User-Independent EMG Gesture Recognition Method Based on Adaptive Learning.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Surface electromyogram (sEMG) signals exhibit high individual variability, limiting the generalizability of gesture recognition models across different users.
• Existing sEMG-based gesture recognition systems often require extensive user-specific training or pre-experimentation, hindering widespread adoption.
• Muscle synergy, reflecting neural control of movement, offers a robust feature for characterizing gestures.

Purpose of the Study:

• To develop an adaptive learning method for cross-user sEMG gesture recognition.
• To eliminate the need for pre-experimentation for new users in sEMG-based systems.
• To enhance the performance and adaptability of gesture recognition models using muscle synergy features.

Main Methods:

• An adaptive learning strategy was proposed, utilizing muscle synergy as the feature vector.
• An adaptive K-nearest neighbor (KNN) algorithm was employed to obtain labels for new users.
• A risk evaluator assessed new user data to update the training set and KNN weights, facilitating adaptation.

Main Results:

• The adaptive learning method achieved average recognition accuracies of 68.04%, 73.35%, and 83.05% on Ninapro databases (DB1 and DB5).
• The system demonstrated user-dependent performance without requiring re-training for new users.
• Performance improved with increased usage frequency, indicating effective adaptation.

Conclusions:

• The proposed adaptive learning method successfully addresses the challenge of individual variability in sEMG signals for gesture recognition.
• The system facilitates widespread implementation of sEMG control by avoiding re-training and improving performance over time.
• This approach enhances the practicality and efficiency of gesture recognition systems for diverse users.