Graph Attention Networks for Detecting Epilepsy From EEG Signals Using Accessible Hardware in Low-Resource Settings.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Technology

Background:

• Epilepsy diagnosis is challenging in low-income countries due to limited neurologists and high costs of diagnostic tools.
• There is a need for accessible and affordable epilepsy detection methods, especially in resource-limited settings.

Purpose of the Study:

• To develop and evaluate a graph-based deep learning framework for epilepsy detection using low-cost Electroencephalography (EEG) hardware.
• To ensure fair, accessible automatic assessment and provide explainability for epilepsy biomarkers.
• To adapt deep learning models for low-fidelity EEG recordings and enable deployment on low-power devices.

Main Methods:

• Modeled Electroencephalography (EEG) signals as spatio-temporal graphs.
• Utilized graph attention networks (GAT) to classify signals and identify interchannel relationships and temporal dynamics.
• Adapted GAT to analyze graph edges for connectivity biomarkers and developed a lightweight architecture for deployment on Raspberry Pi devices.

Main Results:

• Achieved promising epilepsy classification performance.
• Outperformed standard classifiers like random forest and graph convolutional networks in accuracy and robustness.
• Highlighted specific fronto-temporal region connectivity patterns as potential biomarkers.

Conclusions:

• Graph attention networks (GATs) show potential for insightful and scalable diagnostic support for epilepsy in underserved regions.
• The developed framework can pave the way for affordable and accessible neurodiagnostic tools.
• The approach demonstrates the feasibility of using deep learning with low-cost EEG for epilepsy diagnosis in resource-limited settings.