Dynamic Multi-Graph Convolution-Based Channel-Weighted Transformer Feature Fusion Network for Epileptic Seizure...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Electroencephalogram (EEG) based seizure prediction is crucial for closed-loop neuromodulation systems.
• Existing graph convolution methods often overlook dynamic graph structures and use suboptimal feature fusion.
• This limits the performance of current seizure prediction models.

Purpose of the Study:

• To propose a novel network, MB-dMGC-CWTFFNet, for superior patient-specific seizure prediction.
• To address limitations in static graph construction and coarse-grained feature fusion in existing methods.
• To enhance the accuracy and reliability of seizure warning systems.

Main Methods:

• Utilized a multi-branch feature extractor to capture temporal, spatial, and spectral EEG representations.
• Developed a dynamic multi-graph convolution network (dMGCN) for learning deep, dynamic graph structures.
• Implemented a channel-weighted transformer feature fusion network (CWTFFNet) with multi-head self-attention for efficient feature fusion.

Main Results:

• The MB-dMGC-CWTFFNet demonstrated outstanding prediction performance on both public (CHB-MIT EEG) and private (sEEG) datasets.
• Achieved superior results compared to existing state-of-the-art seizure prediction methods.
• Validated the effectiveness of dynamic graph learning and channel-weighted fusion strategies.

Conclusions:

• The proposed MB-dMGC-CWTFFNet offers a significant advancement in patient-specific seizure prediction.
• The network effectively models multi-domain dynamic changes in EEG data for enhanced prediction accuracy.
• This approach provides a promising tool for developing reliable, real-time seizure warning systems.