Towards Cross-Subject EMG Pattern Recognition via Dual-Branch Adversarial Feature Disentanglement.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Inter-subject variability in electromyography (EMG) signals complicates pattern recognition.
• Conventional methods require subject-specific calibration, limiting real-world applicability.
• Developing calibration-free cross-subject EMG pattern recognition is crucial for widespread deployment.

Purpose of the Study:

• To develop an end-to-end deep learning model for calibration-free cross-subject EMG pattern recognition.
• To disentangle EMG features into pattern-specific and subject-specific components.
• To enable robust pattern recognition for new users and facilitate biometric identification.

Main Methods:

• Proposed an end-to-end dual-branch adversarial neural network architecture.
• Implemented feature disentanglement to separate pattern-specific and subject-specific information.
• Evaluated model performance on unseen users in cross-subject scenarios.

Main Results:

• Achieved robust cross-subject EMG pattern recognition without requiring user-specific calibration.
• The proposed model outperformed various baseline methods in generalization to new subjects.
• Successfully demonstrated the utility of disentangled features for both pattern recognition and biometric identification.

Conclusions:

• The developed dual-branch adversarial network offers a novel solution for calibration-free cross-subject EMG pattern recognition.
• Feature disentanglement is an effective strategy for improving generalization across different users.
• The model shows potential for broader applications, including task-invariant biometric systems.