Explainable Deep-Learning Prediction for Brain-Computer Interfaces Supported Lower Extremity Motor Gains Based on...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Rehabilitation Science

Background:

• Predicting motor function recovery in stroke patients is crucial for effective rehabilitation.
• Electroencephalography (EEG) shows promise in linking cortical neural activity to motor recovery.
• Multi-state EEG recordings may offer more accurate predictions than single-state recordings.

Purpose of the Study:

• To develop and evaluate a multi-state fusion neural network for predicting motor recovery in stroke patients undergoing EEG-brain-computer-interface (BCI) rehabilitation.
• To identify key EEG features (power spectral density and functional connectivity) contributing to motor recovery prediction using explainable deep learning.

Main Methods:

• Designed a multi-state fusion neural network combining eyes-closed (EC) and eyes-open (EO) EEG states.
• Applied an explainable deep learning method to analyze EEG power spectral density and functional connectivity.
• Trained and validated the model on EEG data from stroke patients post-BCI rehabilitation.

Main Results:

• The multi-state fusion network achieved 82% prediction accuracy, significantly outperforming single-state models.
• Explainable AI identified critical brain regions (frontal, central, occipital) and frequency bands (delta, alpha, theta, beta) related to motor recovery.
• Specific EEG features like power spectral density in delta/alpha bands and functional connectivity in delta/theta/alpha (EC) and delta/theta/beta (EO) bands were linked to recovery.

Conclusions:

• Multi-state fusion neural networks enhance the accuracy of predicting motor recovery after BCI training in stroke patients.
• This approach reveals underlying neural mechanisms, highlighting specific brain regions and frequency oscillations involved in motor recovery.
• The findings support the use of advanced EEG analysis for personalized stroke rehabilitation strategies.