Unsupervised machine-learning identifies cardio-kidney-metabolic phenotypes and determines residual risk after acute coronary syndromes

Background Residual cardiovascular risk reflects a complex interplay of lipid, inflammatory, thrombotic, and cardio-kidney-metabolic (CKM) pathways. In patients with acute coronary syndromes (ACS), this risk remains substantial despite intensive low-density lipoprotein cholesterol (LDL-C) lowering therapy.1,2 A more holistic understanding of ACS patient phenotypes is required to allow for more targeted interventions.3,4 Purpose To identify ACS phenotypes by machine learning (ML) and determine their individual contributions to residual cardiovascular risk. Methods We analysed 4,787 Swiss statin-treated acute coronary syndrome (ACS) patients from the Special Programme University Medicine Acute Coronary Syndrome (SPUM-ACS) study and 15,283 French statin-treated ACS patients from the obseRvatoire des Infarctus de Côte-d’Or (RICO) study. By using the Davies–Bouldin index and silhouette coefficient, unsupervised ML identified four patient clusters (k = 4) across lipid, inflammatory, and CKM domains. One-year major adverse cardiovascular events (MACE) were compared across phenotypes using Kaplan–Meier and multivariable Cox analyses. Results In SPUM-ACS, four patient clusters were identified by unsupervised ML, namely HDL-, LDL-, HbA1c-, and CRP-creatinine-dominant patient phenotypes. In RICO, similar clusters were identified. Risk of 1-year MACE differed significantly across clusters (log-rank p < 0.001) in both Swiss and French patients, resulting in a 2.45- (adjusted hazard ratio [aHR], 2.45, 95% CI 1.64–3.66) and 2.40-fold (aHR, 2.40, 95% CI 1.60–3.59) higher MACE risk for HbA1c- and CRP-creatinine-dominant phenotypes in Swiss patients, respectively, while LDL- (aHR, 1.07, 95% CI 0.68–1.69) and HDL-dominant (reference) phenotypes showed the best outcomes among all patient clusters, with similar results in French patients (Figure 1). Importantly, the proportion of patients exceeding guideline-recommended risk thresholds differed significantly across the identified clusters, irrespectively whether Swiss or French patients were analyzed (Figure 2). Conclusions ML-derived unsupervised clustering identifies reproducible CKM phenotypes with distinct risk profiles, providing a foundation for phenotype-guided secondary prevention strategies in the contemporary era.