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

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Predicting knee joint trajectory is crucial for controlling intelligent walking-assistive devices.
• Surface electromyography (sEMG) shows promise for decoding motion intentions.
• Intersubject and intrasubject variability pose significant challenges for accurate prediction.

Purpose of the Study:

• To develop a semi-subject-independent deep learning framework for robust knee joint trajectory prediction.
• To address intersubject variability by learning shared cross-subject representations.
• To enable precise control of walking-assistive systems through advanced motion intention decoding.

Main Methods:

• A semi-subject-independent deep learning framework involving pretraining on source subjects.
• Gait kinematic decoupling (GKD) to separate shared motion patterns from subject-specific terms.
• Muscle activation filtering to enhance gait-related neuromuscular information from sEMG signals.

Main Results:

• Achieved state-of-the-art performance on in-house and public datasets.
• Reported average root-mean-square errors (RMSEs) of 3.03° ± 0.49° and 4.49° ± 1.14°.
• Successfully predicted knee angles 50 ms in advance with high accuracy.

Conclusions:

• The proposed framework effectively reduces cross-subject variability in knee trajectory prediction.
• The method enhances the robustness and practicality of intelligent walking-assistive systems.
• This approach supports reliable control by accurately decoding user motion intentions from sEMG.