Characterising the heterogeneity of heart failure with preserved ejection fraction: moving beyond subgroups and distinguishing disease from risk
F Soltani, N Black, J Bradley, N Ko Ko, H Glasse, I Milner, M Schmitt, Z Raisi-Estabragh, S E Petersen, T Mcdonagh, S G Williams, D A Jenkins, A P Morris, N Peek, C A MillerAbstract
Background
Heart failure with preserved ejection fraction (HFpEF) is a complex, heterogeneous syndrome commonly hypothesised to comprise discrete subtypes. Studies aimed at reclassifying HFpEF into more homogenous subgroups have not translated clinically.
Purpose
To apply a range of data science techniques to large, deeply phenotyped cohorts with multi-modal data to simplify HFpEF into interpretable models, enabling visualisation of heterogeneity, assessment of phenotypic drivers of outcome, identification of genetic determinants, and differentiation of HFpEF from associated comorbidities.
Methods
902 patients with heart failure and a left ventricular ejection fraction (LVEF) ≥ 50% were prospectively recruited. Phenotyping included clinical, biochemical, electrocardiography, echocardiography, and cardiac magnetic resonance data. Genotyping was performed using the UK Biobank Axiom array. Clustering algorithms were applied to derive discrete subgroups. A novel dimensionality reduction approach using DDRTree was applied to explore phenotypic variation without enforcing subgroup assignment, deriving a two-dimensional tree structure. Phenotypes and outcomes for each individual were overlaid onto the tree. Genome-wide association studies (GWAS) were conducted to identify variants associated with tree dimensions. Internal validation was conducted using Jaccard indices, and external validation was conducted using an independent HFpEF cohort (n=148). Features distinguishing HFpEF from associated comorbidities were identified by multivariable logistic regression using a comorbidity-matched non-HFpEF cohort (n=902). Sensitivity analyses included patients with mildly reduced ejection fraction (n=1154).
Results
Clustering algorithms failed to identify meaningful clusters (Figure 1). DDRTree revealed a highly stable tree structure (median Jaccard index 0.84), and captured continuous variation in phenotypic profiles (Figure 2). Distinct regions of the tree were characterised by variation in cardiac structure and function, physical status, biomarkers, and comorbidities. Outcomes varied across the tree (median follow up 6.8 years); the risks of cardiovascular death, heart failure hospitalisation, and adverse renal outcome were higher for individuals in the right of the tree, and the risk of hospitalisation for infection/sepsis was higher for individuals in the lower part of the tree. GWAS revealed a locus near CRACD that was associated with tree dimension 1. Findings were reproducible in external and sensitivity analyses. A profile of cardiac and systemic characteristics that distinguish HFpEF from its associated comorbidities were identified.
Conclusion
This study provides a novel approach to understanding the heterogeneity of HFpEF. The findings provide a biologically plausible basis for developing a more personalised approach to HFpEF, and support a shift towards mechanism-focused research in large scale, deeply phenotyped cohorts.Figure 1.Cluster plots.For image description, please refer to the figure legend and surrounding text.Figure 2.Variation in phenotype.For image description, please refer to the figure legend and surrounding text.