Novel deep learning framework with hierarchical temporal feature refinement for enhanced epileptic seizure detection

Abstract

Epileptic seizure detection remains a significant challenge in the field of neurology due to the complexity and variability of electroencephalogram (EEG) signals. Traditional manual methods of analyzing EEG data are time-consuming, subjective, and prone to misdiagnosis, particularly in long-term recordings. To address these challenges, this study proposes a novel hierarchical temporal feature refinement network for epileptic seizure detection (HTFRNet-ESD) framework designed to enhance the accuracy and efficiency of epileptic seizure detection. The methodology comprises three key components: a spatial-temporal signal encoder (STSE) for capturing spatio-temporal attributes, an adaptive temporal dynamics encoder (ATDE) for detailed temporal feature analysis, and a hierarchical temporal refinement unit (HTRU) module for integrating frequency-driven power features. The proposed model leverages advanced deep learning techniques, including convolutional neural networks (CNNs) and transformers, to effectively learn both coarse and fine-grained temporal dynamics of EEG signals. Evaluation on publicly available EEG datasets demonstrates the model’s superior sensitivity, specificity, and overall accuracy compared to existing approaches. The HTFRNet-ESD sets a new benchmark for automated seizure detection, offering a reliable and scalable solution for real-world clinical applications. This work paves the way for improved patient outcomes through timely and precise epilepsy diagnosis.