Brain Network Feature Analysis And Application Based On Graph Convolutional Neural Network

Brain science is the focus of research that countries around the world are competing to pay attention to,and brain science research projects have been launched in succession.Brain network omics research is a hot topic in this field.Analyzing the structure and functional information of brain networks is crucial for understanding the state of the brain.In recent years,deep learning methods have become the key to analyzing brain network topology data,and are also an effective method for analyzing brain networks.However,there are some problems in the research of brain networks based on EEG signals,such as neglecting the correlation between EEG channels,not comprehensively considering the influencing factors of brain network organization,and the weak interpretability of the model.To address the above issues,this paper proposes a graph convolution neural network model based on epoch length,frequency band,and functional connectivity indicators,and applies the model to the recognition and diagnosis of schizophrenia.The main work of this article is as follows:(1)A graph convolutional neural network model(EFC-GCN)based on epoch lengths,frequency bands,and functional connectivity indicators was constructed.The EEG signals of the subjects were divided into different frequency bands and epoch lengths,and on this basis,different functional connectivity indicators were used to construct a brain functional network,establishing associations between various brain regions,fully considering the joint impact of the epoch length,frequency band,and functional connectivity indicators of the EEG signals on the study;Extract the time and frequency domain features of the EEG signal and local features of the brain network to form a node feature matrix.Using the node feature matrix and the brain function network together as graph signals,perform feature extraction and classification recognition through a graph convolution neural network,fully learn the correlation between EEG channels,and obtain the optimal epoch length,frequency band,and functional connectivity indicators for recognition.(2)The EFC-GCN model was applied to the data set of first-episode schizophrenia.The EFC-GCN model was used to identify and diagnose patients with first-episode schizophrenia and normal subjects,and the performance of the model was evaluated using a ten-fold cross validation method.The experimental results show that the model has the best recognition performance with an average accuracy of 90.01% under the Theta frequency band,6s epoch length,and PLV functional connectivity indicators in the firstepisode schizophrenia data set.Experiments have demonstrated the advantages of combining time and frequency domain features with local features of the brain network,as well as the effectiveness of the model proposed in this paper when compared with other baseline models.Meanwhile,the EFC-GCN model was applied to the P50 dataset of schizophrenia to verify its effectiveness in the task state dataset.(3)In order to further explore the pathogenesis and mechanism of first-episode schizophrenia,search for the brain regions that cause the onset of first-episode schizophrenia in patients with schizophrenia,use gradient weighted class activation mapping to identify brain regions that contribute to the graph convolution class,and explore the correlation between the network topology characteristics of significant brain regions and clinical scores.The results showed that the most significant brain regions helpful for classification were located in the parietal lobe region,and there was a significant correlation between the local efficiency,betweenness,and clustering coefficient of the brain regions where electrodes FC2,FCz,and FC1 were located and the general score,providing a valuable reference for clinical medicine.