Cross-Subject Lifelong Learning for Continuous Estimation From Surface Electromyographic Signal

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyographic (sEMG) signals offer a non-invasive method for hand kinematics estimation in human-machine interfaces.
• Existing subject-specific models lack broad applicability, while current cross-subject methods struggle with both new and existing users.
• Challenges include limited generalization and effective user adaptation in cross-subject sEMG analysis.

Purpose of the Study:

• To develop a novel method for robust cross-subject estimation of hand kinematics from sEMG signals.
• To enhance the generalization capabilities of sEMG-based models across diverse users and time scales.
• To address the limitations of current subject-specific and cross-subject approaches in human-machine interfaces.

Main Methods:

• Introduction of the Cross-Subject Lifelong Network (CSLN), a novel lifelong learning approach.
• Maintaining sEMG signal patterns across a varied user population and different temporal scales.
• Utilizing joint and sequential training strategies to evaluate model performance.

Main Results:

• CSLN demonstrates enhanced performance in cross-subject scenarios for sEMG-based hand kinematics estimation.
• The proposed method effectively mitigates catastrophic forgetting during lifelong learning.
• Improved generalization of acquired sEMG signal patterns across individuals and temporal contexts.

Conclusions:

• CSLN offers a significant advancement in developing broadly applicable and adaptable human-machine interfaces.
• Lifelong learning is a viable strategy for improving the efficacy and robustness of cross-subject sEMG models.
• The CSLN model enhances training efficacy and broadens the potential applications of sEMG in human-machine interaction.