【作者】 陈浩；

【导师】 宋兵；

【作者基本信息】 兰州大学 ， 临床医学（专业学位）， 2023， 硕士

【摘要】 背景:急性Stanford A型主动脉夹层(ATAAD)属于极其危重的疾病之一,近年发病率不断上升。由于手术难度大、时间长、涉及体外循环等,术后易发生各种并发症,涉及器官范围广,其中尤以脑部并发症多见。目的:本研究旨在使用机器学习方法构建ATAAD术后脑部并发症的最佳预测模型,为早期识别高风险患者提供支持,并基于SHAP对模型提供解释,探究相关危险因素,为临床提供一定的参考依据。方法:本项研究回顾性纳入2017年1月--2021年9月就诊于兰州大学第一医院心血管外科的ATAAD患者,严格按照纳排标准筛选资料。首先清洗数据、编码及分组,对全部数据标准化处理,以7:3随机划分全部数据为训练集和测试集。在训练集内,对无序多分类变量引入虚拟变量(DV),通过最小绝对收缩和选择算子(LASSO)回归筛选最佳特征,引入过采样和欠采样方式相结合的SMOTEEN算法平衡数据集,基于Logistic回归(LR)、K最邻近(KNN)、随机森林(RF)、梯度提升机(GBM)、支持向量机(SVM)、全连接神经网络(MLP)构建预测模型,使用K折交叉验证或超参数网格搜索进一步调优参数,以准确度(Accuracy)为标准确定最佳参数并代入模型。完成后在测试集内验证各模型性能,绘制各模型受试者工作特征曲线(ROC),采用Brier评分(BS)、准确度(Accuracy)、灵敏度(Sensitivity)、特异度(Specificity)、F1分数(F1 score)以及曲线下面积(AUC)综合评价模型效能,以AUC最大者为最佳模型,进一步行Hosmer-lemeshow检验(HL),绘制决策曲线分析(DCA)明确模型的净收益。采用SHAP(Shapley additive explanations)算法解释模型并可视化。结果:(1)Lasso回归确定的最佳特征集为:术后ALT值、手术总时间、术中冷沉淀的输注量、术后LDH值、头臂干受累及情况、术后尿素值、年龄和高血压。(2)确定的最佳模型为RF模型,AUC值最大,为0.828[95%CI(0.585,0.902)],F1分数(F1 Score)为0.667[95%CI(0.375,0.959)],灵敏度为0.833[95%CI(0.535,0.932)],特异度为0.892[95%CI(0.792,0.992)],BS为0.139[95%CI(0.035,0.244)],综合性能良好。在HL检验中P=0.183,提示RF模型拟合较好。DCA提示在较大的阈值范围内通过RF模型可获得较高净收益。(3)基于SHAP算法解释模型,提示RF模型中对预测结果影响最大的特征为术后ALT水平,其次分别为手术总时间和术中输注冷沉淀。其中术后ALT水平、手术总时间对模型的预测结果起正向贡献,输注冷沉淀对模型起负向贡献。结论:(1)术后ALT值、手术总时间、术后LDH值、头臂干受累及、术后尿素值、年龄和高血压史为ATAAD术后脑部并发症的危险因素,对于排名靠前的相关因素需给予足够重视;(2)术中适量输注冷沉淀可以减少脑部并发症的发生,未来可通过更大样本数据进一步验证;(3)在本研究中,相比Logistic回归,RF算法模型具有更好的预测能力,综合性能更优,提示机器学习算法在数据挖掘与分析中具有独特优势;在本研究中RF模型可及时识别ATAAD术后脑部并发症高风险患者,为早期提供干预、改善患者预后提供支持。更多还原

【Abstract】 Background:Acute type A aortic dissection is one of the most serious diseases with an increasing incidence in recent years.Due to the difficulty of operation,long operation time,involving extracorporeal circulation and so on,postoperative complications are easy to occur,involving a wide range of organs,especially in the cerebral complications.Objective:This study aims to use machine learning method to build the best prediction model of cerebral complications after ATAAD surgery,offer support for early identification of high-risk patients,provide explanations to the model based on SHAP,explore related risk factors,and provide reference for clinical work to a certain extent.Methods:In this study,patients with ATAAD received treatment in the Department of Cardiovascular Surgery of the First Hospital of Lanzhou University from January 2017 to September 2021 were retrospectively included,and the data were screened strictly according to inclusion and exclusion criteria.Firstly,the data were cleaned,coded and grouped,all the data were standardized.All the data were randomly divided into training set and test set by 7:3.In the training set,dummy variables were introduced for unordered multicategorical variables,and the best features were screened by least absolute shrinkage and selection operator regression.A SMOTEEN algorithm combining oversampling and undersampling was introduced to balance data set.Prediction models were constructed based on Logistic regression,K-nearest neighbor,Random forest,Gradient boosting machine,Support vector machine and Multi-layer perception.Further,parameters were tuned using K-folds cross validation or hyperparameter grid search,and the optimal parameters were determined by Accuracy and substituted into the models.Next,the performance of each model was verified in the test set.The receiver operating characteristic curves of all models were plotted.Brier score,Accuracy,Sensitivity,Specificity,F1 score and area under the curve were used to evaluate the comprehensive efficacy of the models.The model with the largest area under the curve was regarded as the best model.Hosmer-lemeshow test was further conducted for the best model,and decision curve analysis was drawn to clarify the net income of the model.Shapley additive explanations algorithm was used to explain the model and the results were visualized lastly.Results:(1)The best features set determined by Lasso regression included: postoperative ALT value,total operative time,intraoperative cryoprecipitate infusion amount,postoperative LDH value,conditions of brachiocephalic trunk artery,postoperative urea value,age and hypertension.(2)The best model was RF model with the maximum AUC value of 0.828[95%CI(0.585,0.902)],and F1 Score,sensitivity,specificity,Brier Score was0.667[95%CI(0.375,0.959)],0.833[95%CI(0.535,0.932)],0.892[95%CI(0.792,0.992)],0.139[95%CI(0.035,0.244)],respectively.The comprehensive performance of RF was great.In Hosmer-lemeshow test,P value was 0.183,suggesting that the RF model was well fitted.DCA suggested that high net income can be obtained through RF model within a larger threshold range.(3)The model was interpreted based on the SHAP algorithm,and the results suggested that the most influential feature in the RF model on the prediction results was the postoperative ALT value,and the next two were the total operative time and the intraoperative cryoprecipitate infusion amount,respectively.The postoperative ALT value and the total operative time contributed positively to the prediction results of the model,and the intraoperative cryoprecipitate infusion amount contributed reversely to the model.Conclusion:(1)Postoperative ALT value,total operative time,postoperative LDH value,involvement of brachiocephalic trunk artery,postoperative urea value,age and hypertension are the risk factors of cerebral complications after ATAAD.Enough attention should be paid to the factors at the top of the list.(2)Appropriate infusion of cryoprecipitate during operation can reduce the occurrence of cerebral complications after ATAAD,which could be further verified by larger sample data in the future.(3)In the data of our center,RF prediction model has better prediction ability and better comprehensive performance than Logistic regression,suggesting that machine learning algorithm has unique advantages in data mining and analysis.In this study,RF model can identify high-risk patients in time and provide support for early intervention and improving patient prognosis.更多还原