文献类型：期刊文献

英文题名：DUFGNet: A dual-stream U-Net framework with frequency-guided channel attention and graph integration for epileptic seizure prediction

作者：Lou, Jionghao[1];Zhang, Jian[2];Li, Zhongmei[1];Chen, Lanlan[1];Feng, Enbo[1]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China

年份：2026

卷号：667

外文期刊名：NEUROCOMPUTING

收录：;EI(收录号：20255019695310);WOS:【SCI-EXPANDED(收录号:WOS:001640547500001)】；

基金：Acknowledgments This work was supported by Shanghai Rising-Star Program (24QA2706100) , Fundamental Research Funds for the Central Universities and the State Key Laboratory of Industrial Control Technology, China (ICT2024A01) and National Natural Science Foundation of China (62376095) .

语种：英文

外文关键词：EEG; Seizure prediction; Two-stream network; Self-supervised learning; Attention mechanism

摘要：Effective prediction of seizures is essential for alleviating patient suffering and facilitating timely medical in tervention. However, achieving reliable prediction remains challenging due to patient-specific variability and the complex evolution of epileptic EEG signals across spatial, temporal, and spectral domains. To address these challenges, we propose DUFGNet, a dual-stream self-supervised framework designed to model comprehensive spatio-temporal representations under conditions of limited labeled data. Specifically, a dual-stream U-Net is em ployed to extract features from two distinct perspectives: global interchannel dependencies and local intrachannel structures. Additionally, we develop a global-local feature integration module based on graph attention. It is coupled with a frequency-guided channel attention module. Together, these modules adaptively fuse multi-scale features across spatial, temporal, and spectral domains. Furthermore, we introduce a multi-task self-supervised learning strategy that incorporates two domain-specific pretext tasks. Extensive experiments on the public CHB-MIT dataset and a private Renji dataset demonstrate that DUFGNet achieves high performance, with accuracies of 97.11 % and 94.11 %, and AUC scores of 0.9955 and 0.9736, respectively. DUFGNet outperforms state-of-the-art supervised and self-supervised baselines using only 0.28 M parameters. Notably, even with just 1 % labeled data under extremely low-label conditions, it maintains over 85 % accuracy, demonstrating strong robustness. These results suggest that DUFGNet offers an accurate, lightweight, and label-efficient solution for seizure prediction, with significant potential for deployment in resource-constrained clinical or wearable settings.