Deep Learning Based Post-stroke Myoelectric Gesture Recognition: From Feature Construction to Network Design.

Area of Science:

• Biomedical Engineering
• Neurorehabilitation
• Machine Learning

Background:

• Robot-assisted rehabilitation enhances stroke patient training intensity and reduces therapist workload.
• Surface electromyography (sEMG) is a potential control source for assistive technologies.
• Accurate hand gesture recognition is crucial for effective post-stroke motor recovery.

Purpose of the Study:

• To investigate the potential of deep learning (DL) for post-stroke hand gesture recognition using sEMG signals.
• To evaluate different sEMG feature domains, data structures, and neural network architectures.
• To assess the impact of post-processing algorithms on recognition accuracy.

Main Methods:

• Collected sEMG signals from eight chronic stroke subjects.
• Evaluated 18 DL models (CNN, CNN-LSTM, CNN-LSTM-Attention) using time, frequency, and wavelet features in 1D and 2D formats.
• Performed intra-subject and inter-subject transfer learning tasks.
• Analyzed Model Voting and Bayesian Fusion post-processing algorithms.

Main Results:

• For intra-subject testing, CNN-LSTM with 2D frequency features achieved the highest accuracy (72.95%).
• For inter-subject transfer learning, CNN-LSTM-Attention with 1D frequency features yielded the highest accuracy (68.38%).
• Frequency features demonstrated a significant advantage over time and wavelet features.
• Post-processing, particularly Model Voting, improved accuracy by up to 2.03%.

Conclusions:

• Deep learning models, especially CNN-LSTM and CNN-LSTM-Attention, are effective for sEMG-based post-stroke hand gesture recognition.
• Frequency domain features are optimal for this application.
• Post-processing algorithms can further enhance the performance of DL models in stroke rehabilitation.