Introduction

Heart failure with preserved ejection fraction (HFpEF), which affects approximately one-half of patients with heart failure, results in substantial morbidity, mortality, and impaired quality of life. Although several pharmacological therapies are known to improve patient outcomes in HFrEF, no pharmacological therapies have been clearly demonstrated to reduce adverse events in HFpEF.

HFpEF likely represents a heterogeneous group of disease processes. This heterogeneity may contribute to the difficulty identifying effective treatments for HFpEF. A wide range of clinical risk factors for HFpEF have been identified, including older age, female sex, history of hypertension, diabetes, obesity, atrial fibrillation (AF), and coronary artery disease, among others (1,2). Patients with HFpEF also have highly variable underlying cardiac structural and functional abnormalities (3). Accordingly, previous studies have proposed that several different phenotypes of HFpEF exist (4), encompassing relatively discrete subgroups (phenogroups) with distinct clinical features (5). These studies demonstrated that HFpEF phenogroups might be linked to important differences in prognosis of disease (4,5). However, few data exist regarding differences in underlying biologic processes or response to therapies between HFpEF phenogroups. Identifying biomarker profiles within each phenogroup may suggest phenogroup-specific mechanisms that can be targeted for therapeutic purposes. Moreover, recent data indicate that clustering techniques informed by standard clinical features can classify patients with HFrEF into phenogroups that exhibit differential responses to spironolactone (6,7). Whether differential responses to spironolactone are present between HFpEF clinical phenogroups is unknown.

In this study, our objective was to use data and frozen plasma samples from the TOPCAT (Treatment of Preserved Cardiac Function Heart Failure with an Aldosterone Antagonist Trial) to examine plasma protein profiles between clinical phenogroups via de novo measurements of multiple biomarkers; examine differences in outcomes and response to spironolactone therapy between the phenogroups; and further characterize relevant cardiac and vascular characteristics of phenogroups using echocardiographic and arterial tonometry data.

Methods

Please see the Online Methods for additional details regarding the study design and statistical analyses.

Data source and study population

We used data and biosamples from TOPCAT, obtained from the National Heart, Lung, and Blood Institute. TOPCAT was a large multicenter international trial evaluating the efficacy of spironolactone therapy in patients older than 50 years of age with symptomatic HFpEF and left ventricular (LV) ejection fractions ≥45% (8). The study was approved by the institutional review board at each participating site, and all participants provided written informed consent.

A total of 3,445 subjects were enrolled across the United States, Canada, Brazil, Argentina, Russia, and Georgia. Because of substantial differences regarding subject recruitment and study implementation in Russia and Georgia (9), we performed subanalyses restricted to participants from the Americas (n = 1,765) when feasible, accounting for statistical power. The primary endpoint of TOPCAT was a composite of cardiovascular death, heart failure hospitalization, or aborted cardiac arrest.

Using available frozen plasma samples from the baseline visit, we measured 49 protein analytes of key disease pathways/mechanisms (Online Tables 1 and 2), using a Luminex Bead-Based multiplexed assay (Bristol Myers-Squibb, Ewing Township, New Jersey). Analytes were chosen a priori to represent a diverse number of physiological processes related to cardiovascular disease and downstream effects (Online Table 1). We also analyzed data from other ancillary substudies, including an echocardiographic study and an arterial tonometry study (see Online Methods). The assay range for each analyte in the Luminex multiplexed assay are shown in Online Table 2.

Assignment of clinical phenogroups

Latent class analysis (LCA) was performed to determine clusters of clinical phenotypes. LCA is a clustering statistical technique that uses finite mixture modeling to classify individuals into mutually exclusive and exhaustive subgroups, maximizing within-group similarities and between-group differences on the basis of multiple observed characteristics in a population (6,7,10,11). Participants were characterized based on age, sex, race, diabetes status, history of AF, obesity, severe heart failure symptoms (New York Heart Association functional class III or IV), and chronic kidney disease (CKD) status. These clinical covariates were selected a priori to incorporate widely available clinical covariates that can be easily obtained in routine practice and taking into account known associations with adverse outcomes in HFpEF (3,4).

The LC models were compared across successive numbers of subgroups (i.e., phenogroups). Several metrics were assessed to determine the optimal number of phenogroups: the parametric bootstrap likelihood ratio (LR) test, Akaike’s Information Criterion (AIC), Bayesian Information Criterion (BIC), and sample-size-adjusted BIC (12,13).

Statistical analyses of clinical outcomes

Patient characteristics were summarized using standard descriptive statistics. Given that we compared multiple biomarkers and echocardiographic/tonometry features of cardiac structure and function across the phenogroups, we also report corrected p values that account for multiple comparisons, based on the number of underlying principal components (14,15). Because biomarkers and echocardiographic parameters were available in only a subset of participants, there was insufficient power to perform these specific comparisons after restricting to participants enrolled in the Americas (16).

To compare outcomes between the phenogroups, Kaplan-Meier curves with log-rank testing for trend were applied to assess for equality of survival distributions for the primary composite endpoint, a composite endpoint of all-cause mortality or hospitalization for heart failure, hospitalization for heart failure alone, and all-cause mortality alone. Cox proportional hazard models were performed to estimate unadjusted and adjusted hazards ratios (HRs) and 95% confidence intervals (CIs). Exploratory models were adjusted for age, sex, and race and Meta-Analysis Global Group in Chronic Heart Failure (MAGGIC) risk score (17). We assessed the effect of spironolactone treatment using analyses stratified by phenogroup and formally tested effect modification using interaction terms of phenogroup membership and spironolactone treatment arm.

Statistical analyses were performed using STATA version 15.0 (StataCorp LP, College Station, Texas), using the LCA Stata Plugin, version 1.2, the LCA Bootstrap, version 1.0 (18,19), and the Matlab statistics and machine learning toolbox (Matlab 2016b, Mathworks, Natwick, Massachusetts).

Results

Clinical characteristics of identified phenotypes

In the overall study population at baseline, the mean age was 69 ± 10 years, with 52% women and 9% black participants. CKD was present in 43% of participants; the overall mean estimated glomerular filtration rate (eGFR) was 65 ± 19 ml/min/1.73 m²; 32% had diabetes mellitus; 55% were obese; 35% had a history of AF; and 10% were smokers.

The optimal number of clinical phenogroups was 3 (bootstrap LR test for 2 vs. 3 classes, p = 0.01; LR test for 3 vs. 4 classes, p = 0.16; AIC, BIC, and sample-size–adjusted BIC also supported a 3-class solution). Online Figure 1 shows the distribution of participants included in the analysis, and Tables 1 and 2 show the clinical characteristics across the phenogroups for the overall study population and restricted to the Americas, respectively. Importantly, characteristics of the clinical phenogroups were similar in the overall cohort and after restricting to participants enrolled in the Americas. However, phenogroup 1 comprised the majority of participants enrolled in Russia/Georgia (53.49%) but only a minority of participants enrolled in the Americas (17.96%) (Online Figure 2).

Phenogroup 1 was composed of younger individuals (mean age 61 ± 6 years) with relatively preserved functional class, the highest prevalence of smoking (24%) among the groups, along with relatively preserved renal function (mean eGFR 76 ± 16 ml/min/1.73 m²) and a low prevalence of diabetes (9%). Phenogroup 2 was characterized by older age (mean age 77 ± 5 years), the highest proportion of women (56%), a high prevalence of atrial fibrillation (49%), and CKD (mean eGFR 58 ± 16 ml/min/1.73 m²) but a low prevalence of diabetes and obesity. Phenogroup 3 exhibited intermediate age (mean age 66 ± 8 years), with a very high prevalence of obesity (98%), diabetes mellitus (88%), and prominently impaired functional class; it also exhibited a high prevalence of CKD (57%), depression (36%), and a higher proportion of participants of black race (21%). All 3 phenotype groups had similar rates of spironolactone treatment assignment and other antihypertensive drug use.

Circulating biomarkers across clinical phenotypes

Differences in the examined biomarkers across the clinical phenotypes are shown in Figure 1, Online Table 3, and Online Figure 3. Phenogroup 1 demonstrated the lowest NT-proBNP levels and much higher levels of metalloproteinase (MMP)-9 (an extracellular turnover marker) compared with the other groups. This phenogroup also exhibited high levels of syndecan-4 (a cell-matrix interaction marker) compared with the other clinical phenogroups. Phenogroup 2 exhibited the highest levels of osteoprotegerin (a regulator of mineral metabolism and tissue calcification), tissue inhibitor of MMP (TIMP)-4 (a marker of extracellular turnover), and inflammatory biomarkers related to the innate immune response (interleukin-8, pentraxin-3, soluble intercellular adhesion molecule-1). Phenogroup 3 exhibited the highest levels of biomarkers of tumor necrosis factor (TNF)-mediated inflammation (TNF-alpha, soluble TNF receptors type 1 and 2), intermediary metabolism (fatty-acid binding protein 4, fibroblast growth factor [FGF]-21 and GDF-15), YKL-40 (liver fibrosis), plasma renin, kidney injury (cystatin C and kidney injury molecule-1), mineral metabolism/calcification (FGF-23 and osteoprotegerin), angiogenesis (angiopoietin and vascular endothelial growth factor A), and tissue remodeling (sST2).

Echocardiographic parameters and large-artery stiffness across clinical phenotypes

Echocardiographic findings across the clinical phenotypes in the 935 subjects who took part in the echocardiographic ancillary study are shown in Figure 2 and Online Table 4. Phenogroup 1 exhibited the least concentric LVs, with largest LV cavities, and the lowest absolute and relative wall thickness, LA volumes, and mitral septal and lateral E/e′ ratios (both septal and lateral). This group also exhibited the lowest values of resistive arterial load (systemic vascular resistance), pulsatile arterial load (total arterial compliance), and large-artery stiffness (carotid-femoral pulse wave velocity). Phenogroup 2 demonstrated a distinct pattern characterized by small concentric LVs with the lowest LV mass among the groups, the largest left atria, the lowest mitral annular tissue velocities, the stiffest large arteries (highest carotid-femoral pulse wave velocity), and the highest pulsatile and resistive arterial load. Finally, phenogroup 3 exhibited a distinct pattern of concentric LV hypertrophy, with the highest values of LV wall thickness, LV mass, and LV mass indexed for height (but not of LV mass indexed for body surface area), and E/e′ ratios. This phenogroup exhibited relatively low values of resistive arterial load but high pulsatile arterial load indexed for body size (total arterial compliance index).

Prognostic relationship between clinical phenotypes and patient outcomes

Compared with phenogroup 1 (younger, with mild symptoms and normal LV geometry), participants in phenogroup 2 (older, with stiff arteries and small ventricles) exhibited a significantly higher risk of the combined primary endpoint (HR: 2.17; 95% CI: 1.76 to 2.68) and heart failure hospitalization alone (HR: 3.07; 95% CI: 2.28 to 4.12), with phenogroup 3 (older obese with diabetes and LVH), demonstrating the highest risk of both outcomes (primary endpoint HR: 3.44; 95% CI: 2.79 to 4.24; heart failure hospitalization HR: 5.91; 95% CI: 4.42 to 7.89) (Figure 3, Online Table 5). Both phenogroup 2 and phenogroup 3 demonstrated similarly increased risk of combined all-cause mortality or heart failure hospitalization (phenogroup 2 HR: 2.59; 95% CI: 2.04 to 3.28; phenogroup 3 HR: 3.04; 95% CI: 2.36 to 3.91) and all-cause mortality (phenogroup 2 HR: 2.36, 95% CI: 1.89 to 2.95; phenogroup 3 HR: 2.26; 95% CI: 1.77 to 2.87). In exploratory analyses adjusting for age, sex, and race (Online Table 5), phenogroup 2 and phenogroup 3 both exhibited an increased risk of each of the endpoints compared with phenogroup 1. In analyses adjusting for the MAGGIC risk score, phenogroup 3 continued to be significantly associated with increased risk of the primary endpoint, combined all-cause mortality or heart failure hospitalization, and heart failure hospitalization alone.

Spironolactone randomized therapy was associated with a more pronounced reduction in the risk of the primary endpoint (p for interaction = 0.016) and heart failure hospitalization (p for interaction = 0.007) in phenogroup 3. In stratified analyses, randomized spironolactone treatment was associated with a significantly lower risk of the primary endpoint (HR: 0.75; 95% CI: 0.59 to 0.95), and heart failure hospitalization (HR: 0.69; 95% CI: 0.53 to 0.90) in phenogroup 3 but did not appear to substantially benefit the other phenogroups (Table 3, Figure 4).

Analyses restricted to the americas

In analyses restricted to individuals enrolled in the Americas, 3 clinical phenotypes of HFpEF were also identified (bootstrap LR test for 2 vs. 3 classes, p = 0.01; LR test for 3 vs. 4 classes, p = 0.24), with very similar clinical characteristics across the groups as observed in the overall study population (Table 2).

In Cox proportional hazards models, phenogroup 3 consistently demonstrated poorer prognosis than phenogroup 1 (Online Table 6, Online Figure 4); phenogroup 2 demonstrated higher risk of the composite endpoint of all-cause mortality or hospitalization for heart failure (HR: 1.63; 95% CI: 1.19 to 2.23) and all-cause mortality (HR: 1.68; 95% CI: 1.24 to 2.27) compared with phenogroup 1 but no difference in the primary endpoint and hospitalization for heart failure.

Similar to the overall cohort, in analyses stratified by phenogroup, assignment to the spironolactone treatment arm was associated with significantly lower risk of the primary endpoint (HR: 0.74; 95% CI: 0.56 to 0.97, p for interaction = 0.031) and hospitalization for heart failure (HR: 0.70, 95% CI: 0.52 to 0.95, p for interaction = 0.023) in phenogroup 3 but not in the other phenogroups (Table 3, Online Figure 5).

Discussion

In this analysis of TOPCAT, we identified 3 phenogroups of HFpEF based on standard clinical features using LCA and characterized circulating biomarker profiles, cardiac and vascular phenotypes, outcomes, and response to spironolactone therapy between these phenotypic groups (summarized in Figure 5 and the Central Illustration). We demonstrate important differences in the levels of multiple biomarkers, suggesting distinct underlying pathophysiologic processes in these phenogroups. We also found pronounced differences in large-artery stiffness, pulsatile and resistive arterial load, and echocardiographic parameters of LV and LA structure and function among the groups. Importantly, we identified the phenogroup that best responded to spironolactone randomized therapy in TOPCAT. Our findings support the existence of distinct clinical HFpEF phenogroups and contribute to a better understanding of the heterogeneity across these phenogroups with regard to biomarker profiles, cardiac/arterial structure and function, prognosis, and response to spironolactone.

Comparison with HFpEF phenogroups identified in previous studies

Previous studies have identified subgroups of clinical characteristics in HFpEF that were associated with differences in clinical outcomes (4,5). In a post hoc analysis of data from the I-PRESERVE (Irbesartan in Heart Failure with Preserved Ejection Fraction Study) and CHARM (Candesartan in Heart failure: Assessment of Reduction in Mortality and morbidity)-Preserved study (4), Kao et al. used latent class analysis to identify 6 clinical subgroups of patients with HFpEF. Two of the subgroups, predominantly composed of obese patients with diabetes and older individuals with the highest prevalence of AF, were consistent with our phenotypes 3 and 2, respectively, with regard to clinical characteristics and prognosis. There were no detailed echocardiographic or biomarker data available in this previous study (only LV ejection fraction and n-terminal B-type natriuretic peptide). Although we identified a different number of subgroups by LCA, this study only used unadjusted BIC to determine the optimal number of subgroups, which has been demonstrated to be inferior to the bootstrap LR test used in the current study (corroborated with AIC, BIC, and sample-size–adjusted BIC) (12,13).

In a previous single-center study of 397 patients with HFpEF, Shah et al. (5) used hierarchical modeling to identify 3 phenotypes characterized by younger age with lower B-type natriuretic peptide, the least concentric LVs, smallest LA volumes, and lowest E/e′ ratio; higher prevalence of diabetes and obesity, with large LV mass; and older age with the highest prevalence of AF and LA dilation. The authors found a similarly elevated risk of cardiac hospitalization, hospitalization for heart failure, death, and a composite of the 3 individual endpoints among their second and third phenotypes compared with the first phenotype. The phenogroups found in our study are in general consistent with the findings of Shah et al. (5). Our study was further strengthened by the use of data from a multicenter international study population with robust biomarker data to facilitate exploration of underlying pathophysiologic differences and differences in treatment response across the clinical phenotypes.

Phenogroup 1 (younger, with mild symptoms)

Previous evidence suggests that younger individuals with confirmed HFpEF have similar LV filling pressures and prevalence of LV hypertrophy to older individuals with HFpEF (20), which we did not observe. Thus, it is unclear if phenogroup 1 comprised individuals with genuine HFpEF or if it was confounded by individuals with noncardiac causes of dyspnea. This hypothesis is supported by the observation that phenogroup 1 had the highest prevalence of smoking and highest levels of MMP-9 among the phenogroups. MMP-9 is involved in respiratory-tract remodeling and is elevated in patients with asthma and chronic obstructive pulmonary disease (21). Of note, phenogroup 1 had the highest rates of enrollment in Russia and Georgia. Our findings are thus consistent with previous work questioning the validity of HFpEF diagnoses among Eastern European participants of TOPCAT (9). Although the proportion of patients in this phenogroup in the Americas was small, the results were similar after restricting to subjects enrolled in the Americas. Our findings emphasize the importance of careful echocardiographic assessment and pulmonary function testing when evaluating individuals for possible HFpEF (22).

Phenogroup 2 (older, with stiff arteries, small LVs and AF)

Phenogroup 2 was characterized by the oldest mean age, highest prevalence of women and AF, and intermediate E/e′ ratios. This phenogroup is consistent with the phenotype of HFpEF in older adults reported by Tromp et al (23). We found that this phenogroup also exhibited normotrophic concentric LV remodeling, LA enlargement, and large-artery stiffness. Phenogroup 2 demonstrated the highest levels of several markers of innate immunity, which have been previously associated with vascular injury as well as aging (24). This group also exhibited the highest levels of osteoprotegerin, which is a regulator of tissue calcification that has been linked with increased large-artery stiffness, independent of atherosclerotic disease (25). Aortic stiffness is associated with faster wave transit time from the heart to reflection sites and back to the aorta (26). These premature wave reflections increase the mid-to-late systolic LV workload, which is, in turn, associated with diastolic dysfunction (27). In addition, pulsatile LV load, aortic stiffness, and AF are each associated with the increased LA remodeling and dysfunction (28), which is consistent with the marked LA enlargement in this phenogroup.

Phenogroup 3 (obese, diabetic, with advanced symptoms)

Phenogroup 3 exhibited the worst overall prognosis, even after adjustment for comorbid diseases and the lowest risk associated with spironolactone therapy versus placebo. This phenogroup also exhibited the highest levels of biomarkers in multiple pathways, including high renin, TNF-alpha–mediated inflammation, markers of renal injury/dysfunction, dysregulated intermediary metabolism, liver fibrosis, angiogenesis and FGF-23 (a regulator of mineral metabolism that also promotes LV hypertrophy) (29). These elevated biomarkers are consistent with known links of metabolic dysregulation in obesity with inflammation, elevated renin-angiotensin-aldosterone system activity, nonalcoholic fatty liver disease (resulting in hepatic fibrosis), and CKD (30). Individuals with obesity have impaired natriuresis due to upregulation of renin-angiotensin-aldosterone system activity, which is exacerbated by HFpEF, likely driving the increased fluid retention and resulting hospitalizations observed in phenogroup 3. Related to these mechanisms, in HFrEF, mineralocorticoid antagonist therapy yields more pronounced reduction in adverse outcomes in obese compared with nonobese persons (31). This pathophysiologic evidence is consistent with the lower risk of adverse outcomes we observed in patients in the spironolactone treatment arm in phenogroup 3. This phenogroup also exhibited the highest prevalence of depression. Depression and HF have also been associated with overlapping underlying biologic mechanisms, including sympathetic activation (32) and elevated proinflammatory cytokines such as interleukin-6 and TNF-alpha (33). Therefore, the additive effects of inflammation among patients with both HF and depression may affect severity and progression of disease.

Study strengths

Strengths of our study include the inclusion of a well-characterized multicentric HFpEF cohort, the use of multiple biomarkers, the use of detailed phenotypes of cardiac and arterial structure/function, and the application of robust and well-established clustering techniques.

Study limitations

We did not have plasma samples, echocardiographic or tonometry data from all TOPCAT trial participants, and we had to restrict aspects of the study to subpopulations with available samples or data. Although Luminex assays allow for the measurement of multiple biomarkers, assay-specific limits of detection are not necessarily equivalent to clinical assays.

Conclusions

Our study demonstrates meaningful differences in biomarkers as well as cardiac and arterial structure across 3 distinct, clinically identifiable phenogroups: phenogroup 1 (younger, with mild symptoms), phenogroup 2 (older, with stiff arteries, small LVs, and AF), and phenogroup 3 (obese, diabetic, with advanced symptoms). The 2 latter phenotypes, which constitute genuine high-risk HFpEF, exhibit distinct abnormalities in biomarkers, cardiac/arterial structure and function, and differential response to spironolactone therapy. In contrast, phenogroup 1 represents a low-risk group, which may not represent genuine HFpEF and may be confounded by lung disease, which, in turn, explains geographic differences in TOPCAT. Our findings provide important insights into potential driving factors behind differences in prognosis and response to treatment across the clinical phenogroups. In the absence of clear therapies to improve prognosis in HFpEF, separation of individuals into clinically identifiable subgroups can help to identify patients who have the potential to benefit most from targeted interventions. Specifically, phenogroup 3 had the poorest prognosis and exhibited a reduction in adverse outcomes associated with randomized treatment with spironolactone. Given the heterogeneity of risk factors and outcomes in patients with HFpEF, future trials should focus on different interventions in distinct phenogroups of patients with HFpEF.

Footnotes