A machine learning decision-support tool for the management of patients presenting to the emergency department with chest pain

Introduction The management of patients presenting to the Emergency Department (ED) with chest pain is resource intensive and time critical. Timely identification of acute coronary syndromes (ACS) is essential for improving outcomes, while safe early triage is central to effective use of healthcare resources. Conventional risk-stratification tools depend on serial troponin testing and ECGs. In practice, this leads to prolonged ED stays and diagnostic lag due to time taken for investigations due to workflow pressures. There is a gap for a tool to support clinical decision making at the point of ED contact, before definitive diagnostic information is available. Purpose We developed and validated a workflow-aware System for Machine-Assisted Recommendation of Treatment in Acute Coronary Syndrome (SMART-ACS). We designed the tool to support sequential clinical decision-making support. First to support early triage (at initial assessment before ECG and troponin), second for definitive pathway selection (prediction of revascularisation). Methods We used a retrospective registry of 23,488 consecutive patients presenting with chest pain to three EDs serving a diverse population within the UK (2021–2023) for model development. Machine Learning XGBoost models were trained for two tasks: Triage (no treatment for ACS, medical management, or invasive management) and Revascularisation. Models were configured with two input stages: a "baseline" model using variables available within 10 minutes of arrival (demographics, medical history and observations) and a "full" model incorporating subsequent ECG and laboratory data as inputs. We externally validated the models on an independent Swedish cohort (n=1,200). Results The baseline triage model (without ECG or troponin input data) achieved a micro-AUROC of 0.843 (95% CI: 0.827 - 0.860) on internal validation and maintained good discrimination on external validation AUROC 0.805 (95% CI: 0.724 - 0.880). For definitive management and support with referral for invasive management, the full-variable Revascularisation model achieved an AUROC of 0.947 (95% CI: 0.932–0.962), significantly outperforming the conventional HEART score benchmark (AUROC 0.634; 95% CI: 0.622 - 0.647). External validation on the Swedish cohort demonstrated robust generalisability of the full Revascularisation model (AUROC 0.888). Conclusion The SMART-ACS framework provides good discriminative performance while aligning with real-world clinical workflows and providing management recommendations. The framework has potential to support earlier triage and resource allocation. Prospective evaluation of the models is needed to test real world clinical utility and compare model recommendations against clinician judgement. Model outputs are aligned with clinically meaningful decisions made at different stages of ED assessment. This approach offers a practical pathway to support early clinical decision-making and timely interventionModel performance (internal validation)Revascularisation model