【作者】 刘玲玲；

【导师】 许光旭；

【作者基本信息】 南京医科大学 ， 康复医学与理疗学（专业学位）， 2023， 博士

【摘要】 背景:脑卒中是我国第三大死亡原因,深静脉血栓形成(DVT)是卒中后常见致命性并发症。目前尚无可靠的专门针对脑卒中患者下肢DVT的风险评估量表及功能预防措施。因此,建立专门针对卒中后下肢DVT的风险模型至关重要。近年来,人工智能(AI)有望帮助临床医生确定标准化的诊断方法,优化风险预测,提供精准医疗。目的:本研究旨在基于人工智能算法,首先从患者不同功能层面来开发和验证卒中后下肢DVT的风险评估模型,并与传统Padua风险评估模型进行性能比较;然后从患者可干预的功能角度出发,建立一个急性卒中后下肢DVT的风险预后模型。通过阐明模型的特征重要性,构建在线网页计算器预测平台,来完善脑卒中专科DVT风险评估体系。方法:在第一部分中,回顾性收集2015年12月1日至2019年2月28日的脑卒中患者作为开发队列,前瞻性收集2022年10月1日至2023年5月31日的脑卒中患者作为外部验证队列。根据是否存在血栓将患者分为DVT组和非DVT组,并进行数据预处理。采用共线性诊断及最小绝对收缩和选择算子(LASSO)筛选特征变量。选择逻辑回归(LR)、支持向量机(SVM)、随机森林(RF)、决策树(DT)、神经网络(NN)、极限梯度提升(XGBoost)、朴素贝叶斯(NB)及K最近邻(KNN)八种人工智能算法分别对训练集及外部验证集进行测试,十倍交叉验证后,使用ROC曲线、AUC、PR曲线、PRAUC、准确率、灵敏度、特异度分别对模型进行综合评估,应用临床决策曲线(DCA)来验证模型的临床适用性。引入Shapley加性解释(SHAP)来解释最优AI算法模型,并根据预测模型中特征贡献程度进行排序。基于最优AI算法,将不同的特征集模型与Padua评分模型进行比较,最后将模型应用于临床。在第二部分中,为评估急性卒中后下肢DVT的预后风险,回顾性纳入2015年12月1日至2023年5月31日期间入院的140例发病15天内的脑卒中患者。通过利用共线性诊断和LASSO回归技术进行变量筛选。采用梯度提升机(GBM)、随机生存森林(RSF)、广义线性模型(GLM)确定急性卒中后DVT的最佳生存模型。采用一致性指数(C指数)、累积/动态AUC指标(C/D AUC)和综合Brier评分(IBS)将最优模型与传统的Cox比例风险模型进行比较。为更深入了解观测值与模型输出之间的关系,通过Surv SHAP(t)进一步分析最优模型。最后,将生存预测模型应用于临床。结果:1.在第一部分研究中,最终纳入620名符合条件的患者。所有患者根据数据集分为训练集(417人)和外部验证集(203人)。训练集中,28.78%(120/417)是DVT患者;外部验证集中,38.92%(79/203)是DVT患者。经共线性诊断和LASSO回归分析后,共有10个变量被确定为显著预测因子:包括年龄、卒中类型、性别、糖尿病史、饮酒史、步行能力、Brunnstrom分期(下肢)、抗血小板药物服用史、凝血酶原时间和D-二聚体。RF模型在模型训练集和外部验证集上均具有最高的AUC(0.74/0.73)、PRAUC(0.58/0.58)、准确率(0.75/0.77)、灵敏度(0.78/0.80),其特异度在训练集和外部验证集中表现良好(0.92/0.92)。此外,DCA分析显示,RF模型对临床净效益的适用性最高,提示RF模型是预测卒中后DVT的最优AI算法模型,最具有临床适用性。SHAP模型解释性分析显示,D-二聚体对DVT的影响最为显著,其次是年龄、Brunnstrom分期(下肢)、凝血酶原时间、步行能力。基于RF算法,本研究构建了包含不同数量特征集的四个模型。综合分析表明,仅依赖Padua评分的模型性能最差。性能良好的特征集模型以在线网络计算器的形式呈现,以便于临床应用。2.在第二部分研究中,共收集急性脑卒中患者140例。采用共线性诊断、LASSO回归后选取年龄、性别、糖尿病史、Brunnstrom分期(下肢)、站立平衡5个变量进行预测建模。Ranger随机生存森林模型在IBS(0.04)、C/D AUC(0.52)及C指数(0.87)方面表现最优。变量重要性排序-时间依赖折线图及Surv SHAP(t)图分析显示,年龄、Brunnstrom分期(下肢)是DVT发生的主要贡献变量。部分依赖图及Ceteris Paribus图表明,年龄增长和较低的Brunnstrom分期(下肢)与DVT的发生风险增加有关。风险预后模型以网络计算器形式可视化,以指导临床应用。结论:本研究基于人工智能算法,建立了专门针对卒中后下肢DVT的风险评估和预后模型。随机森林模型及Ranger随机生存森林模型综合性能表现最好,而传统Padua评分模型不足以准确评估卒中后DVT的风险。年龄、Brunnstrom分期(下肢)是脑卒中后DVT风险预测的主要贡献变量。更多还原

【Abstract】 Background:Stroke is the third leading cause of death in China,and Deep Vein thrombosis(DVT)is a common vascular disease after stroke,which can lead to fatal complications.Currently,there is a lack of reliable risk assessment scales and effective preventive measures for lower limb DVT in stroke patients.Consequently,it is imperative to develop risk models specifically for lower extremity DVT after stroke.In recent times,the utilization of artificial intelligence(AI)is expected to help clinicians establish standardized diagnostic approaches,enhance risk prediction accuracy,and facilitate the delivery of personalized medicine.Objectives:This study seeks to develop and validate a risk assessment model for lower limb DVT after stroke based on artificial intelligence algorithms.The model will be compared to the traditional Padua risk assessment model.Additionally,a prognostic model for the risk of lower limb DVT after acute stroke will be established,focusing on the intervenable limb function of patients.To facilitate the implementation of this risk assessment system by elucidating the significant predictive features of the model and constructing an online webpage calculator prediction platform.Methods:For the initial segment,stroke patients were retrospectively gathered from December1,2015 to February 28,2019,constituting the development cohort.Additionally,stroke patients were prospectively collected from October 1,2022 to May 31,2023,forming the external validation cohort.These patients were classified into DVT and non-DVT groups based on the presence or absence of thrombus,and subsequent data preprocessing was conducted.Collinearity diagnostics and the least absolute shrinkage and selection operator(LASSO)were employed for screening feature variables.A total of eight AI algorithm models,namely logistic regression(LR),support vector machine(SVM),random forest(RF),decision tree(DT),neural network(NN),extreme gradient boosting(XGBoost),Bayesian(NB),and K-nearest neighbor(KNN),were chosen for evaluating the training set and the external validation set,respectively.Following the implementation of ten-fold cross-validation,an extensive evaluation of the models was conducted,encompassing the analysis of ROC curves,AUC,PR curves,PRAUC,accuracy,sensitivity,and specificity.Additionally,the clinical decision curves(DCA)were employed to appraise the clinical applicability of the models.Shapley’s additive explanation(SHAP)is introduced to explain feature importance according to the degree of contribution in the prediction model.Finally,based on the optimal AI algorithm,different feature set models are compared with the Padua scoring model to select the best feature set model.In the second part,to evaluate the prognostic risk of DVT in acute stroke patients,a total of 140 consecutive stroke patients admitted within 15 days of symptom onset,from December 1,2015 to May 31,2023,were retrospectively included in this study.Variable selection was conducted through the utilization of collinearity diagnostics and the LASSO regression technique.The optimal survival model for assessing the risk of DVT following acute stroke was determined through the utilization of the gradient boosting algorithm(GBM),random survival forest(RSF),and generalized linear model(GLM).The comparison between the performance of the optimal model and the conventional Cox regression model was conducted using the consistency index(C index),cumulative/dynamic AUC(C/D AUC),and integrated Brier score(IBS).To gain a deeper understanding of the association between observations and model outputs,the optimal model was further analyzed through Surv SHAP(t)analysis.Results:(1)In the first part of the study,620 eligible patients were eventually included.All patients were divided into a training set(417 participants)and an external validation set(203 participants)based on the data set.In the training set,28.78%(120/417)were DVT patients;in the external validation set,38.92%(79/203)were patients with DVT.Following the examination of collinearity diagnostics and LASSO regression analysis,a total of 10 variables were identified as significant predictors.These variables encompassed age,stroke type,gender,history of diabetes,alcohol history,walking ability,Brunnstrom stage(lower limb),history of antiplatelet aggregation drugs,prothrombin time,and D-dimer.The RF algorithm demonstrates superior performance in various evaluation metrics,including AUC(0.74/0.73),PRAUC(0.58/0.58),accuracy(0.75/0.77),and sensitivity(0.78/0.80),in both the model training set and test set.Additionally,its specificity(0.92/0.92)exhibits commendable results in both the training and test sets.Furthermore,the DCA analysis reveals that the RF model exhibits the highest applicability to net clinical benefit,thereby establishing it as the optimal artificial intelligence algorithm for predicting DVT after stroke.The interpretative analysis of the SHAP model revealed that D-dimer exerted the most significant influence on DVT,followed by age,Brunnstrom stage(lower limb),prothrombin time,and walking ability.Utilizing the RF algorithm,we constructed four distinct models incorporating varying sets of features,the model solely reliant on the Padua score exhibited the poorest performance.Consequently,we present the well-performing feature set model in the form of an online web calculator.(2)In the second part of the study,140 acute stroke patients were recruited.Five variables of age,gender,history of diabetes,Brunnstrom staging(lower limb),and standing balance were selected for prediction modeling using collinearity diagnostics and LASSO regression.The Ranger randomized survival forest model demonstrated superior performance,exhibiting optimal results in terms of the IBS(0.04),C/D AUC(0.52),and C-index(0.87).The analysis of time-dependent feature importance and Surv SHAP(t)plot revealed that age and Brunnstrom staging(lower limb)were the primary contributing variables.Furthermore,the examination of partial dependence survival profiles and ceteris paribus plots indicated that advancing age and lower Brunnstrom staging(lower limb)were associated with an increased risk of developing DVT.The risk prognostic model was effectively visualized as a network calculator.Conclusions:This study utilized AI algorithms to establish a risk assessment and prognosis model for lower limb deep vein thrombosis following stroke.The random forest model and Ranger random survival forest model demonstrated superior overall performance in this study,whereas the Padua scoring model proved inadequate for accurately assessing the risk of post-stroke DVT.Notably,age and Brunnstrom stage(lower limb)emerged as significant contributing variables in the prediction of post-stroke DVT.更多还原