An Effective Semi-Subject-Independent sEMG-Based Learning Framework to Continuously Predict Knee Joint Trajectory

Predicting knee joint trajectory is critical for controlling intelligent walking-assistive devices, with surface electromyography (sEMG) emerging as a promising modality for motion intention decoding. However, accurate and continuous prediction remains challenging because both intersubject and intrasubject variability must be addressed simultaneously. To tackle this problem, this article proposes a semi-subject-independent deep learning framework that is pretrained on source subjects to learn shared cross-subject representations and then calibrated with only a few trials from an unseen target subject before testing on that subject's held-out trials. The framework contains two complementary components. First, gait kinematic decoupling (GKD) separates knee trajectory prediction into a shared normalized motion pattern and subject-dependent amplitude and offset terms, thereby reducing cross-subject label variability. Second, muscle activation filtering uses physiological activation priors to suppress motion-irrelevant sEMG components and enhance gait-related neuromuscular information. Experiments on both in-house and public datasets show state-of-the-art performance, with average root-mean-square errors (RMSEs) of $3.03^{\circ }~\pm ~0.49^{\circ }$ and $4.49^{\circ }~\pm ~1.14^{\circ }$ , respectively, while predicting knee angles 50 ms in advance. These results suggest that the proposed framework can support robust and practical control of intelligent walking-assistive systems.