Se In Sung

Sungkyunkwan University School of Medicine, Korea

Abstract Title:

Implementation of an Extra Trees-Based Model for Predicting Hemodynamic Significance of Patent Ductus Arteriosus in Very Low Birth Weight Infants

Biography:

Se In Sung, MD, PhD, is a prominent clinician-scientist and Professor in the Department of Pediatrics at Samsung Medical Center, Sungkyunkwan University School of Medicine. His research focuses on neonatology, specifically the cardiovascular health and neurological outcomes of extremely preterm infants. Dr. Sung is at the forefront of integrating digital health and artificial intelligence into the NICU, aiming to develop predictive models that enhance bedside decision-making. He has authored numerous high-impact publications and leads several clinical trials dedicated to improving the survival and quality of life for VLBW infants.

Research Interests:

This research details the development of a computational monitoring framework designed to estimate the hemodynamic severity of patent ductus arteriosus (HSPDA) in very low birth weight (VLBW) infants using routinely collected clinical data. Utilizing a decade-long retrospective dataset from 1,033 neonates (3,074 clinical observations), we evaluated multiple machine-learning architectures to support real-time clinical monitoring in the NICU. The Extra Trees algorithm demonstrated superior performance, achieving an Area Under the Curve (AUC) of 0.829, with a balanced sensitivity of 0.767 and specificity of 0.766. To address the dynamic nature of ductal shunting, we introduced a Transition-Optimized Threshold (TOT) that improves the detection of shifting hemodynamic states. Explainable AI analysis via SHAP values highlighted diastolic blood pressure, fluid intake, and arterial pH as the most significant predictors, aligning with established pathophysiological indicators of ductal steal and systemic hypoperfusion. This AI-driven surveillance system provides a non-invasive, accessible alternative for assessing PDA severity, particularly in resource-limited settings where immediate echocardiographic expertise may be unavailable. By integrating multimodal clinical variables including vital signs and laboratory results, the model supports clinicians in optimizing conservative management and personalizing the timing of interventions for premature infants. This tool bridges the gap between complex physiological data and actionable bedside insights.