Machine learning models based on ultrasound parameters and clinical data for predicting reversed ductus arteriosus flow in fetuses with Ebstein’s anomaly

Background: Ebstein’s anomaly (EA) is a rare congenital heart defect characterized by downward displacement of the tricuspid valve, which can lead to severe right heart dysfunction and adverse perinatal outcomes. Retrograde ductus arteriosus flow (RDAF) is an important hemodynamic marker of disease severity and is associated with a higher risk of fetal hydrops, intrauterine demise, and poor postnatal prognosis. Early and accurate identification of fetuses at risk for RDAF is essential for timely intrauterine intervention and perinatal management. We aimed to develop multiple machine learning (ML) models based on ultrasound parameters and clinical data to predict the risk of RDAF in fetuses with EA, compare their predictive performance, and provide reference for early intrauterine intervention and perinatal management.

Methods: A total of 117 fetuses diagnosed with EA at Beijing Anzhen Hospital between January 2011 and July 2023 were retrospectively enrolled and randomly divided into a training set (n=82) and a validation set (n=35) at a 7:3 ratio. Thirty-eight ultrasound parameters and seven clinical variables were collected. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression, and independent risk factors for RDAF were identified by multivariable logistic regression. Six prediction models were constructed using the selected features: logistic regression (LR), Gaussian naïve Bayes (GNB), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), and k-nearest neighbor (KNN). Model discrimination was assessed by receiver operating characteristic (ROC) curves, area under the curve (AUC), sensitivity, and specificity. Model calibration was evaluated using Brier score, and clinical utility was assessed by decision curve analysis (DCA).

Results: Gestational age, maternal age, right atrial diameter, the ratio of pulmonary valve annulus to aortic valve annulus (PA/AO), and pulmonary valve peak velocity were identified as independent risk factors for RDAF (p<0.05). All six models demonstrated good predictive performance in both the training and validation sets, with the XGBoost model achieving the best overall performance. It yielded the highest sensitivity and specificity, the lowest Brier score, and the greatest net benefit on DCA.

Conclusions: ML models based on ultrasound and clinical data can effectively predict the risk of RDAF in fetuses with EA. Among them, the XGBoost model showed the strongest discrimination, best calibration, and good generalizability, providing a reliable tool to support intrauterine intervention planning and perinatal management.

Dr. Xianfeng Guo is an ultrasound medicine specialist specializing in fetal congenital heart disease diagnosis and cardiac function assessment. He holds an MD and is Associate Chief Physician and Master’s Supervisor. Dr. Guo has led multiple research projects and published extensively. He is currently Director of the Ultrasound Department at Shanghai Seventh People’s Hospital, a visiting scholar at London’s Royal Free Hospital, and recognized as a Young Top Innovative Talent, Outstanding Youth, and Academic Leader in Shanghai’s Pudong District. He also serves on the Fetal Cardiology Committee of the Chinese Maternal and Child Health Association.