EEG-based Epilepsy Detection System

Epilepsy is a common neurological disease that occurs at all ages.Epilepsy not only causes physical pain to the patient,but also places a huge burden on the lives of the patient and his or her family.At present,epilepsy detection is still achieved by medical staff observing electroencephalogram(EEG).However,this process is time-consuming and prone to fatigue,which will bring a huge workload to medical staff.Therefore,automatic detection of epilepsy becomes particularly important.In the current research of epilepsy detection,although many feature extraction methods have been proposed,these features can only show part of the valid information of the signal,for example,time domain features can only show local information,while frequency domain features can only show global information.At the same time,it is difficult to find the right combination of features for epilepsy detection.Therefore,it becomes particularly important to automatically learn important features for epilepsy detection from EEG signals.In contrast,current researchers have focused on the EEG data of each channel,thereby neglecting the information on the graph structure that exists between channels.Based on the public CHB-MIT data set,this thesis focuses on the current feature selection difficulties and the classification model ignoring the graph structure information in the EEG signal.The following research work is carried out on the problem of epilepsy detection:(1)A data pre-processing scheme was designed for the CHB-MIT dataset.(2)Explore the network structure of the auto-encoder and its effectiveness.In this thesis,comparative experiments with different structures of auto-encoders are performed for the 18 channels used.Subsequently,this thesis uses a support vector machine as a classifier to compare the presence or absence of auto-encoder features to verify the effect of auto-encoder.(3)This thesis proposes a new network structure,the model is mainly composed of high-order graph neural network and simple graph convolutional network.In this process,the complex graph structure relationship is first integrated through a high-order graph neural network,and then graph convolution is used to combine the graph structure and features to finally achieve classification.This thesis uses specific patient experiments and whole experiments to verify the model in this article.Comparing with other experiments,the model proposed in this paper has achieved better experimental results in terms of sensitivity and specificity.