急性意识障碍不同预后患者fMRI海马网络的功能差异和机器学习预测模型构建

Functional differences of fMRI hippocampal network and construction of machine learning prediction model in patients with different outcomes of acute disturbance of consciousness

目的探讨不同预后的急性意识障碍患者功能磁共振成像(fMRI)海马网络的功能差异,构建支持向量机(SVM)机器学习预测模型并验证。方法回顾性分析2022年9月—2023年7月南京医科大学第一附属医院神经外科收治并完成fMRI检查的43例急性意识障碍(aDOC)患者的临床资料,所有患者出院后均随访3个月。采用昏迷恢复量表修订版(CRS-R)评估受试者随访期间的意识状态,并剔除影像数据不合格的患者,最终纳入37例患者,其中自发性脑出血患者19例,创伤性脑损伤患者18例。根据随访CRS-R评分将患者分为脱离微小意识状态组(eMCS组,n=13)和慢性意识障碍组(pDOC组,n=24)。比较两组患者的临床资料,基于MATLAB平台对影像fMRI数据进行海马功能网络FC值分析,机器学习采用MATLAB内部SVM代码并采用留一法进行交叉验证,采用受试者工作特征(ROC)曲线展示预测效能。结果两组患者临床资料比较,差异无统计学意义(P>0.05);eMCS组的格拉斯哥评分[(9.0±1.8)比(6.0±2.1)分]和全面无反应性量表评分[13.00(11.00,13.00)比10.00(8.25,11.75)分]均高于pDOC组,差异有统计学意义(t=3.67,Z=-3.24;P<0.01)。fMRI中血氧水平依赖(BOLD)的对比序列海马网络分析结果显示,两组患者在双侧楔前叶(t=4.632,P<0.005,TFCE校正)和右侧舌回(t=3.940,P<0.005,TFCE校正)脑区中存在活动差异。基于fMRI数据海马网络FC值全部差异脑区构建SVM模型的ROC曲线结果显示,曲线下面积(AUC)为0.85,敏感度为0.69,特异度为0.83;基于差异脑区中楔前叶构建SVM模型的AUC为0.88,准确度为81.08,敏感度为0.86,特异度为0.83;基于差异脑区中右侧舌回构建SVM模型的AUC为0.75,准确度为70.27,敏感度为0.77,特异度为0.71。结论不同预后的急性意识障碍患者的fMRI海马网络在双侧楔前叶和右侧舌回脑区的活动度存在差异,基于这些脑区差异可以构建机器学习模型用于精准预测aDOC患者预后。

Objective To explore the functional differences of hippocampal network in patients with different outcomes of acute disturbance of consciousness(aDOC)by functional magnetic resonance imaging(fMRI),construct and verify the support vector machine(SVM)machine learning prediction model.Methods Clinical data of 43 patients with aDOC admitted to the Department of Neurosurgery of the First Affiliated Hospital of Nanjing Medical University and completed fMRI examination from September 2022 to July 2023 were retrospectively analyzed.All patients were followed up for three months after discharge.The revised version of the Coma Recovery Scale(CRS-R)was used to evaluate the consciousness of the subjects during follow-up,and patients with unqualified imaging data were excluded.Finally,37 patients were included,including 19 patients with spontaneous cerebral hemorrhage and 18 patients with traumatic brain injury.According to the follow-up CRS-R score,patients with aDOC were divided into emergence from minimally conscious state(eMCS)group(n=13)and prolonged disorders of consciousness(pDOC)group(n=24).The clinical data of two groups of patients were compared.Based on the MATLAB platform,fMRI data were analyzed for hippocampal functional network FC values.The internal SVM code in MATLAB was used for machine learning,while leave-one-out was used for cross validation.Receiver operating characteristic(ROC)curve was adopted to demonstrate predictive performance.Results There was no statistically significant difference in clinical and demographic data between the two groups of patients(P>0.05).There were statistically significant differences in Glasgow Coma Scale score[(9.0±1.8)vs(.6.0±2.1)]and Full Outline of Unresponsiveness Scale score[13.00(11.00,13.00)vs.10.00(8.25,11.75)]between eMCS group and pDOC group(t=3.67,Z=-3.24;P<0.01).The comparative sequence hippocampal network analysis of blood oxygen level dependent(BOLD)in fMRI showed that there were statistically significant differences in brain activity between the two groups of patients in the bilateral anterior cingulate cortex(t=4.632,P<0.005,TFCE corrected)and the right lingual gyrus(t=3.940,P<0.005,TFCE corrected).The ROC curve of the SVM model based on the differences in FC values of the hippocampal network in all brain regions using fMRI data showed that the area under the ROC curve(AUC)was 0.85,the sensitivity was 0.69,and the specificity was 0.83.The AUC,accuracy,sensitivity,and specificity of the SVM model based on the anterior cingulate cortex in different brain regions were 0.88,81.08,0.86,and 0.83,respectively.The AUC,accuracy,sensitivity,and specificity of the SVM model based on the right lingual gyrus in different brain regions were 0.75,70.27,0.77,and 0.71,respectively.Conclusions There are differences in the activity of fMRI hippocampal networks in the bilateral anterior cingulate cortex and right lingual gyrus of aDOC patients with different outcomes.Based on these differences in brain regions,the machine learning model can be constructed to accurately predict the outcome of aDOC patients,which provides ideas and targets for exploring the recovery mechanism and treatment of aDOC.