Early Prediction of Critical Care Interventions From Pediatric Emergency Department Triage.

BACKGROUND AND OBJECTIVES: Most pediatric emergency departments (EDs) in the United States use Emergency Severity Index (ESI) system to triage patients. Because the 5-level classification provides limited risk stratification, this study aims to improve patient prioritization by developing an operationally useful model that predicts risk of critical care interventions using only information available during ED triage. METHODS: We conducted a retrospective study at a large urban academic pediatric ED from 2016 to 2024. We developed predictive models using 6 machine learning (ML) algorithms. Models were evaluated on Average Precision and tradeoff between sensitivity and positive predictive value (PPV). We performed a counterfactual analysis to assess potential clinical effects of risk predictions on timeliness of evaluation for critical care patients. RESULTS: Among 886 183 ED visits, 26 721 (3.0%) received critical care interventions. The neural network had the highest Average Precision of 0.6 (95% CI 0.59-0.61). The model could identify 88% (87%-89%) of patients who received critical care interventions with PPV of 32% (31%-32%). Supplementing ESI with these risk predictions would have increased the proportion of critical care patients being timely evaluated by physicians from 23.3% to 75.0% for ESI 3 patients. Similarly, improvements would have been achieved for other ESI levels. CONCLUSION: We developed models capable of quickly identifying ED pediatric patients at risk of requiring critical care interventions without causing alarm fatigue. Potential improvements in time-to-pediatrician for at-risk patients suggest utility of our ML-support triage framework in improving patient care and safety in pediatric ED.