Machine learning-based prediction of hemodynamically significant patent ductus arteriosus in preterm neonates: a pioneering insight.

Background: Hemodynamically significant patent ductus arteriosus (hPDA) in premature infants is a common congenital cardiac anomaly associated with substantial morbidity and mortality. Traditional diagnostic methods like echocardiography face challenges such as expertise requirement and inconsistent accessibility. This study investigates the efficacy of the Random Forest machine learning model in predicting hPDA in premature infants, aiming to provide a non-invasive, objective, and reliable alternative. Methods: This retrospective study analyzed data from 657 premature infants hospitalized between 2014 and 2019. Patients were categorized into hPDA and asymptomatic PDA (aPDA) groups. The Random Forest classification model, implemented in JASP software, utilized prenatal, natal, and postnatal clinical data, including gestational week, birth weight, and the need for resuscitation at birth. Model performance was assessed using metrics such as accuracy, Area Under the Curve, F1 score, Matthews Correlation Coefficient, recall, precision, and feature importance. Results: The Random Forest model demonstrated strong predictive performance, achieving a test accuracy of 91.7%, an AUC of 0.950, an F1 score of 0.923, and an MCC of 0.775. Notably, the recall for the hPDA group was 100%. Gestational week, birth weight, and the need for resuscitation at birth were identified as the most significant predictors. The model also revealed complex relationships, showing variables deemed statistically insignificant by classical methods (e.g., gender, 5th-minute APGAR score, oligohydramnios) to be significant within the Random Forest framework. Conclusions: The Random Forest model effectively predicts hPDA risk in premature infants, offering superior predictive power compared to classical statistical analyses. This approach has the potential to enhance early detection, facilitate timely interventions, and support personalized treatment strategies, thereby improving patient outcomes. Further validation through large-scale, multi-center prospective studies is essential for its integration into clinical practice.