Early Prediction of Critical Care Interventions From Pediatric Emergency Department Triage

Area of Science:

• Pediatric Emergency Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• The Emergency Severity Index (ESI) is widely used in US pediatric emergency departments for patient triage.
• The 5-level ESI classification offers limited risk stratification, potentially delaying care for critically ill children.
• There is a need for improved methods to identify high-risk pediatric patients during emergency department (ED) triage.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting the risk of critical care interventions in pediatric ED patients.
• To create an operationally useful model using only information available at triage.
• To improve patient prioritization and timeliness of evaluation for critical care.

Main Methods:

• Retrospective study at a large urban academic pediatric ED (2016-2024).
• Development and evaluation of 6 ML algorithms for predicting critical care interventions.
• Model performance assessed using Average Precision and the trade-off between sensitivity and positive predictive value (PPV).
• Counterfactual analysis to evaluate the clinical impact of risk predictions on evaluation timeliness.

Main Results:

• Among 886,183 ED visits, 3.0% received critical care interventions.
• A neural network model achieved the highest Average Precision (0.6).
• The best model identified 88% of critical care patients with a PPV of 32%.
• Integrating ML risk predictions could increase timely evaluation for ESI 3 patients from 23.3% to 75.0%.

Conclusions:

• Developed ML models can rapidly identify pediatric ED patients at risk for critical care interventions.
• These models can enhance triage accuracy without causing 'alarm fatigue'.
• The ML-support triage framework shows potential for improving patient care and safety in pediatric EDs.