Does increasing model complexity improve HEART-based chest pain risk stratification? A multicentre development and...

The History, ECG, Age, Risk factors and Troponin (HEART) score is widely used for emergency department (ED) chest pain risk stratification, but its fixed-point structure may not optimally represent relationships between its components and major adverse cardiac events (MACE). We evaluated whether increasingly flexible modelling approaches improve prediction when restricted to the same five HEART components.

We conducted a retrospective multicentre cohort study of 47 853 adult ED patients with chest pain and documented HEART scores across six EDs from 2021 to 2023. Data from five EDs (n=39 587) were used for model development and one geographically held-out ED (n=8266) was used for validation. The conventional HEART score was compared with categorical cross-validated logistic regression (CVLR) and extreme gradient boosting (XGBoost) for predicting 30-day MACE. Performance was assessed using the area under the receiver operating characteristic (AUROC), the area under the precision-recall curve (AUPRC), the Brier score, paired bootstrap comparisons and clinical performance at comparable-sensitivity operating points.

In the held-out validation cohort, 267 patients (3.23%) experienced 30-day MACE. AUROCs were similar for HEART (0.855), CVLR (0.857) and XGBoost (0.856; pairwise p>0.05). In contrast, both data-driven models demonstrated significantly higher AUPRC than HEART: CVLR (0.237 vs 0.187; Δ=0.050, p<0.001) and XGBoost (0.244 vs 0.187; Δ=0.057, p<0.001) with no significant difference between CVLR and XGBoost (p=0.305). At comparable-sensitivity operating points, HEART, CVLR and XGBoost classified 63.74%, 59.39% and 61.58% of patients as low risk, respectively, with corresponding low-risk MACE rates of 0.53%, 0.51% and 0.57%.

Data-driven reweighting of HEART components improved precision-recall performance and risk ranking compared with fixed HEART scoring. However, XGBoost provided no incremental benefit over categorical CVLR, and neither approach expanded low-risk classification at comparable sensitivity.