Hybrid Attention Network for Epileptic EEG Classification.

Area of Science:

• Medical Technology
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning has advanced automatic seizure detection from electroencephalography (EEG).
• Existing methods often fail to fully utilize spatial-temporal information across EEG channels.
• Cross-patient seizure detection remains a significant challenge compared to patient-specific approaches.

Purpose of the Study:

• To propose a novel Hybrid Attention Network (HAN) for enhanced automatic seizure detection.
• To effectively extract and integrate spatial-temporal features from EEG signals.
• To address the challenges of both patient-specific and cross-patient seizure detection.

Main Methods:

• Developed a Hybrid Attention Network (HAN) combining Graph Attention Network (GAT) for spatial features and Transformer for temporal features.
• Employed an attention mechanism to capture complex spatial-temporal correlations in EEG data.
• Utilized a focal loss function to handle class imbalance inherent in EEG seizure detection datasets.

Main Results:

• The HAN model demonstrated significant efficacy in both patient-specific and patient-independent seizure detection tasks.
• Experimental validation was performed on the public CHB-MIT EEG database.
• The proposed approach successfully addressed limitations in excavating spatial-temporal information.

Conclusions:

• The Hybrid Attention Network (HAN) offers a robust solution for automatic seizure detection from EEG.
• HAN effectively leverages spatial-temporal correlations, improving detection accuracy.
• The model shows promise for real-world clinical applications, including challenging cross-patient scenarios.