An EEG-Based Seizure Prediction Model Encoding Brain Network Temporal Dynamics.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Informatics

Background:

• Epileptic seizure prediction using electroencephalography (EEG) is challenged by complex temporal dynamics of brain networks.
• Metastability, characterized by recurring neural activity patterns, offers a promising avenue for understanding pre-seizure brain states.

Purpose of the Study:

• To develop a patient-independent seizure prediction model by integrating physiological priors of brain network dynamics into deep learning.
• To fuse consistent epileptic network processes across subjects into a shared latent space for enhanced prediction.

Main Methods:

• Constructing metastable transition patterns to identify recurrent network states.
• Employing adversarial feature learning and variational autoencoder (VAE) for latent space embedding.
• Utilizing Maximum Mean Discrepancy (MMD) to reduce patient-specific variations.

Main Results:

• The proposed model achieved improved Area Under the Curve (AUC), sensitivity, and specificity on the CHB-MIT dataset compared to existing methods.
• Demonstrated enhanced performance in patient-independent seizure prediction by leveraging fused network dynamics.
• Successfully integrated brain network-based physiological priors with deep learning for EEG representation.

Conclusions:

• Combining brain network metastability with deep learning offers a novel strategy for EEG-based seizure prediction.
• The developed method enables reliable patient-independent seizure forecasting by capturing complex brain network variations.
• This approach advances the field of neurological disorder prediction through advanced computational techniques.