Enhancing Motor Imagery Classification with Residual Graph Convolutional Networks and Multi-Feature Fusion.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Stroke rehabilitation increasingly utilizes motor imagery (MI)-based brain-computer interface (BCI) systems.
• Analyzing electroencephalogram (EEG) signals from stroke patients presents significant challenges in accuracy and efficiency.

Purpose of the Study:

• To develop an advanced framework for improved EEG classification in MI-based BCI for stroke patients.
• To enhance the accuracy and efficiency of EEG signal analysis for stroke rehabilitation.

Main Methods:

• Proposed a novel M-ResGCN framework integrating modified S-transform (MST) for time-frequency feature extraction and self-attention into a residual graph convolutional network (ResGCN).
• Derived spatial EEG features using the absolute Pearson correlation coefficient (aPcc) to construct a brain network's adjacency matrix, reflecting channel connectivity.
• Applied the framework to EEG data from 16 stroke patients and 16 healthy subjects.

Main Results:

• Achieved the highest classification accuracy of 94.91% with a Kappa coefficient of 0.8918.
• Demonstrated significant improvements in classification quality and robustness across tests and subjects.
• 10x10-fold cross-validation yielded average accuracy of 94.38% and F1 scores of 94.36%.

Conclusions:

• The proposed M-ResGCN framework effectively enhances EEG signal analysis and feature encoding for MI-based BCI.
• Brain networks constructed using aPcc accurately reflect overall brain activity, validating its utility in EEG analysis.
• The method offers a promising approach for real-time applications in stroke rehabilitation BCI systems.