Time-stratified HEART score discrimination and calibration in emergency chest pain: a retrospective cohort study

We compared HEART score discriminative performance and calibration for 30-day major adverse cardiac events (MACE) between off-hours and on-hours emergency department chest pain presentations.

A single-center retrospective cohort study was conducted at Istanbul Medipol Mega University Hospital from January 2020 through December 2025, enrolling 2,800 patients (off-hours n = 2,133; on-hours n = 667). Off-hours comprised evenings, nights, weekends, and public holidays. The primary outcome was 30-day MACE, defined as acute myocardial infarction, unplanned coronary revascularization, cardiac arrest, or all-cause death. Discrimination was compared by the DeLong method; an absolute between-group AUC difference below 0.08, set a priori, was used as the margin of clinical similarity. Calibration was assessed by the Hosmer-Lemeshow test, observed-to-expected ratio, and calibration slope.

30-day MACE occurred in 14.4% of 2,800 patients. The HEART score AUC was 0.830 (95% CI 0.805-0.855) in off-hours and 0.821 (95% CI 0.778-0.865) in on-hours patients; ΔAUC was + 0.009 (95% CI -0.041 to + 0.058; p = 0.736), with the confidence interval contained within the a priori margin. At the rule-out threshold (HEART ≥ 4), off-hours sensitivity was 95.7% (95% CI 92.8-97.5%) and negative predictive value (NPV) was 97.7% (95% CI 96.2-98.7%). No differential shift effect was detected (interaction OR 1.039; 95% CI 0.867-1.247; p = 0.676). Both groups demonstrated systematic overestimation of absolute MACE risk relative to Backus 2013 probabilities (off-hours calibration intercept - 0.352; 95% CI -0.495 to -0.208; p < 0.001), with no between-group difference (p = 0.998).

The HEART score showed clinically similar discrimination for 30-day MACE regardless of presentation time, supporting its use across different presentation times in tertiary-care settings with comparable infrastructure, pending prospective multicenter confirmation. Local recalibration of absolute risk estimates should be considered before probability-guided triage protocols are implemented in comparable settings.