Using Machine Learning and Feature Importance to Identify Risk Factors for Mortality in Pediatric Heart Surgery

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Congenital heart defects (CHDs) are common congenital malformations and a leading cause of mortality from birth defects.
• Pediatric cardiac surgery for CHDs carries significant mortality risks.
• Identifying predictive factors for mortality is crucial for improving patient outcomes.

Purpose of the Study:

• To identify risk factors for mortality following surgery for congenital heart defects (CHDs) in pediatric patients.
• To apply machine learning (ML) models for predicting mortality after CHD surgery.
• To utilize model explainability tools to understand ML predictions.

Main Methods:

• Retrospective monocentric study of 1302 pediatric patients (<18 years) undergoing CHD-related cardiac surgery (2011-2020).
• Application of Random Survival Forest (RSF) and eXtreme Gradient Boosting (XGB) algorithms to model mortality.
• Validation of models on an independent holdout test dataset (40% of cohort).
• Feature importance assessment using SHapley Additive exPlanations (SHAP) and SurvSHAP(t).

Main Results:

• Both RSF and XGB models demonstrated strong predictive performance (C-indices of 0.85 and 0.79, respectively).
• Machine learning models identified maximum serum creatinine levels within 72 hours post-surgery as a critical predictor of mortality.
• Model explainability tools (SHAP, SurvSHAP(t)) confirmed the importance of early postoperative serum creatinine.

Conclusions:

• Machine learning methods effectively identify mortality risk factors in pediatric CHD surgery patients.
• Early postoperative serum creatinine is a significant modifiable risk factor for mortality.
• The analytical workflow provides a foundation for federated learning in medical research.