From data to diagnosis: An innovative approach to epilepsy prediction with CGTNet incorporating spatio-temporal features

Abstract

Epilepsy affects around 50 million people globally, causing significant burdens. While many methods predict seizures, current models struggle with handling spatiotemporal features and balancing accuracy with computational efficiency.This paper introduces a novel deep learning architecture called CGTNet, which is composed of a multi-scale convolutional network, gated recurrent units (GRUs), and Sparse Transformers. It is specifically designed for analyzing elec-troencephalogram (EEG) data to predict epileptic seizures. CGTNet enhances the ability to extract spatiotemporal features from EEG signals, demonstrating its exceptional performance in seizure prediction through rigorous evaluation on the renowned CHB-MIT and SWEC-ETHZ EEG datasets. The model achieved an accuracy of 98.89%, sensitivity of 98.52%, specificity of 98.53%, an AUROC value of 0.97, and an MCC value of 0.975 on these datasets. These results not only highlight the technical innovations of CGTNet but also validate the immense potential of deep learning in processing medical signals. Our research provides an effective new tool for early detection and continuous monitoring of epilepsy, laying the foundation for advancing healthcare with artificial intelligence technology.