Research And Application Of Epileptic EEG Signal Classification Method Based On Graph Convolutional Neural Networ

As a chronic neurological disorder,epilepsy brings great difficulties and burdens to patients and their families.The classification of epileptic EEG signals is an important medical application technology that can help doctors diagnose and treat epilepsy more accurately,as well as predict the likelihood of epileptic seizures,thereby helping patients improve their quality of life.In recent years,with the development of graph convolutional neural networks,an increasing number of researchers have started to pay attention to analyzing and processing problems using graph data.Electroencephalogram(EEG)signals can be represented as graph structures,where nodes represent electrode channels and edges represent the relationship between channels.However,most existing methods for epileptic EEG signal classification fail to adequately consider the mutual relationships between channels,leading to suboptimal results.Therefore,we conducted the following work:(1)A multi-feature multi-relationship graph convolutional neural network-based classification method for epileptic EEG signals is proposed.The method adopts the idea of graph theory and transforms EEG data into a graph structure by using multiple features and relationships to construct the graph.Through frequency domain analysis,time-frequency analysis and nonlinear dynamics analysis of EEG signals,several features are extracted as node features of the model input.Spatial similarity and spectral similarity between EEG channels are extracted and fused as the edge relationship matrix among the nodes of the overall graph.Graph convolutional neural networks(GCN)are used to update node features by aggregating neighboring EEG channel node information for EEG classification.The experimental results show that the classification evaluation indexes of Accuracy,Precision,Recall,F1-Score and AUC are 0.866,0.910,0.819,0.862 and 0.900,which are superior to the traditional machine learning and convolutional neural network models.Moreover,compared with the graph convolutional neural network with single feature and single relation,it also has obvious improvement.(2)The topology of the graph constructed from EEG signals is generally defined and calculated under the guidance of expert knowledge,and it is not changed during the training process of the GCN classification model.This approach cannot guarantee that the graph topology reflects the real correlation between EEG channels.To address this problem,a novel epileptic EEG classification method based on iterative deep graph learning(IDGL)called E-IGCN is proposed.This method employs multi-head graph attention mechanism for learning node similarity measures and improves the cosine similarity used in IDGL,which has insufficient expression capability.It iteratively optimizes the graph structure and the parameters of the GCN to find the optimal graph structure and epileptic EEG classification performance.Experimental results on the TUEP dataset show that,compared to the control groups of ordinary GCN and IDGL,E-IGCN obtains an optimized graph structure and significantly improves the classification performance of epileptic EEG.Furthermore,the effectiveness of the E-IGCN model is further validated on the TUEP single-polarity montage dataset and the TUAB and MPI LEMON joint dataset.