Automatic Seizure Detection And Research Based On Graph Convolutional Network

Epilepsy is a neurological disease caused by abnormal discharge of central neurons in the brain.When seizures occur,people will suddenly lose consciousness and convulse.If rescue measures are not taken in time,it will endanger people’s lives.With the progress of medical equipment and the rapid development of machine learning,the automatic seizure detection methods based on EEG has gradually replaced the traditional manual diagnosis.According to the latest neuroscience research,there exits an interaction mechanism between different brain regions.The exiting seizure detection algorithms has considered the characteristics of multi-channel EEG signals in the time dimension,but ignored the spatial relationship between EEG channels.Therefore,exploring the temporal and spatial correlation of EEG signals is the first key research content of this thesis.In addition,clinical medicine and existing EEG datasets show that the total duration of EEG signal acquisition during epileptic seizures is much shorter than the total duration of normal EEG signals,resulting in the problem of sample imbalance.Unbalanced samples will cause the overall detection results of the model to be biased towards normal EEG signals,resulting in the model being unable to accurately detect seizures.Most of the existing seizure detection algorithms are trained based on balanced EEG samples,and have achieved ideal detection results,but their clinical generalization performance still needs to be improved.Therefore,exploring the problem of imbalanced EEG signal classification is the second focus of this paper.Based on the above two research focuses,the proposed work applies GCN(Graph Convolution Network)and focal loss to conduct experiments on the public EEG dataset of CHB-MIT(Children’s Hospital Boston--Massachusetts Institute of Technology),which deeply excavates and discusses the temporal and spatial correlation of unbalanced EEG signals.LGCN(Linear Graph Convolution Network)combined with focal loss is proposed to model unbalanced EEG data.The proposed model considers the temporal and spatial correlation between EEG channels.The temporal correlation refers to the time-domain characteristics contained in the selected EEG channels in the same time period,and the spatial correlation refers to the spatial topological relationship between channels calculated according to the time-domain characteristics.The spatial topology is defined as a graph structure.The nodes in the graph represent the selected EEG channels,and the edges represent the spatial relationship between channels.The characteristics of the nodes are updated through the edge index.For the class imbalance problem,this method applies focal loss to evaluate and update the overall performance of the model.In this thesis,the effectiveness of the proposed method is verified by stratified ten-fold cross validation experiments.In the detection of 23 subjects,the average detection sensitivity of LGCN reaches 96.95%,and the average detection AUC(area under the curve)reaches 97.07%.In order to obtain deeper and advanced EEG features,this thesis continues to deepen the depth of the model based on the first work,and introduces residual network and attention mechanism,proposing the model of AGRN(Attention-based Graph ResNet)with focal loss for seizure detection.This method models unbalanced EEG signals.The constructed depth graph convolution layer further excavates the hidden spatial and temporal information in the brain area.The proposed residual block accelerates the flow of information,and the introduced mechanism enhances model interpretability and enables feature optimization.This study verifies the effectiveness of the method through experiments.The average detection sensitivity and average detection AUC of 23 subjects has reached 97.93% and 98.56%,respectively.In summary,this thesis applies LGCN and AGRN combined with focal loss for seizure detection.Moreover,based on the proposed automatic seizure detection scheme,this thesis carries out cross validation experiment on the published 10-20 international EEG dataset CHB-MIT.The experimental part of this thesis also verifies the feasibility of GCN model combined with focal loss,which provides a scientific reference value for the research of automatic seizure detection algorithm and practical application of clinical medicine.