Graph-based Brain Connectivity Analysis for Seizure On-set Prediction

Abstract

Epilepsy, a neurological disorder characterized by recurrent seizures, affects millions globally. Accurate prediction of seizures, particularly in the preictal state, can significantly enhance patient care by enabling timely interventions. This project aims to develop a robust, real-time seizure prediction system using Graph Neural Networks (GNNs) and Brain Network Transformers (BNTs) to analyze brain connectivity from electroencephalogram (EEG) data. Unlike traditional methods that overlook the dynamic connectivity between brain regions, our graph-based approach models these intricate relationships, offering superior accuracy and explainability critical for clinical applications. The proposed system processes raw EEG signals into graph representations, where nodes correspond to brain regions and edges reflect functional connectivity. These graphs are analyzed using GNNs and BNTs, leveraging their ability to capture both local and global patterns in brain connectivity. The model is optimized for deployment on GPU-enabled Nvidia Jetson devices, ensuring real-time inference on live EEG data. Additionally, a mobile application, developed using React Native Expo, will visualize brain activity graphs and provide timely notifications for pre-ictal states, enhancing accessibility and usability for caregivers and medical professionals. This project employs open-source EEG datasets [1] for training and validation, emphasizing portability, cost-effectiveness, and scalability. By integrating real-time data acquisition, graph-based analysis, and user-friendly mobile interfaces, the system has the potential to transform epilepsy management, particularly in resource-constrained settings. Through its focus on brain connectivity and cutting-edge graph-based techniques, this solution represents a significant advancement in seizure prediction technology.