基于机器学习构建DSA影像学模型对缺血性脑卒中大血管病变病因分类的评估价值  

The evaluation value of DSA imaging model based on machine learning in the etiological classification and prediction of ischemic stroke macrovascular lesions

作　　者：高立源 杨闯 柳国斌[3] 樊炜静 黄仁燕 游洋 Gao Liyuan;Yang Chuang;Liu Guobin;Fan Weijing;Huang Renyan;You Yang(First Clinical College,Liaoning University of Traditional Chinese Medicine,Shenyang 110847,China;Intervention Center,the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine,Shenyang 110847,China;Department of Vascular Surgery,Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine,Shanghai 200021,China)

机构地区：[1]辽宁中医大学第一临床学院,沈阳110847 [2]辽宁中医药大学附属医院介入中心,沈阳110847 [3]上海中医药大学附属曙光医院血管外科,上海200021

出　　处：《中国临床实用医学》2025年第1期1-6,共6页China Clinical Practical Medicine

基　　金：国家自然科学基金(82274528);辽宁省科学技术计划项目(2023-MS-231)。

摘　　要：目的基于机器学习构建数字减影血管造影(DSA)影像学模型,探讨其对缺血性脑卒中大血管病变病因分类的评估价值。方法选取辽宁中医药大学附属医院介入中心2021年7月至2024年7月收治的53例接受DSA下动脉血管再通治疗的缺血性脑卒中患者,均为大动脉粥样硬化型、心源性栓塞型等大血管病变,男21例,女32例,年龄(61.5±8.3)岁,年龄范围为45~78岁。采用随机数字表法按照7∶3的比例将53例患者随机分为训练组(n=37)与内部验证组(n=16)。另选取辽宁中医药大学附属医院介入中心2024年7—11月收治的20例缺血性脑卒中患者作为外部验证组,男8例,女12例,年龄(63.0±5.0)岁,年龄范围为53~73岁。提取训练组与内部验证组患者治疗前、治疗后的DSA影像组学特征,假定"栓塞"为正样本、"狭窄"为负样本。通过Spearman相关性分析降低特征间的冗余度,而后进行Wilcoxon秩和检验,筛选P值较小的前5个特征。训练组将筛选出的特征数据通过深睿医疗多模态平台在全局超参模式下采用三种机器学习模型进行训练和调优,并通过内部验证组的特征数据进行模型性能分析,筛选出效能最佳模型。导出最终建立的模型并采用外部验证组患者DSA影像进行外部验证。结果依据训练组数据建立的模型中,SVM、LR、BernoulliNB模型鉴别病因的受试者操作特征(ROC)曲线下面积分别为0.94(95%CI:0.83~1.00)、0.87(95%CI:0.74~0.99)、0.80(95%CI:0.66~0.95)。内部验证结果显示,SVM模型鉴别病因的效能最佳,ROC曲线下面积为0.89(95%CI:0.73~1.00),故采用SVM模型作为最终模型。SVM模型解释及可视化分析结果显示,"原始-形态特征-延长率-开通后"、"原始-形态特征-短轴长度-开通前"这两种特征的数值越大越倾向于病因为正样本,即"栓塞";反之则更倾向于病因为负样本,即"狭窄"。外部验证结果显示,SVM模型鉴别病因的ROC曲线下面积为0.83(95%CI:0.64~1.00),�ObjectiveTo construct a digital subtraction angiography(DSA)imaging model based on machine learning,and to explore the value of the model in classifying and evaluating the etiology of macrovascular lesions in ischemic stroke.MethodsA total of 53 patients with ischemic stroke who received artery recanalization treatment under DSA in the Intervention Center of the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine from July 2021 to July 2024 were selected,and all of them had macrovascular lesions such as atherosclerosis and cardiogenic embolization,with 21 males and 32 females aged(61.5±8.3)years old,the age ranging from 45 to 78 years old.The selected 53 patients were randomly divided into the training group(n=37)and the internal validation group(n=16)by random number table method with a ratio of 7∶3.In addition,20 patients with ischemic stroke admitted to the Intervention Center of the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine from July 2024 to November 2024 were selected as the external verification group,including 8 males and 12 females,aged(63.0±5.0)years old,the age ranging from 53 to 73 years old.The DSA imaging features before and after the treatment of patients in the training group and the internal verification group were extracted.The embolization was assumed to be a positive sample and the stenosis was assumed to be a negative sample.Spearman correlation analysis was used to reduce the redundancy among the features,and then Wilcoxon rank sum test was used to screen the first 5 features with small P-values.The selected feature datas of the training group were used to train and tune the three machine learning models through the Shengrui Medical multi-modal platform in the global hyper-parametric mode,the model performance was analyzed through the feature datas of the internal verification group to select the model with the best performance.The final model was exported and the DSA images of patients in the external verification group were used fo

关 键 词：数字减影血管造影 缺血性脑卒中 机器学习 病因分类评估 

分 类 号：R743.3[医药卫生—神经病学与精神病学]