Derivation and validation of a clinical prediction score for ICU utilization at trauma intake

Document Type

Journal Article

Publication Date

1-23-2026

Journal

The American journal of emergency medicine

Volume

102

DOI

10.1016/j.ajem.2026.01.047

Keywords

AI; ICU; Machine learning; NTDB; Resource allocation; Resource utilization; TQIP; Trauma

Abstract

BACKGROUND: Delayed admission to the intensive care unit (ICU) after trauma can lead to tripling of in-hospital mortality. Accurate ICU resource prediction at initial trauma assessment can help appropriately target resources and transfers for these patients and improve timeliness of ICU care delivery. Existing models for predicting ICU need in trauma, often have low generalizability or accuracy. This study aims to bridge this gap by deriving and externally validating a sensitive, temporally stable, registry-based prediction model for identifying trauma patients requiring ICU care during emergency department (ED) stabilization. METHODS: Trauma encounters from the U.S. National Trauma Data Bank (NTDB) between 2015 and 2021 were used to predict ICU care at index hospitalization. Predictor variables were derived from clinical, demographic, procedural, and diagnostic data. Exclusion criteria included patients under 17, encounters with missing ED lengths of stay, encounters from non-ACS Level 1 or 2 centers, and transfers with incomplete initial stabilization data. A logistic regression (LR) model was trained on data from 2015 to 2020 with five-fold cross-validation and calibrated for a 99% sensitivity threshold. The model was subsequently tested on 2021 data withheld from initial training and validation. Fairness analysis was conducted to measure model performance across race, sex, and insurance status. RESULTS: Of 16,877,474 trauma encounters in the NTDB, 14.3% met inclusion criteria. The logistic regression model achieved an AUC of 0.855 and PR-AUC of 0.705 during internal validation, translating to 99.2% sensitivity and 8.6% specificity. When applied to the 2021 test set, the model maintained 99.0% sensitivity and 9.9% specificity. Model discrimination was slightly higher for Medicare patients, White patients, and male patients. However, the model's sensitivity declined to 73% in non-ACS-verified Level I/II centers. SIGNIFICANCE: This study demonstrates that for trauma patients presenting to ACS-verified Level I or II centers, the need for intensive care can be predicted with very high sensitivity using routinely available clinical variables. The model provided a generalizable framework for early ICU triage, reducing unnecessary activations while ensuring an extremely low rate of under-triage. Prospective integration of this model into trauma workflows could significantly improve ICU resource mobilization without compromising patient safety.

Department

Emergency Medicine

Link to Full Text

Share

COinS