Machine Learning to Predict Interstage Mortality Following Single Ventricle Palliation: A NPC-QIC Database Analysis

Area of Science:

• Pediatric Cardiology
• Machine Learning in Medicine
• Congenital Heart Disease Research

Background:

• Single ventricle heart disease (SVHD) palliation involves high interstage mortality risk.
• Optimizing risk prediction is crucial for improving outcomes between surgical stages.

Purpose of the Study:

• To apply advanced machine learning algorithms for predicting interstage mortality in SVHD.
• To identify key clinical features associated with interstage mortality after Stage I palliation.

Main Methods:

• Retrospective analysis of 3267 patients undergoing Stage I palliation (2008-2019).
• Utilized logistic regression, random forest, gradient boosting, XGBoost, and LightGBM models.
• Trained models on 180 clinical features to predict interstage mortality.

Main Results:

• Identified 208 interstage deaths (6.4%). Digoxin use at discharge was associated with lower mortality risk.
• Blalock-Taussig-Thomas shunt increased risk compared to Sano conduit.
• Non-modifiable risks: female sex, lower gestational age, lower birth weight. Post-operative risks: unplanned catheterization, severe atrioventricular valve insufficiency.

Conclusions:

• Machine learning effectively identified modifiable and non-modifiable risk factors for interstage mortality.
• Despite identifying key factors, predictive model performance was modest, indicating potential unmeasured confounders.