Heart failure (HF) with preserved ejection fraction (HFpEF) is a multi-organ, systemic syndrome that involves multiple cardiac and extra-cardiac pathophysiological abnormalities [1,2,3]. Because it is a heterogeneous syndrome and resistant to a “one-size-fits-all” approach, HFpEF has proven to be very difficult to treat [4]. By definition, HF is heterogeneous because it is the end result of a wide variety of cardiovascular diseases and risk factors. However, patients with HF with reduced ejection fraction (HFrEF) respond in a more homogeneous manner to therapies [5], whereas HFpEF patients do not. For these reasons, several research groups have been working on combining deep phenotyping with innovative analytical strategies to classify HFpEF into therapeutically homogeneous subtypes [6,7,8,9,10,11].

Shah et al. initially used a form of machine learning, unsupervised model-based clustering, of deep phenotypic data in HFpEF (n = 397), which they termed “phenomapping” [6], and found that from a data-driven perspective, HFpEF is very heterogeneous. In addition, they identified multiple unique HFpEF “pheno-groups” that have differences in clinical characteristics, biomarkers, cardiac structure/function, pathophysiology, and outcomes (validated in a separate group of 107 HFpEF patients). Subsequently there has been a proliferation of such studies throughout the field of HF [12,13,14,15,16,17,18,19].

Despite the novelty of these studies, machine learning analyses [20, 21] are only the starting point for further investigation into HFpEF subtypes, just an initial step on the way to the ultimate goal of unraveling the pathobiology underlying these subtypes so that we can develop effective therapeutics. In addition, at present it is unclear whether machine learning strategies will be the optimal way to identify (1) mechanisms underlying specific types of HFpEF and (2) therapeutically homogeneous HFpEF subtypes. It is likely that a combination of techniques (clinical, pathophysiological, hemodynamic, -omics, exercise testing, and machine learning) will be helpful in the future to identify HFpEF subtypes. Here, we provide a framework for understanding the phenotype-based approach to HFpEF by reviewing (1) the historical context of HFpEF (i.e., diastolic HF); (2) the current paradigm of comorbidity-induced inflammation and endothelial dysfunction as stressors on multiple organs, including the heart; (3) various ways of sub-phenotyping HFpEF, with examples of how these have been leveraged toward potential therapeutics; (4) comorbidity-based classification and treatment of HFpEF; (5) machine learning approaches to classifying HFpEF; (6) examples from HFpEF clinical trials that have informed future phenotype-based HFpEF treatments; and (7) the future of phenomapping (machine learning and other advanced analytics) for classification of HFpEF.

Historical context of HFpEF

Chronic HF remains a deadly clinical syndrome associated with considerable loss of quality life and high socioeconomic burden. Dyspnea on exertion and exercise intolerance are the leading symptoms but a myriad of clinical manifestations from the early stages of the HF syndrome (e.g., fatigue) to late-stage HF (e.g., loss of appetite, cachexia) can occur. Having emerged as a common and growing type of HF, HFpEF (defined most recently by international consensus as HF with EF ≥ 50% [22]) currently accounts for 50% of all HF and is projected to grow in proportion of HF over the next few decades. Patients with HFpEF generally do not respond to strategies known to improve the prognosis in patients with HF and reduced EF (HFrEF). The reasons proposed for this include a variety of different underlying pathomechanisms (that may not be very responsive to neurohormonal antagonists, in particular), a much greater heterogeneity in therapeutic response, and a higher prevalence of non-cardiac comorbidities compared with HFrEF. It is now clear that HFpEF is not an early form of HFrEF [23]; systematic longitudinal studies have shown that the transition from HFpEF to HFrEF is rare [24]. The lack of prognostic benefit in HFpEF from current HFrEF strategies suggests novel disease mechanisms but also questions the classic concept of myocardial injury as the main driver of disease progression in HFpEF.

In the classic mechanistic concept of HFpEF, the heart was the primary source of the syndrome, with left ventricular (LV) hypertrophy (LVH) and diastolic dysfunction as the main drivers, triggered by systemic hypertension, with contributions from risk factors such as coronary artery disease (CAD). This mechanistic concept arose from early studies in the 1970s of diastolic dysfunction in the cardiac catheterization laboratory, and later by various case series showing impaired filling of the LV in patients with HF and a normal EF [25]. The field was heavily influenced by hypertrophic cardiomyopathy (HCM); indeed, HFpEF was viewed as a forme fruste of HCM in that it was similar pathophysiologically but did not meet criteria for HCM because there was no family history, no genetic abnormalities, and a cause of LVH (hypertension) was present in the majority of patients. The lack of dedicated HFpEF programs also likely impaired this stage of HFpEF understanding because clinicians were not seeing the full extent of HFpEF in the community; rather, it was often patients with severe HFpEF or specific cardiomyopathies (e.g., cardiac amyloidosis, HCM) that were presenting to HF clinics with “diastolic HF” [26]. Furthermore, over the past 50 years, we have witnessed an epidemiological transition from uncontrolled hypertension and smoking (major risk factors for LVH) and high rates of CAD to a better control of these risk factors and an explosion of morbid obesity, diabetes, atrial fibrillation, coupled with an aging population. Whether HFpEF itself has changed due to these epidemiological transitions, or whether it was always the case, it is now clear that HFpEF is not simply cardiac-centric but instead a systemic syndrome, with multiple involved organs; not only the heart but the lungs, liver, adipose tissue, skeletal muscles, and kidneys are all variably involved in individual patients with HFpEF [27].

A primary cardiac insult is the trigger for HFrEF (e.g., myocardial infarction, myocarditis, genetic abnormality, chemotherapy). In HFpEF, however, cardiomyocyte cell death and severe loss of contractility at rest are not commonly observed. Instead, cardiac function in HFpEF is characterized by impaired cardiac filling and altered diastolic properties (stiffness) of the LV, which result in increased LV filling pressure, congestion, and dyspnea on exertion. The underlying pathomechanisms are still not completely resolved, but a leading theory is that myocardial injury in HFpEF is not primary but rather secondary to comorbidity-induced stress on the endothelium [28].

Comorbidity–inflammation–endothelial dysfunction paradigm of HFpEF

According to a leading conceptual framework of HFpEF, largely based on preclinical disease models, systemic inflammation, triggered by arterial hypertension and metabolic disease states such as diabetes and obesity, leads to coronary microvascular inflammation and dysfunction, subendocardial ischemia, and altered cardiomyocyte mechanics and metabolism (as reviewed in [29]). Thus, HFpEF is considered a result of a multisystem disorder and is strongly associated with advanced age, which may reflect cumulative effects of an increasing number and duration of systemic comorbidities.

Clinical evidence for heterogeneity in patients classified as HFpEF

The multitude of comorbidities that lead to HFpEF, coupled with the multi-organ, systemic nature of the HFpEF syndrome, contribute to its heterogeneity. Patients with HFpEF share signs and symptoms of HF and echocardiographic features or biomarker evidence of elevated left atrial pressure with preserved LVEF as the main criteria within the current HF classification [30]. However, clinical evidence from prospective trials and observational studies suggests heterogeneity in pathophysiological triggers and clinical presentation with relevance for further diagnostics and therapy. For several decades HCM has been a model for diastolic dysfunction related to structural heart disease, but HCM is found in only a minority (~5%) of HFpEF patients. Based on findings in HCM studies, LVH was considered to be a major contributor to HFpEF because hypertension is a very common comorbidity and has been considered a trigger for HFpEF. Yet, the association between the degree of LVH and diastolic dysfunction in clinical studies is weak, and LVH is only found in about half of the patients with HFpEF [31], which suggests that even myocardial remodeling is heterogeneous in HFpEF.

Recently, HFpEF related to cardiac amyloidosis with accumulation of misfolded proteins in the extracellular matrix has moved into focus, and even though cardiac amyloidosis is likely to be present in a small but relevant fraction of HFpEF patients (~5%, [32]), it is an example of how there may be subtypes of HFpEF that have unique features on cardiac imaging, a specific confirmatory diagnostic approach, and targeted therapy. The common form of HFpEF has been associated with metabolic disease such as obesity or diabetes. Indeed, seminal studies in human myocardium from HFpEF patients have suggested that diabetes mellitus, which is present in almost half of HFpEF patients (45%, [33]), may induce specific alterations in myocardial passive stiffness not observed in non-diabetic HFpEF tissue [34, 35]. In this context, HFpEF is now considered one manifestation of diabetic cardiomyopathy [36] with a higher prevalence of myocardial hypertrophy and fibrosis and a worse prognosis. Chronic kidney disease (CKD; defined as estimated glomerular filtration rate < 60 ml/min/1.73 m2), present in approximately 50% of HFpEF patients [37], may be another discriminator. For instance, in a recent Japanese study, patients with moderate CKD but not patients with manifest LVH profited from renin–angiotensin–aldosterone system inhibitors [38]. Varying definitions for the cut-off for a preserved EF in randomized trials (between 40 and 50%) have added to the perceived variability also between HFpEF cohorts. For instance, the fraction of patients with ischemic heart disease was higher in studies with a lower EF cut-off (≥ 40%) as in the CHARM-Preserved (65% of patients) and EMPEROR-Preserved (36% of patients) trials compared with studies with a higher EF cut-off (≥ 45%) as in the PARAGON (23% of patients with ischemic heart disease) and PARAMOUNT (21% of patients) trials [39,40,41,42].

Methods for classification of HFpEF

A variety of methods for classification of HFpEF have been proposed, and each may be useful clinically for selecting appropriate therapies and designing targeted clinical trials. These include clinical (etiological, or primary comorbidity driving HFpEF in a particular patient), pathophysiological (primary pathophysiology driving HFpEF in a particular patient), myocardial, type of clinical presentation, and data-driven approaches to the classification of HFpEF (Table 1). Examples of methods for identification of specific HFpEF subtypes, and how these strategies can lead to novel therapies, are discussed next.

Table 1 Classification schemes for heart failure with preserved ejection fraction (HFpEF). Reproduced with permission from [82]
Full size table

Clinical subtype: left atrial myopathy

Shah et al. have shown that left atrial (LA) myopathy in HFpEF, defined by reduced LA reservoir strain, is associated with increased pulmonary vascular resistance, decreased peak VO2, and poor outcomes [43]. Recently, they defined a unique LA myopathy phenotype in HFpEF in the PROMIS-HFpEF observational study by taking the residuals of linear regression of LA reservoir strain and LV longitudinal strain to define a continuous phenotype that ranges from LA-preserved, to LA/LV-balanced, to LA-predominant (LA myopathy) phenotypes (Fig. 1a; [44]). The LA myopathy phenotype is associated with better LV systolic and diastolic function but worse hemodynamics (lower stroke volume, increased pulmonary artery pressure). In a proteomic analysis in the PROMIS study (with external validation), LA myopathy was associated with increased natriuretic peptides and several novel circulating proteins, which were independent of AF (Fig. 1b, c). Interatrial shunt devices and procedures [45,46,47,48,49,50,51,52,53,54,55] offer a therapeutic option for HFpEF-LA myopathy and therefore could be targeted to these patients if better identified. In addition, automated deep learning algorithms of electrocardiograms and echocardiograms could be trained to identify the LA myopathy HFpEF subtype for further evaluation in clinical trials.

Fig. 1
figure 1

Left atrial myopathy phenotype of heart failure with preserved ejection fraction (HFpEF). Top panel: Scatterplot of left atrial (LA) vs. left ventricular (LV) longitudinal strain showing a correlation between the two phenotypes and deviation from the correlation representing different LA/LV phenotypes of HFpEF, including the “LA predominant” (i.e., LA myopathy) phenotype, defined as LA reservoir strain lower than expected for any given value of LV longitudinal strain. Middle panel: Volcano plot of proteins associated with LA myopathy. Proteins in red on the right represent those in which higher levels are significantly associated with increased LA myopathy, and on the left represent those in which higher levels are significantly associated with reduced LA myopathy. Bottom panel: Venn diagram showing that while three proteins overlap between LA myopathy and atrial fibrillation, several others appear to be specific for the disproportionate LA myopathy phenotype (proteins identified using a proteomic analysis in the PROMIS-HFpEF study with validation in the Northwestern University HFpEF cohort). Reproduced with permission from Patel RB et al. [44]

Full size image

Genetic subtype

Phase 3 HFpEF clinical trials such as TOPCAT and PARAGON have shown that patients with LVEF > 65%, especially men, do not respond favorably [39, 56] and are particularly resistant to conventional HF therapeutics. Thus, there may be a unique hypercontractile HFpEF phenotype that may or may not be associated with unrecognized cardiomyopathy variants. A recent CHARM trial analysis found that on whole-exome sequencing, 20 of 767 (2.6%) HFpEF patients had pathogenic or likely pathogenic cardiomyopathy rare variants, which was only slightly lower than the rate found in HFrEF (3.5%), and both were much higher than in the general population [57]. While 2.6% may seem trivial, it translates into nearly 80,000 patients in the United States alone given the estimated HFpEF prevalence (over 3 million) in the United States, and is similar to the frequency of transthyretin amyloid cardiomyopathy (~3–4% of HFpEF [58]), for which there is an approved treatment that lowers mortality [59]. In HFpEF, both hypercontractile and hypocontractile (despite preserved global LVEF) subtypes exist, and these HFpEF phenotypes could be treated with myosin inhibitors or activators, respectively, both of which are currently in development [60, 61]. Further investigation of the frequency of rare cardiomyopathy variants in HFpEF and use of deep learning algorithms to train on the electrocardiograms and echocardiograms of HFpEF associated with these variants will be essential to further our understanding of the genetic subtypes of HFpEF.

Molecular subtype: plasminogen activator inhibitor-1

Plasminogen activator inhibitor (PAI)-1 is one of the only biomarkers found in cross-cohort collaboration studies to be uniquely associated with incident HFpEF but not HFrEF [62]. It has been implicated in aging, senescence, visceral adiposity, and impaired metabolism [63,64,65]. Activation of PAI‑1 may represent a unique HFpEF subtype characterized by accelerated aging, inflammation, and presence of metabolic comorbidities. Indeed, PAI‑1 is a key protein secreted by metabolically unhealthy visceral adipose tissue, which is found in the vast majority of HFpEF patients. Preclinical studies have identified PAI‑1 as a promoter of senescence (Fig. 2). Clinical deep phenotyping studies in an Old Order Amish kindred found that a rare variant in the gene that encodes for PAI‑1, SERPINE1, is associated with PAI‑1 deficiency, increased insulin sensitivity, and longevity [65]. Taken together, these data suggest a central role for PAI‑1 in a potential “accelerated aging” HFpEF phenotype. Studies on the associations between PAI‑1 levels and multi-omics were conducted with HFpEF patients, and these associations could be compared with those in HFrEF and controls. Identification of a unique HFpEF phenotype with elevated PAI‑1 levels is relevant because of the availability of a novel PAI‑1 inhibitor, which is currently being tested in a Phase 2 trial in COVID-19 (clinicaltrials.gov NCT04634799) and could be repurposed for HFpEF. Determining molecular pathways associated with increased PAI‑1 in HFpEF could also identify novel targets, and pathophysiologic abnormalities associated with PAI‑1 in HFpEF could inform endpoints for Phase 2 trials.

Fig. 2
figure 2

Potential role of plasminogen activator inhibitor‑1 (PAI‑1) as a molecular mechanism underlying heart failure with preserved ejection fraction (HFpEF). Both comorbidity-induced inflammation (via reactive oxygen species [ROS]-induced cellular stress) and visceral adipocytes (common in HFpEF) induce upregulation of PAI‑1 in the circulation. PAI‑1 and related proteins (e.g., insulin-like growth factor binding protein‑3 [IGFBP3]) result in cellular senescence and accelerated aging, which may be important risk factors for the development of HFpEF

Full size image

Targeting primary comorbidities in HFpEF

The different clinical manifestations of HFpEF in clinical trials and registries, as well as the failure of generalized medical approaches such as renin–angiotensin–receptor signaling cascade blockers to improve the prognosis in HFpEF cohorts, have emphasized the need for a more differentiated therapeutic strategy. A proposed treatment strategy suggests the identification of leading comorbidities based on the clinical phenotype, which would then guide therapeutic interventions [66]. This approach is in line with current European Society of Cardiology (ESC) guideline recommendations to treat etiologies and cardiovascular and non-cardiovascular comorbidities in HFpEF [67]. However, whether specific treatment of such selected subgroups improves the prognosis in HFpEF remains to be determined. For instance, treating arterial hypertension can reduce the risk for HF, including HFpEF [68]. However, while lowering elevated systolic blood pressure in patients with manifest HFpEF was associated with reduced hospitalization rates it did not lower all-cause or cardiovascular mortality [69].

The majority of HFpEF patients have increased pulmonary artery pressure (pulmonary hypertension [PH]), although in most this is typically pulmonary venous hypertension; however, there are some HFpEF patients who have combined post- and pre-capillary PH. Treatment using specific drugs approved for pulmonary artery hypertension in unselected patients with HFpEF has yielded mixed results and is currently not recommended [70]. Stimulators of the soluble guanyl cyclase (enhancing cGMP-signaling) have been effective in patients with primary pulmonary artery hypertension (riociguat [71]) and also in HFrEF (vericiguat [72]). Yet, vericiguat demonstrated neutral results in HFpEF patients in the VITALITY-HFpEF [73] and SOCRATES-Preserved trials [74]; however, both trials did not differentiate patients without versus with PH.

Dietary restriction in obese patients improves diastolic function [75], and a reduction in calorie intake by ~400 kcal/day increases exercise capacity and quality of life in elderly obese HFpEF patients [76]. Again, the impact on prognosis has yet to be established. Ongoing and future trials of drugs that result in significant weight loss (e.g., GLP‑1 receptor agonists) will hopefully answer the question of whether weight loss improves outcomes in HFpEF. In summary, in patients with HFpEF, treatment of contributing risk factors may or may not be sufficient to treat the cardiac and systemic alterations that they have accumulated from these risk factors.

Machine learning approaches to the classification of HFpEF

As mentioned earlier, multiple studies have now applied data-driven approaches (“phenomapping”) to HFpEF classification. These studies have used machine learning analytic techniques similar to our original phenomapping study and have shown that the HFpEF subtypes identified have different clinical features and outcomes (Fig. 3; [7]). Nonetheless, the study by Shah and colleagues was only the initial “proof of concept” that HFpEF is a truly heterogeneous syndrome (as shown in the heatmap in Fig. 3, left panel), warranting larger-scale, multicenter investigations to advance the science of identifying novel HFpEF subtypes and treatment targets using next-generation phenomics. Importantly, machine learning analysis of deeply phenotyped HFpEF cohorts is only the starting point for further investigation into HFpEF subtypes. While several HFpEF phenomapping studies have been published and have agreed on up to five common HFpEF sub-phenotypes (Fig. 4; [77]), additional research is required to determine the pathophysiological and molecular mechanisms underlying each HFpEF subtype.

Fig. 3
figure 3

Phenomapping of heart failure with preserved ejection fraction (HFpEF). Left panel: hierarchical clustering heatmap of 397 patients with HFpEF (columns) and 67 continuous variables (features, rows) demonstrating the heterogeneity of HFpEF. Red indicates increased and blue indicates decreased values. Middle panel: principal components analysis showing the clear differentiation of the three identified pheno-groups. Right panel: Kaplan–Meier curves for survival free of cardiovascular hospitalization or death among the three pheno-groups. PC principal component, CV cardiovascular. Reproduced with permission from Shah SJ et al. [7]

Full size image
Fig. 4
figure 4

Protein clusters identified by weighted coexpression network analyses in the PROMIS-HFpEF derivation cohort and the Northwestern University validation cohort. The inflammation cluster (turquoise) mediated the association between comorbidity burden and markers of elevated left atrial pressure in heart failure with preserved ejection fraction (HFpEF) and differentiated HFpEF from controls in an external cohort. a Adjacency network map of circulating proteins color-coded by cluster assignment by hierarchical clustering-based nearness or coexpression of proteins. For clarity of presentation, only nodes (proteins) that were assigned to a cluster are shown (n = 159/248); the remaining proteins lie on the outer edges of the network map. b Overrepresented, nonredundant pathways in each cluster with false discovery rate corrected p values. c Detailed network maps of proteins in the three clusters that were representative of inflammation (i.e., overrepresentation of ≥2 inflammatory pathways). Node size reflects intracluster connectivity (i.e., the sum of weighted edges [correlations] with all other proteins in the cluster). Node color density reflects the strength of cluster membership. Edge thickness and transparency reflect the adjacency of proteins according to weighted coexpressions. d Adjacency network map of circulating proteins in the Northwestern patients with HFpEF in the validation cohort. Clusters with most significant overlap were assigned the same color as the corresponding cluster in the PROMIS-HFpEF cohort. e Adjacency network map of circulating proteins in the Northwestern control patients in the validation cohort. Cluster preservation was tested against the Northwestern patients with HFpEF; clusters with significant overlap were assigned the same color as the corresponding cluster in the Northwestern patients with HFpEF. FDR false discovery rate, PROMIS-HFpEF Prevalence of Microvascular Dysfunction in Heart Failure With Preserved Ejection Fraction, WCNA weight coexpression network analysis. Reproduced with permission from Sanders-van Wijk S et al. [9]

Full size image

As an example, since the initial HFpEF phenomapping study [7], Shah et al. have probed further into the three identified HFpEF subtypes by investigating underlying disease mechanisms and designing and conducting targeted clinical trials in these subtypes, as shown in Table 2. In addition, once HFpEF subtypes have been identified, analytical approaches such as natural language processing and supervised machine learning can be used to assist with automated identification of patients within specific HFpEF subtypes, which will assist with enrollment in future targeted HFpEF clinical trials and targeting of specific therapies to HFpEF subtypes. For example, natural language processing of unstructured electronic health record data has been used to automate the identification of eligible patients for an HFpEF clinical trial (PARAGON; [78]). In addition, large-scale machine learning analyses of national electronic health record claims data have been used to develop a prediction model for amyloidogenic transthyretin cardiomyopathy to improve/automate clinical recognition of this treatable HFpEF subtype [79]. Deep learning machine learning models have also been used for the automated identification of amyloidogenic transthyretin cardiomyopathy using electrocardiography and echocardiography [80].

Table 2 Examples of going beyond initial HFpEF phenomapping studies: ongoing follow-up studies of the original three HFpEF pheno-groups identified by Shah et al. [7]
Full size table

Various types of -omics data can be analyzed with data reduction techniques with subsequent identification of overrepresented biological pathways, which can provide insight into underlying biological mechanisms of HFpEF and its subtypes. This has been done using principal components analysis and weighted coexpression network analyses ([9]; Fig. 5). These techniques reduce high-dimensional data into a low-dimensional data space, summarized by eigenvalues. Eigenvalues can then be used in conventional multivariable regression and mediation analyses to determine associations and perform causal inference with particular HFpEF subtypes.

Fig. 5
figure 5

Main heart failure with preserved ejection fraction sub-phenotypes identified across phenomapping studies. NP natriuretic peptide, CV cardiovascular, T2DM type 2 diabetes mellitus, LV left ventricular, CVP central venous pressure, PAH pulmonary arterial hypertension. Reproduced with permission from Galli E et al. [77]

Full size image

Despite the rapid rise of machine learning and -omics studies in HF, including HFpEF, there are several challenges to these types of studies that must be considered when evaluating their clinical utility in the classification (sub-phenotyping) of HFpEF and in understanding underlying molecular mechanisms. Table 3 lists key steps in performing machine learning analyses relevant to classification of HFpEF. Table 4 lists various potential problems with these types of studies (e.g., lack of external validation, single timepoint assessments, disease progression bias, cohort bias, feature bias, and publication bias) and offers potential solutions for overcoming these limitations.

Table 3 Key steps in the derivation and validation of HFpEF subtypes using unsupervised machine learning (phenomapping) analyses
Full size table
Table 4 HFpEF phenomapping analyses: possible problems and potential solutions
Full size table

Examples from HFpEF clinical trials that have informed future phenotype-based treatment

Several ongoing and recently completed HFpEF clinical trials have taken advantage of sub-phenotyping of HFpEF in order to target therapeutics to particular HFpEF subtypes in an attempt to improve chances for HFpEF clinical trial success. For example, in the EMBARK-HFpEF trial (open-label proof-of-concept study of mavacamten [myosin inhibitor] in HFpEF [NCT04766892]), only patients with LVEF > 60%, LV hypertrophy, and elevated biomarkers (elevated natriuretic peptides or troponin) are allowed entry into the trial. These HFpEF patients are most similar to patients with genetic/familiar forms of HCM and therefore are hypothesized to be the patients most likely to benefit (and least likely to have adverse effects) from mavacamten therapy.

In the SERENADE trial (NCT03153111), which is a randomized controlled trial of macitentan (a dual endothelin receptor A and B antagonist) in patients with HFpEF, elevated natriuretic peptides, cardiac remodeling (LVH or LA enlargement), and evidence of pulmonary vascular disease (invasive defined elevation in pulmonary vascular resistance or diastolic pulmonary gradient; or elevated pulmonary artery pressure with evidence of right ventricular dysfunction) are allowed entry into the trial. These patients are hypothesized to be those who will benefit most from macitentan. To further select patients who may benefit, SERENADE included a 4-week placebo run-in phase (to ensure clinical stability) and a 5-week macitentan run-in phase (to exclude patients who develop early fluid retention in response to macitentan). Only patients who qualify for the trial and make it through these two run-in phases were randomized in the trial.

In the recently completed REDUCE LAP-HF II trial ([81]; a phase 3, pivotal, multicenter trial of 626 HFpEF patients randomized 1:1 to an atrial shunt device vs. sham control), all patients were required to have documented evidence of elevated pulmonary capillary wedge pressure (> 25 mm Hg during exercise) on their screening right heart catheterization study. Furthermore, patients with evidence of greater than mild right ventricular dysfunction, greater than mild tricuspid regurgitation, elevated pulmonary vascular resistance (PVR > 3.5 WU), or other reasons to suspect poor outcomes with the device were excluded. Thus, REDUCE LAP-HF II was an enrichment trial (a type of precision medicine trial; [82]). Nevertheless, despite the careful patient selection process, the overall trial results were neutral. However, further post hoc analyses demonstrated that patients with peak exercise pulmonary vascular resistance < 1.74 WU appeared to benefit from the device (Fig. 6; [83]). Thus, future precision medicine trials of atrial shunt devices and procedures in HFpEF may benefit from only including those patients with the ability to vasodilate the pulmonary vasculature during exertion.

Fig. 6
figure 6

Beneficial effects of interatrial shunt device therapy in patients with heart failure with preserved ejection fraction (HFpEF) with peak exercise pulmonary vascular resistance (PVR) < 1.74 WU enrolled in the REDUCE LAP-HF II trial. Top panel: In the sham control group, there was no association between peak exercise PVR and change in health status (Kansas City Cardiomyopathy Questionnaire overall summary score [KCCQ-OSS]) from baseline to 12 months (all patients improved approximately 10 points). However, in the atrial shunt device-treated patients, those with peak exercise PVR < 1.74 WU improved to a greater extent with the device compared with sham, whereas those with peak exercise PVR ≥ 1.74 WU did worse with the device compared with sham. Middle panel: Only HFpEF patients with peak PVR < 1.74 WU experienced a significant improvement in NYHA functional class in the trial. Bottom panel: Patients without latent pulmonary vascular disease (PVD; i.e., patients with peak exercise PVR < 1.74 WU) and no pacemaker at baseline had the highest win ratio, lowest HF event rate, and greatest improvement in health status in response to the atrial shunt device (compared with sham control). KCCQ-OSS Kansas City Cardiomyopathy Questionnaire, WU Wood units, NYHA New York Heart Association, LA left atrial. Reproduced with permission from Borlaug BA et al. [83]

Full size image

Phenomapping using advanced modalities

Current phenomapping approaches for HFpEF are largely based on clinical data available from demographics and routine clinical parameters such as medical history, medication history, routine blood tests, as well as electrocardiographic and echocardiographic read-outs [7, 15]. Future HFpEF phenomapping studies ideally will be planned prospectively and have already started at various centers across the world (e.g., the US National Institutes of Health HeartShare Study [www.HeartShareStudy.org]). In addition, advanced imaging, e.g., machine learning-based (magnetic resonance) image analysis and deep-learning algorithms for echocardiographic data may provide incremental prognostic value in HFpEF [84, 85]. Metabolic profiling quantifying blood serum metabolites by liquid chromatography–tandem mass spectrometry and proton-nuclear magnetic resonance spectroscopy and evaluation of their response to therapy (e.g., exercise training) has also been used to characterize HFpEF patients [86, 87]. An interesting approach is also to combine medical history and multi-omics analyses (e.g., serum metabolome and gut microbiome in cardiometabolic disease patients) to compute the response to drugs in patient subgroups [88].

Conclusion

Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous, systemic, multi-organ syndrome that is rising in prevalence and is associated with high morbidity and mortality. While certain therapeutics (e.g., sodium-dependent glucose contransporter‑2 inhibitors) may be beneficial in the large majority of HFpEF patients, most treatments are not well suited for a one-size-fits-all approach. Unsupervised machine learning (phenomapping) combined with -omics analyses is a growing approach to the classification of HFpEF and likely represents the future of HFpEF precision medicine. Nevertheless, investigators and clinicians should understand the potential limitations of such approaches and should augment their initial phenomapping studies of HFpEF with follow-up studies to identify underlying molecular mechanisms with the hope of conducting successful precision medicine trials in the future.