Phenotypic clustering of myocardial infarction complicated by out-of-hospital cardiac arrest using unsupervised machine learning.

Abstract

Patients with myocardial infarction (MI) complicated by out-of-hospital cardiac arrest (OHCA) represent a heterogeneous population with variable outcomes. Data-driven approaches may help uncover clinically meaningful subgroups to improve risk stratification and guide management. We applied an unsupervised machine learning analysis using k-means clustering to a prospective cohort of 478 patients admitted after OHCA related to MI. Candidate variables included demographics, cardiovascular risk factors, cardiac arrest characteristics, admission laboratory data, hemodynamic parameters, and coronary angiography findings. In-hospital outcomes included all-cause mortality, bleeding, and stent thrombosis. Ninety-day all-cause mortality was also assessed. Three clusters were identified. Cluster 1 (n = 260, 54%) included younger patients with few comorbidities, predominantly shockable rhythms, and favorable hemodynamics. Cluster 2 (n = 118, 25%) included older patients with hypertension, diabetes, and prior coronary artery disease. Cluster 3 (n = 100, 21%) was characterized by severe cardiogenic shock, high lactate, low factor V, reduced left-ventricular ejection fraction, and frequent use of extracorporeal membrane oxygenation. Adverse in-hospital events, including all-cause mortality, bleeding, and stent thrombosis, were most frequent in cluster 3. Ninety-day mortality differed across groups: 22.5% in cluster 1, 53.0% in cluster 2, and 77.2% in cluster 3 (p < 0.001). Compared with cluster 1, hazard ratios for mortality at 90 days were 2.97 (95% CI, 2.07-4.26) in cluster 2 and 6.75 (95% CI: 4.74-9.60) in cluster 3. Unsupervised machine learning identified three phenotypes among patients with MI-related OHCA associated with distinct outcomes. This phenotypic classification may facilitate personalized management and refined prognostic assessment in this high-risk population.