Lightweight seizure detection with channel attention and multilayer gated recurrent unit for cross-patient EEG analysis

Abstract

Real-time epileptic seizure detection from multichannel EEG signals remains a challenging clinical problem due to high inter-patient variability, severe class imbalance, and complex multi-scale ictal dynamics. Existing deep learning approaches predominantly rely on computationally intensive architectures such as Bidirectional GRU (BiGRU), increasing model complexity. This paper proposes SE-TCN-MGRU, a novel lightweight framework integrating three modules: a Squeeze-and-Excitation Network (SE-Net) for adaptive channel-wise recalibration; a Temporal Convolutional Network (TCN) with dilated causal convolutions for multi-scale temporal feature extraction; and an enhance Multilayer GRU (MGRU) with residual connections and Layer Normalization as a parameter-efficient, unidirectional replacement for BiGRU. A strict cross-patient LOOCV protocol with an independent validation strategy is applied on the CHB-MIT dataset to evaluate the proposed model. SE-TCN-MGRU achieves an accuracy 95.82%, sensitivity of 88.72%, specificity of 97.24%, F1-score of 87.83%, MCC of 86.17%,ROC-AUC of 0.9659 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm$$\end{document} 0.0659, and PR-AUC of 0.9314 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm$$\end{document} 0.1133. The proposed framework contains only 345,330 trainable parameters, corresponding to a 47.8% reduction compared with the reproduced SE-TCN-BiGRU baseline, while maintaining competitive seizure detection performance. These results demonstrate strong cross-patient generalization and computational efficiency, highlighting the potential of SE-TCN-MGRU for real-time wearable EEG-based seizure monitoring. This work supports United Nations Sustainable Development Goal 3 (SDG-3: Good Health and Well-being) by promoting accessible and automated monitoring of neurological disorders in epilepsy patients globally. Not applicable. This study is not a clinical trial.