Use of Machine Learning to Predict Hospital Admission for EMS-Treated Infants After a Suspected BRUE

OBJECTIVES: This study explored the use of different applied machine learning (ML) classification algorithms to predict hospital admission for infants treated by emergency medical services (EMS) after a suspected brief resolved unexplained event (BRUE). METHODS: Data from a large regionalized pediatric care system were obtained for infants in which paramedic suspected a BRUE and who were transported between July 2017 and February 2021. After data pre-processing, a random 80%/20% split for training and testing was performed. First, a random forest ML classification model was used to identify and select the most important variables influencing the prediction of hospital admission. Then, multiple ML-based models and a statistical model were trained with this subset of variables and evaluated the performance of each to predict hospital admission. Model performance characteristics including the area under the receiver operator curve (AUROC) were reported. RESULTS: A total of 508 infants were included; 300 (59%) were admitted and 76 (15%) required critical care. The most important variables in predicting hospital admission were age, history of bystander interventions (ie, cardiopulmonary resuscitation and back blows), presence of past medical history, and a normal appearing examination. In the prediction of hospital admission, the support vector machine model achieved the highest AUROC of 0.85, with a sensitivity of 0.88 (95% CI: 0.80-0.96) and specificity of 0.71 (95% CI: 0.57-0.85). The predictive performance of the extreme gradient boosting, RF, and logistic regression models were similar (AUROC: 0.83 to 0.84). CONCLUSIONS: The applied ML models demonstrated good predictive performance for hospital admission for EMS-treated infants with a paramedic suspected BRUE. ML and statistical models had similar predictive performance.