Risk Stratification Models in Cardiovascular Emergencies

Abstract

Cardiovascular emergencies -- acute myocardial infarction (AMI), acute heart failure (AHF), aortic dissection, pulmonary embolism (PE), and malignant arrhythmias -- account for nearly 40% of emergency department presentations that result in death or ICU admission, yet the accuracy with which clinicians triage these patients remains inconsistent. Established risk scores like HEART, GRACE, TIMI, and the Wells criteria were developed decades ago using logistic regression on relatively small cohorts, and their discriminative performance in contemporary, heterogeneous emergency populations has come under increasing scrutiny. This study compared five risk stratification approaches -- conventional scoring systems, enhanced logistic regression with expanded biomarkers, gradient-boosted machine learning models, deep learning on ECG waveforms, and multi-modal fusion integrating clinical data, biomarkers, ECG, and imaging -- across four cardiovascular emergency types in a retrospective cohort of 8,462 patients presenting to four Italian emergency departments between January 2017 and December 2021. The primary endpoint was 30-day major adverse cardiovascular events (MACE). We developed a Cardiovascular Risk Stratification Score (CRSS) integrating discrimination (C-statistic), calibration, reclassification improvement, clinical utility (decision curve analysis), and implementation feasibility. Multi-modal fusion achieved the highest CRSS (0.928), with a C-statistic of 0.91 for MACE prediction and net reclassification improvement of 0.18 over conventional scores. Gradient-boosted models (CRSS 0.906) offered the best trade-off between accuracy and implementability. These results suggest that integrating multiple data streams substantially improves cardiovascular emergency risk stratification, though simpler ML models may be more practical for widespread adoption.