基于距离相关功能连接网络的机器学习模型在精神分裂症诊断中的价值

Machine learning model of functional connectivity network based on distance correlation:diagnosing value in schizophrenia

目的探讨基于距离相关的功能连接(FC_(D))网络的机器学习模型能否更准确地识别精神分裂症病人,以实现精神分裂症的精准诊断。方法前瞻性纳入性别、年龄相匹配的精神分裂症病人103例和健康对照103例,均收集静息态功能MRI数据。采用自动解剖标记(AAL)模板将全脑分为116个脑区,分别构建基于Pearson相关的功能连接(FC_(P))网络和FC_(D)网络,FC_(P)、FC_(D)网络中均有6670条连接。采用独立样本t检验分别对精神分裂症病人和健康对照的FC_(P)和FC_(D)网络中的每条连接进行比较分析,以年龄、性别、头动参数作为协变量,进行Bonferroni多重比较。采用MVPANI软件包进行多变量模式分析,基于FC_(P)和FC_(D)网络特征以及两者融合特征构建线性支持向量机(SVM)分类模型,采用置换检验评价分类模型的诊断准确度。计算各个模型的分类诊断敏感度和特异度,应用受试者操作特征曲线下面积(AUC)评估模型的预测效能。结果与健康对照相比,精神分裂症病人各脑区间FC_(P)及FC_(D)网络均可见广泛失连接。基于FC_(P)、FC_(D)网络特征及两者融合特征建立的SVM分类模型对精神分裂症病人和健康对照的分类准确度分别为76.4%、82.6%、84.7%(均P<0.05),对应的AUC分别为0.86、0.88、0.91。结论与FC_(P)网络特征相比,基于FC_(D)网络特征的机器学习模型能够更准确地诊断精神分裂症病人,基于两者融合特征的机器学习模型能够进一步提高预测性能。

Objective To investigate whether the machine learning model of functional connectivity(FC)network based on the distance correlation can identify the schizophrenia patients accurately,so as to achieve accurate diagnosis of schizophrenia.Methods In this study,we prospectively enrolled 103 schizophrenia patients and 103 age-and gender-matched healthy controls,and collected the resting state fMRI data of patients and healthy control.The whole brain was divided into 116 brain regions by automatic anatomical labeling(AAL)templates to construct the functional connectivity based on Pearson correlation(FC_(P))and distance correlation(FC_(D))algorithms,respectively.There were 6670 connections in both FC_(P)and FC_(D),respectively.A comparation of each connection in the FC_(P)and FC_(D)networks of schizophrenia patients and health controls was performed using independent samples t-test with age,gender,and head movement parameters as covariates.Bonferroni multiple comparisons were also conducted.Multi-variate pattern analysis was carried out by the MVPANI package.The linear support vector machine(SVM)classification model was constructed based on FC_(P),FC_(D),and their fusion network features,and the permutation test was used to evaluate the diagnostic accuracy of the classification model.The sensitivity and specificity of the classification diagnosis for each model were calculated,and the predictive efficacy was evaluated by the area under the receiver operator characteristic curve(AUC).Results Compared with healthy controls,extensive disconnections were identified in the FC_(P)and FC_(D)networks among the brain regions in schizophrenia patients.The classification accuracy of FC_(P),FC_(D),and their fused network feature model for discriminating schizophrenia patients and healthy controls was 76.4%,82.6%,and 84.7%,respectively(all P<0.05).The corresponding area under the curve of FC_(P),FC_(D),and their fused network feature model was 0.86,0.88,and 0.91,respectively.Conclusion Compared with FC_(P)network feature,the machine learning model based on the FC_(D)network feature is more accurately in the diagnosis of schizophrenia,and the fused network feature model of FC_(P)and FC_(D)can further improve the predictive performance.