Abstract

Epilepsy is a chronic neurological disorder characterized by recurrent seizures caused by abnormal electrical activity in the brain, for which timely and accurate automated detection is essential for effective clinical management. Manual Electroencephalography (EEG) interpretation is labor-intensive and subject to inter-expert variability, highlighting the need for automated, reliable, and clinically deployable diagnostic systems. This study proposes a hybrid Bidirectional Long Short-Term Memory-Relevance Vector Machine (BiLSTM-RVM) framework that combines a Bidirectional Long Short-Term Memory (BiLSTM) network with a Relevance Vector Machine (RVM) for automated epileptic seizure detection from Electroencephalography (EEG) signals. The proposed framework explicitly addresses four key challenges: inter-subject generalization, model interpretability, cross-database robustness, and resistance to signal degradation. Evaluated on the University of California, Irvine (UCI) Epileptic Seizure Recognition Dataset, the model achieves an accuracy of 99.57% and an AUC of 99.97%. A Leave-One-Subject-Out (LOSO) cross-validation strategy is employed to evaluate inter-subject generalization, while cross-database assessment on the Children’s Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) Scalp EEG Database confirms strong robustness across diverse acquisition conditions and patient populations. Furthermore, Gaussian noise sensitivity analysis is conducted to assess model robustness under simulated real-world EEG artifacts. SHapley Additive exPlanations (SHAP) are employed to provide temporal interpretation of model decisions, enabling clinically meaningful identification of EEG segments associated with seizure activity. Overall, the results demonstrate that the proposed BiLSTM-RVM framework provides an accurate, generalizable, interpretable, and noise-resilient solution for computer-assisted epilepsy diagnosis.