Efficient Multi-View Graph Convolutional Network with Self-Attention for Multi-Class Motor Imagery Decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram-based motor imagery (MI-EEG) decoding is crucial for brain-computer interfaces (BCI).
• Current deep learning methods struggle to fully utilize topological brain region information, limiting classification performance.

Purpose of the Study:

• To propose a novel Multi-View Graph Convolutional Attention Network (MGCANet) with a residual learning structure for enhanced multi-class MI decoding.
• To improve the classification accuracy of MI-EEG signals by leveraging brain region topology and adaptive feature fusion.

Main Methods:

• Developed a multi-view graph convolution spatial feature extraction method utilizing brain region topological relationships.
• Implemented an adaptive weight fusion (Awf) module to merge features from different brain views.
• Incorporated a self-attention mechanism for feature selection to capture global dependencies in EEG signals.

Main Results:

• The proposed MGCANet achieved a mean accuracy of 78.26% on the BCIC IV 2a dataset and 73.68% on the OpenBMI dataset.
• Demonstrated significantly superior classification performance compared to existing representative methods.
• Verified the effectiveness of the multi-view graph convolution, adaptive weight fusion, and self-attention mechanisms.

Conclusions:

• The MGCANet model offers a significant advancement in MI-EEG decoding for BCI applications.
• The proposed approach effectively utilizes topological brain information and adaptive feature fusion for improved accuracy.
• This research provides novel perspectives and a robust framework for future MI decoding studies.