Supplementary Material for: A BNP-BASED MACHINE LEARNING MODEL FOR EARLY HEMODYNAMIC SYMPTOMATIC PDA PREDICTION

Introduction. Spontaneous ductal closure is common in preterm populations, however a subset of infants develops a hemodynamically significant PDA (hsPDA), which has been associated with adverse outcomes. The objective was to develop and internally validate a predictive model for (hsPDA) in preterm infants using a machine learning approach.
Methods. A prospective cohort study including infants born at <33 weeks of gestation. B-type natriuretic peptide (BNP) levels within the first 120 hours, gestational age, birth weight, and surfactant use were used to train a Random Forest classifier. The outcome was hsPDA diagnosed by standardized echocardiography. Model performance was assessed using stratified 5-fold cross-validation.
Results. Sixty-seven infants were included; 46.3% had hsPDA. The Random Forest model achieved an AUC of 0.86, outperforming logistic regression using BNP alone (AUC 0.82). BNP was the strongest predictor (48% importance), followed by gestational age, birth weight, and surfactant use.
Conclusion. A machine learning-based model combining BNP with clinical variables showed high accuracy in predicting hsPDA. The accompanying calculator may assist clinicians in early risk stratification, though external validation is required before clinical implementation.