Federated Meta‐Analysis of <scp>HEART</scp> Score Performance for Emergency Department Chest Pain

OBJECTIVES: Multicenter evaluation of emergency department (ED) risk stratification tools is often limited by barriers to patient-level data sharing. We used the HEART score as a clinical use case to evaluate whether a federated diagnostic meta-analytic approach yields performance estimates comparable to those obtained from centralized patient-level analysis for predicting 30-day major adverse cardiovascular events (MACE30). METHODS: We conducted a retrospective, multicenter observational study across six EDs between January 1, 2020, and December 31, 2023. Adult patients presenting with chest pain who had a documented HEART score were included. MACE30 was defined as acute myocardial infarction, coronary revascularization, or all-cause mortality. The diagnostic performance of the HEART score was evaluated using a federated bivariate random-effects meta-analysis based on site-level 2 × 2 tables, yielding pooled sensitivity, specificity, and hierarchical summary receiver operating characteristic (HSROC) estimates. These results were compared with performance metrics derived from centralized patient-level analysis. RESULTS: Among 57,906 ED encounters with documented HEART scores, MACE30 occurred in 2.2%. In federated meta-analysis, the HEART score demonstrated high specificity and negative predictive value, consistent with its intended rule-out function, with modest between-site variability. The pooled HSROC area under the curve was 0.759 (95% CI, 0.646-0.831). Centralized patient-level analysis yielded similar discrimination (AUROC 0.785; 95% CI, 0.776-0.794). Differences between federated and centralized estimates were small and clinically modest, reflecting preservation of site-level heterogeneity and variation in outcome prevalence across EDs. CONCLUSIONS: Federated diagnostic meta-analysis produced HEART score performance estimates closely aligned with those obtained from centralized patient-level data. This approach enabled scalable, privacy-preserving multicenter evaluation of ED risk-stratification tools while accommodating heterogeneity across practice settings.