An Idiopathic Generalized Epilepsy Classification Model Based On Graph Convolution Network

Epilepsy is a disorder of repetitive,sudden,and transient central nervous system dysfunction caused by excessively synchronized firing of a large number of neurons in the brain.In addition to the great physical and psychological burden of recurrent seizures,prejudice and discrimination from society and the general public worsen the living conditions of patients with epilepsy.Generalized tonic-clonic seizures(GTCS)is a typical subtype of generalized epilepsy,which manifests as widespread abnormal electrical signal transmission in both hemispheres of the brain during seizures,without obvious metabolic abnormalities and brain structural lesions.Today,the rapid development of imaging technology provides multimodal imaging information for epilepsy research,and with the help of reasonable and effective data analysis methods,we can gradually deepen our understanding of the pathological mechanism of GTCS.At the same time,the fusion of deep learning and multimodal image features will also help us further improve the diagnostic efficiency of GTCS.In this work,we first use the graph signal processing(GSP)to perform graph-frequency analysis on the brain networks of GTCS and normal controls.We use harmonic energy as a bridge to combine spectral and spatial domains to explore global and local brain signal abnormalities between GTCS patients and normal controls.The results shows that the total energy of low harmonic frequency in GTCS group is significantly higher than that of normal controls,and the brain regions causing global energy abnormalities are widely distributed in both the left and right hemispheres of the brain.These areas are mainly located in the precentral gyrus,prefrontal cortex,cingulate belt,precuneus,middle temporal gyrus,central sulcus cover and insula,etc.Through the division of Yeo-7 network,it is found that these abnormal areas are mainly located in the sensorimotor network,control network,ventral attention network,default network related to body sensation,motor control,and advanced cognition.These global or local abnormalities,can be used as potential image features of GTCS.Then we take advantage of the graph convolution network(GCN)and fused the above image features to construct a prediction model for distinguishing GTCS from normal controls.In addition,we also propose a local receptive field and K-nearest neighbor(K-NN)algorithm to optimize the graph structure of network and further improve the prediction performance of GCN.Compared with the prediction effect of the traditional convolution network we find that the prediction performance of the GCN with graph optimization has been significantly improved.The optimal accuracy of the model is 76%,the sensitivity is 77.2% and the area under the curve is 80%.Compared with the traditional neural network,the three evaluation indicators have increased by8%,9.2% and 11.5% respectively.In summary,this study provides a reference for us to search for the biological markers and imaging features of GTCS,and can also provide a new perspective for explaining the physiological mechanism of epilepsy.At the same time,a new prediction framework based on graph convolutional network is also provided for the clinical diagnosis of GTCS.