Enhancing seizure prediction using a DC-SA-EBiLSTM framework with self-attention mechanism

Accurately predicting seizures remains challenging. With advances in smart medical technology, EEG-based monitoring has become essential. This study aims to improve prediction accuracy using a hybrid framework that models multiscale EEG characteristics.

EEG signals are decomposed into multiple sub-bands using the Discrete Wavelet Transform, and representative time-frequency and nonlinear features are extracted. These features are fed into a channel-centric model integrating depthwise separable convolution, self-attention, and an enhanced bidirectional long short-term memory network (DC-SA-EBiLSTM). The architecture integrates depthwise separable convolution for local spatial feature extraction, multi-head self-attention for global inter-channel dependencies, and an enhanced BiLSTM for channel-wise sequence modeling. The proposed method was evaluated on the CHB-MIT dataset using a 10-fold cross-validation protocol. An event-level leave-one-seizure-event-out validation was also conducted to assess alarm-based prediction performance.

The proposed approach achieved an average accuracy of 95.89%, sensitivity of 96.70%, specificity of 95.48%, and AUC of 99.02%. In the event-level validation, the model achieved an event sensitivity of 95.96%, an average false alarm rate of 0.316 FPR/h, and a mean early warning time of 30.52 min.

The DC-SA-EBiLSTM framework effectively captures local and global inter-channel dependencies and provides a feature-driven approach for patient-specific preictal state prediction, showing potential for EEG-based seizure prediction.