A Causal Learning-Based sEMG Disentanglement Framework for Multi-Posture Domain Generalization.

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Surface electromyography (sEMG) systems show high accuracy in controlled settings but struggle with real-world posture variations.
• Personalized biases from posture changes degrade the performance of sEMG-based human-computer interaction (HCI) models.
• Existing sEMG datasets often lack sufficient posture variation samples.

Purpose of the Study:

• To develop a robust sEMG-based HCI model that generalizes well across different user postures.
• To disentangle posture-invariant pattern components from sEMG signals for improved recognition.
• To address the challenge of posture variability in real-world sEMG applications.

Main Methods:

• Treated sEMG signals as a combination of pattern and posture components using causal encoders.
• Disentangled these components into separate latent spaces for analysis.
• Developed a high-density sEMG (HD-sEMG) dataset with 16 subjects across four common HCI postures.
• Trained a robust pattern recognition model leveraging posture-invariant features.

Main Results:

• Achieved an average accuracy of 90.3% across four generalization tasks.
• Demonstrated superior performance compared to existing domain generalization models.
• Successfully extracted posture-invariant pattern components from sEMG signals.

Conclusions:

• The proposed method effectively improves the generalization capabilities of sEMG-based HCI systems.
• The approach mitigates performance degradation caused by user posture variations.
• This work provides a foundation for more reliable and robust real-world sEMG applications.