Related Experiment Video
Updated: Jul 7, 2026
Creation of Patient-Specific Silicone Cardiac Models with Applications in Pre-surgical Plans and Hands-on Training
Published on: February 10, 2022
An interpretable machine learning model for preoperative risk stratification in paediatric congenital heart disease
Yu-Zhen Zhang1, Yong-Yu Si2, Jie Ouyang2
1Department of Anesthesiology, Kunming Children's Hospital, Kunming, 650100, Yunnan, China.
Background:
Congenital heart disease (CHD), one of the most common birth defects, poses challenges to preoperative risk stratification due to its anatomical complexity and developmental vulnerability. Existing tools inadequately predict critical outcomes, including mortality and ventilator dependence. We developed a machine learning-based clinical tool to enable precise preoperative risk assessment.
Methods:
This retrospective dual-centre cohort study analysed 1,363 CHD surgical cases (1,033 from Kunming Children's Hospital and 330 from Fuwai Yunnan Cardiovascular Hospital), with 78 preoperative variables collected. Feature selection using the Boruta algorithm and LASSO regression identified key predictors. Nine machine learning models, including Random Forest and XGBoost, were constructed to predict a primary composite outcome and four secondary adverse events. Model performance was assessed using AUC and F1-score with threshold optimisation, and external validation was performed using the independent cohort.
Results:
Eighteen predictors were selected. For the primary outcome, Random Forest achieved the highest AUC (0.861) and F1-score (0.631). Secondary outcomes showed divergent performance: mortality prediction demonstrated excellent discrimination (AUC 0.948), while prolonged hospital stay was predicted moderately (AUC 0.714). SHAP analysis revealed outcome-specific drivers: weight and RACHS-1 category dominated primary outcome risk, whereas mortality was associated with diastolic interventricular septum thickness and the AST/platelet ratio. Mechanical ventilation dependency correlated strongly with the monocyte-lymphocyte ratio.
Conclusion:
Five interpretable Random Forest models were developed and deployed as a web-based tool for preoperative risk stratification in CHD surgery. External validation demonstrated stable predictive performance, and future multicentre prospective studies are planned to further refine model generalisability and clinical utility.
