Evaluating the predictive power of a machine learning to predict the need for neonatal resuscitation

Abstract Background: Developing a system that accurately predicts the necessity for resuscitation would be valuable to patients and healthcare providers. Hence, this study aims to predict neonatal resuscitation and affecting factors by applying machine learning. Methods: The design employed for this study was a retrospective cohort study. Data for all deliveries is gathered from the electronic health record system at a tertiary Hospital, in Bandar Abbas, Iran, between January 2020 and December 2022. Women with a single cephalic pregnancy were included, while fetal malformations were used as exclusion criteria. Twenty-eight potential factors were initially selected as a feature selection. The input data were used to train eight machine learning models. In our evaluation, we utilized accuracy, the area under the curve (AUC) of the receiver operating characteristic, precision, and recall to assess the performance. Results: During the study period, 230 (7.5%) newborns required resuscitation. The likelihood of requiring resuscitation was higher for preterm newborns, babies born to multiparous women, and those delivered via cesarean section. Conversely, mothers who received support from a doula during labor had reduced odds of neonatal resuscitation. Conditions such as preeclampsia, hypothyroidism, fetal distress, intrauterine growth retardation, and lower fetal weight were found to be linked with an increased likelihood of neonatal resuscitation. Additionally, it was observed that male newborns required more frequent resuscitation. The area under the curve (AUC) for each model turned out to be: Deep learning feed-forward (0.90), random forest classification (0.87), XGBoost classification (0.85), decision tree classification (0.85), permutation classification - knn (0.80), linear regression (0.79), light gradient-boosting (0.75), and logistic regression (0.72). All eight models showed a high accuracy ranging between 0.72-0.87. However, random forest classification performed best with AUC: 087, accuracy: 0.87, precision: 0.84, and recall: 0.90. Conclusions Employing a clinical database and multiple machine learning algorithms to assess the requirement for neonatal resuscitation shows potential benefits. Further prospective research involving intrapartum clinical attributes is necessary to enhance prediction accuracy