基于图卷积神经网络的精神分裂症识别研究  

Recognition of schizophrenia based on graph convolutional neural network

作　　者：林萍[1] 朱耿 李斌[3] 周宇星 徐信毅 李晓欧[1,2] LIN Ping;ZHU Geng;LI Bin;ZHOU Yuxing;XU Xinyi;LI Xiaoou(College of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093;College of Medical Instruments,Shanghai University of Medicine&Health Sciences,Shanghai 201318;Shanghai Yangpu District Mental Health Center,Shanghai 200093)

机构地区：[1]上海理工大学健康科学与工程学院,上海200093 [2]上海健康医学院医疗器械学院,上海201318 [3]上海市杨浦区精神卫生中心,上海200093

出　　处：《北京生物医学工程》2025年第1期26-31,48,共7页Beijing Biomedical Engineering

基　　金：上海市科委地方院校能力建设项目(22010502400);上海市杨浦区科学技术委员会、卫生健康委员会科研项目(YPM202114);上海健康医学院精神卫生临床研究中心项目(20MC2020005)资助。

摘　　要：目的精神分裂症(schizophrenia,SZ)患者存在工作记忆、信息处理、选择性学习等方面的认知障碍,临床上仍由医生经量表进行评估诊断。本文提出了一种不依赖人工特征的基于脑功能连接与图卷积神经网络(graph convolution neural network,GCN)的精神分裂症辅助诊断方法,以实现对精神分裂症的自动分类。方法由于脑网络图与图数据的天然相似性,本文从42例精神分裂症患者和29例健康对照者(healthy control,HC)的强化学习任务中获取事件相关电位(event-related potential,ERP),以电极为节点,使用相位滞后指数构建功能连接矩阵,结合节点特征构造脑网络图数据,输入图卷积神经网络模型进行训练分类。结果GCN模型下使用功率谱密度作为节点特征时,SZ与HC的分类准确率、精确率、F1分数和特异性分别为84.21%、75%、85.71%、70%。与选择原始脑电(electroencephalogram,EEG)向量作为节点特征相比准确率提高了6.43%。与使用随机森林分类器相比,GCN模型提高了3.18%的准确率。结论本文运用图神经网络对脑电信号进行分类,实验结果表明,GCN可以有效识别SZ患者,实现对SZ患者的自动分类。图结构下节点特征的选择相对于传统机器学习模型对分类的准确率有显著提升,且效果更优。Objective Patients with schizophrenia(SZ)suffer from cognitive deficits in working memory,information processing,and selective learning,which are still clinically diagnosed by doctors assessed by scales.In this paper,we propose an auxiliary diagnosis method for schizophrenia based on brain functional connectivity and graph convolution neural network(GCN)without relying on artificial features to realize the automatic classification of schizophrenia.Methods Due to the natural similarity between brain network graphs and graph data,in this paper,we obtained event-related potential(ERP)from a reinforcement learning task with 42 schizophrenia patients and 29 healthy controls(HC),constructed functional connectivity matrices using phase lag indices with electrodes as nodes,and constructed brain network graph data by combining node features,which were inputted into a graph convolutional neural network model for training classification.Results The classification accuracy,precision,F1 score and specificity of SZ and HC when using power spectral density as node features under the GCN model were 84.21%,75%,85.71%and 70%,respectively.The accuracy was improved by 6.43%compared to choosing the original electroencephalogram(EEG)vector as the node feature.The GCN model also improved the accuracy by 3.18%compared to using the random forest classifier.Conclusions In this paper,graph neural network is used to classify EEG signals,and the experimental results show that GCN can effectively recognize SZ patients and realize automatic classification of SZ patients.And the selection of node features under the graph structure has a significant improvement on the classification accuracy relative to the traditional machine learning model,and the effect is better.

关 键 词：脑功能连接 图神经网络 脑电图 精神分裂症 

分 类 号：R318.04[医药卫生—生物医学工程]