Learning graph in graph convolutional neural networks for robust seizure prediction.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) are effective for epilepsy treatment, relying on accurate seizure prediction algorithms.
• Subtle and patient-specific preictal signal patterns pose challenges for robust seizure prediction.
• Existing Graph Convolutional Neural Network (GCNN) models struggle with seizure prediction due to reliance on general prior graphs.

Purpose of the Study:

• To develop a novel approach for automatically learning patient-specific graphs for improved seizure prediction.
• To address the limitations of current GCNN models in capturing diverse preictal period mechanisms.

Main Methods:

• Proposed a Joint Graph Structure and Representation Learning Network (JGRN) for data-driven, patient-specific graph learning.
• Implemented a global-local graph convolutional neural network within JGRN to jointly learn graph structures and weights.
• Optimized learned graph and feature representations for the specific task of seizure prediction using intracranial electroencephalogram (iEEG) signals.

Main Results:

• The JGRN model significantly outperformed standard Convolutional Neural Network (CNN) and GCNN models.
• Performance improvements were particularly pronounced when dealing with subtle preictal features.
• Experimental results demonstrate the effectiveness of JGRN in enhancing seizure prediction accuracy.

Conclusions:

• The proposed JGRN approach achieves robust seizure prediction performance.
• This method shows potential for broader application in brain-computer interface (BCI) challenges.
• JGRN offers a promising solution for personalized and accurate seizure prediction in epilepsy management.