Enhancing Epileptic Seizure Prediction with Advanced Spatio-Temporal-Spectral Architectures and SGD Optimization

The Advanced Spatio-Temporal-Spectral Optimization Network (ASTS-OptNet), a novel framework, significantly enhances epileptic seizure prediction by conducting an in-depth analysis of the preictal phase in electroencephalography (EEG). Traditional approaches in this domain often face challenges in predictive accuracy due to limited analysis of EEG's comprehensive dimensions and reliance on intra-subject training. ASTS-OptNet overcomes these limitations by integrating a multi-aspect attention layer that adjusts dimension weighting, thereby improving the representation of EEG data across spatial, temporal, and spectral domains. Additionally, the incorporation of SpatioTemporal Graph Convolution (STGC) optimizes the spatial relationship modeling between EEG electrodes. The network's efficiency is further enhanced by employing a stochastic gradient descent (SGD) optimizer, facilitating effective training across diverse subjects. ASTS-OptNet's innovative architecture has demonstrated exceptional performance, achieving a 97.21% sensitivity rate and significantly reducing false predictions to 0.068 per hour on the CHB-MIT scalp EEG dataset. This marks a substantial advancement over existing seizure prediction models, positioning ASTS-OptNet as a leading-edge solution in the realm of EEG-based seizure prognostication and a significant contribution to clinical epilepsy care.