基于机器学习模型探索节点综合拓扑属性在精神分裂症研究中的应用价值

Value of nodal integrated topological attributes based on machine learning model in identifying schizophrenia

目的基于机器学习模型探索节点综合拓扑属性(NITA)在精神分裂症研究中的应用价值。方法选择新乡医学院第二附属医院精神科自2022年1月至2023年8月收治的56例首发精神分裂症患者(患者组)及向社会招募的56例健康志愿者(对照组)为研究对象,采集其功能MRI数据,对数据进行预处理后构建大脑网络,并基于图论的方法提取全局和节点拓扑属性指标作为训练特征。将受试者分为训练组(46∶46)和测试组(10∶10),在训练组中分别使用两种特征训练随机森林(RFC)、支持向量机(SVM)和梯度提升树(XGboost)模型,计算出各模型的准确率、召回率、F1值及受试者工作特征曲线下面积(AUC)以进行性能评估,并结合测试组表现分析各模型的泛化能力,筛选出表现优秀的拓扑属性指标。将所选拓扑属性指标通过主成分分析算法降至一维,并用该新维度特征再次分别训练上述模型,根据训练组和测试组的表现筛选出适配模型。对患者组与对照组各脑区新维度特征进行统计学分析并结合错误发现率(FDR)校正,筛选出差异有统计学意义的脑区新维度特征,并再次用其训练适配模型。结果训练组中各机器学习模型使用节点拓扑属性指标的分类模型的准确率、召回率、F1值、AUC均高于使用全局拓扑属性指标的分类模型,测试组中SVM模型使用节点拓扑属性指标的分类模型表现出稳定的泛化性(准确率为75.00%,召回率为100.00%,F1分数为0.80,AUC为0.92)。将节点拓扑属性指标降维并命名新维度特征为NITA。依据SVM模型使用NITA的分类模型在训练组中的验证结果(准确率为77.00%、召回率为72.00%、F1值为0.76、AUC为0.86)及测试组中的泛化表现(准确率为66.67%、召回率为83.33%、F1值为0.71、AUC为0.61),选择SVM作为分类NITA的适配模型。患者组与对照组间右侧额中回腹外侧区、左侧额下回背侧区、右侧中央前回尾腹侧区、左侧颞上回喙部、右侧梭状回中央区、右侧顶下小叶前背侧区及左侧枕极皮层NITA的差异有统计学意义(P<0.05,FDR校正)。将上述脑区NITA作为特征训练得出最优模型FDR-NITA-SVM,其在训练组中准确率为93.74%、召回率为98.00%、F1值为0.94、AUC为0.96,测试组中准确率为83.33%、召回率为66.67%、F1值为0.80、AUC为0.92。结论NITA可能是诊断精神分裂症的潜在影像标志物。NITA异常脑区是精神分裂症患者大脑网络间信息交流和整合的关键节点。

Objective To explore the value of nodal integrated topological attributes(NITA)based on machine learning model in identifying schizophrenia.Methods A total of 56 patients with first-onset schizophrenia admitted to Department of Psychiatry,Second Affiliated Hospital of Xinxiang Medical University from January 2022 to August 2023 and 56 healthy volunteers recruited from community were selected.Functional MRI data were collected,and brain functional networks were constructed after preprocessing.Global and nodal topological attributes were extracted using graph theory as training features.Participants were divided into training set(46 schizophrenia patients and 46 heathy volunteers)and testing set(10 schizophrenia patients and 10 heathy volunteers).Random Forest Classifier(RFC),Support Vector Machine(SVM),and Gradient Boosting Tree(XGBoost)models were fitted to global and nodal topological attributes in the training set to calculate the accuracy,recall rate,F1 value,and area under receiver operating characteristic curve(AUC)of each model.Generalization ability was analyzed based on the performance of testing set,and excellent topological attributes were screened out.Selected topological attributes were reduced to one-dimensional features through principal component analysis,and then fitted to the above models,and feature-adapted model was selected based on the performances of training and testing sets.Statistical analysis of the new dimensional features of each brain region of schizophrenia patients and heathy volunteers was performed.Combined with false discovery rate(FDR),new dimension features with significant differences were selected and fitted with the adapted model.Results In the training set,machine learning models using node topological attributes achieved higher accuracy,recall rate,F1 scores,and AUC compared with those using global topological attributes.In the test set,the SVM model using node topological attributes showed stable generalizability(accuracy=75.00%,recall rate=100.00%,F1 score=0.80,AUC=0.92).The node topological attribute metrics were down-dimensionally named NITA.Based on validation results of SVM model using NITA in the training set(accuracy of 77.00%,recall of 72.00%,F1 value of 0.76,AUC of 0.86)and performance in the testing set(accuracy of 66.67%,recall of 83.33%,F1 value of 0.71,AUC of 0.61),SVM was selected as the adapted model.NITA in the right middle frontal gyrus ventrolateral area,left inferior frontal gyrus dorsal area,right precentral gyrus caudal ventrolateral area,left superior temporal gyrus rostral area,right fusiform gyrus lateroventral area,right inferior parietal lobule rostrodorsal area,left occipital polar cortex showed significant difference between patients and volunteers(P<0.05,FDR-corrected).The optimal model(FDR-PCAN-SVM)obtained via NITA being trained on corresponding brain area reached an accuracy of 93.74%,recall rate of 98.00%,F1 value of 0.94,and AUC of 0.96 in the training set and accuracy of 83.33%,recall rate of 66.67%,F1 value of 0.80,and AUC of 0.92 in the testing set.Conclusion NITA may serve as a potential image biomarker for schizophrenia identification;brain regions with abnormal NITA is key nodes in information exchange and integration within the brain networks in schizophrenia patients.