20^th Annual Feigenbaum Lecture: Echocardiography for Precision Medicine—Digital Biopsy to Deconstruct Biology

Abstract

Heart failure with preserved ejection fraction (HFpEF) is a complex, heterogeneous syndrome in need of improved classification given its high morbidity and mortality, and few effective treatment options. HFpEF represents an ideal setting to examine the utility and feasibility of a precision medicine approach. This article (based on the 20th annual Feigenbaum Lecture, presented at the 2019 American Society of Echocardiography Scientific Sessions) describes the utility of echocardiography as a “digital biopsy” and how deep quantitative echocardiographic phenotyping, coupled with machine learning, can be used to identify novel HFpEF phenotypes. The cellular and ultrastructural basis of abnormal speckle-tracking echocardiography (STE)-based measurements of cardiac mechanics can provide a window into cardiomyocyte calcium homeostasis. STE-based measurements of longitudinal strain can thus inform the extent of myocardial involvement in patients with HFpEF, which may help to determine responsiveness to cardiac-specific HF medications. However, classifying the complex, systemic, multi-organ nature of HFpEF appropriately likely requires more advanced methods. Using unsupervised machine learning, HFpEF can be classified into 3 distinct pheno-groups with differing clinical and echocardiographic characteristics, and outcomes: (1) BNP deficiency syndrome; (2) extreme cardiometabolic syndrome; and (3) right ventricle-cardio-abdomino-renal syndrome. Each can be probed to determine their biological basis. The goal of improved classification of HFpEF is to match the right patient with the right treatment, with the hope of improving the track record of HFpEF clinical trials. This article emphasizes the central role of echocardiography in advancing precision medicine, and illustrates the integration of basic, translational, clinical, and population research in echocardiography with the goal of better understanding the pathobiology of a complex cardiovascular syndrome.

Overview

In medicine, we take advantage of heuristics—short cuts that our brains use to arrive at a diagnosis in the setting of complex clinical scenarios.¹ However, these shortcuts leave us susceptible to diagnostic errors and incomplete understanding of disease.^1,2 We may be anchored on a specific diagnosis, unable to see the correct one, and we may only think of diagnoses that are available to us based on our recent memories and experiences. This was the case for patients suffering from heart failure with preserved ejection fraction (HFpEF). Nearly 35 years ago, in one of the early descriptions of the syndrome, Topol and colleagues described the hypertensive hypertrophic cardiomyopathy of the elderly,³ which fit nicely with the concept of diastolic dysfunction at a time when 2D, M-mode, and Doppler echocardiography were rapidly gaining popularity and widespread use. Thus, for 2 decades we were anchored to the available diagnosis of “diastolic heart failure”: a small left ventricular (LV) cavity, thickened LV walls, and an upward and leftward shifted LV end-diastolic pressure - volume relationship, resulting in diastolic dysfunction and elevated LV filling pressures.⁴ This lack of complete understanding of the HFpEF syndrome may have been a major reason for the disappointing track record of Phase 3 HFpEF clinical trials (CHARM-Preserved,⁵ I-PRESERVE,⁶ PEP-CHF,⁷ TOPCAT⁸, and PARAGON-HF [clinical trial acronym definitions are listed in Appendix A, available at www.onlinejase.com]).

In the early 2000s, it became clear to some that the classic descriptions of diastolic HF did not fit many patients.^9,10. Often the LV cavity was normal in size, and many patients did not have overt LV hypertrophy (LVH).^11,12 Furthermore, additional pathophysiologic abnormalities (including LV systolic dysfunction despite a preserved ejection fraction) were increasingly identified.^13,14 Thus, the term HFpEF was born, and the focus turned to better defining and characterizing the HFpEF syndrome by examining it holistically. It was against this backdrop that I started a dedicated HFpEF clinical and research program at Northwestern University, the first of its kind, in 2007.¹⁵ The model of this type of HFpEF program has since been emulated and expanded upon throughout the United States and across the world. This has coincided with improved enrollment in clinical trials and enhanced characterization of the HFpEF syndrome.

From the start, I was convinced that the heterogeneity of the HFpEF syndrome was why we were failing to find adequate treatments for it, and I was determined to identify novel methods of classifying the HFpEF syndrome. I felt that the only way to free us from our heuristic biases regarding diastolic HF was to try to observe and manage as many HFpEF patients as possible, in order to ascertain the full breadth of the types of patients diagnosed with the syndrome. Therefore, we used a systematic search query of inpatients hospitalized at Northwestern Memorial Hospital¹⁶ which identified patients with any of the following criteria: (1) the words “heart failure” written somewhere in the inpatient notes; (2) an elevated B-type natriuretic peptide (BNP > 100 pg/ml); or (3) 2 or more doses of intravenous diuretics. We narrowed the daily list to those patients with an EF > 50% who met Framingham criteria for HF. We then offered to follow-up with the patients in our newly created HFpEF clinic. Each patient was enrolled into a prospective observational study () with the goal of classifying the HFpEF syndrome thereby reducing its heterogeneity, and all patients who met inclusion/exclusion criteria were offered enrollment into the only major HFpEF trial being conducted at that time (TOPCAT).⁸

The Northwestern HFpEF Program quickly amassed patients with the HFpEF syndrome and offered a view of these patients that differed from prior investigations. These patients were not referred specifically for management of HFpEF. Rather, ours was a collection of “all comers” who were identified during HF hospitalization at an urban academic hospital and had a preserved EF. We noted that often times, the diagnosis of HFpEF ended without an investigation into its etiology or underlying dominant pathophysiology. Furthermore, the experience of seeing a large number of HFpEF patients solidified our notion that it is a heterogeneous syndrome in need of improved classification.

We now recognize HFpEF as a systemic syndrome that involves not only the heart but the lungs, the liver, the visceral adipose tissue, the kidney, and the skeletal muscle.^17,18 It still involves hypertensive remodeling and ventricular and vascular stiffening; however, sedentary lifestyle, poor fitness, and obesity and metabolic stress are now known to play major roles in the development of a global loss of cardiac, vascular, and peripheral reserve, culminating in the HFpEF syndrome.¹⁹ It is the complexity of this systemic, multi-organ syndrome; the heterogeneity of underlying pathophysiologies, etiologies, and type of clinical presentation; and the lack of a favorable response to uniform treatment (such as neurohormonal therapy for HF with reduced ejection fraction [HFrEF]) that has made the management of HFpEF so challenging.

Thus, we are currently faced with a problem. Imagine a world in which we treated all types of cancer the same—we would be failing miserably! And yet that is exactly what we have been trying to do for patients with HFpEF, using a one-size-fits-all approach in which putative therapies may be beneficial in some patients, neutral in some, and harmful in others. In the future, by using advanced imaging analysis, along with blood, urine, and/or tissue analysis, we can hopefully better classify HFpEF and provide targeted treatment, which may result in an improved track record for clinical trials. In oncology, targeted treatment is already a possibility—tissue is often readily available, and thus we can better understand the precise histopathological subtype and the underlying somatic mutation(s) that are driving a particular cancer and thereby provide targeted treatment. However, in HF, particularly HFpEF, obtaining myocardial tissue is challenging. Therefore, we must turn to other techniques, such as echocardiography, to perform a “digital biopsy” of the heart in order to improve our diagnostic precision.

Can we use echocardiography to perform a “digital biopsy”?

Echocardiography has long been used in animals of all types in order to characterize cardiac pathophysiology in vivo, especially serially during the development of disease or in response to interventions.^20,21 When done properly, with maintenance of physiological heart rates and only minimal sedation, it can be a powerful research tool. Each new echocardiography modality or tool developed for humans ultimately makes its way to the study of animals. Such was the case for strain imaging techniques such as speckle-tracking echocardiography (STE). The earliest descriptions of STE in animal models were proof-of-concept studies to validate STE (against sonomicrometry, for example²²), and to show that the technique could be successfully performed in mice and rats, where it clearly identified fibrotic areas of the myocardium.^23,24 At this time it was hypothesized that abnormal regional or global STE were indicative of focal fibrosis or diffuse interstitial fibrosis, respectively.

However, in humans STE appeared to be detecting myocardial disease very early in patients with risk factors for HF, at a time where we would not predict there would be significant myocardial fibrosis. Furthermore, it was known that abnormal excitation-contraction (EC) coupling (i.e., the process by which electrical activation leads to cardiomyocyte contraction and relaxation²⁵) occurs early in response to stressors such as hypertension.²⁶ Thus, we hypothesized that STE would correlate with abnormalities in both EC coupling and the ultrastructural feature of cardiomyocytes that is critical for efficient cardiomyocyte calcium homeostasis: transverse (T)-tubules. T-tubules are invaginations in the cell membrane that penetrate into the cardiomyocyte interior that allow for the efficient coupling of the action potential with calcium homeostasis (Figure 1A).²⁷ Lined with multiple ion channels, and in close proximity to ryanodine receptors and the sarcoplasmic reticulum, T-tubules are essential for normal EC coupling and the orderly and efficient release of calcium into the cardiomyocyte cytoplasm, which initiates cardiomyocyte contraction, followed by the reuptake of calcium back into the sarcoplasmic reticulum, with subsequent cardiomyocyte relaxation.

Figure 1. Ultrastructural and Cellular Basis for the Development of Abnormal Myocardial Mechanics During the Transition from Hypertension to HFpEF.

(A) Schematic of T-tubules in relation to L-type calcium channels, ryanodine receptors, and the sarcoplasmic reticulum. Reproduced with permission from Bers DM. Nature 2002; 415:198–205. (B) Decline in LV longitudinal strain during progression of hypertensive heart disease to HFpEF in the SHR compared to WKY controls. (C) Representative confocal microscopy images obtained from WKY rats and SHRs at different ages using di-4-ANEPPS staining; these images demonstrate T-tubule disruption (arrows), which increases in severity with aging in the SHR. (D) The relationship between T-tubule organization and Ca²⁺ release in a 9-month WKY rat and SHR are shown. There is a significantly delayed calcium release and reuptake in the SHR (arrows). (E) Scatterplot of T-tubule organizational index (a marker of T-tubule disruption) with LV longitudinal strain. (F) Relationship between age and diffuse myocardial fibrosis in the WKY and SHR. ATP = adenosine triphosphate; Ca = calcium; CTRL = control; echo = echocardiography; HFpEF = heart failure with preserved ejection fraction; HTN = hypertension; LV = left ventricular; LVH = left ventricular hypertrophy; Na = sodium; NCX = sodium-calcium exchanger; PLB = phospholamban; RyR = ryanodine receptor; SHR = spontaneous hypertensive rat; SR = sarcoplasmic reticulum; WKY = Wistar-Kyoto rat.

We sought to study T-tubules, cardiomyocyte calcium homeostasis, and cardiac mechanics during the development of HFpEF.²⁸ Therefore, we combined STE with confocal microscopy in a rat model of hypertension, the spontaneously hypertensive rat (SHR).²⁸ The SHR is a model of polygenic hypertension that mimics common forms of human hypertension. Rats with SHR develop early severe hypertension, followed by LVH at ~7–8 months of age, asymptomatic LV diastolic dysfunction at ~10–12 months of age, and overt clinical HFpEF around 16–18 months of age. By studying the SHR serially until the development of HFpEF and comparing to non-hypertensive Wistar-Kyoto (WKY) control rats, we could determine the ultrastructural and cellular basis for abnormal STE. STE was done in these rats using techniques similar to those used in humans. A commercially available echocardiography machine (Vivid 7, GE Medical Systems) with a linear transducer (i13L, GE Medical Systems) was used by sonographers with expertise in human and animal echocardiography to image the SHR and WKY rats. At various time points rats were sacrificed, and their hearts were perfused ex vivo on a Langendorff apparatus. Confocal microscopy was used to perform imaging of T-tubules as well as calcium imaging within the cardiomyocytes.

The results of these detailed studies demonstrated that longitudinal LV strain deteriorated progressively in the SHR, with the onset of decline beginning during the development of LVH when compared to WKY non-hypertensive controls (Figure 1B). On confocal microscopy it was apparent very early (around the time of LVH development) that T-tubules were beginning to be disrupted, a finding that was very apparent by the time of HFpEF development (Figure 1C). These changes in T-tubules were accompanied by inefficient and delayed calcium release and reuptake (Figure 1D). In the SHR, there was a direct correlation between T-tubule organizational index and LV longitudinal strain (Figure 1E) such that as T-tubules became increasingly disrupted (i.e., lower T-tubule organizational index), absolute LV longitudinal strain values were lower, indicating worse cardiac mechanics. On histological analysis of myocardial fibrosis in the SHR and WKY rats, we found that while significant myocardial fibrosis was present reliably in the SHR, it occurred late in the disease development process (around the time of overt clinical HFpEF), well after the onset of abnormalities in T-tubules and STE-based indices (Figure 1F). Thus, in the SHR, LV longitudinal strain on STE provides a window into cardiomyocyte calcium homeostasis, EC coupling, and the health of T-tubules well before the onset of myocardial fibrosis, and therefore can serve as a “digital biopsy” to assess the health of cardiomyocytes.

Can speckle-tracking echocardiography be used to resolve the heterogeneity of HFpEF?

Given the ability of STE to detect early changes in cardiomyocyte calcium homeostasis, it is not surprising that LV longitudinal strain has been found to be impaired in HFpEF compared to hypertensive controls.²⁹ However, there is heterogeneity in the amount of myocardial dysfunction among patients with HFpEF, which may influence response to treatment. Thus, we can leverage STE as a way to determine the extent of myocardial involvement in patients with HFpEF. Consider the spectrum of patients with HF symptoms, elevated LV filling pressures, and a preserved EF (Figure 2A). On one end there are non-cardiac causes of fluid overload (e.g., morbid obesity with plasma volume expansion, renal failure, liver failure), and on the other end there is a prototypical myocardial form of HFpEF such as an infiltrative cardiomyopathy. However, most HFpEF patients reside somewhere in the middle of these 2 extremes. STE can be used to determine the extent of myocardial involvement in a patient with HFpEF, which may have diagnostic and therapeutic implications (Figure 2B). For example, in a patient with very little myocardial involvement, treatment may consist of diuresis and identification of non-cardiac causes of fluid overload, whereas a patient with significant myocardial involvement may benefit from treatment with cardio-centric medications such as neurohormonal antagonists. In the event that STE is not available, careful inspection of tissue Doppler imaging tracings can provide a wealth of information regarding the presence and pattern of myocardial involvement (Figure 2C).

Figure 2. The Utility of Speckle-Tracking Echocardiography and Tissue Doppler Imaging for Sub-phenotyping HFpEF by Extent of Cardiac Involvement.

(A) Most patients with HFpEF fall on a spectrum between non-cardiac and cardiac causes of elevated LV filling pressures. (B) Abnormal strain (particularly LV longitudinal strain) can be used to determine the extent of myocardial involvement in patients with HFpEF. (C) If strain parameters are not available, tissue Doppler imaging can also be used to determine the extent of myocardial involvement in HFpEF. Reductions in systolic (s’), early diastolic (e’), and late diastolic (a’) velocities, prolongation of isovolumic contraction time (IVCT) and isovolumic relaxation time (IVRT), and reduction in ejection time, are all signs of a sick myocardium. HFpEF = heart failure with preserved ejection fraction; LV = left ventricular.

In addition to diagnosing the extent of myocardial involvement, STE can also help phenotype HFpEF beyond simple EF categorization.³⁰ In patients with HFpEF and reduced LV longitudinal strain, the differential diagnosis includes contractile dysfunction, maladaptive hypertrophy, coronary microvascular dysfunction, myocardial fibrosis, and infiltrative cardiomyopathy. In patients with HFpEF and preserved LV longitudinal strain, primary left atrial (LA) dysfunction, vascular stiffening, abnormal ventricular-arterial coupling, primary right ventricular (RV) dysfunction, and extracardiac volume overload should be considered as possible underlying causes. The pattern of abnormal LV longitudinal strain, and the response of strain parameters to stressors such as exercise, can also be helpful in phenotyping HFpEF. Bullseye patterns of segmental STE can diagnose specific disease states that present as HFpEF (such as cardiac amyloidosis and hypertrophic cardiomyopathy), whereas lack of augmentation of LV longitudinal strain with exercise is a marker of poor contractile reserve in the setting of HFpEF.^31,32

Can machine learning be used to resolve the heterogeneity of HFpEF?

The experience of seeing a large volume and breadth of HFpEF patients in the Northwestern HFpEF program led us to consider a variety of classification systems such as phenotyping by extent of myocardial involvement (as described above) and classifying by predominant pathophysiological abnormality, etiology, and type of clinical presentation.¹⁷ However, while simple to implement, none of these classifications were sophisticated enough to embrace the full complexity of our HFpEF patients. We enrolled approximately 400 HFpEF patients into a prospective observational registry within which each patient underwent comprehensive echocardiography under a standardized protocol after which we performed deep quantitative phenotyping of all echocardiograms in a blinded fashion within our core laboratory.^16,33 The wealth of echocardiographic data, combined with quantitative physical exam characteristics, laboratory data, and electrocardiographic data, provided us high-resolution phenotyping, but combining the high density of data in conventional statistical models in order to discover novel HFpEF subtypes proved difficult. Thus, we began performing exploratory machine learning analyses on our data. We hypothesized that machine learning would allow us to perform iterative clinical discovery, with novel biological insights. We planned to use machine learning to find patterns within the data, thereby defining novel, mutually exclusive HFpEF subtypes. By examining these newly defined subgroups within HFpEF from a clinical standpoint, we humans could then test new hypotheses regarding the groups which would then hopefully lead to novel biological discoveries.

There are several types of machine learning analyses, but 3 broad categories are unsupervised learning, supervised learning, and deep learning.³⁴ Unsupervised learning attempts to find patterns in unlabeled data.^35,36 Supervised learning works by using labeled data in a training dataset so that it can effectively learn to label data (e.g., diagnoses, outcomes) in other datasets. Deep learning is useful for application to very high-density datasets such as raw imaging data; while it is often used to learn how to label data (i.e., supervised learning), it can also be used in an unsupervised fashion to identify features within data. For our purposes of finding a novel classification system for HFpEF, unsupervised machine learning was the most suitable, and we termed this approach “phenomapping” because it was based strictly on phenotypes (and not genetic data such as gene expression analyses).³³ Therefore, we first began by performing hierarchical clustering analyses on our data in 397 HFpEF patients. The resultant cluster heatmap (Figure 3A), demonstrated that the HFpEF patients in our dataset (all of whom had a prior HF hospitalization, preserved EF > 50%, and elevated LV filling pressures) were highly heterogeneous. We subsequently moved to a Bayesian machine learning analysis (model-based clustering) and found that the most parsimonious solution was 3 pheno-groups, which were quite different on principal component analysis (Figure 3B) and in terms of outcomes (Figure 3C). Importantly, we showed that our phenomapping model was predictive of outcomes above and beyond the Meta-Analysis Global Group in Chronic (MAGGIC) HF risk score and BNP, and we validated our model prospectively in 107 additional HFpEF patients.

Figure 3. Phenomapping for Novel Classification of Heart Failure with Preserved Ejection Fraction.

(A) Hierarchical clustering heatmap (“pheno-map”) of 397 HFpEF patients. Each row is a quantitative phenotype; each column is an individual patient; red = increased, blue = decreased. (B) Principal component analysis demonstrating clear separation of the 3 pheno-groups. (C) Kaplan-Meier curves for the 3 HFpEF pheno-groups showing survival free of cardiovascular hospitalization or death. Figure panels reproduced with permission from Shah SJ, et al. Circulation 2015; 131:269–279. CV = cardiovascular; PC1 = principal component #1; PC2 = principal component #2; PC3 = principal component #3.

Although each of the 3 HFpEF pheno-groups included patients with a prior history of HF hospitalization, preserved EF, LA enlargement, and elevated LV filling pressures, the groups were quite different in terms of their clinical, electrocardiographic, and echocardiographic characteristics (Figure 4). Pheno-group #1, the “BNP deficiency syndrome”, had the least cardiac remodeling and dysfunction, and the lowest BNP levels. Pheno-group #2, the “extreme cardiometabolic syndrome”, had the highest prevalence of diabetes and the most severely impaired myocardial relaxation. Pheno-group #3, the “RV-cardio-abdomino-renal syndrome”, had the most severe cardiac and electrical remodeling (including significant RV enlargement and dysfunction), and renal dysfunction. No single phenotypic variable was able to easily classify patients into a particular pheno-group, and these groups were not meant to be rigid classifications (indeed, membership in a particular HFpEF pheno-group simply meant that the probability of that pheno-group was highest based on each patient’s phenotypic characteristics). Our HFpEF phenomapping classification was not meant to be definitive; rather, its purpose was to highlight potential differences among the 3 pheno-groups in order to inform future investigations of underlying pathobiology.

Figure 4. HFpEF Pheno-groups.

Example echocardiograms (apical 4-chamber view) for each of the 3 pheno-groups. Reproduced with permission from Shah SJ. J Cardiovasc Transl Res 2017; 10:233–244. BNP = B-type natriuretic peptide; LA = left atrium; LVEF = left ventricular ejection fraction; PCWP = pulmonary capillary wedge pressure; RV = right ventricular.

Pheno-group #1: BNP deficiency syndrome

In caring for a large number of HFpEF patients (in whom we often perform invasive hemodynamic testing) it became apparent early in our experience that BNP was below the diagnostic threshold for abnormal (e.g., < 100 pg/ml for BNP) in a large proportion of HFpEF patients. In our Northwestern HFpEF Program cohort, we found that 29% of patients with elevated pulmonary capillary wedge pressure (PCWP) confirmed invasively had a BNP < 100 pg/ml.³⁷ Not surprisingly, these patients were more obese and younger than those with BNP > 100 pg/ml. Obesity is well known to be associated with lower BNP levels due to increased natriuretic peptide clearance (via NPR-C receptors on adipocytes), and reduced BNP production.³⁸ In addition, regardless of obesity, for any given level of volume overload, patients with HFpEF will have a lower BNP level compared to HFrEF due to lower diastolic wall stress in HFpEF (wall stress = pressure × radius / wall thickness; while diastolic pressure is elevated, the radius of the LV is normal or small, and the wall thickness is normal to increased). Thus, HFpEF in general is a BNP deficient state with lower than necessary BNP secretion by cardiomyocytes due to insufficient wall stress, thereby resulting in inadequate vasodilation and natriuresis. Interestingly, natriuretic peptides have lipolytic effects,³⁹ promote “browning” of white adipocytes to increase energy expenditure,⁴⁰ and are associated with lower levels of visceral adiposity.⁴¹ Therefore, BNP deficiency results in inadequate lipolysis, resulting in increased visceral adiposity and accumulation of white (unhealthy) adipocytes.⁴² Thus, BNP deficiency results in more visceral adiposity, which results in even lower circulating BNP due to increased adipocyte-mediated BNP clearance.

We now know that in addition to obesity and low wall stress, natriuretic peptide deficiency in the setting of HFpEF can be due to a wide variety of other reasons, including genetic causes⁴³ (polymorphisms in the NPPA and NPPB genes for atrial natriuretic peptide and BNP, respectively); insulin resistance;⁴⁴ increased androgenicity in women;⁴⁵ and African ancestry.⁴⁶ Thus, these BNP deficient individuals are susceptible to increased plasma volume, higher blood pressure, and increased adiposity, all of which lead to the development and pathogenesis of the HFpEF syndrome. The increased adiposity has other ramifications: chest wall, pericardial, and epicardial fat can all result in heightened pericardial constraint and ventricular interdependence,⁴⁷ and can mimic constrictive pericarditis, as shown in Figure 5.

Figure 5. Example of the BNP Deficiency Syndrome Due to Morbid Obesity.

Images from a 74-year-old woman with HFpEF, a body mass index of 46 kg/m², and a BNP of 28 pg/ml with evidence of volume expansion plus pericardial constraint due to excessive epicardial and pericardial fat. (A) Apical 4-chamber view demonstrating epicardial and pericardial fat (arrows), as well as a prominent diastolic septal bounce (video 1). (B) Cardiac magnetic resonance demonstrates a normal pericardium but severe epicardial and pericardial fat overlying the right ventricle (arrows), as well as a prominent diastolic septal bounce (video 2). (C) Simultaneous right and left heart catheterization demonstrating discordance of the LV and RV pressure tracings at peak inspiration, which indicates the presence of diastolic ventricular interdependence. LV = left ventricle; RV = right ventricle.

The BNP deficiency syndrome is difficult to diagnose because of the lack of BNP elevation and the presence of obesity, which can hide the diagnosis and lead clinicians astray. These patients require careful examination of the echocardiogram (and often invasive hemodynamic testing) to make the diagnosis. The diagnosis should be considered whenever the BNP is below the diagnostic threshold for HF in a patient with signs and symptoms of HFpEF, especially if the patient is morbidly obese. The presence of a diastolic septal bounce or other signs of ventricular interdependence with otherwise fairly normal cardiac structure and function (except for LA enlargement) is also a clue to this type of HFpEF (as long as pericardial disease is excluded). Treatment consists of diuresis and weight loss (through diet, medications, or surgery). Potential novel treatments include medications that increase natriuretic peptides (e.g., neprilysin inhibitors), Sodium-glucose cotransporter-2 (SGLT-2) inhibitors (which result in diuresis and weight loss), and pericardiotomy.^48,49

Pheno-group #2: Extreme cardiometabolic syndrome

In 2013, Paulus and Tschöpe proposed a novel paradigm (the “Paulus hypothesis”)⁵⁰ for the development of HFpEF, whereby comorbidities (e.g., hypertension, obesity, diabetes, chronic kidney disease, etc.) create a pro-inflammatory milieu which results in systemic inflammation and dysfunction of the endothelium in multiple organs, including the heart, lungs, kidneys, and skeletal muscle. In the heart, coronary endothelial inflammation and dysfunction results in coronary microvascular dysfunction. In addition, there is impaired endothelium-cardiomyocyte signaling, which results in reduced cyclic guanosine monophosphate (cGMP) and protein kinase G, leading to hypophosphorylation and stiffening of titin, the major molecular spring in cardiomyocytes. In addition, chronic low-grade inflammation results in migration of leukocytes into the myocardial interstitium, with secretion of transforming growth factor (TGF)-β which promotes collagen deposition by myofibroblasts and subsequent diffuse myocardial fibrosis. Both mechanisms (cardiomyocyte stiffening and myocardial fibrosis), which are present in varying degrees in individual HFpEF patients, can cause the increased LV diastolic stiffness that is characteristic of the HFpEF syndrome.^51,52

The extreme cardiometabolic syndrome phenotype of HFpEF most closely appears to follow the aforementioned molecular paradigm of HFpEF, but this paradigm still needs to be validated and further explored. We sought to examine the Paulus hypothesis indirectly by investigating it in population-based studies. In the Hypertension Genetic Epidemiology Network (HyperGEN), Cardiovascular Health Study (CHS), and Multi-Ethnic Study of Atherosclerosis (MESA), all cardiovascular population-based studies funded by the National Institutes of Health, we have performed speckle-tracking analysis on > 13,000 echocardiograms (including digitization of archival echocardiograms⁵³ in HyperGEN and CHS), thereby allowing us to examine the associations of various HFpEF risk factors with abnormal cardiac mechanics.

Specific to the Paulus hypothesis, we found that in at-risk individuals prior to the development of HFpEF, increasing comorbidity burden, higher waist:hip ratio (a marker of visceral adiposity), and greater degrees of low-grade albuminuria (a marker of widespread endothelial dysfunction) are all associated with worse LV systolic longitudinal strain and lower e’ velocities, reflecting the myocardial substrate (both systolic and diastolic dysfunction) for HFpEF.^54–56 Furthermore, in the Prevalence of Coronary Microvascular Dysfunction in HFpEF (PROMIS-HFpEF) study, we prospectively enrolled 202 patients with HFpEF from 5 centers across the world and performed coronary flow reserve (CFR) testing with adenosine transthoracic Doppler echocardiography of the mid-to-distal left anterior descending coronary artery (Figure 6).⁵⁷ In PROMIS-HFpEF (the largest prospective study of coronary microvascular dysfunction in HFpEF to date), we found that coronary microvascular dysfunction (defined as CFR < 2.5) is present in 75% (95% CI 69–81%) of HFpEF patients. Furthermore, we found that impaired CFR was associated with higher BNP levels, greater magnitude of albuminuria, worse RV systolic function, and greater systemic endothelial dysfunction.

Figure 6. Comparison of Coronary Doppler Tracings and Right Ventricular Indices in HFpEF Patients in the Presence and Absence of Coronary Microvascular Dysfunction.

The patient without coronary microvascular dysfunction (left panel) had a normal coronary flow reserve (2.88), whereas the patient with coronary microvascular dysfunction (right panel) had a reduced coronary flow reserve (1.63). The lower coronary flow reserve in the patient with coronary microvascular dysfunction was associated with lower tricuspid annular plane systolic excursion and worse right ventricular free wall strain (bottom of each panel), as well as lower reactive hyperemia index (1.65 vs. 2.05), worse left ventricular global longitudinal strain (7.8% vs. 13.2%), and worse left atrial reservoir strain (6.7% vs. 15.8%). Reproduced with permission from Shah SJ, et al. Eur Heart J 2018; 39:3439–3450. CFR = coronary flow reserve; CMD = coronary microvascular dysfunction; LV = left ventricular; NTproBNP = N-terminal B-type natriuretic peptide; RHI = reactive hyperemia index; RV = right ventricular; RVFW = right ventricular free wall; TAPSE = tricuspid annular plane systolic excursion.

Together, the data from population-based studies and PROMIS-HFpEF clinical physiological study provide support for the Paulus hypothesis; however, these are associative studies and not definitive. We and others are currently performing additional basic science and clinical studies to further examine the Paulus hypothesis. For example, proteomic analyses with a larger number of inflammatory, endothelial, and cardiovascular markers can be examined using a suite of techniques to cluster related proteins that represent biological pathways. These clusters of proteins can then be related to comorbidity burden, abnormal cardiac structure/function in HFpEF, and, by using mediation analyses, causal inferences can be made. Furthermore, recent developments in rodent models for HFpEF suggest that we are getting closer to more relevant “extreme cardiometabolic” pre-clinical models for HFpEF that can be probed for disease mechanisms, as was done recently in a 2-hit HFpEF mouse model (high fat diet plus N(ω)-nitro-L-arginine methyl ester [L-NAME]), which showed that nitrosative stress was important in the development of cardiometabolic HFpEF.⁵⁸ The development of animal models that better recapitulate the HFpEF syndrome open the door for novel investigation of inflammation in HFpEF (e.g., macrophages, which likely play a central role in HFpEF pathogenesis in response to cardiometabolic stress).^59,60

From an echocardiographic standpoint, assessment of cardiac mechanics (both with STE and tissue Doppler imaging) is of significant value in the evaluation of the extreme cardiometabolic phenotype of HFpEF. These highly reproducible echocardiographic measures can be related to comorbidities, circulating markers, and genetics in large population-based studies, but they can also be utilized in detailed clinical physiological studies like PROMIS-HFpEF, and in animal models, as demonstrated above in the SHR. From a treatment standpoint, along with diuresis, addressing the comorbidities that drive this type of HFpEF is essential. In addition, as shown in the TOPCAT trial, spironolactone may be especially useful in these extreme cardiometabolic syndrome patients who have only mild elevations of BNP.⁶¹ Finally, SGLT-2 inhibitors, neprilysin inhibitors, and drugs in development that improve coronary microvascular dysfunction or enhance cGMP (e.g., nitrites, sGC stimulators, and phosphodiesterase-9 [PDE9] inhibitors) may ameliorate this type of HFpEF.⁶²

Pheno-group #3: Right ventricle – cardio-abdomino-renal syndrome

Pheno-group #3, the RV-cardio-abdomino-renal syndrome type of HFpEF, is the highest risk and most difficult to manage. While many investigations of this type of HFpEF patient center around treatment to improve RV function and decreased RV afterload by vasodilating the pulmonary vasculature, there has been relatively less attention to the downstream effects of chronic venous congestion. It is now well known that in the setting of HF (regardless of HFrEF or HFpEF), chronic venous congestion is likely a greater driver of renal dysfunction and worse outcomes than reduced cardiac output.^63–65 This is especially the case in HFpEF when cardiac output at rest is often within normal limits. Chronic venous congestion leads to splanchnic congestion, which can have dramatic effects on the gut (Figure 7).^66,67 Enterocytes that line the intestines operate in a hypoxic environment to begin with; chronic venous congestion likely enhances an anerobic environment in these enterocytes, thereby leading to intracellular acidosis. In order to combat this phenomenon, we hypothesize that there is activation of the sodium-hydrogen exhanger-3 (NHE3), the major ion channel responsible for sodium and fluid absorption in the gut. Persistent activation of NHE3, which we suspect occurs in patients with right-sided HF and significant venous congestion, would then create a low pH (due to increased H⁺) in the gut lumen, and worsen fluid retention and volume overload (due to enhanced Na⁺ and fluid absorption). These changes in the gut lumen could result in microbial dysbiosis, with subsequent reduction in short-chain fatty acids (which can increase gut permeability); and increased Trimethylamine N-oxide (TMAO), which could worsen systemic inflammation. These changes in turn could result in increased susceptibility to infections, sepsis, and cachexia. Worsening fluid overload could exacerbate venous congestion, leading to worse renal venous congestion, worsening renal function, and increased susceptibility to HF exacerbation and death.

Figure 7. Conceptual Diagram of the Relationship Between Right-Sided Heart Failure, Splanchnic Congestion, the Intestinal Microenvironment, and Adverse Outcomes.

Top panel: Patients in HFpEF pheno-group #3 have right heart failure that then leads to venous congestion (as indicated by a dilated inferior vena cava), with subsequent splanchnic congestion. Bottom panel: It is well known that right-sided HF and venous congestion is associated with increased risk of renal failure and death; however, the exact mechanisms between these observations are less well known. Here we propose a potential pathophysiological mechanism to explain these associations. Right-sided heart failure causes venous congestion, which leads to splanchnic congestion. Venous congestion of the splanchnic vasculature causes reduced blood flow to the gut enterocytes, causing hypoxia and anaerobic metabolism in these cells. These conditions induce upregulation and increased activation of NHE3, inducing increased sodium absorption and decreased luminal pH at the brush border resulting in worsened volume overload, and a microbial dysbiosis. Microbial dysbiosis may result in an increased bacterial secretion of TMAO, reduced production of SCFAs, and worsening of systemic inflammation. Reproduced with permission from Polsinelli V, et al. Curr Opin Support Palliat Care 2019;13:24–30. HF = heart failure; Na⁺ = sodium; NHE3 = sodium-hydrogen exchanger-3; RA = right atrium; RV = right ventricle; SCFAs = short-chain fatty acids; TMAO = trimethylamine-N-oxide.

From an echocardiographic standpoint, quantitative phenotyping of right heart structure and function, coupled with estimation of central venous pressures, could provide additional insight into gut and renal pathophysiology in these patients. Besides treatment with diuretics, there are novel medications in development such as tenapanor, which blocks NHE3 in the gut.⁶⁸ These patients may also benefit from inotropes like levosimendan, which is currently being tested in a clinical trial () as a way to enhance RV function in HFpEF. Diuretic titration with implantable hemodynamic sensors (e.g., CardioMEMS) can also be very helpful in these patients with HFpEF, RV failure, and cardiorenal syndrome.⁶⁹ Finally, carefully designed clinical trials of pulmonary vasodilators (e.g., endothelin receptor antagonists [] or prostacyclin analogues []) to target selected patients with HFpEF and combined post- and pre-capillary pulmonary hypertension and RV failure are currently underway.

How can we bring phenomapping to the clinic?

Our study of machine learning-based phenomapping highlights a novel approach to enhance classification of HFpEF, with the ultimate goal of providing targeted therapeutics in a precision medicine framework. However, our approach requires detailed quantitative analysis of echocardiograms by a central core lab, which is not feasible in routine clinical care. In the future, deep learning and computer vision techniques will likely supplant the need for human quantitation of echocardiograms, though there will still be a need for error-checking findings from a computer algorithm. Nevertheless, the advent of fully automated echocardiographic measurements and even disease diagnosis and tracking via machine learning⁷⁰ suggests that phenomapping could be done more quickly and at a larger scale in the near future (Figure 8).

Figure 8. Proposed Workflow for Phenomapping in the Clinic: Deep Phenotyping, Deep Learning of Echocardiography Images, and Statistical Learning of Derived Features to Assign Pheno- Groups.

(A) History and physical exam, laboratory, electrocardiography, and echocardiography data and images are collected, after which deep machine learning is used to automate the identification and quantitation of echocardiographic features. These features are then analyzed with statistical learning to assign pheno-groups after which patients undergo targeted treatments and/or enrollment into targeted clinical trials. Reproduced with permission from Shah SJ. J Cardiovasc Transl Res 2017; 10:233–244. (B) Example of deep machine learning and computer vision techniques that can be used for deep learning step shown in panel A. t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of echocardiographic view classification. The t-SNE plot depicts the successful grouping of test images corresponding to 23 different echocardiographic views (“no occlusions” means the view is complete, with visualization of all chambers normally seen in that view; “occluded” refers to zoomed in views of particular chambers [i.e., A4c.LA occluded = zoomed in view with the left atrium not in the view, as shown in the corresponding echocardiographic image shown in the panel next to that grouping]). Each color represents one of the 23 unique echocardiographic views (listed to the right of the figure). On the t-SNE plot, and the distance between the groupings is indicative of the performance of the deep learning algorithm for view identification (the farther the groups are from each other, the more accurate the performance of the algorithm). Once views are accurately identified by the deep learning algorithm, cardiac chambers can be identified and quantitated by machine learning and computer vision techniques. Reproduced with permission from Zhang J, et al. Circulation 2018; 138:1623–1635. A2c = apical 2-chamber view; A3c = apical 3-chamber view; A4c = apical 4-chamber view; A5c = apical 5-chamber view; AoV = aortic valve; ECG = electrocardiography; Echo = echocardiography; LA = left atrium; LV = left ventricle; MV = mitral valve; PapMuscle = papillary muscle; PLAX = parasternal long axis view; PSAX = parasternal short axis view; RV = right ventricle.

Matchmaking for targeted therapeutics in HFpEF

Although we eagerly await the results of ongoing Phase 3 trials of the angiotensin receptor-neprilysin inhibitor sacubitril/valsartan^71,72 and SGLT-2 inhibitors⁷³ (large-scale trials that essentially still utilize a “one-size-fits-all” approach), in the future, we believe that there is still a major unmet need to classify HFpEF into therapeutically homogenous subtypes based on advanced image analysis, machine learning, exercise phenotyping, liquid biopsy (i.e., circulating biomarkers), and even cardiac biopsy, thereby leading to phenotype-specific treatment. While we are not there yet, we have begun to realize a precision medicine framework for HFpEF within our Northwestern HFpEF Program. All patients enrolled in our HFpEF program undergo detailed echocardiographic phenotyping, as well as exercise echocardiography and invasive hemodynamic testing when possible. These tests allow us to specify the particular HFpEF subphenotype (Figure 9), which we then target towards specific clinical trials. This type of approach has been labeled an “umbrella design” in the literature⁷⁴ and has mainly been used in oncology trials thus far. For example, within a lung cancer clinic, histology and genomic analysis of excised tumor tissue could identify specific molecular pathways that are then targeted with specific treatments tested in separate clinical trials. In a similar fashion, by targeting specific subtypes of HFpEF based on phenotype, we can attempt to match the right patient to the right treatment and its associated clinical trial.⁷⁵

Figure 9. Umbrella Design for HFpEF Clinical Trials.

Once the diagnosis of HFpEF is confirmed, patients are enrolled undergo deep echocardiographic phenotyping (and additional testing, such as exercise hemodynamics), after which patients are directed to the most appropriate HFpEF clinical trial based on their underlying etiology or pathophysiology. It is important to note that the inclusion/exclusion criteria for these trials are not necessarily mutually exclusive. Thus, currently the assignment of patients into particular trials is based on clinical judgment, in addition to the inclusion/exclusion criteria. BP = blood pressure; CpcPH = combined post- and pre-capillary pulmonary hypertension; HCM = hypertrophic cardiomyopathy HFpEF = heart failure with preserved ejection fraction; LAP = left atrial pressure; PVR = pulmonary vascular resistance; RAP = right atrial pressure; RV = right ventricular. Clinical trial acronym definitions are listed in Appendix A, available at www.onlinejase.com.

Consider the “LA-predominant” phenotype of HFpEF.^76,77 It is now clear that there are some patients with HFpEF who have significant LA dysfunction and normal right heart function. On echocardiography, these HFpEF patients are characterized by significant LA enlargement, an interatrial septum that bows from left to right, and normal size and collapsibility of the inferior vena cava. On invasive hemodynamic testing, these patients have significant elevation in PCWP, particularly with exercise, and have normal or near-normal RA pressure with no significant elevation of pulmonary vascular resistance (PVR) and preserved cardiac output. In these patients, we have been developing a novel treatment—an InterAtrial Shunt Device (Corvia Medical) (Figure 10A)—that allows dynamic decompression of the overloaded LA into the larger reservoir of the right atrium, great veins, and hepatic veins.^78–80 The idea for this novel device-based therapy for HFpEF came from the Lutembacher syndrome, first described more than 100 years ago, in which a pre-existing congenital atrial septal defect occurs in a patient with acquired mitral stenosis, which results in a reduction in LA pressure and leads to less symptoms and greater exercise tolerance.⁸¹

Figure 10. InterAtrial Shunt Device for the Treatment of HFpEF: Concept and Results of the REDUCE LAP-HF I Randomized Controlled Trial.

(A) The Corvia InterAtrial Shunt Device creates an interatrial shunt that unloads the left atrium by shunting blood from the higher pressure left atrium to the lower pressure right atrium. (B) Pulmonary capillary wedge pressure during exercise hemodynamic testing: baseline versus 1-month post-randomization, stratified by treatment group. P values were calculated using paired t tests (within-group comparisons of baseline versus 1-month values). *P<0.05; **P<0.01. (C) Cumulative 12-month incidence of heart failure events requiring intravenous diuretic treatment, stratified by treatment group. Panels A and B reproduced with permission from Feldman T, et al. Circulation 2018; 137:364–375. Panel C reproduced with permission from Shah SJ, et al. JAMA Cardiol 2018; 3:968–977. HF = heart failure; IASD = InterAtrial Shunt Device; LA = left atrial; PCWP = pulmonary capillary wedge pressure; RA = right atrial.

In open-label and sham-controlled randomized controlled trials (REDUCE LAP-HF⁸² and REDUCE LAP-HF I,^83,84 respectively [clinical trial acronym definitions are listed in Appendix A, available at www.onlinejase.com]), we have shown that the InterAtrial Shunt Device is associated with a reduction in exercise PCWP (Figure 10B), appears to be safe, and may be associated with improved outcomes (Figure 10C). A major factor in the success of these trials has been the key role of echocardiography in phenotyping and selecting patients to enrich the trials with the LA-predominant phenotype of HFpEF. All patients undergo detailed echocardiographic analysis to exclude right heart dysfunction or significant tricuspid regurgitation; those that qualify then undergo invasive hemodynamic screening with exercise right heart catheterization to confirm (1) exercise PCWP > 25 mmHg, and (2) the absence of significant elevations in right atrial pressure or PVR.

Concluding remarks

In an ideal world, we would perform myocardial biopsy and detailed echocardiographic and other imaging analyses to determine a subtype of HFpEF, after which its molecular basis could be ascertained. This would then lead to targeted drug development based on the underlying molecular mechanism. Novel imaging techniques that allow non-invasive diagnosis would then be used to enroll patients into clinical trials that would demonstrate improved symptoms, exercise capacity, and outcomes in patients taking the study drug compared to placebo. Some may believe that this idealized scenario is merely a fantasy and precision medicine a fallacy that is based on hype. However, the aforementioned scenario recently came to fruition for patients with amyloidogenic transthyretin cardiomyopathy (ATTR-CM),⁸⁵ which often falls under the umbrella of HFpEF, misdiagnosed as the more common, “garden variety” form of HFpEF. Over the past 10 years, the dream of targeted therapeutics in HFpEF has come true, first through an observational registry of ATTR patients, followed by a phase 3 clinical trial of a transthyretin stabilizer (tafamidis)⁸⁶ in which we tested whether prevention of disassociation of the transthyretin tetramer into monomers (the rate-limiting step in amyloid fibril formation in ATTR-CM) occurs. We now have a highly sensitive and specific method on echocardiography to diagnose cardiac amyloidosis (relative apical sparing of longitudinal strain);⁸⁷ a highly specific confirmatory test for ATTR-CM (bone scintigraphy such as technetium pyrophosphate scanning);⁸⁸ and a proven treatment (tafamidis) which was shown to decrease all-cause mortality and prevent declines in exercise capacity and functional status in the ATTR-ACT randomized controlled trial.⁸⁶ The successes of diagnosing ATTR-CM gives us hope that we can use echocardiography-based precision medicine in the future for clinical syndromes such as HFpEF.⁸⁵

On the occasion of the 20^th anniversary of the Feigenbaum Lecture, it is appropriate to return to Dr. Harvey Feigenbaum and others who were pioneers in the development of echocardiography.⁸⁹ At its dawn, echocardiography opened up a new and exhilarating era of bedside diagnosis, allowing us to visualize the heart and make diagnoses with our eyes and minds. Today we are entering a new phase, one in which we can use echocardiography as a digital biopsy, accompanied by machine learning and data driven analytics, to see the unseen and provide unprecedented insight into the biology of heart disease.

In summary, echocardiography can be used as a tool to enhance precision medicine for complex cardiovascular syndromes. In addition, STE can serve as a “digital biopsy” that provides a window into the earliest manifestations of abnormal cardiomyocyte calcium homeostasis and deranged EC coupling. Furthermore, machine learning plus deep phenotyping of echocardiography can allow for the novel classification of HFpEF, which, in turn, may lead to novel biological insights. Finally, in the near future, echocardiography-based precision medicine should lead to more successful targeted treatment for complex, difficult-to-treat cardiovascular syndromes.

Supplementary Material

1. Echocardiographic apical 4-chamber view from a 74-year-old woman with HFpEF, obesity, and low B-type natriuretic peptide, with pericardial constraint due to excessive epicardial and pericardial fat.

The video demonstrates epicardial and pericardial fat, as well as a prominent diastolic septal bounce, consistent with ventricular interdependence.

2. Cardiac magnetic resonance 4-chamber cine from a 74-year-old woman with HFpEF, obesity, and low B-type natriuretic peptide, with pericardial constraint due to excessive epicardial and pericardial fat.

The video demonstrates epicardial and pericardial fat, as well as a prominent diastolic septal bounce, consistent with ventricular interdependence. There is no evidence of pericardial thickening, and there was no pericardial enhancement.