Patient-Specific Seizure Prediction Using a Hybrid Transformer Based on Multi-Dimensional Attention

EEG signals can be used to predict seizures in epilepsy patients. An effective seizure prediction system can send out early warning signals to alert patients or medical personnel to take appropriate measures in time to assist in diagnosis and treatment. However, the main challenge is characterizing preictal EEG features effectively, and current seizure prediction performance is unsatisfactory. This paper presents MDA_Trans, a hybrid Transformer model that combines CNN and Transformer to extract and fuse EEG features from different dimensions. The proposed Multi-Dimensional Attention (MDA) is incorporated to improve the learning ability of potential features. The MDA_Trans enhances feature extraction by focusing on significant patterns in EEG data across multiple dimensions. It was evaluated on the CHB-MIT dataset to demonstrate its effectiveness in seizure prediction, achieving a sensitivity of 91.164%, a false prediction rate of 0.108, and an AUC score of 0.927. These results indicate a high sensitivity and a low false prediction rate compared to the baseline model.