HD-sEMG Feature Decomposition via Muscle Synergy and Dissimilarity Metric Learning for Robustness Against Unknown...

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Human-Computer Interaction

Background:

• Traditional myoelectric pattern recognition (MPR) systems struggle with recognizing novel gestures, leading to performance degradation.
• Existing MPR methods often fail to reject unknown gestures, limiting their real-world applicability.
• Robustness in myoelectric control is crucial for advanced prosthetic limb functionality.

Purpose of the Study:

• To develop a robust MPR framework that enhances intra-class compactness and improves open-set rejection.
• To enable simultaneous classification of known gestures and rejection of unknown gestures.
• To improve the overall reliability and performance of myoelectric interfaces.

Main Methods:

• Decomposition of high-density surface electromyography (HD-sEMG) time-domain features into pattern-specific and pattern-variant components.
• Integration of dissimilarity metric learning with classification to create a unified model.
• Definition of pattern-specific decision boundaries based on anomaly scores for gesture classification and rejection.

Main Results:

• Achieved over 99% classification accuracy for known gestures and >98% rejection accuracy for unknown gestures on intra-session experiments.
• Maintained over 77% open-set recognition accuracy under challenging inter-session conditions.
• Demonstrated substantial performance improvement over existing open-set MPR methods.

Conclusions:

• The proposed MPR framework effectively enhances intra-class compactness and open-set rejection performance.
• Combining muscle synergy decomposition with dissimilarity metric learning significantly improves the robustness of myoelectric interfaces.
• This approach offers a promising solution for more reliable and versatile prosthetic control.