Predicting the Need for Cardiopulmonary Resuscitation in Preterm Infants in the Delivery Room Using Machine Learning Models: Analysis of a Korean Neonatal Network Database

Abstract

Background: This study aimed to develop a specialized model for predicting the stages of neonatal resuscitation for preterm infants using prospectively collected data on very-low birth-weight infants in South Korea. Methods: A prospective cohort study was conducted using the Korean Neonatal Network database, including neonates weighing < 1,500 g. Overall, 9,684 infants were included, and external validation was performed using data of 71 infants collected from Jeonbuk National University Hospital. Logistic regression, random forest, and eXtreme Gradient Boosting (XGB) were the machine learning models employed. Results: The final models particularly in predicting the need for “endotracheal intubation or higher” performed well, with the XGB ensemble algorithm showing the best performance (area under the receiver operating characteristic curve, 0.91; area under the precision-recall curve, 0.86; and accuracy, 0.85). The most influential variables affecting the performance of the predictive models in the ensemble algorithm were gestational age and birth weight. Conclusion: The developed predictive model enabled the early identification of the need for neonatal resuscitation in preterm infants. When used as a clinical decision support system in neonatal intensive care units and delivery rooms, it is expected to not only facilitate efficient staffing by healthcare professionals but also increase resuscitation procedure success rates.