Predictive performance of vital sign–based prediction models for in-hospital cardiac arrest: a systematic review and meta-analysis

Abstract Background In-hospital cardiac arrest (IHCA) is a major patient-safety event often preceded by detectable clinical deterioration. Vital sign–based cardiac arrest prediction models have been developed to identify hospitalized patients at increased risk of IHCA, but their predictive performance has not been quantitatively synthesized across studies. Methods This PRISMA-based systematic review and meta-analysis was registered in PROSPERO. Ovid Embase, Ovid MEDLINE, and the Cochrane Central Register of Controlled Trials (CENTRAL) within the Cochrane Library were searched for peer-reviewed studies evaluating vital sign–based prediction models for IHCA in hospitalized general ward patients. Eligible models used routinely collected bedside vital signs as core predictors; models relying on laboratory values, imaging data, or other non-bedside-measured clinical variables as core predictors were excluded. The primary performance measure was the area under the receiver operating characteristic curve (AUROC); area under the precision-recall curve (AUPRC) was extracted and summarized when reported. Risk of bias was assessed using PROBAST + AI (Prediction model Risk of Bias Assessment Tool + Artificial Intelligence). Results Of 10,645 records identified, six studies were included. The included studies evaluated Vital sign–based cardiac arrest prediction models, including Cardiac Arrest Risk Triage (CART), the deep learning–based early warning score (DEWS), and the deep learning–based cardiac arrest risk score (DeepCARS), and were published between 2012 and 2024. Three studies reported cardiac arrest-specific AUROC estimates suitable for the primary meta-analysis; the exploratory pooled AUROC was 0.89 (95% CI, 0.88–0.91; I² = 92%). Overall risk of bias was low in two studies, high in three, and unclear in one. Conclusions Vital sign–based cardiac arrest prediction models showed potentially useful discrimination for predicting IHCA; however, the pooled AUROC should be interpreted as an exploratory summary because only three studies contributed to the primary meta-analysis, heterogeneity was substantial, and half of all included studies were judged to have a high overall risk of bias. Standardized outcome definitions, external validation, and prospective evaluation of calibration, alert burden, and clinical utility are needed before broad clinical implementation.