Early prediction of intensive care unit admission in emergency department patients using machine learning

Background: The timely identification and transfer of critically ill patients from the emergency department (ED) to the intensive care unit (ICU) is important for patient care and ED workflow practices. Objective: We aimed to develop a predictive model for ICU admission early in the course of an ED presentation. Methods: We extracted retrospective data from the electronic medical record and applied natural language processing and machine learning to information available early in the course of an ED presentation to develop a predictive model for ICU admission. Results: We studied 484 094 adult (≥18 years old) ED presentations, amongst which direct admission to the ICU occurred in 3955 (0.82%) instances. We trained machine learning in 323 678 ED presentations and performed testing/validation in 160 416 (70 546 for testing and 89 870 for validation). Although the area under the receiver operating characteristics curve was 0.92, the F1 score (0.177) and Matthews correlation coefficient (0.257) suggested substantial imbalance in the dataset. The strongest weighted variables in the predictive model at the 30-min timepoint were ED triage category, arrival via ambulance, quick Sequential Organ Failure Assessment score, baseline heart rate, and the number of inpatient presentations in the prior 12 months. Using a likelihood of ICU admission of more than 75%, for activation of automated ICU referral, we estimated the model would generate 2.7 triggers per day. Conclusions: The infrequency of ICU admissions as a proportion of ED presentations makes accurate early prediction of admissions challenging. Such triggers are likely to generate a moderate number of false positives.