Hybrid Deep Learning Network with Convolutional Attention for Detecting Epileptic Seizures from EEG Signals

Rapid detection of epileptic seizures from electroencephalography (EEG) signals is vital for timely treatment. However, accurately identifying seizures automatically remains challenging due to the diversity of EEG data. Recently, deep learning approaches have shown promising performance. This research presents a hybrid deep neural network combining convolutional neural network layers to automatically extract spatial-temporal patterns, bidirectional long short-term memory layers to model long-term temporal relationships, and a convolutional block attention module to focus on informative features. We evaluated this model using a benchmark EEG dataset for classifying seizures into two groups. Our model achieved a peak accuracy of 97.70% and F1-score of 97.70%, surpassing state-of-the-art techniques. The attention module improved the extraction of valuable features from EEG channels and time segments related to seizures. We provide visualizations of EEG events that the model emphasizes during seizure and non-seizure periods. This hybrid model with channel and temporal attention enhances the efficacy and interpretability of EEG-based seizure detection, further enabling patient-specific seizure characterization. The model shows promise for assisting clinicians in safely identifying seizures from multichannel EEG data.