Predicting clinical deterioration in the hospital: the impact of outcome selection

Abstract

Background: Clinical deterioration of ward patients can result in intensive care unit (ICU) transfer, cardiac arrest (CA), and/or death. These different outcomes have been used to develop and test track and trigger systems, but the impact of outcome selection on the performance of prediction algorithms is unknown.

Methods: Patients hospitalized on the wards between November 2008 and August 2011 at an academic hospital were included in the study. Ward vital signs and demographic characteristics were compared across outcomes. The dataset was then split into derivation and validation cohorts. Logistic regression was used to derive four models (one per outcome and a combined outcome) for predicting each event within 24h of a vital sign set. The models were compared in the validation cohort using the area under the receiver operating characteristic curve (AUC).

Results: A total of 59,643 patients were included in the study (including 109 ward CAs, 291 deaths, and 2638 ICU transfers). Most mean vital signs within 24h of the events differed statistically, with those before death being the most deranged. Validation model AUCs were highest for predicting mortality (range 0.73-0.82), followed by CA (range 0.74-0.76), and lowest for predicting ICU transfer (range 0.68-0.71).

Conclusions: Despite differences in vital signs before CA, ICU transfer, and death, the different models performed similarly for detecting each outcome. Mortality was the easiest outcome to predict and ICU transfer the most difficult. Studies should be interpreted with these differences in mind.