Unified comparison of machine learning paradigms for blood transfusion prediction in pediatric congenital heart surgery

<h2>Summary</h2> Blood transfusion prediction studies in pediatric cardiac surgery have employed direct regression, two-stage, and multi-class classification paradigms, but the fundamentally different outputs of these paradigms have prevented direct head-to-head comparison. We developed a dual-mean absolute error (MAE) composite metric that converts all predictions to continuous transfusion volume estimates, enabling unified comparison across paradigms. Using data from 3,342 pediatric patients undergoing congenital heart surgery, we evaluated six imputation methods, three class-balancing strategies, and fifteen machine learning algorithms. Iterative regression imputation achieved the highest performance (AUC = 0.864), and synthetic minority oversampling technique (SMOTE) combined with undersampling improved F1 by 37% (RBC) and 19.5% (plasma). The two-stage approach yielded the best composite scores for RBC (0.791; Neural Network + Extra Trees) and plasma (0.755; AdaBoost + KNN), whereas multi-class LightGBM was optimal for platelets (0.652; AUC=0.960). SHAP analysis identified cardiopulmonary bypass as the dominant predictor for RBC/plasma and Aristotle score for platelets. The framework supports evidence-based, blood-product-specific model selection.