A Study on the Epileptic Seizure Detection using Deep Neural Network and its practical application

Abstract

Epileptic seizure detection has been studied for a long period of time for diagnosis and treatment of patients, but the accuracy of seizure detection has not been achieved to the practical application level. Therefore, in order to detect seizure during the EEG monitoring, it is necessary to analyze it with the visual inspection of medical staff, which requires too much time, so the research for automatic seizure detection algorithm is needed. Conventional epileptic seizure detection method has been studied with the signal processing methods such as frequency band separation, Fourier transform, and wavelet transform. However, since the EEG characteristics of the epileptic patients vary from person to person and the spatial and spectral features of the rhythmic pattern of ictal EEG are also different from each other, seizure detection with the signal processing method cannot achieve high detection accuracy. Since the deep learning has been applied to the epileptic seizure detection, seizure detection using the deep neural network has achieved high seizure detection accuracy compared to conventional seizure detection methods. However, CNN is used for extracting spectral features in EEG signal, and RNN is used to consider temporal correlation in ictal EEG. In case of using only one of CNN or RNN, there is a limit to extract spectral and temporal features from ictal EEG simultaneously. Therefore, we propose a seizure detection algorithm that considers both temporal and spectral features to get high sensitivity and specificity. In this study, the convolutional recurrent neural network (CRNN) which combines CNN and RNN, is used for the proposed algorithm to consider the characteristics of the ictal EEG. CRNN has the advantage of extracting both spectral and temporal features from the EEG signal. Moreover, EEG signals were transformed into spectral signals through short-time Fourier transform (STFT) using windows of various sizes. With the spectral signals of various resolutions, it has the advantage to analyze both the temporal and spectral features of the ictal EEG more precisely than the conventional seizure detection algorithms for analyzing spectral signals of a fixed resolution. The proposed neural network was trained using the CHB-MIT Scalp EEG database and the EEG dataset of Seoul National University Hospital. The proposed method can achieve high seizure detection accuracy without modifying the algorithm for various datasets. As a result of the simulation, the proposed model has shown a high seizure detection accuracy of 94.3% for CHB-MIT EEG dataset and 99.2% for SNUH EEG dataset, respectively.