Prehospital Risk Stratification Using Unsupervised Machine Learning in <scp>STEMI</scp>

Abstract

BACKGROUND: ST-elevation myocardial infarction (STEMI) exhibits substantial clinical heterogeneity complicating prehospital risk stratification. Traditional risk assessment tools often fail to capture the complexity of this condition. Machine learning offers opportunities to identify complex clinical patterns not readily apparent during prehospital care. AIM: To identify distinct phenotypes in STEMI patients using unsupervised machine learning algorithms based on prehospital parameters, and to determine their association with short-term mortality and cardiovascular outcomes. METHODS: Prospective multicenter observational cohort study including adult patients with prehospital STEMI code activation transported by emergency medical services from January 2022 to August 2025. Only EMS-transported patients were included; those who self-presented to the emergency department were excluded. Prehospital variables, including demographic, clinical, and procedural data, were used for clustering. Factor Analysis of Mixed Data and a two-step clustering: hierarchical clustering (exploring structure and number of clusters) and k-means (clustering assigning patients to phenotypes). A Random Forest classifier with SHapley Additive exPlanations values was used to identify variables contributing to cluster assignment. The primary outcome was 30-day all-cause mortality, assessed through follow-up records. RESULTS: Among 744 patients (median age, 65 years; 76.3% male) unsupervised clustering identified three distinct phenotypes: Phenotype-1 (70.3%) characterized by hemodynamic stability, vessel locations, Killip class I presentation (70.6%), and favourable laboratory parameters; Phenotype-2 (24.3%) presented higher comorbidity burden and metabolic derangements; and Phenotype-3 (5.4%) exhibiting profound hemodynamic instability, severe respiratory failure, out-of-hospital cardiac arrest with return of spontaneous circulation (87.5%), Killip class IV presentation (67.5%), and marked metabolic derangements. The 30-day mortality rates were: 3.4% in Phenotype-1, 22.1% in Phenotype-2, and 75.0% in Phenotype-3. CONCLUSIONS: Three clinically distinct STEMI phenotypes were identified with markedly different mortality risks and treatment requirements during prehospital care. Phenotypes derived from readily available prehospital parameters may facilitate early risk stratification, optimize triage decisions, and guide individualized therapeutic strategies.

Ähnliche Arbeiten