Epileptic Seizure Prediction using Multi-Channel Raw EEGs with Convolutional Neural Network

Subject Areas : Journal of Computer & Robotics

Jamal  Nazari ¹ (Faculty of Electrical, Biomedical and Mechatronics Engineering, Qazvin , Branch, Islamic Azad University, Qazvin, Iran)
Ali  Motie Nasrabadi ² (Department of Biomedical Engineering, Faculty of Engineering, Shahed University, Tehran, Iran)
Mohamad Bager  Menhaj ³ (Department of Electrical Engineering, Amirkabir University of Technology, Tehran, Iran)
Somayeh  Raiesdana ⁴ (Faculty of Electrical, Biomedical and Mechatronics Engineering, Qazvin , Branch, Islamic Azad University, Qazvin, Iran)

Received: 2023-03-07 Accepted : 2023-03-18 Published : 2023-06-01

Keywords: EEG, epilepsy, Seizure Prediction, ConvNet,

Abstract :

Epileptic seizure prediction has been one of the interesting topics among researchers in recent years. Recent evidence suggests that, in many seizures, changes in the preictal signal begin minutes before the ictal begins, raising hopes of predicting the seizure onset before it occurs. Convolutional neural network (ConvNet) is a powerful computational tool with deep learning capacity which is able to detect complex structures in data. In this study, we employed a ConvNet and a set of techniques to make optimal use of the existing data for an end-to-end learning. Multi-channel non-invasive raw EEGs from the CHB-MIT database were used for training of the proposed model. The proposed method resulted in sensitivity of 92.05% and false prediction rate of 0.073/h with the cross-validation approach in distinguishing preictal and ictal. We obtained a 10-minute seizure prediction horizon that is relatively higher than the values obtained in other researches. This longer time period can give the patient more opportunity for preventive actions. Seizure occurrence period was computed nearly 20 minutes which lets the patient wait less for the seizure to occur and this in turn makes him have less anxiety. Furthermore, a feature map visualizing method was employed in the present work to decode the employed deep network and to understand how it learns and what it learns when trying to solve the seizure prediction task. By investigating feature maps of the used ConvNet’s middle layer, we observed that the proposed network retains most of the beta and gamma band properties in layers.