Abstract
<title>Abstract</title> <p>Background Cardiac arrest in the emergency department may be preceded by measurable physiological deterioration. We developed a 24-hour risk prediction strategy that uses an in-hospital model when prehospital data are unavailable and an expanded model when such data are present. Methods This single-centre study comprised model development with 1,000-bootstrap internal validation and temporal evaluation of the locked models in a later, non-overlapping cohort. Prespecified logistic models used information available before a defined index time. Evaluation included discrimination, Brier score, calibration, prespecified threshold performance, risk stratification and decision-curve analysis. Results The development cohort included 2,354 patients and 171 cardiac arrests (7.3%), and the temporal-validation cohort included 1,187 patients and 107 arrests (9.0%). In temporal evaluation, the routing strategy had an AUC of 0.839 (95% CI 0.798–0.879), a Brier score of 0.0628, a calibration intercept of 0.104 (95% CI − 0.123 to 0.332) and a calibration slope of 0.965 (95% CI 0.797–1.132). Its AUC exceeded that of M0 by 0.033 (95% CI 0.017–0.050). Observed arrest rates were 2.7%, 8.6% and 35.2% in the low-, intermediate- and high-risk groups. At the 5% threshold, sensitivity was 81.3% and negative predictive value was 97.3%; at the 15% threshold, specificity was 89.3% and positive predictive value was 35.2%. Conclusions The locked routing strategy retained good discrimination and acceptable calibration in a later temporal cohort and produced clinically distinct risk groups. Prehospital systolic blood pressure, Glasgow Coma Scale score and malignant ventricular arrhythmia added predictive information beyond early in-hospital variables. Multicentre external validation and direct comparison with existing warning approaches are required before clinical use. Trial registration: Not applicable.</p>