[A model based on the graph attention network for epileptic seizure anomaly detection].

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Context:

• Epilepsy seizure detection algorithms often suffer from overfitting and poor generalization.
• Manual labeling of electroencephalogram (EEG) data and data imbalance present significant challenges.
• Existing methods struggle with the complex temporal and spatial patterns of epileptic seizures.

Purpose:

• To propose an unsupervised learning method for epileptic seizure detection using a joint Graph Attention Network (GAT) and Transformer framework (GAT-T).
• To address limitations of existing algorithms, including reliance on manual labeling and data imbalance.
• To develop a method that adaptively learns channel correlations and captures temporal information for improved EEG analysis.

Summary:

• The proposed GAT-T method integrates a GAT encoder for adaptive channel correlation learning and a 1D convolution decoder for temporal information capture.
• Predicted EEG values are generated by combining encoder and decoder outputs, leading to a collective anomaly score for seizure detection.
• The system determines a detection threshold, demonstrating effective epileptic seizure detection capabilities.

Impact:

• GAT-T achieved average performance exceeding 90% (or 99%) with 0.25s (or 2s) time segments, effectively detecting epileptic seizures.
• The channel association probability matrix can aid clinicians in initial screening of the epileptogenic zone.
• Ablation experiments confirmed the significance of individual modules within the GAT-T framework, supporting its clinical utility.