Addressing Multi-Day Generalizability for EMG-Based Gesture Recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Robotics

Background:

• Myoelectric prostheses rely on electromyographic (EMG) signals for control.
• Inter-day performance degradation is a major challenge due to muscle fatigue and electrode shift.
• Current deep learning methods often necessitate extensive daily retraining, limiting practical application.

Purpose of the Study:

• To develop a novel transfer learning framework for myoelectric prostheses.
• To address the issue of significant inter-day performance degradation.
• To reduce the need for extensive daily retraining and improve long-term stability.

Main Methods:

• Utilized a pre-trained convolutional neural network (CNN) for robust feature extraction.
• Employed a fine-tuned linear discriminator for efficient daily adaptation.
• Evaluated the framework on the Kanoga dataset.

Main Results:

• Demonstrated superior inter-day generalizability compared to state-of-the-art methods.
• Achieved enhanced long-term stability in prosthesis control.
• Required substantially less daily calibration data for adaptation.

Conclusions:

• The proposed transfer learning framework effectively mitigates inter-day performance degradation in myoelectric prostheses.
• This approach offers a more practical and user-friendly solution for real-world applications.
• The method shows significant promise for improving the reliability and usability of advanced prosthetic devices.