Phenotypic clustering of dilated cardiomyopathy patients highlights important pathophysiological differences

Abstract

Aims: The dilated cardiomyopathy (DCM) phenotype is the result of combined genetic and acquired triggers. Until now, clinical decision-making in DCM has mainly been based on ejection fraction (EF) and NYHA classification, not considering the DCM heterogenicity. The present study aimed to identify patient subgroups by phenotypic clustering integrating aetiologies, comorbidities, and cardiac function along cardiac transcript levels, to unveil pathophysiological differences between DCM subgroups.

Methods and results: We included 795 consecutive DCM patients from the Maastricht Cardiomyopathy Registry who underwent in-depth phenotyping, comprising extensive clinical data on aetiology and comorbodities, imaging and endomyocardial biopsies. Four mutually exclusive and clinically distinct phenogroups (PG) were identified based upon unsupervised hierarchical clustering of principal components: [PG1] mild systolic dysfunction, [PG2] auto-immune, [PG3] genetic and arrhythmias, and [PG4] severe systolic dysfunction. RNA-sequencing of cardiac samples (n = 91) revealed a distinct underlying molecular profile per PG: pro-inflammatory (PG2, auto-immune), pro-fibrotic (PG3; arrhythmia), and metabolic (PG4, low EF) gene expression. Furthermore, event-free survival differed among the four phenogroups, also when corrected for well-known clinical predictors. Decision tree modelling identified four clinical parameters (auto-immune disease, EF, atrial fibrillation, and kidney function) by which every DCM patient from two independent DCM cohorts could be placed in one of the four phenogroups with corresponding outcome (n = 789; Spain, n = 352 and Italy, n = 437), showing a feasible applicability of the phenogrouping.

Conclusion: The present study identified four different DCM phenogroups associated with significant differences in clinical presentation, underlying molecular profiles and outcome, paving the way for a more personalized treatment approach.

Keywords: Clustering; Dilated cardiomyopathy; Machine learning; Pathophysiology.

Graphical abstract

Figure 1

Summary of the aims and study design including data processing steps, survival analysis, and application.

Figure 2

Four mutually exclusive phenogroups as determined by hierarchical clustering of principal component using phenotypical information as input. The most distinct clinical characteristics are listed per phenogroup. Variables with an asterisk are key parameters to distinguish the phenogroups, as selected by supervised decision tree modelling (A). Characteristic plots of the four proposed phenogroups including their most representative clinical variables. The over- or underrepresentation of a variable within a cluster was analysed by v-test within the hierarchical clustering of principal component function, based on the hypergeometric distribution. A positive value indicates overrepresentation of this variable in the applicable phenogroup, a negative value indicates underrepresentation of the corresponding variable (B).

Figure 3

Analysis of RNA-sequencing data of endomyocardial biopsies from dilated cardiomyopathy patients. Principal component analysis (PCA) of RNA-sequencing data divided on phenogroup (PG). Principal component 1 shows strong division of PG4 and the others (A). Venn diagram of the number of significant differentially expressed genes in the comparison between two corresponding PG (FDR < 0.01 + fold change >1.5) (B). Significantly enriched Kyoto Encyclopaedia of Genes and Genomes pathways (P-value < 0.05) in the comparison between the PG (Table 4) (C).

Figure 4

Event-free survival stratified by phenogroup. Kaplan–Meier curves for the combined outcome of life-threatening arrhythmias, cardiovascular death, heart transplantation, or left ventricular assist device implantation stratified by phenogroup.

Figure 5

A simplified adaptation of the decision tree model shows the clinical parameters which are the core of the phenogroups. The black box indicates that the model cannot accurately place the patient in one of the four phenogroups (see Supplementary material online, Figure S5 for exact accuracy).

Figure 6

Event-free survival stratified by phenogroup as determined by the supervised decision tree model. Kaplan–Meier curves for the combined outcome of life-threatening arrhythmias, cardiovascular death, heart transplantation, or left ventricular assist device implantation stratified by phenogroup in the application cohort (Madrid and Trieste).

Take home figure

This study identified four phenotypic clusters based on patterns in clinical data which had a unique underlying molecular profile identified by RNA-sequencing on cardiac biopsies of patients from each phenogroup. A classifier derived from the initial clustering could be succesfully applied on an external cohort placing every DCM patient in a specific phenogroup.