Machine Learning Algorithms to Predict Mortality of Neonates on Mechanical Intubation for Respiratory Failure

Background: Early identification of critically ill neonates with poor outcomes can optimize therapeutic strategies. We aimed to examine whether machine learning (ML) methods can improve mortality prediction for neonatal intensive care unit (NICU) patients on intubation for respiratory failure. Methods: A total of 1734 neonates with respiratory failure were randomly divided into training (70%, n = 1214) and test (30%, n = 520) sets. The primary outcome was the probability of NICU mortality. The areas under the receiver operating characteristic curves (AUCs) of several ML algorithms were compared with those of the conventional neonatal illness severity scoring systems including the NTISS and SNAPPE-II. Results: For NICU mortality, the random forest (RF) model showed the highest AUC (0.939 (0.921-0.958)) for the prediction of neonates with respiratory failure, and the bagged classification and regression tree model demonstrated the next best results (0.915 (0.891-0.939)). The AUCs of both models were significantly better than the traditional NTISS (0.836 (0.800-0.871)) and SNAPPE-II scores (0.805 (0.766-0.843)). The superior performances were confirmed by higher accuracy and F1 score and better calibration, and the superior and net benefit was confirmed by decision curve analysis. In addition, Shapley additive explanation (SHAP) values were utilized to explain the RF prediction model. Conclusions: Machine learning algorithms increase the accuracy and predictive ability for mortality of neonates with respiratory failure compared with conventional neonatal illness severity scores. The RF model is suitable for clinical use in the NICU, and clinicians can gain insights and have better communication with families in advance.