Self-Recalibrating Surface EMG Pattern Recognition for Neuroprosthesis Control Based on Convolutional Neural Network.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Rehabilitation Technology

Background:

• Surface electromyography (sEMG) pattern recognition is crucial for neuroprosthetic control of upper limbs.
• Real-world prosthetic performance degrades due to the non-stationary nature of sEMG signals.
• Existing systems often require manual retraining, limiting long-term usability.

Purpose of the Study:

• To develop a self-recalibrating classifier for robust sEMG-based hand movement classification.
• To maintain stable neuroprosthetic control performance over time without user intervention.
• To improve the long-term adoption and effectiveness of upper limb prosthetics.

Main Methods:

• Utilized a convolutional neural network (CNN) with dimension-reduced sEMG spectrograms as input.
• Implemented a self-recalibration mechanism that automatically updates the classifier using recent prediction data.
• Evaluated the system on the NinaPro database, including data from intact and amputee subjects.

Main Results:

• The self-recalibrating CNN classifier demonstrated significant accuracy improvements: ~10.18% for intact subjects (50 movements) and ~2.99% for amputee subjects (10 movements).
• Achieved higher absolute performance and greater accuracy gains compared to an unrecalibrated classifier over five testing sessions.
• Outperformed support vector machine (SVM) classifiers in terms of absolute performance, improvement, and training efficiency.

Conclusions:

• The proposed self-recalibrating CNN system effectively addresses sEMG signal non-stationarity for improved neuroprosthetic control.
• Automatic recalibration enhances classification accuracy and stability, reducing the need for user retraining.
• This technology shows promise for facilitating the long-term adoption of advanced prosthetic devices for amputees.