Subgrouping patients with ischemic heart disease by means of the Markov cluster algorithm

Abstract

Background: Ischemic heart disease (IHD) is heterogeneous with respect to onset, burden of symptoms, and disease progression. We hypothesized that unsupervised clustering analysis could facilitate identification of distinct and clinically relevant multimorbidity clusters.

Methods: We included IHD patients who underwent coronary angiography (CAG) or coronary computed tomography angiography (CCTA) between 2004 and 2016 and used the earliest procedure as the index date. Patient health records were obtained from the Danish National Patient Registry, the Danish National Prescription Registry, and two in-hospital laboratory database systems. Genetic data were obtained from the Copenhagen Hospital Biobank. Using registered pre-index diagnosis codes (n = 3046), patients were clustered by application of the Markov Cluster algorithm. Multimorbidity clusters were then characterized using Cox regressions (new ischemic events, non-IHD mortality, and all-cause mortality) and enrichment analysis to explore both risks and phenotypical characteristics.

Results: 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) are identified. Comparing each cluster to the 30 others, seven clusters (9,590 patients) have significantly higher or lower risk of new ischemic events (five and two clusters, respectively). A total of 18 clusters (35,982 patients) have higher or lower risk of death from non-IHD causes (12 and six clusters, respectively), and 23 clusters have a statistically significant higher or lower risk for all-cause mortality. Cardiovascular or inflammatory diseases are commonly enriched in clusters (13). Distributions for 24 laboratory test results differ significantly across clusters. Polygenic risk scores are increased in a total of 15 clusters (48.4%).

Conclusions: Based on prior disease profiles, unsupervised clustering robustly stratify patients with IHD in subgroups with similar clinical features and outcomes.

Fig. 1. Graphical overview of the study.

Steps A–C: Construction of patient-specific disease-frequency vectors by assembling the ICD-10 codes registered in the electronic health records. Using the date of the first CAG/CCTA as the index date, only ICD-10 codes registered before the index date are included. Steps D and E: Patient-specific vectors are embedded using SVD, and a patient similarity matrix is constructed from the cosine of the angle between the embedded vectors. Step F: Application of the MCL algorithm to obtain clusters of patients with specific patterns of multimorbidity. Step G: Characterization of the resulting clusters to examine their risk of three pre-defined outcomes and phenotypic characteristics defined from laboratory, medication and genetic data. CAG coronary angiography, CCTA coronary computed tomography angiography, ICD-10 International Statistical Classification of Diseases and Related Health Problems 10th Revision, IHD ischemic heart disease, MCL Markov Cluster, SVD singular value decomposition.

Fig. 2. Flowchart: data sources and study population.

NPR The Danish National Patient Registry. IHD ischemic heart disease (ICD-10 codes I20–I25). CAG coronary arteriography, CCTA coronary computed tomography angiography, ICD-10 International Statistical Classification of Diseases and Related Health Problems 10th Revision.

Fig. 3. Overview of quantitative characterization of multimorbidity clusters.

Each box in all panels represents a cluster (n = 31). In all panels, boxes are ranked by the number of patients in a specific cluster (decreasing going right and down). Yellow and green panels display cluster size (panel A) and  mean age at index (panel B). Panel C displays the distribution of patient sex in each cluster. C Panels D–F display the dependent variables in the survival analyses. Panel D: New ischemic events. Panel E: Non-IHD mortality. Panel F: All-cause mortality. Asterisks mark statistical significance with adjusted P-value < 0.05. For exact P-values, see Table 2.

Fig. 4. Graphical summary of study results.

Boxes represent clusters, and clusters are grouped according to similarity with respect to results from the survival analysis. Boxes located bottom right of all groups represent the results from the survival analysis. Orange boxes: Risk of secondary ischemic events. Green boxes: Risk of death from non-IHD causes. Symbols to the left of the boxes that represent clusters indicate characteristic findings from integrated analysis of medication data, laboratory test results, and genetic data, respectively. IHD ischemic heart disease.