Related Experiment Video
Updated: Apr 2, 2026
Analysis of Congenital Heart Defects in Mouse Embryos Using Qualitative and Quantitative Histological Methods
Published on: March 10, 2020
Explainable machine learning model for identifying key risk factors in congenital heart disease prediction using
Yunyan Wang1,2, Tiezheng Li2, Jiajia Chen2
1Heart Center, Children's Hospital, Zhejiang University School of Medicine, Zhejiang Key Laboratory of Neonatal Diseases, National Clinical Research Center for Children and Adolescent's Health and Diseases, Hangzhou, 310052, China.
Background:
Current neonatal congenital heart disease (CHD) screening strategies face significant challenges in low-income or underdeveloped regions due to a shortage of experienced physicians. Therefore, we aim to develop a cost-effective method to identify high-risk populations, supplementing neonatal screening.
Methods:
A multicenter case-control study was conducted in three hospitals in Zhejiang, China, from September 2022 to July 2024. Pregnant women were surveyed using a self-designed questionnaire, and newborns were diagnosed using echocardiography. We utilized three steps to establish a prediction model for neonatal CHD. Initially, LASSO regression was used to screen variables. Subsequently, five representative machine learning (ML) algorithms, including SVM, XGBoost, Random Forest (RF), Logistic Regression (LR) and LightGBM, were applied to establish the CHD risk prediction models. Finally, the Shapley Additive exPlanation (SHAP) method and logistic regression were adopted to identify key risk factors for CHD.
Results:
A total of 1,633 mother-infant pairs were included, with 437 infants diagnosed with CHD. Among the five machine learning models, XGBoost showed the best performance, with an AUC of 0.724 on the internal test set and 0.706 on the external validation cohort. The internal AUCs for RF, SVM, LightGBM, and LR were 0.673, 0.647, 0.672, and 0.649, respectively. The key risk factors for CHD were identified as COVID-19 infection within the three months prior to the last menstrual period, gestational medication use and viral infections during pregnancy. The odds ratios (ORs) with 95% confidence intervals (CIs) were 8.31 (4.86-14.72), 2.24 (1.65-3.03), 1.49 (1.13-1.97), respectively.
Conclusions:
Our CHD prediction model is non-invasive, cost-effective, and demonstrates robust performance, making it a valuable supplement to routine neonatal screening for identifying newborns at high risk of CHD. The ML model may contribute to guiding better clinical risk assessment and decision making for neonatal CHD.
Clinical Trial Registration:
Not applicable.
