Machine learning models based on contrast-transthoracic echocardiography and transesophageal echocardiography combined with clinical and laboratory indicators for predicting patent foramen ovale-associated stroke
10.13929/j.issn.1003-3289.2025.09.013
- VernacularTitle:经胸右心声学造影及经食管超声心动图联合临床及实验室指标机器学习模型预测卵圆孔未闭相关卒中
- Author:
Xiaoke ZENG
1
;
Yali XU
1
;
Yuan LIU
1
;
Hao ZUO
1
;
Chun LI
1
Author Information
1. 陆军军医大学第二附属医院超声科,重庆 400037
- Publication Type:Journal Article
- Keywords:
stroke;
foramen ovale,patent;
machine learning
- From:
Chinese Journal of Medical Imaging Technology
2025;41(9):1517-1521
- CountryChina
- Language:Chinese
-
Abstract:
Objective To develop the value of machine learning(ML)models based on contrast-transthoracic echocardiography(cTTE)and transesophageal echocardiography(TEE)combined with clinical and laboratory indicators for predicting patent foramen ovale-associated stroke(PFO-AS).Methods Totally 313 patients with PFO diagnosed with cTTE and TEE were retrospectively enrolled.Among them,65 cases were found complicated with ischemic stroke and confirmed as PFO-AS(PFO-AS group),and the rest 248 cases without ischemic stroke were classified as non-PFO-AS group.The patients were divided into training set(n=219,including 48 cases of PFO-AS and 171 cases of non-PFO-AS)and test set(n=94,including 17 cases of PFO-AS and 77 cases of non-PFO-AS)at the ratio of 7∶3.Univariable and multivariable logistic regression(LR)were used to analyze clinical and laboratory indicators as well as cTTE and TEE parameters in training set to screen independent predictive factors of PFO-AS.ML models,including LR,K-nearest neighbor(KNN),support vector machine(SVM),random forest(RF),decision tree(DT),back propagation neural network(BPNN)and gradient boosting machine(GBM)were constructed,and the predictive efficacy of the models for predicting PFO-AS was evaluated,then the optimal model was selected.Results Patient's age＞49-69 years,with smoking history,plasma albumin≥43.8 g/L,significant right-to-left shunt at rest shown on cTTE and complicated atrial septal aneurysm shown on TEE were all independent predictors of PFO-AS,which were used to construct ML models.The area under the curve(AUC)of LR,KNN,SVM,RF,DT,BPNN and GBM models in training set was 0.779-0.853,while in test set was 0.730-0.877.RF model had relatively high and comparable sensitivity,specificity and AUC in both training and test sets,also higher precision and smaller Brier score in test set,hence was regarded as the optimal ML model.Conclusion RF model based on cTTE and TEE combined with clinical and laboratory indicators could be used to effectively predict PFO-AS.
