MCAN: Cross-domain self-supervised attention network based on multiscale EEG feature learning for epileptic seizure detection

• Published in: Knowledge-based systems Vol. 327; p. 114199
• Main Authors: Hong, Tingxuan, Li, Desheng, Chang, Yuan, Wang, Xiangqing, Cai, Ziliang, Wang, Rongfei, Zhang, Xiaochen, Liu, Xiaoya, Yang, Chunxiao, Yu, Shengyuan, Liu, Shuang, Ming, Dong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 09.10.2025
• Subjects: Electroencephalogram; Epileptic seizure detection; Multiscale feature learning; Self-supervised learning; Multiscale feature learning; Electroencephalogram; Epileptic seizure detection; Self-supervised learning
• ISSN: 0950-7051
• DOI: 10.1016/j.knosys.2025.114199

•Proposes MCAN for automatic EEG-based seizure detection under real-world class imbalance and distribution shifts.•Designs cross-domain hybrid self-supervised learning to capture temporal, spatial, and spectral EEG structures.•Develops multiscale feature learning to model hierarchical spatiotemporal EEG dynamics with diverse receptive fields.•Introduces seizure-guided self-attention with a sparse adjacency matrix to capture seizure-related neuronal synchrony.•Demonstrates an AUC of 0.914 and F1-score of 0.709, showing potential for real-world clinical applications. Epilepsy is a common neurological disorder that severely affects patient safety and quality of life. Electroencephalography (EEG) is crucial for detecting epileptic seizures. However, the manual annotation of seizures during long-term EEG monitoring is labor-intensive, error-prone, and highly dependent on clinical expertise. Furthermore, real-world EEG-based seizure detection is also challenging due to distribution shifts across subjects and datasets, along with the significant class imbalance between seizure and normal segments. To address these challenges, we propose a cross-domain hybrid self-supervised attention network (MCAN) for the automatic detection of seizures. The network offers the following key contributions: Firstly, a cross-domain hybrid self-supervised learning strategy is designed to capture the temporal dynamics of EEG signals while simultaneously preserving the spatial distributions across electrodes and the spectral characteristics of neural oscillations. Secondly, a multiscale feature learning module is developed to model the hierarchical spatiotemporal dynamics of EEG signals through diverse receptive fields, thereby reducing the model’s dependency on subject-specific features. Thirdly, we propose a self-attention mechanism guided by a sparse electrode adjacency matrix to effectively capture seizure-related neuronal synchrony. Extensive experiments were conducted using real-world clinical epilepsy monitoring datasets and three publicly available datasets to evaluate the performance of our MCAN. The results demonstrate that MCAN consistently outperforms baseline methods in seizure detection across multiple datasets. Notably, it achieves an area under the receiver operating characteristic curve of 0.914 and an F1-score of 0.709, highlighting its potential for seizure monitoring applications.