Phenomapping in heart failure with preserved ejection fraction: insights, limitations, and future directions

Abstract

Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous entity with complex pathophysiology and manifestations. Phenomapping is the process of applying statistical learning techniques to patient data to identify distinct subgroups based on patterns in the data. Phenomapping has emerged as a technique with potential to improve the understanding of different HFpEF phenotypes. Phenomapping efforts have been increasing in HFpEF over the past several years using a variety of data sources, clinical variables, and statistical techniques. This review summarizes methodologies and key takeaways from these studies, including consistent discriminating factors and conserved HFpEF phenotypes. We argue that phenomapping results to date have had limited implications for clinical care and clinical trials, given that the phenotypes, as currently described, are not reliably identified in each study population and may have significant overlap. We review the inherent limitations of aggregating and utilizing phenomapping results. Lastly, we discuss potential future directions, including using phenomapping to optimize the likelihood of clinical trial success or to drive discovery in mechanisms of the disease process of HFpEF.

This article is part of the Spotlight Issue on Heart Failure

1. Introduction

Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous entity caused by a complex interaction between non-cardiac and cardiac factors. While data on HFpEF pathophysiologic mechanisms are evolving, contemporary studies support the hypothesis that comorbidities and risk factors result in abnormalities of cardiac and non-cardiac structure and function, and thereby lead to symptoms, exercise intolerance, morbidity, and mortality.^1,2 HFpEF probably represents multiple biological phenotypes (e.g. obese/metabolic, right heart dysfunction/pulmonary vascular disease, older hypertensive and vascular aging, left atrial myopathy) with significant morphological and functional heterogeneity.^3–9 This variation in mechanisms and manifestations is considered a primary contributor to the modest or neutral treatment effects observed in most HFpEF trials to date, falling short of meeting primary outcomes, aside from recent sodium-glucose cotransporter-2 (SGLT-2) inhibitor trials.

Therefore, the development and understanding of various phenotypes within HFpEF has become a priority, and phenomapping has emerged as a tool to achieve this goal. Phenomapping is the application of statistical learning techniques to data to categorize patients into distinct subgroups (Figure 1). Shah et al. published novel findings of phenomapping in HFpEF in 2015, utilizing a range of detailed data inputs and statistical learning techniques to produce three unique phenogroups.¹ In the past few years since this seminal study, there has been a significant increase in phenomapping efforts in HFpEF, using various data sources, data elements, and statistical techniques. This review summarizes methodologies and key findings from these studies (Table 1, Supplementary material online, Table S1), including conserved discriminating factors and phenotypes in HFpEF across studies. We also review the inherent limitations of aggregating and utilizing phenomapping results and note that phenomapping has had limited implications for clinical care and clinical trials because currently described phenotypes are not always reliably identified in each study and may have significant overlap. We discuss future directions including the potential application of phenomapping results in mechanistic and clinical trial research and clinical practice as well as unmet needs to be addressed by future phenomapping efforts.

Figure 1.

Phenomapping to date in HFpEF—Individual study structure, data inputs, potential goals, gaps, and outcomes.

Table 1.

Summary of selected HFpEF phenomapping studies and data inputs

Study	Derivation		Validation		Data inputs								# of groups identified	Differential outcomes by group demonstrated
	Source	n	Source	n	Clinical	Basic Labs	Imaging	Select Biomarker	Large-scale ‘Omics	ECG	Exercise data	Invasive Hemodynamics
Shah et al.1	Single-centre/clinical	397	External	107	✓	✓	✓	✓NP	–	✓	–	–  a	3	✓
Kao et al.10	I-PRESERVE trial	4113	External	3203	✓	✓	–	–	–	–	–	–	6	✓
Przewlocka-Kosmala et al.11	Single-centre/clinical	177b	n/a	n/a	–	–	✓	✓Galactin-3	–	–	✓	–	3	✓ c
Cohen et al.12	TOPCAT trial	1765 d	n/a	n/a	✓	–	–	–	–	–	–	–	3	✓
Segar et al.13	TOPCAT Americas	654	Int/ext	1113/198	✓	✓	✓	✓NP	–	✓	–	–	3	✓
Hedman et al.14	Multicentre registry	320	n/a	n/a	✓	✓	✓	✓NP	–	–	–	–	6	✓
Schrub et al.15	Multicentre registry	356	n/a	n/a	✓	✓	✓	–	–	–	–	–	3	–
Stienen et al.16	Multicentre registry	392	n/a	n/a	–	–	–	–	✓	–	–	–	2	✓
Harada et al.17	Single-centre/clinical	350	Internal	133	✓	✓	✓	–	–	–	–	–	4	✓
Arevalo-Lorido et al.18,e	Multicentre registry	1934	n/a	n/a	✓	✓	–	–	–	–	–	–	7	✓
Sabbah et al.19	Multiple trialsf	301	n/a	n/a	✓g	–	–	–	✓	–	–	–	3	✓
Uijl et al.20	Multicentre registry	6909	External	2153	✓	✓	–	–	–	–	–	–	5	✓
Gu et al.21	Single-centre/clinical	970	External	290	✓	✓	✓	✓NP	–	–	–	–	3	✓
Casebeer et al.22	Clinical/claims	1515	n/a	n/a	✓	–	–	–	–	–	–	–	3
Nouraei et al.23	Single-centre/clinical	197	n/a	n/a	✓	✓	✓	–	–	–	–	–	6	✓
Wu et al.24	Multigenerational registry	125	n/a	n/a	–	–	–	–	✓	–	–	–	2	✓
Woolley et al.25	Multicentre registry	429	n/a	n/a	–	–	–	–	✓	–	–	–	4	✓
Hahn et al.26	Single-centre/clinical	38	n/a	n/a	–	–	–	–	✓	–	–	–	3	✓
Jones et al.27	Single-centre/clinical	21 h	n/a	n/a	–	–	✓	–	–	–	–	✓	3	–
Fayol et al.28	Single-centre/clinical	928	n/a	n/a	✓	✓	✓	✓NP	–	–	–	–	3	✓

Completed in 216 patients and not formally included in phenomapping technique.

177 HFpEF patients and 51 asymptomatic controls.

Outcomes evaluated after combining 2 of the 3 phenogroups given similarities observed by authors.

TOPCAT Americas cohort; full TOPCAT cohort also analysed in this study.

Similar, smaller study from same author/year not included separately.

Three Heart Failure Clinical Research Network (HFN) HFpEF trials—RELAX (Phosphodiesterase-5 Inhibition to Improve Clinical Status and Exercise Capacity in Diastolic Heart Failure), NEAT (Isosorbide Mononitrate in Heart Failure with Preserved Ejection Fraction), and INDIE (Inorganic Nitrite Delivery to Improve Exercise Capacity in Heart Failure With Preserved Ejection Fraction).

Obesity status only

Along with 10 HFrEF patients.

2. Methodology and terminology

To identify relevant published data, we searched MEDLINE for articles published between January 1994 and December 2021 (see the Appendix 1 for the search strategy, and Appendix 2 for related phenomapping studies excluded from the primary analysis). We manually searched reference lists of relevant reviews, including studies and background data, to find pertinent citations that our searches might have missed. For this review, mutually exclusive subgroups of patients are described as clusters or phenogroups within HFpEF based on the predominant approach used in the literature. Summarized data across studies resulted in subgroups of patients with shared characteristics but not necessarily mutually exclusive (i.e. overlapping features) from other subgroups; these groups are termed phenotypes within HFpEF (Figure 2). These overlapping phenotypes may be more consistent with the clinical reality encountered in practice but this remains to be proven definitively and degree of overlap in aggregate phenomapping results will affect the utility of findings as discussed further in Limitations and Future Directions below.

Figure 2.

Mutually exclusive phenogroups by individual study and overlapping phenotypes resulting from data summarization.

3. Data sources and data elements

Table 1 shows the wide range of data sources and data elements used for phenomapping in HFpEF. Importantly, the quality and depth of the phenogroups identified through these approaches are as good (or bad) as the input variables used to separate the clusters. Several early and recent studies have focused on readily available clinical data, including demographics, vital signs, comorbidities, and routine laboratory data.^{10,12,18,20,22,29} Many studies have also used features of cardiac structure and function from echocardiography, ranging from basic diastolic function measures to complex non-invasive haemodynamic estimations.^{1,11,13–15,17,21,23,27} Select biomarkers, including B-type natriuretic peptide (BNP), N-terminal proBNP (NT-proBNP), and galectin-3, have been added to the clustering analyses of several studies,^{1,11,13,14,21}. More recent studies have used large-scale ‘omics data alone to classify patients.^{16,19,24–26,30,31} To date, only a few studies have included response to exercise data¹¹ (despite exercise intolerance representing a primary symptom of chronic HFpEF) and invasive haemodynamics²⁷ as phenomapping inputs, and none have utilized raw image data (e.g. echocardiographic or magnetic resonance imaging DICOM data) as input. Given the evolution of ejection fraction cutoffs and the addition of heart failure with mildly reduced ejection fraction (HFmrEF) over time, ‘HFpEF’ phenomapping studies to date have utilized datasets with varied EF cutoffs. Given increasing signals for therapy responsiveness across EF phenotypes,^32–34 future phenomapping studies may benefit from broader inclusion of HF across EF spectrum as discussed further in the section ‘Future directions’.

Notably, there may be significant differences between the utilization of routine clinical data as source input data for phenomapping and the details observed through subsequent deeper assessment of these identified patient groups (or vice-versa). For instance, Cohen et al. focused on simple, widely accessible clinical data to produce their phenomapping process but then determined proteomic differences among the identified HFpEF phenogroups.¹² Wooley et al., on the other hand, utilized proteomics data alone (363 proteins) in their phenomapping technique and then assessed a range of clinical and outcome data across the identified proteomic-based HFpEF phenogroups.²⁵ These approaches may result in similar or different phenogroups at times, and generalizability will vary based on the patient population tested and the methodological techniques.

Additionally, the ability to develop hypothesis-generating results regarding causality, prediction, or effect modification depends, in part, on phenomapping design. For example, Flint et al.³⁵ used previously established HFpEF phenogroups by Kao et al.¹⁰ to study differences in treatment response to spironolactone in the Treatment of Preserved Cardiac Function Heart Failure with an Aldosterone Antagonist (TOPCAT) trial from the Americas. The study found significant differences in biological response to spironolactone (including changes in potassium, creatinine, and blood pressure), clinical outcomes, and patient-reported outcomes—all features best studied across established phenotypes instead of including as phenomapping inputs.

Data sources utilized in phenomapping also determine the utility, limitations, and generalizability of each study’s findings. Phenomapping efforts to date have utilized a variety of data sources across the USA, Europe, and Asia, including single-centre clinical cohorts or multicentre registries such as Karolinska Rennes (KaRen) Study, the Swedish Heart Failure Registry (SwedeHF), the Metabolic Road to Diastolic HF (MEDIA-DHF), and the Asian Sudden Cardiac Death in Heart Failure (ASIAN-HF) study, and the Systems Biology Study to Tailored Treatment in Chronic Heart Failure (BIOSTAT-CHF) study. Others have used trial data, namely TOPCAT (full and Americas cohorts), Irbesartan in Heart Failure with Preserved Ejection Fraction Study (I-PRESERVE), Candesartan in Heart failure: Assessment of Reduction in Mortality and morbidity (CHARM)-Preserved, and Prevalence of Microvascular Dysfunction in HFpEF (PROMIS-HFpEF) trials. Importantly, several studies have included validation cohorts, either internal (within the derivation cohort) or external (in a separate cohort compared to the original dataset)—Shah et al.,¹ derivation n = 397, external validation n = 107; Seger et al.,¹³ derivation n = 654, internal validation n = 1113, external validation n = 198; Uijl et al.,²⁰ derivation n = 6909, external validation n = 2153; and others. Yet, similar to that of risk model development, many phenomapping analyses have not included robust validation due in part to dataset limitations (e.g. access to or availability of data elements).

4. Analytical strategies

Analytical approaches to phenomapping broadly fit along the spectrum of statistical learning techniques from latent class analysis (LCA) to unsupervised statistical learning strategies and dimensionality reduction methods such as principal component analysis. LCA, for instance, is a commonly employed, iterative, model-based approach to identify clusters using specific combinations of multiple characteristics and is based on estimates of the probability of each participant being a member of each latent class (cluster), similar to other statistical learning techniques.^36,37 Recent phenomapping studies in HFpEF have largely relied on unsupervised statistical learning techniques (i.e. agglomerative hierarchical clustering, K-means clustering, spectral clustering), which learn the intrinsic structure within a dataset and aim to find patterns in data.¹ Supervised techniques (i.e. support vector machines), on the other hand, utilize labelled training data in order to predict specified outcomes or identify given labels in unlabelled data, and, therefore, may be more useful if/when phenogroups are established.^1,38

Each unsupervised statistical learning approach has strengths and limitations, impacting characteristics and relative size of phenotypes identified. For instance, hierarchical, model-based clustering divides data into categories, then subcategories, and can then merge clusters for final phenotypes, compared to partitioning around medoids (PAM), a nonhierarchical method that assigns ‘k’ random entities to be medoids (k-medoids, similar to k-means);²³ there is some evidence that k-medoids algorithms may have higher cluster accuracy compared with other techniques for mixed variable data,³⁹ and this approach has produced a different set of distinct HFpEF phenogroups compared to other approaches.²³

Incorporating Bayesian information criterion analysis is an important consideration for clustering studies. This will penalize increased model complexity (i.e. higher number of phenogroups) and thereby help produce the most parsimonious solution, improving generalizability to other datasets; this approach is termed ‘regularization’.⁴⁰ Generalizability is degraded by degree of overfitting (including modelling inherent noise in the training dataset), which limits the model performance in external datasets (termed low bias, high variance); bias and variance should be minimized in modelling efforts to balance this trade-off.³⁸ Importantly, there is no established gold standard among clustering techniques. Some studies that have applied different methodologies to the same data in parallel (i.e. LCA vs. k-means) arrived at different results, both in the distribution of patients and the number of clusters.³⁷ Other studies have used combined statistical learning approaches; this combination of technique strengths and sensitivity analyses with alternative approaches may prove most accurate in defining generalizable, reproducible HFpEF clusters.¹

5. Key discrimating factors

Before reviewing the specific phenotypes consistently observed across studies, we summarize key considerations around specific variables assessed in prior analyses. Phenomapping work in HFpEF has identified numerous key discriminating factors that differentiate phenotypes in terms of characteristics and outcomes (see Supplementary material online, Table S1). These factors overlap with influential characteristics of HFpEF identified in non-phenomapping work, and it should be noted that the advanced statistical learning techniques of phenomapping were not necessary to demonstrate many of these findings in HFpEF; many of these findings are logical conclusions from general knowledge of HFpEF and disease processes in general. For instance, age and sex are prognostic markers, common to many disease states, which have played an influential role in HFpEF clustering. In comparing phenomapped clusters, ‘older’ clusters are generally >75 years of age on average (cohort range from 60s–80 s years old) and are more commonly associated with hypertension, chronic kidney disease (CKD), atrial fibrillation (AF), female sex, and worse hospitalization/mortality outcomes.^{1,11,12,14,17,20,23} While hypertension was common across all phenotypes (and therefore a less prominent discriminator), large-artery stiffness (as measured by pulse wave velocity or other metrics) and concentric remodeling with a relatively small left ventricular (LV) cavity (compared to other phenotypes including normal LV volumes and concentric hypertrophy cavities) were found to be distinctly more common in older, more female cohorts.^12,26,27

Comorbidities, including AF, CKD, type II diabetes mellitus, and obesity, play an important role in clustering results (see Supplementary material online, Table S1).^{1,10,12–17,20,21,24–26} While coronary artery disease (CAD) is well known to be common and impactful in HFpEF (and was common in full cohorts of the analysed studies), the presence of this pathophysiology did not commonly discriminate mutually exclusive phenogroups in these studies (which may be due to lack of granularity in CAD severity, microvascular disease, etc.). Still, it is clear that CAD is an important and influential component of HFpEF pathophysiology and optimal care, even if it does not strongly distinguish mutually exclusive phenogroups. Among laboratory parameters, BNP/NT-proBNP were the most commonly identified differentiating laboratory measures; only one study identified metabolic markers including glucose and total bilirubin to have higher discriminatory power.¹³ While lower BNP levels were linked to obesity to a certain extent, particularly low BNP levels were closely associated with younger patients with variable burden of obesity and better outcomes than other clusters.^{1,10,12,14,17,19,25} Echocardiography was useful in distinguishing phenotypes using LV features (LV hypertrophy, LV mass index, LV cavity size), non-invasive RV haemodynamics (pulmonary arterial systolic pressure—PASP, tricuspid annular plane systolic excursion—TAPSE), and left atrial dimensions (left atrial volume index—LAVi).

It should be noted that these characteristics discriminate phenogroups typically assessed at a single timepoint in the included studies. Still, it is plausible that these characteristics could also represent discriminating features of stage of disease, as discussed in other reviews.⁴¹ For instance, age and arterial stiffness or pulmonary hypertension and RV dysfunction may represent features of advanced, progressive HFpEF, instead of a mechanistically different phenotype. These longitudinal intricacies of phenotypes and the potential for diagnosis bias (i.e. length or lead time bias) may become clearer through further research with multiple, longitudinal datapoints (with or without phenomapping methodology).

6. Identified phenotypes

Phenomapping studies have identified between two and six clusters/phenogroups with shared features of several phenotypes among 20 studies (see Supplementary material online, Table S1). Eleven of the 20 included studies resulted in three phenogroups, the mean number of phenogroups among studies was 3.75, and the median number of phenogroups was 3; therefore, we chose to focus on the three most commonly identified phenogroups (and include details on up to four phenogroups in Supplementary material online, Table S1), while acknowledging that it is not currently clear how many distinct phenotypes are present within the HFpEF population or if it will be possible to identify truly distinct phenotypes through future work.

Three phenotypes repeatedly arise in the reviewed phenomapping literature and seem to represent significant subgroups of HFpEF with additional support from non-phenomapping HFpEF literature. While the degree of precise concordance/agreement between clusters across studies is low, cohorts with features of each of these three phenotypes arose in, at least, 5–10 studies each (see Supplementary material online, Table S1); while further concordance would be ideal, this degree of evidence from studies with widely disparate data inputs and statistical techniques across a wide range of regions and HFpEF populations with variable quality/rigor of study designs may be considered hypothesis-generating. Further, there is additional support for the features, outcomes, and potential treatment responsiveness of these phenotypes from the non-phenomapping literature, as included below. Still, the proposed phenotypes from this review will require significant additional exploration and validation before reliable use in clinical or trial settings is possible.

For this review, phenotypes were sorted by relative size (largest to smallest percentage of HFpEF cohorts, Supplementary material online, Table S1) and termed ‘older, vascular aging’ phenotype, ‘metabolic, obese’ phenotype, and ‘relatively younger, natriuretic peptide (NP) deficiency’ phenotype (Figure 3).

Figure 3.

Consort diagram of proposed HFpEF phenotypes and insights from phenomapping studies to date.

6.1. ‘Older, vascular aging’ phenotype

An ‘older, vascular aging’ phenotype appears to represent ∼30–50% of HFpEF patients by many phenomapping studies (i.e. most of those identifying three clusters)^{1,11,12,14,16} and as low as ∼15% of some cohorts (i.e. those with >3 clusters or, potentially, lower severity-of-illness populations such as Wu et al.²⁴). This cluster was first identified by Shah et al., and its defining, shared characteristics across studies are older age, prominent arterial stiffness/hypertension, CKD, and adverse outcomes.^{1,10–12,14,16,17,20,24} This phenotype has also been identified in phenomapping analysis of a large Asian HFpEF/HFrEF cohort²⁹ and two other non-phenomapping studies stratifying patients with HFpEF according to age.^42,43 Importantly, there is some evidence that this phenotype has significant overlap with the ‘metabolic, obese’ phenotype of HFpEF (which would limit its clinical utility as a distinct cluster entity) as a few phenomapping studies identify clusters that have prominent features of aging, arterial stiffness, and renal disease (‘older, vascular aging’ phenotype) along with a heavy burden of metabolic inflammatory disease (the ‘metabolic, obese’ phenotype) such as Segar et al.¹³

Patients from the ‘older, vascular aging’ phenotype often had more cardiac hypertrophy and patterns of concentric remodelling,^{1,12,14,26,27} but smaller hearts by left ventricular end-diastolic volume (LVEDV) were not a defining feature in every ‘older, vascular aging’ phenotype cluster.¹ This may reflect that there is a sub-cluster of patients with relatively small volume, concentrically remodelled hearts and less severe outcomes, or may reflect the overlap between phenotypes across studies.

Additional characteristics of elevated PASP, sometimes associated with right ventricular (RV) dysfunction, and left atrial myopathy/failure seem to represent subsets of this phenotype in particular.^1,16 Notably, the cardiorenal, RV dysfunction, and LA myopathy components of this phenotype appear to drive increased symptoms of congestion and dyspnea on exertion, commonly noted as worse NYHA functional class, lower peak oxygen consumption (VO₂), and increased hospitalization/mortality rates in phenomapping studies.^{1,10,11,16,17} Lastly, several studies have noted clusters of patients with ‘older, vascular aging’ phenotype-like physiology as ‘HFrEF-like’ compared to HFrEF cohorts (i.e. stiffer, dilated hearts with lower contractility and RV dysfunction) with associated adverse outcomes.^26,27

‘Older, vascular aging’ phenotype clusters have demonstrated increased inflammation^11,12,16, immune system activity (particularly innate),^12,16 metabolism (particularly mineral metabolism and tissue calcification, such as osteoprotegerin).^12,16 Specifically, inflammatory biomarkers highlighted include galectin-3,¹¹ tumor necrosis factors (TNF) and TNF receptor activity,¹⁶ and those associated with innate immunity (interleukin-8, pentraxin-3, soluble intercellular adhesion molecule-1).¹² Wooley et al. also highlights inflammatory pathways, including TNF and immune regulation processes in a cluster similar to the ‘older, vascular aging’ phenotype, albeit with a greater burden of DM.²⁵

Notably, in the first human myocardial tissue phenotyping study to date, Hahn et al. derived phenotypes of HFpEF from myocardial transcriptomics, comparing human HFpEF myocardial tissue to HFrEF and controls. In this study, 3 HFpEF subgroups were identified based on myocardial transcriptomics and, of these, a phenotype that was older with increased pulmonary vascular resistance, increased LVMI, and increased NT-proBNP was associated with the worst clinical outcomes of the three groups. This subgroup was transcriptionally closest to HFrEF, whereas the other two groups identified were transcriptionally distinct from HFrEF.²⁶

6.2. ‘Metabolic, obese’ phenotype

A ‘metabolic, obese’ phenotype appears to represent 25–30% of HFpEF patients in most phenomapping studies and represents an increasingly recognized ‘metabolic’ subtype of HFpEF.^{1,12,13,19,25} In this group of patients, it is proposed that the inflammatory, metabolic milieu of comorbidities, led by obesity and DM, contributes to the pathophysiology of HFpEF, as originally outlined by Paulus and Tschope⁴⁴, and later expanded upon by Paulus and Zile.⁴⁵ Along with its burden of obesity and DM, this group also demonstrates heavy overlap with the ‘older, vascular aging’ phenotype, including frequent CKD and elevated RVSP as well as higher rate of adverse outcomes.^13,25 These patients are typically slightly younger than the ‘older, vascular aging’ phenotype and older than the ‘relatively younger, NP deficiency’ phenotype. Additionally, phenomapping analysis of the TOPCAT trial (full and Americas-only cohorts) indicates that this phenotype may respond favourably to spironolactone.¹² Furthermore, natriuretic peptide-specific analysis indicates that the ‘obesity phenotype’ may benefit from spironolactone therapy, particularly younger, obese patients with relatively low BNP (overlap with the ‘relatively younger, NP deficiency’ phenotype).⁴⁶ Randomized trial evidence also has shown that caloric restriction and aerobic exercise both separately improve peak oxygen consumption in an older, obese, HFpEF cohort very similar to this phenotype, and that these effects are additive.⁴⁷

The non-phenomapping literature in HFpEF has importantly identified that regional adiposity and abdominal visceral and/or epicardial adipose tissue may be critical to the pathophysiology and outcomes of patients with HFpEF, compared to simply total body mass index, in USA and European cohorts.^48–51 In Asia, this cardiometabolic phenotype of HFpEF is particularly common in countries in Southeast Asia like Singapore.²⁹ Interestingly, Asian data also suggest that the phenotypic expression of HFpEF may be driven mainly by diabetes and cardiometabolic-inflammatory factors, rather than obesity per se, since cluster analysis revealed a ‘lean diabetic’ phenotype of HFpEF in the same Asian regions, with similar characteristics despite a normal (or even low) body mass index.²⁹ In these lean diabetic patients with HFpEF, it is postulated that, rather than body weight per se, it is the distribution of fat, with increased visceral and epicardial adiposity, that may be playing a role in the pathophysiology of HFpEF.^52–54

In terms of biomarkers of pathophysiology, a ‘metabolic, obese’ phenotype appears to be largely defined by inflammatory proteins and pathways. Notably, TNF-α (and related receptors) and growth differentiation factor 15 (GDF15) were more heavily expressed in cohorts of this phenotype compared to others.^12,19,25 Several studies also found fibroblast growth factor-23 (FGF23), a mediator of mineral metabolism, to be more highly expressed in ‘older, vascular aging’ phenotype and/or ‘metabolic, obese’ phenotype cohorts, namely in patients with CKD in these groups.^12,14 Cohen et al. also identified markers of renal injury/dysfunction (i.e. cystatin C), dysregulated metabolism, liver fibrosis, and angiogenesis in this ‘metabolic, obese’ phenotype cohort.¹² In addition to TNF-α, Sabbah et al. found elevated expression of other inflammatory mediators (IL-6, interleukin-6; VCAM-1, vascular cell adhesion molecule-1; ICAM-1, intercellular adhesion molecule-1; INFγ, interferon-gamma; MCP-1, monocyte chemoattractant protein-1) and pro-fibrotic mediators (CITP, C-telopeptide for type I collagen; PIIINP, procollagen III n-terminal peptide; IGFBP7, insulin-like growth factor-binding protein-7; and, to a lesser extent, GAL-3, galectin-3) in a ‘pan-inflammatory’ cohort consistent with a ‘metabolic, obese’ phenotype.¹⁹ Lastly, in the Hahn et al. transcriptomics study, the ‘metabolic, obese’ phenotype identified had a distinct transcriptomic signature compared to the ‘older, vascular aging’ phenotype.²⁶

6.3. ‘Relatively younger, natriuretic peptide deficiency’ phenotype

The ‘relatively younger, NP deficiency’ phenotype appears to be characterized by relatively younger age (∼60–70 in several studies, up to ∼75 years of age), lower risk of adverse outcomes, and relatively low natriuretic peptides. This group often represents 15–35% of the HFpEF phenomapping studies,^1,12,14,19 but is described in cohorts as large as 40–45% in some studies.^17,21,25 Importantly, whether all of this group represents ‘true’ HFpEF is unclear and strict application of a HFpEF definition in future phenomapping studies as well as in observational and clinical studies will be critical to clarify this.

The pathophysiology of relatively low natriuretic peptide levels in this HFpEF phenotype (reported as 29% of patients with elevated pulmonary capillary wedge pressure in one HFpEF cohort⁵⁵) has been described in detail in other reviews.⁵⁶ Briefly, while HFpEF patients overall demonstrate lower BNP/NT-proBNP levels compared to their HFrEF counterparts due to lower diastolic wall stress, this younger, often obese (‘relatively younger, NP deficiency’ phenotype, overlapping with the ‘metabolic, obese’ phenotype) cohort is particularly defined by low natriuretic peptides.⁵⁶ Obesity is well known to be related to lower natriuretic peptide levels through reduced production and increased clearance;⁵⁷ there is a further circular nature to this relationship as low natriuretic peptide levels are associated with increased visceral adiposity and therefore even lower circulating natriuretic peptide levels.^56,58 Additionally, there may be contributions to low BNP/NT-proBNP levels from genetic determinants, including variations in natriuretic peptide genes or in race-related genetic ancestry,^59,60 sex-specific androgen levels and hormone therapy,⁶¹ and insulin resistance.⁶² Exercise intolerance is another central feature of HFpEF patients present across the full spectrum of HFpEF including in this ‘relatively younger, NP deficiency’ phenotype as demonstrated by Sabbah et al. and others, even though these patients tend to have lower NT-proBNP and lower risk of adverse outcomes compared to other phenotypes.¹⁹ This profile of HFpEF patients with lower/normal natriuretic peptides has been highlighted by a recent invasive haemodynamic study, which showed this group of patients was younger and had better outcomes than the higher NP-cohort.⁶³

There is evidence that patients of a ‘relatively younger, NP deficiency’ phenotype may respond favorably to spironolactone from a phenomapping analysis of TOPCAT Americas (no significant interaction effect across phenogroups, but trend toward nominally lower risk of outcomes in contrast to other phenogroups)¹³ and other post-hoc TOPCAT Americas analyses.⁴⁶ Two studies stratifying patients with HFpEF according to age categories in ASIAN-HF, TOPCAT Americas, CHARM-Preserved and I-Preserve demonstrated that younger patients were men, predominantly obese, had lower natriuretic peptides and better outcomes than older patients.^42,43 Importantly, in the study performed in ASIAN-HF, younger patients demonstrated worse outcomes than age- and sex-matched hypertensive controls, suggesting that these patients truly have a phenotype of HF.⁴² Together, these data suggest that the ‘young HFpEF’ phenotype is a ‘true’ subgroup that might be responsive to treatment. Still, a complete understanding of this phenogroup is developing, and it is unclear whether it serves as a precursor to other phenogroups and also whether it may be partially comprised of patients without ‘true’ HFpEF as shown in TOPCAT (full cohort) analyses as well.¹²

Regarding biomarker profiles, patients of a ‘relatively younger, NP deficiency’ phenotype demonstrated the lowest NT-proBNP and soluble ST2 levels.^12,14,26 Additional upregulated and downregulated proteins varied by study. Sabbah et al. studied an obese, but otherwise ‘non-inflammatory’, group and found elevated levels of C-reactive protein (CRP) and serum amyloid A (SAA) but otherwise lower levels of inflammatory and pro-fibrotic mediators compared to a pan-inflammatory, group consistent with a ‘metabolic, obese’ phenotype.¹⁹ Compared with the other phenogroups, Cohen et al. found higher levels of syndecan-4, a cell-matrix interaction marker, and metalloproteinase (MMP)-9, an extracellular turnover marker (and noted a potential association between this latter protein levels and rates of pulmonary disease overlying in this phenogroup).¹² Woolley et al. noted down-regulation of proteins in this cluster including those associated with TNF-α activity, viral protein interactions with cytokines, and regulation of cardiac hypertrophy.²⁵ Lastly, Hedman et al. noted an up-regulation of NF-κ-B essential modulator (NEMO) in this cluster of HFpEF patients, which has been proposed to be important in cardiac function by animal model studies.¹⁴

7. Gaps and limitations

Phenomapping studies in HFpEF have contributed to the understanding of pathophysiology and epidemiology in HFpEF, but have been limited in several respects. More than half of these studies were single-centre or post-hoc analyses from a single clinical trial dataset, although there is an increasing evidence base from multicentre registries. The heterogeneous nature of these populations and study techniques limit the ability to generalize findings across studies. Further, there is, at times, a lack of rigorous evaluation for HFpEF mimickers (i.e. amyloidosis)⁶⁴ hiding in these data before phenomapping techniques, which could cloud phenotype results. Perhaps most importantly, there remains the inherent difficulty of applying these insights to the substantial number of real-world individual patients that straddle several clusters. Additionally, phenomapping studies to date have included very limited haemodynamic and exercise response data, and to date, only a single study that has described tissue-based molecular phenotyping from human myocardial tissue. Furthermore, in syndrome where skeletal muscle is implicated, no large studies have applied phenomapping to detailed skeletal muscle physiology data (i.e. changes in tissue composition and function, including increased adipose infiltration, decreased capillary density, and decreased O₂ diffusive transport⁵), an increasingly recognized contributor to exercise intolerance in HFpEF.² Further, studies to date have been largely unable to account for socioeconomic factors which may be driving or confounding characterization of phenogroups. Lastly, phenomapping studies have been limited by failure to assess data longitudinally; trends in many of the dominant demographic, laboratory, biomarker, echocardiographic, and haemodynamic features of HFpEF may be highly informative in refining phenotypes, such as whether the early-stage ‘relatively younger, NP deficiency’ phenotype progresses to a ‘metabolic, obese’ phenotype over time, or remains a unique cohort.

Furthermore, it could be argued that phenomapping results to date do not sufficiently support further utilization in HFpEF. For instance, one might identify findings from the reviewed phenomapping studies that were already well established without phenomapping methodology (i.e. association of age with arterial stiffness and renal insufficiency, or association between reduced natriuretic peptides and obesity/better outcomes). Additionally, the overlap between phenotypes as data is aggregated and summarized across multiple studies could be viewed as a ‘regression to the mean’ that dilutes the utility of phenomapping. For instance, the ‘older, vascular aging’ phenotype and the ‘metabolic, obese’ phenotype have overlapping features as seen in biomarker/proteomics data and the pervasive nature of CKD in both of these HFpEF phenotypes. Similarly, a ‘metabolic, obese’ phenotype and a ‘relatively younger, NP deficiency’ phenotype share some features and it seems that studies may support the use of spironolactone in either of these groups (although this may reflect that the optimal cohort for spironolactone in HFpEF is the overlap of these two groups). In future work, intentional modeling of overlapping phenotypes could be considered and may better reflect the clinical reality of significant overlap; however, such findings would be highly dependent on methodology choices and may prove less actionable for investigation in translational/clinical research if degree of overlap outweighs unique features of phenotypes. Lastly, while utilizing phenotype data in a trial protocol may aim to target a more specific subgroup of patients that are more likely to respond to intervention (as done in the REDUCE-LAP HF trials), this will likely result in increased resources in the short term to identify a sufficient sample of eligible patients; the gains of reduction in necessary sample size (i.e. higher response rate) may not outmatch this resource burden, and it remains unclear whether this will result in better outcomes for HFpEF trials as underscored by the neutral results of REDUCE-LAP HF II.⁶⁵ Advances in identification and recruitment of patients (i.e. electronic health record screening programs/alerts, statistical learning techniques, e-consent, etc.) may be necessary to pair alongside phenotype-based trial criteria to overcome, as least, the issues of enrollment efficiency. Further, these phenotype-guided results to date could be contrasted with the recent success of sodium-glucose cotransporter-2 (SGLT2) inhibitors across the full range in HFmrEF/HFpEF, although this should not be interpreted as evidence for HFpEF as a homogeneous syndrome. Overall, these intrinsic limitations of phenomapping and mixed results to date in HFpEF serve as valid concerns. Still, traditional strategies for understanding pathophysiology and testing interventions in HFpEF as a single disease phenotype have proven challenging with modest efficacy results, and therefore, alternative strategies (phenomapping or otherwise) should continue to be considered.

8. Future directions

8.1. Guidance for future phenotyping and phenomapping studies in HFpEF

Phenomapping work and results to date in HFpEF should be considered as hypothesis-generating and a starting point for further exploration as opposed to a final conclusion on subtypes of HFpEF. Future phenomapping work in HFpEF should ideally build upon existing work with utilization of best practices and (i) external validation of established results, or (ii) novel data inputs and/or analytical techniques (Figure4). While there is no established gold standard among clustering techniques, Shah et al. have recently outlined a clear framework for considering statistical learning studies and best practices if statistical learning methodologies are chosen.³⁸ Notably, the utilization of derivation and external validation cohorts to establish and confirm proposed phenogroups is critical. Further, additional validation of proposed cohorts and reproducibility across several dataset cohorts remains a gold standard goal that has not yet been achieved. Validation of results can also come through connection to mechanisms of disease in physiology-based, deeply phenotyped studies of smaller sample size (without the need for phenomapping/statistical learning). These steps should make phenomapping clusters more actionable in clinical and trial settings. Additionally, as mentioned previously, there is little longitudinal data in HFpEF phenotyping research and studies (with or without phenomapping methodology) that utilize multiple, longitudinal datapoints will be useful to assess different stages of disease as compared to unique phenotypes over time. Lastly, it must be noted that there will remain a fundamental tension between optimizing reproducible phenogroups with limited overlap for targeted evaluation and treatment (i.e. precision medicine) and the conventional strength of large, adequately powered randomized, controlled trials, which are increasingly supported by pragmatic features. Both of these approaches are different ends of the methodology spectrum and components of both will likely be required to support progress in the HFpEF field.

Figure 4.

Future Directions—Phenotyping and Phenomapping Studies in HFpEF.

8.2. Techniques and data

Ideally, retrospective novel phenomapping studies should continue to enhance characterization of the HFpEF population, specifically through historically application of underutilized data elements such as invasive haemodynamics, advanced imaging techniques such as cardiac magnetic resonance imaging, raw image DICOM data, detailed exercise data (i.e. cardiopulmonary testing +/− invasive or non-invasive haemodynamics), microvascular coronary pathophysiology, and skeletal muscle physiology. Further, meta-analyses utilizing individual participant data (IPD) may improve discrimination and reproducibility of phenogroups. Continued advancement in statistical learning techniques should improve integration of complex multi-modality data in a clinically meaningful way for each HFpEF patient. Alternatively, given the ongoing debate on the reliance on EF to phenotype HF due to its lack of reliable correlation with pathophysiology, outcomes, and in some cases, treatment response,⁶⁶ it may become more fruitful to perform phenomapping analyses across the spectrum of HF (including HFrEF and HFpEF) as performed by Tromp et al.,²⁹ Horiuchi et al.,⁶⁷ and Gevaert et al.⁶⁸ Otherwise, future studies focusing on specific EF phenotypes should utilize now well-established consensus definitions of HFrEF (EF ≤ 40%), HFmrEF (EF 41–49%), and HFpEF (EF ≥ 50%).^69,70 Further, distinguishing patients with similar phenotype profiles as described in this review but who do not develop HFpEF remains a significant unmet need for further pathophysiologic, outcomes, and/or phenomapping investigation. Additionally, as described above, goals of phenomapping can take many forms, and therefore, planned application of phenomapping moving forward (i.e. risk stratification, mechanistic drug development, clinical trial design, or direct clinical application) will be critical to designing efficient and targeted analyses. Lastly, the utilization of prospective trial structures such as in the ongoing PACIFIC-PRESERVED study should prove useful in the development of reproducible HFpEF phenotypes.

8.3. Utilization/application of HFpEF phenomapping results

Since phenomapping in HFpEF is a relatively new and evolving field, there has been relatively little, large-scale implementation or utilization in clinical practice or research. Utilization considerations will be largely driven by setting (i.e. laboratory, clinical trial, clinical practice, etc.) and objectives, but, across all domains, there will remain tension between phenomapping approaches with simplicity (i.e. few, easily attainable inputs) that may be more easily applied broadly and those with complex clustering approaches (i.e. numerous inputs and/or less readily available ‘omics data).

Given the hypothesis-generating nature of HFpEF phenomapping studies and lack of consistent reproducibility, the phenomapping results and proposed phenogroups described in this review will require further exploration in mechanistic evaluations, observational studies, and clinical trials in order to allow researchers to iterate and test additional hypotheses regarding phenotypes. If phenogroups can be rigorously validated, one of the first natural applications of phenomapping one might consider is to define high- and low-risk subsets of the large, heterogenous HFpEF population. Many of the presented phenomapping studies include evaluation of outcomes across identified clusters; from this work, it is clear that the ‘older, vascular aging’ phenotype and the ‘metabolic, obese’ phenotype have a higher rate of adverse hospitalization/mortality outcomes as compared to the ‘relatively younger, NP deficiency’ phenotype, and this trend is more reliable for the ‘older, vascular aging’ phenotype.^{1,10–14,16,17,26} Phenomapping results can be applied for risk stratification but application for this goal is quite limited compared to risk assessments that can be applied across the full cohort of HFpEF patients in order to advance risk prediction beyond standard scores such as the Meta-Analysis Global Group in Chronic Heart Failure (MAGGIC) risk score. Discarding the goal of creating distinct clusters (i.e. discarding strict phenomapping design) in order to focus on the prediction of risk alone may prove more useful for this goal and studies have achieved this using techniques such as supervised support vector machine (SVM) algorithms (as in a component of Shah et al.’s seminal analysis¹), deep biomarker investigation (as in Chirinos et al.³⁰) and/or non-clustering statistical learning (as done by Angraal et al.,⁷¹ for instance).

Phenomapping results may be better suited for application in mechanistic research and drug development. For instance, Hahn et al. identified key molecular pathways either upregulated or downregulated from HFpEF myocardial tissue, and then, within HFpEF, identified top distinguishing pathways between three phenotypes identified from transcriptomics-derived clustering.²⁶ Multi-omics analyses from biological specimens such as this study offer insights into therapeutic targets and future directions for drug development. Additionally, development of accurate large animal models of HFpEF has been challenging, in part due to the difficulty in encapsulating all of the features of clinical, human HFpEF.^72,73 Focusing animal model development on specific phenotypes or subtypes (as, for example, in the metabolic + hypertensive stress murine model of Schiattarella et al.⁷⁴) could streamline the process and allow for study of pathophysiology and intervention that could then be tested in a pre-specified and identified phenotype of HFpEF. For instance, one model could focus on simulating arterial stiffness, LV remodeling, and renal insufficiency (and testing for innate immunity and inflammatory processes) in an ‘older, vascular aging’ phenotype model, while another could focus on simulating cardiometabolic inflammatory disease with obesity and diabetes (and testing for prominent inflammatory pathways) in a ‘metabolic, obese’ phenotype model with then clear targets for clinical application of findings. This has shown potential, for instance, in the investigation of glucagon-like peptide receptor agonists and SGLT2 inhibitors in a cardiometabolic mouse model phenotype of HFpEF.⁷⁵ Just as importantly, phenomapping has the potential to uncover new patterns of disease and phenotypes in HFpEF that reveal novel mechanisms of disease, but this remains an ideal at this time with lack of definitive evidence to support that phenomapping will achieve this in HFpEF.

In terms of clinical research, results from synthesized phenomapping work in HFpEF would ideally be applied to clinical trial design, but there is insufficient data to support pursuing this currently and it is unclear if phenomapping will lead towards clinically useful phenotypes in HFpEF. Historically, HFpEF trials have relied on a combination of age, ejection fraction, NYHA functional class, recent HF hospitalization, elevations of NT-proBNP, diuretic use, and/or indices of diastolic dysfunction by echocardiography in attempts to identify ‘true’, symptomatic, and/or decompensated HFpEF without specific phenotype evaluation.⁷⁶ Clustering/phenomapping analyses balance the interest in precision, individualized medicine in HFpEF with the need to group similar patients in order to generate enough of a sample size to study interventions (i.e. often resulting in ∼3 clusters of HFpEF instead of dozens or more). Therefore, testing an intervention in a phenotype-defined group (by demographics, comorbidities, echocardiographic and/or haemodynamic parameters, or even biomarkers beyond NT-proBNP) could prove more efficient. The REDUCE LAP-HF trials are some of the best examples of this approach to date; in this set of trials, haemodynamics (namely, pulmonary capillary wedge pressure, PCWP and central venous pressure, CVP) were utilized in an effort to target the predominant LA myopathy/enlargement phenotype.^77–79 Unfortunately, recent results from REDUCE LAP-HF II were disappointingly neutral, failing to lend support to a specific-phenotype-guided clinical trial approach in HFpEF (at least with current phenotype understanding and ability to implement).⁶⁵ Results further suggested that more detailed selection based on the presence of latent pulmonary vascular disease, unmasked by exercise, may be needed to identify patients most likely to benefit from the treatment. Other ongoing trials including the STEP HFpEF DM (semaglutide) and SUMMIT (tirzepatide) studies may provide more insight into the feasibility of targeting a different specific phenotype within HFpEF—the obese/cardiometabolic phenotype.

Lastly, phenomapping results would ideally be applied in the current clinical care paradigm. Currently, it is not logistically feasible to deeply phenotype each individual patient in the clinical setting with the complex algorithms and data components utilized in phenomapping, nor is it yet proven that deep phenotyping is clinically necessary for every HFpEF patient. Advances in statistical learning and artificial intelligence systems may facilitate individual patient-level deep phenotyping on a large-scale in the future,⁵⁶ but importantly, this must be demonstrated as clinically useful and cost-effective in order to consider this investment and implementation. Still, developing a sense of where each individual patient falls in the overlapping spectrum of HFpEF (Figure 3) could be quite useful. For instance, older patients with evidence of heavy burden of arterial stiffness and CKD (‘older, vascular aging’ phenotype) could be expected to suffer from a heavy rate of adverse outcomes^{1,11,16,17,20,23} and closer monitoring in the outpatient, ‘stable’ setting may be warranted. Similarly, early clinical identification of patients with features of ‘metabolic, obese’ phenotype and optimization of inflammatory/metabolic comorbidities is critical and may help slow progression of this phenotype (although this has not yet been rigorously proven). Maintaining a low threshold for invasive haemodynamics/exercise testing in patients reflective of the ‘relatively younger, NP deficiency’ phenotype is warranted given the diagnostic challenge in this group of young patients with often disproportionate exercise intolerance and obesity/deconditioning overlap. Even more specifically, increased consideration of spironolactone, particularly in the overlap between the ‘metabolic, obese’ and ‘relatively younger, NP deficiency’ phenotypes (i.e. young, obese, volume overloaded, low NT-proBNP) may be warranted, compared to older, non-obese, high NT-proBNP (‘older, vascular aging’ phenotype-like) patients, but this will require further prospective evaluation, particularly in an evolving pharmacologic paradigm (i.e. SGLT-2i utilization). Further, identifying specific subtypes of HFpEF such as those with obesity/cardiometabolic phenotype has implications for current clinical care, namely targeting for enrollment in targeted trials such as SUMMIT or STEP HFpEF DM. Lastly, phenomapping results have added some insight to a common cohort of HFpEF patients with pulmonary disease complicating their apparent HFpEF diagnosis (as described in Cohen et al. for instance¹²), for which cardiopulmonary exercise testing and/or invasive haemodynamics⁸⁰ should be considered early to distinguish limitations that are pulmonary rather than cardiac in nature or similarly driven by both systems (either in parallel or through interaction such as long-standing HF leading to compromise of pulmonary function and diffusion).

9. Conclusions

Phenomapping, the application of statistical learning techniques to categorize patients into distinct, mutually exclusive subgroups, has produced a variety of phenogroups among HFpEF cohorts. Summarizing these studies results in, at least, three, overlapping phenotypes in the heterogenous HFpEF population (an ‘older, vascular aging’ phenotype; a ‘metabolic, obese’ phenotype; and a ‘relatively younger, natriuretic peptide deficiency’ phenotype) with some unique characteristics, outcomes, and biomarker profiles. The continued study and utilization of these phenotypes could have significant implications for moving the field forward in translational research, trial design, and clinical practice, but this remains to be proven. Phenomapping results in HFpEF should be considered as hypothesis-generating and will require further validation in mechanistic and outcomes research, as well as in the next phase of phenomapping studies in HFpEF.

Supplementary material

Supplementary material is available at Cardiovascular Research online.

Funding

Funded in part by: R01AG045551; R01AG18915; P30AG021332; U24AG059624, U01HL160272 (Kitzman)

Data availability

There are no new data associated with this article.