Hand Gesture Intention Detection Using sEMG and Transfer Learning in Stroke Survivors

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning

Background:

• Surface electromyography (sEMG) offers intuitive control for assistive devices in motor-impaired individuals.
• Accurate sEMG classification is challenging in stroke patients due to altered neuromuscular patterns, especially with severe impairments.

Purpose of the Study:

• To investigate the efficacy of transfer learning for classifying sEMG signals in stroke patients with varying impairment levels.
• To enhance the performance of wearable assistive technology control systems for individuals post-stroke.

Main Methods:

• A transfer learning approach was employed, pretraining a model on healthy sEMG data (Ninapro DB5) and adapting it to stroke patient data.
• Ten stroke patients performed hand gestures across various arm postures using a Myo armband.
• Multiple transfer learning configurations were tested by adjusting frozen layers.

Main Results:

• The best transfer learning model achieved 93.6% overall accuracy across all stroke participants.
• Specific accuracies included 84.32% for severe, 97.42% for moderate, and 99.07% for low impairments.
• Training solely on stroke data yielded a significantly lower accuracy of 46%.

Conclusions:

• Transfer learning significantly enhances sEMG pattern recognition in stroke patients, even with limited individual data.
• Leveraging healthy sEMG data is crucial for developing robust and adaptive control systems for post-stroke rehabilitation.
• This methodology supports inclusive wearable technology design for diverse motor impairment levels.