A Novel Bilateral Data Fusion Approach for EMG-Driven Deep Learning in Post-Stroke Paretic Gesture Recognition

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Stroke survivors often experience motor impairments, necessitating effective neurorehabilitation tools.
• Electromyography (EMG) signals offer a viable method for detecting residual motor function.
• Deep learning models show promise for analyzing complex biological signals like EMG.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model for hand gesture recognition using EMG signals in subacute stroke patients.
• To assess the model's performance over time and across different numbers of gesture classes.
• To introduce and validate a novel bilateral data fusion technique to improve model performance with limited data.

Main Methods:

• A one-dimensional convolutional long short-term memory (CNN-LSTM) neural network was designed for EMG signal processing.
• Data from 25 subacute stroke patients were collected twice, forming datasets A and B.
• A bilateral data fusion approach incorporating non-paretic limb EMG signals was implemented.

Main Results:

• The CNN-LSTM model achieved classification accuracies between 81.69% and 88.36% across datasets and gesture classes.
• Bilateral data fusion significantly improved sensitivity, specificity, accuracy, and F1-scores.
• Classification accuracy for a three-gesture subset increased by over 5% with data fusion.

Conclusions:

• The hybrid CNN-LSTM model with bilateral data fusion demonstrates effective hand gesture recognition in stroke patients.
• This approach offers a promising, feature-engineering-free method for neurorehabilitation applications.
• The findings highlight the potential of advanced machine learning for personalized stroke recovery.