Patient-Specific Seizure prediction from Scalp EEG Using Vision Transformer

Abstract

Epilepsy is a chronic disorder of the central nervous system. Accurate prediction of seizures using the patient's scalp EEG signal is of great importance in clinical practice. This paper proposed a personalized seizure prediction model based on Vision Transformer. First, the raw EEG signal of each patient EEG for CHB-MIT was filtered and the preictal and interictal periods were extracted for labelling. Then, the processed EEG signal was transformed into a two-dimensional spectrograms by means of the short-time Fourier transform(STFT). Finally, the processed two-dimensional spectrograms are fed into the Vision Transformer model to complete the feature extraction and classification prediction of specific epileptic EEG signals. The results showed that epilepsy prediction using Vision Transformer was best the chb21 patient (94.6% accuracy, 98.6% Recall, 89.8% Specificity, 90.5% Precision and an AUC value of 0.989).