Prediction of Epileptic Seizures based on EEG Signals using Deep Neural Networks an‎d Attention Mechanism

9/9/2026 12:00:00 AM

Epilepsy is one of the most preva‎lent neurological disorders, with a significant proportion of patients remaining drug-resistant an‎d experiencing unpredictable seizures. The sudden onset of seizures not only imposes severe physical an‎d psychological risks but also drastically reduces quality of life. Therefore, developing intelligent EEG-based seizure prediction systems has become a crucial challenge in computational neuroscience. In this study, we designed a three-branch deep learning framework that integrates Convolutional Neural Networks (CNNs) for capturing local spatial patterns, Graph Convolutional Networks (GCNs) for modeling topological relationships among EEG channels, an‎d Tangent-space representations for embedding covariance features in a Riemannian geometry space. EEG data were obtained from the CHB-MIT database, including 19 patients, 528.1 hours of recordings, an‎d 99 valid seizures. Preprocessing involved defining 15-minute preictal intervals prior to each seizure an‎d interictal segments away from seizures (excluding 30 minutes before an‎d after). Signals were segmented into 5-second windows with 50% overlap, an‎d undersampling was applied to balance the classes. The framework was eva‎luated under three scenarios. In patient-specific 5-fold validation, the mean (±SD across patients) was Accuracy = 97.37 ± 1.35%, Sensitivity = 97.44 ± 1.44%, Specificity = 97.61 ± 1.34%, Precision = 97.63 ± 1.38%. In generalized training with LOOCV, macro averages were Accuracy = 85.81 ± 4.92%, Sensitivity = 80.47 ± 6.53%, Specificity = 91.15 ± 2.94%, Precision = 80.92 ± 5.24%. In the event-based LOSO eva‎luation, 97 out of 99 seizures were correctly predicted (Sn = 98.48%, AUC = 0.951, FPR/h = 0.049). These findings demonstrate that the integration of CNN, GCN, an‎d Tangent-space branches, combined with rigorous preprocessing an‎d data balancing, can deliver a highly accurate, stable, an‎d generalizable framework for seizure prediction. The proposed approach holds strong potential for future deployment in clinical systems an‎d wearable EEG-based monitoring devices. In future research, the use of other types of attention mechanisms an‎d related techniques, as well as different graph networks an‎d different graph construction methods, can also be considered. In conclusion, the proposed three-branch deep learning framework represents an important step forward in EEG-based seizure prediction, bridging methodological advances in deep learning with the pressing clinical need for reliable seizure forecasting. With further refinement an‎d validation, it holds significant potential for improving patient outcomes an‎d enabling proactive, personalized epilepsy management.

پيش بيني تشنج صرعي , صرع , يادگيري عميق , مكانيسم توجه , الكتروانسفالوگرام

seizure prediction , deep learning , seizure , attention mechanism , cross attention , graph convolutional network

Ali Hedayati

Abbas Erfanian Omidvar