Epileptic Seizure Prediction Using Spatiotemporal Feature Fusion on EEG.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is vital for epilepsy analysis.
• Current Convolutional Neural Network (CNN) methods struggle to capture multi-channel EEG's spatiotemporal features for seizure prediction.
• Effective preictal state identification is challenging.

Purpose of the Study:

• To propose an end-to-end model for epileptic seizure prediction.
• To extract spatial relationships and temporal correlations from multi-channel EEGs.
• To improve the accuracy of seizure prediction by integrating Graph Attention Network (GAT) and Temporal Convolutional Network (TCN).

Main Methods:

• An end-to-end model combining GAT and TCN was developed.
• Low-pass filtered EEG signals were processed through GAT for spatial feature extraction.
• TCN was employed to capture temporal features, enabling spatiotemporal correlation acquisition.
• The model was evaluated on the CHB-MIT database.

Main Results:

• Segment-based accuracy: 98.71%
• Segment-based specificity: 98.35%
• Segment-based sensitivity: 99.07%
• Segment-based F1-score: 98.71%
• Event-based sensitivity: 97.03%
• False Positive Rate (FPR): 0.03/h

Conclusions:

• The proposed GAT-TCN model achieves superior performance in seizure prediction.
• Fusion of EEG spatiotemporal features enhances prediction accuracy.
• The model eliminates the need for manual feature engineering.