摘要
目的 基于多参数MRI构建急性脑卒中血管内治疗后出血转化风险的可解释模型。材料与方法 回顾性分析在我院就诊的急性脑卒中患者病例274例。根据治疗后24 h的CT或者MRI评估患者出血转化情况。应用PyRadiomics软件提取1143个弥散加权成像特征和1143个灌注加权成像特征,并构建影像组学评分(radiomics score,Radscore)。通过SHapley Additive exPlanations选择模型开发的最佳特征。应用6种不同的机器学习分类器[梯度提升分类器、随机森林(random forest,RF)、极限梯度提升(eXtreme gradient boosting,XGB)、自适应提升、高斯朴素贝叶斯和逻辑回归]构建出血转化风险可解释预测模型。通过受试者工作特征(receiver operating characteristic,ROC)曲线和决策曲线分析(decision curve analysis,DCA)评估机器学习模型的预测效能。结果 经特征筛选降维后共筛选出15个与急性脑卒中出血转化高度相关的特征。5个差异有统计学意义的临床变量[包含年龄、发病至MRI检查时间、入院美国国立卫生研究院卒中量表(National Institute of Health Stroke Scale,NIHSS)评分、糖尿病史、房颤史]及Radscore被纳入机器学习模型中,其中RF模型的预测效能最好,其AUC达0.928。当临界值为0.844时,其准确度为85.5%、敏感度为83.0%、特异度为88.2%。DCA显示RF模型在预测急性脑卒中出血转化风险方面具有较好的净收益。结论 多参数MRI影像组学联合临床特征的RF可解释模型可更为准确地预测急性脑卒中血管内治疗后出血转化风险,为临床早期干预治疗提供指导。
Objective:To develop an interpretable model to predict the risk of hemorrhagic transformation after endovascular treatment in acute stroke,utilizing multiparameter MRI.Materials and Methods:A retrospective analysis was conducted on 274 patients who presented with acute stroke at our hospital.The assessment of hemorrhagic transformation in these patients was performed using CT or MRI 24 hours post-treatment.Utilize the Py Radiomics software to extract 1143 features from diffusion-weighted imaging and an additional 1143 features from perfusion-weighted imaging,and develop a radiomics score(Radscore) based on these extracted features.Utilize SHapley Additive ex Planations(SHAP) to identify the most pertinent features for model development.Develop an interpretable prediction model for assessing the risk of bleeding conversion by employing six distinct machine learning classifiers:gradient boosting classifier,random forest(RF),eXtreme gradient boosting(XGB),adaptive boosting,Gaussian naive Bayes,and logistic regression.Assess the predictive performance of these machine learning models using receiver operating characteristic(ROC) curves and decision curve analysis(DCA).Results:Following feature screening and dimensionality reduction,15 features demonstrating a strong correlation with the transformation of acute ischemic stroke bleeding were identified.Five clinical variables with statistical differences(age,time from onset to MRI examination,NIHSS score on admission,history of diabetes,and history of atrial fibrillation) and radscore were incorporated into the machine learning model.Among the models evaluated,the RF model exhibited the highest predictive performance,achieving an area under the curve(AUC) of 0.928.When the critical value is set at 0.844,the model demonstrates an accuracy of 85.5%,a sensitivity of 83.0%,and a specificity of 88.2%.DCA indicates that the RF model provides a substantial net benefit in predicting the risk of hemorrhagic transformation in cases of acute stroke.Conclusions:The interpretable RF model,which integrates multiparameter MRI radiomics with clinical features,enhances the accuracy of predicting the risk of hemorrhagic transformation following mechanical thrombectomy in acute ischemic stroke.This model offers valuable guidance for early clinical intervention and treatment.
作者
于慧华
姜亮
彭明洋
耿文
殷信道
周春艳
YU Huihua;JIANG Liang;PENG Mingyang;GENG Wen;Yin Xindao;ZHOU Chunyan(Department of Radiology,Nanjing First Hospital,Nanjing Medical University,Nanjing 210006,China)
出处
《磁共振成像》
北大核心
2025年第4期19-24,共6页
Chinese Journal of Magnetic Resonance Imaging
基金
国家自然科学基金项目(编号:82202128)。
关键词
卒中
出血转化
磁共振成像
影像组学
机器学习
stroke
hemorrhagic transformation
magnetic resonance imaging
radiomics
machine learning
