CACBH-Net: A CNN-Attention BiLSTM Hybrid Deep Learning Model for Enhanced Detection of Epileptic Seizure

Reliable and early identification of epileptic seizures plays a vital role in improving clinical outcomes and supporting real-time therapeutic responses. Electroencephalogram (EEG) signals, despite being rich in brain activity information, are nonlinear and complex, which makes manual analysis both labor-intensive and prone to subjectivity. To address this challenge, we propose CACBH, a combined deep learning based architecture that aims at integrating CNN and BiLSTM. CNN is useful for capturing spatial representations, and BiLSTM units are effective for temporal sequence modeling, and an Attention mechanism for emphasizing clinically significant patterns in the EEG signal. The input signals are processed using bandpass filtering (0.5–40 Hz), z-score normalization, and segmented into 10-second intervals. The model’s training and testing is based on the CHB-MIT Scalp EEG dataset using a stratified data split that preserves the ratio of seizure to non-seizure events. Results show that CACBH achieves an impressive accuracy of 95.03%, surpassing traditional CNN and LSTM approaches. The attention layer improves interpretability by focusing on seizure-relevant features, while the overall architecture provides consistent detection performance. Evaluation through confusion matrices and classification metrics further validates the system’s reliability. Overall, CACBH shows promise as a deployable and scalable solution for automated seizure monitoring in both hospital-based EEG systems and portable health devices.