Abstract Body: Background
Artificial intelligence early warning scores are increasingly being used in hospital wards to identify patients at risk of clinical deterioration, but few have been specifically validated for use in the emergency department despite the fact that many admitted patients are boarded there. Prolonged emergency department boarding is common and associated with increased morbidity and mortality. eCARTv5 is a gradient boosted machine learning model that was recently FDA cleared for use on medical-surgical ward patients to predict impending ICU transfer or death but it’s utility in emergency department patients is unknown.

Aim
To validate eCARTv5 in an ED boarding population.

Methods
We conducted a retrospective analysis of prospectively collected data in consecutive ED boarding patients at 7 hospital campuses in a single health system between October 3, 2023 and May 31, 2024. Area under the receiver operating characteristic curve (AUROC) was calculated for the primary outcome of death or ICU transfer within 24 hours of score as well as death within 24 hours. Descriptive statistics were used to compare patients who ever had an elevated eCARTv5 score (≥93) while boarding and those who did not. A frequency distribution was constructed using the final eCARTv5 score in the ED and hospital mortality or any ICU transfer.

Results
Of the 65,799 ED boarders included in the study, 5,543 (8.4%) experienced an outcome during the hospitalization. A total of 3,918 (6.0%) patients were directly transferred to the ICU from the ED, 1,293 (2.0%) were ever transferred to the ICU from the ward, and 1,061 (1.6%) died during the hospitalization, with higher incidence for all outcomes among elevated eCARTv5 patients (Table 1). Patients were boarded in the ED for a median of 3.6 hours (IQR 1.4-13.5). The median final eCARTv5 score in the ED was higher for patients who ever experienced an outcome during their hospitalization (95 [IQR 87-98]) compared to those who did not (72 [IQR 52-86]) (Figure). The accuracy for predicting deterioration and mortality were comparable to previously published findings on eCARTv5’s performance in ward hospitalized patients (AUROC 0.833 [95% CI 0.832-0.834] and AUC 0.921 [95% CI 0.918-0.924]), respectively; Table 2).

Conclusion
eCARTv5, which was trained and validated on hospitalized ward patients, has the potential to be used on admitted patients boarding in the ED to monitor for risk of deterioration.