ASA-STGCN: Adaptive Sparse Awareness-Spatiotemporal Graph Convolutional Network for Multi-Class Motor Imagery EEG...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Graph Convolutional Networks (GCNs) show potential for electroencephalogram (EEG) signal classification by modeling brain connectivity.
• Over-smoothing in GCNs leads to feature homogenization and reduced classification discrimination.
• Existing methods struggle to effectively capture complex spatial and temporal dynamics in EEG data.

Purpose of the Study:

• To develop an advanced GCN model for improved motor imagery EEG classification.
• To overcome the over-smoothing problem in GCNs for EEG signal analysis.
• To enhance the model's ability to capture spatial and temporal features for better brain-computer interface (BCI) applications.

Main Methods:

• Proposed an Adaptive Sparse Awareness-Spatiotemporal Graph Convolutional Network (ASA-STGCN).
• Incorporated adaptive sparse graph convolution and attention mechanisms.
• Utilized a Graph Sparse Convolutional Network (GSCN) for feature selection and a Graph Node Neighborhood Awareness Layer (GNNAL) for topological relationship reinforcement.
• Employed a Multi-scale Temporal Convolution Module (MTCM) to capture temporal dependencies.

Main Results:

• Achieved high classification accuracies: 97.2%±3.4% (binary) and 83.6%±4.9% (four-class) on BCIC-IV-2a.
• Attained 96.6%±3.1% (binary) accuracy on BCIC-III IVa and 83.41%±4.3 (binary) on OpenBMI.
• Demonstrated the effectiveness of the ASA-STGCN in overcoming GCN over-smoothing and improving EEG classification.

Conclusions:

• The ASA-STGCN model significantly enhances motor imagery EEG classification performance.
• The proposed methods effectively address GCN over-smoothing and improve feature discrimination.
• The model shows promise for neurorehabilitation and clinical brain-computer interface applications.