep learning models for joint moment estimation faces several challenges, such as the requirement for large datasets, model generalizability, and interpretability. To address these challenges, this paper presents a physics-informed neural network (PINN)-based framework for knee moment prediction, designed using the sit-to-stand process of fifteen healthy individuals. The proposed model integrates Hill-type muscle dynamics and enforces constraints on model outputs through a physics-informed composite loss function. Using five-channel lower-limb sEMG signals as input, the model predicts knee joint moments and provides additional information on muscle activations and forces, offering biomechanical insights. The primary contribution of this design strategy lies in its capability to predict knee joint moments from sEMG signals while providing physiologically interpretable intermediate outputs. Our model achieved an average coefficient of determination (R²) of 0.919 ± 0.059 and an average root mean square error (RMSE) of 4.438 ± 3.136 Nm in leave-one-subject-out cross-validation across fifteen subjects. This work introduces a PINN design strategy that improves knee moment estimation accuracy while maintaining biomechanical plausibility within the evaluated sit-to-stand task.