SPERTL: Epileptic Seizure Prediction using EEG with ResNets and Transfer Learning

Abstract

Epilepsy is a chronic condition that causes repeat unprovoked seizures and many epileptics either develop resistance to medications and/or are not suitable candidates for surgical solutions. Hence, these recurring unpredictable seizures can have a severely negative impact on quality of life including an elevated risk of injury, social stigmatization, inability to take part in essential activities such as driving and possibly reduced access to healthcare. A predictive system that informs patients and caregivers about a potential upcoming seizure ahead of time is not only desirable but an urgent necessity. In this paper, we contribute by designing and developing patient-specific epileptic seizure (ES) prediction models using only electroencephalography (EEG) data with residual neural networks (ResNets) and transfer learning (TL) - (SPERTL). We train our proposed model on EEG data from 20 patients with a seizure prediction horizon (SPH) of 5 minutes and use the validation data to plot precision-recall curves for selecting the best thresholds. Testing on unseen data shows our model outperforms the state-of-the-art methods by achieving the highest average sensitivity of 88.1%, specificity of 92.3%, and accuracy of 92.3%. Our results also demonstrate the proposed model is less susceptible to false positives while maintaining a high positive prediction rate.