Subgrouping multimorbid patients with ischemic heart disease by means of unsupervised clustering: A cohort study of 72,249 patients defined by 3,046 diagnoses

Background: There are no methods for classifying multimorbid patients with ischemic heart disease (IHD), although such methods might be clinically useful due to the marked differences in presentation and disease-course. Methods: A population-based cohort study from a Danish secondary care setting of patients with IHD (2004-2016) and subjected to a coronary angiography (CAG) or coronary computed tomography angiography (CCTA). Data sources were The Danish National Patient Registry, in-hospital laboratory data, and genetic data from Copenhagen Hospital Biobank. Comorbidities included diagnoses assigned prior to presentation of IHD. Patients were clustered my means of the Markov Clustering Algorithm based on the entire spectrum of registered multimorbidity. The two prespecified outcomes were: New ischemic events (including death from IHD causes) and death from non-IHD causes. Patients were followed from date of CAG/CCTA until one of the two outcomes occurred or end of follow-up, whichever came first. Biological and clinical appropriateness of clusters was assessed by comparing risks (estimated from Cox proportional hazard models) in clusters and by phenotypic and genotypic enrichment analyses, respectively. Findings: In a cohort of 72,249 patients with IHD (mean age 63.9 years, 63.1% males), 31 distinct clusters (C1-31, 67,136 patients) were identified. Comparing each cluster to the 30 others, eight clusters (9,590 patients) had statistically significantly higher (five clusters) or lower (three clusters) risk of new ischemic events; 18 clusters (35,982 patients) had a higher (11 clusters) or lower (seven clusters) risk of death from non-IHD causes. All clusters at increased risk of new ischemic events, associated with risk of death from non-IHD causes as well. Cardiovascular or inflammatory diseases were the commonly enriched in clusters (13), and distributions for 24 laboratory test results differed significantly across clusters. Polygenic risk scores for atrial fibrillation and diabetes were increased in x and y clusters respectively. Conclusions: Clustering of patients with IHD based on comorbidities identified subgroups of patients with significantly different clinical outcomes. This novel approach may support differentiation of treatment intensity dependent on expected outcomes.