DynSeizureGAT: Multi-Band Dynamic Graph Attention Network for Interpretable Seizure Detection and Analysis of...

Area of Science:

• Neuroscience
• Machine Learning
• Epilepsy Research

Background:

• Drug-Resistant Epilepsy (DRE) seizure detection is challenging due to dynamic epileptic discharge propagation.
• Traditional methods struggle with comprehensive brain network feature representation and dynamic learning.
• Existing models lack interpretability concerning seizure mechanisms.

Purpose of the Study:

• To propose a novel multi-band dynamic graph attention network (DynSeizureGAT) for precise and interpretable DRE seizure detection and analysis.
• To address limitations in representing evolving brain network features and model interpretability.
• To develop a model that aligns with seizure propagation mechanisms.

Main Methods:

• Constructing a seizure network sequence using multi-band directed transfer function matrices and epileptic index node features.
• Integrating a dynamic graph attention module for adaptive weighting of spatial scales.
• Employing spatial-spectral-temporal attention mechanisms for enhanced ictal and interictal state characterization.

Main Results:

• Achieved high seizure detection performance on a public Stereotactic Electroencephalography (SEEG) dataset (OpenNeuro): 94.6% accuracy, 93.4% sensitivity, 96.4% specificity.
• Successfully quantified and visualized the importance of frequency bands and dynamic abnormal connectivity patterns.
• Demonstrated strong dynamic propagation feature learning capabilities aligned with seizure mechanisms.

Conclusions:

• DynSeizureGAT offers a promising approach for precise and interpretable DRE seizure detection.
• The model's interpretability aids in understanding seizure propagation and potentially localizing the epileptogenic zone.
• This method enhances the analysis of dynamic brain network features in epilepsy.