Clinical Risk Prediction in Acute Coronary Syndrome

Abstract

Acute coronary syndrome encompasses a spectrum of ischaemic myocardial emergencies -- from unstable angina through non-ST elevation myocardial infarction to ST-elevation myocardial infarction -- in which rapid, accurate risk stratification determines treatment urgency, intervention timing, and resource allocation. Established scores including GRACE, TIMI, and HEART provide population-level discrimination but show substantial performance variability across patient subgroups defined by age, sex, presenting symptoms, and comorbidity burden. Machine learning risk prediction models that incorporate high-dimensional clinical, electrocardiographic, biochemical, and imaging features offer a route to more personalised and accurate stratification. We prospectively evaluated five risk prediction frameworks across 38,642 ACS presentations at 16 European cardiac centres in Rome, Paris, Copenhagen, Zurich, and Warsaw between 2016 and 2023. Frameworks ranged from GRACE score application through multimodal deep learning integrating 12-lead ECG, troponin kinetics, echocardiographic strain, and clinical features. Performance was benchmarked using our ACS Risk Prediction Score (ARPS), integrating 30-day MACE prediction accuracy, calibration, net clinical benefit, equity across sex and age, and prospective clinical outcome impact. The multimodal deep learning framework achieved the highest ARPS (0.928), improving 30-day MACE C-index from 0.74 (GRACE) to 0.89 and identifying 24.2% of GRACE low-risk patients as ML high-risk -- a reclassification with confirmed prognostic significance. Women and patients above 80 years showed the largest absolute accuracy gains from ML reclassification, addressing the systematic underestimation of ACS risk in these groups by conventional scores