Abstract

A patient specific epileptic seizure(ES) prediction model using only electroencephalography (EEG) data with residual neural networks (ResNets) and transfer learning (TL) techniques (i.e., SPERTL) is provided. One exemplary provided model was trained on EEG data from 23 patients with a seizure prediction horizon (SPH) of 5 minutes and used the validation data to plot precision-recall curves to aid in selecting preferred thresholds. Testing on unseen data shows the provided model outperforms related art methods by achieving the highest average sensitivity of 88.1%, specificity of 92.3%, and accuracy of 92.3%. Results also demonstrate the proposed model is less susceptible to false positives while maintaining a high positive prediction rate.