An Attention-Based Hybrid Deep Learning Approach for Patient-Specific, Cross-Patient, and Patient-Independent Seizure Detection

Automatic detection of epilepsy plays a crucial role in diagnosing and treatment of patients, while most current methods rely on patient-specific models and have shown promising results, which is not suitable for clinical application, especially when new patient data are used for diagnosis in EEG epileptic seizure detection (ESD). Therefore, the proposed study introduces a novel hybrid deep learning approach consisting of a one-dimensional convolutional neural network (1D CNN), a Multi-Long Short-Term Memory Network (MLSTM) with a multi-attention layer (MAT) for patient-specific, cross-patient, and patient-independent seizure detection. The 1D CNN model extracts spatial features, while the MLSTM extracts temporal features from segmented EEG data. Moreover, the MAT layer conducts feature fusion and identifies relevant patterns. Experiments conducted using the CHB-MIT EEG dataset confirm our method's superiority over other sibling and state-of-the-art methods by an average of 2% in classification accuracy, recall, specificity, and G.mean of using patient-specific, cross-patient, and patient-independent seizure detection, demonstrating a robust and effective framework in EEG ESD.