摘要
目的开发一种基于CT灌注成像(CTP)技术的机器学习模型来预测急性缺血性卒中机械取栓治疗不良预后的风险。方法回顾性收集2020年1月至2024年6月于郑州市中心医院接受机械取栓术的462例急性缺血性脑卒中患者的相关影像学及临床资料。所有患者均随访90 d,采用m RS评分评估预后,其中175例患者预后不良(>2分),287例预后良好(0~2分)。将数据集按8∶2的比例划分为训练集(n=370)和验证集(n=92)。采用基于梯度提升的XGBoost算法作为变量筛选工具,对基线数据、实验室检查数据及影像学特征进行特征选择。分别以极限梯度提升(XGBoost)、逻辑回归(LR)、高斯朴素贝叶斯(GNB)、支持向量机(SVM)及轻量梯度提升机(LightGBM)五种机器学习算法构建急性缺血性卒中机械取栓治疗不良预后的预测模型。采用准确性、灵敏度、特异性、曲线下面积(AUC)、查准查全曲线(PR)、校准曲线和决策(DCA)曲线评价模型的有效性。结果与预后良好组相比,预后不良组患者男性占比、高血压、糖尿病、房颤病史比例、吸烟饮酒史及术后脑出血发生率、梗死核心体积及胆固醇、甘油三酯、NIHSS评分、同型半胱氨酸、肌酐、NT-proBNP、D2聚体、纤维蛋白原水平-α角、预测纤溶指数增高,而缺血半暗带体积、侧支循环良好占比、血块强度-最大振幅及血凝块力学强度-G值则低于预后良好组,差异均具有统计学意义(P<0.05)。XGboost机器学习筛选结果显示,NIHSS评分、总胆固醇、梗死核心体积、半暗带体积是急性缺血性卒中机械取栓患者不良预后的贡献度最高变量。在多种机器学习模型比较中,XGBoost模型对不良预后的预测能力最好,XGBoost模型在训练集的AUC为0.941(95%CI:0.920~0.962),验证集的AUC为0.875(95%CI:0.770~0.979),PR曲线也显示XGBoost模型的优异性能,在训练集的PR曲线下面积为0.961(95%CI:0.954~0.968),验证集的PR曲线下面积为0.910(95%CI:0.887~0.934)。结论构建了一种基于XGboost算法的预测急性缺血性卒中机械取栓治疗后不良预后的机器学习模型,该模型可有效帮助临床医生做出临床决策和实施个性化的治疗措施。
Objective To develop a machine learning model based on computed tomography perfusion(CTP)imaging for predicting the risk of poor outcomes following mechanical thrombectomy(MT)in patients with acute ischemic stroke(AIS).Methods Relevant clinical and imaging data of 462 AIS patients who underwent MT at Zhengzhou Central Hospital from January 2020 to June 2024 were retrospectively collected.All patients were followed up for 90 days,and functional outcomes were assessed using the modified Rankin Scale(mRS);poor outcomes were defined as mRS>2(n=175),while favorable outcomes were defined as mRS 0-2(n=287).The dataset was split into a training set(n=370)and a validation set(n=92)in a 8∶2 ratio.The gradient-boosting-based XGBoost algorithm was employed for feature selection from baseline characteristics,laboratory findings,and CTP imaging parameters.Five machine learning algorithms—extreme gradient boosting(XGBoost),logistic regression(LR),Gaussian naive bayes(GNB),support vector machine(SVM),and light gradient boosting machine(LightGBM)—were utilized to construct predictive models for poor outcomes after MT.Model performance was evaluated using accuracy,sensitivity,specificity,area under the receiver operating characteristic curve(AUC),precision-recall(PR)curve,calibration curve,and decision curve analysis(DCA).Results Compared with the favorable outcome group,patients with poor outcomes exhibited significantly higher proportions of males and histories of hypertension,diabetes mellitus,atrial fibrillation,smoking,and alcohol consumption(all P<0.05).They also had higher rates of postoperative intracranial hemorrhage,larger infarction core volumes,and elevated levels of total cholesterol,triglycerides,National Institutes of Health Stroke Scale(NIHSS)scores,homocysteine,creatinine,NT-proBNP,D-dimer,fibrinogenα-angle,and predicted fibrinolytic indices(all P<0.05).Conversely,the poor outcome group had smaller volumes of ischemic penumbra,lower rates of good collateral circulation,reduced clot strength(maximum amplitude),and lower clot elastic modulus(G-value)compared with the favorable outcome group(all P<0.05).Feature importance analysis using XGBoost identified NIHSS score,total cholesterol,infarction core volume,and penumbra volume as the top predictors of poor outcomes in AIS patients undergoing MT.Among the evaluated models,XGBoost demonstrated superior predictive performance,with an AUC of 0.941(95%CI:0.920-0.962)in the training set and 0.875(95%CI:0.770-0.979)in the validation set.The PR curve further confirmed its excellence,with areas under the curve of 0.961(95%CI:0.954-0.968)and 0.910(95%CI:0.887-0.934)in the training and validation sets,respectively.Conclusion An XGBoost-based machine-learning model was constructed to predict poor outcomes after mechanical thrombectomy in acute ischemic stroke.This model can effectively assist clinicians in clinical decision-making and in formulating individualized treatment measures.
作者
薛晓娟
杨苗
吴玥
苏慧
董帅珂
肖新广
XUE Xiao-juan;YANG Miao;WU Yue;SU Hui;DONG Shuai-ke;XIAO Xin-guang(Department of Medical Imaging,Zhengzhou Central Hospital,Zhengzhou 450000,Henan,CHINA)
出处
《海南医学》
2025年第11期1610-1617,共8页
Hainan Medical Journal
基金
河南省郑州市医疗卫生领域科技创新指导计划项目(编号:2024YLZDJH253)。
