Biomarker Phenotypes in Heart Failure with Preserved Ejection Fraction Using Hierarchical Clustering: A Pilot Study

Heart failure with preserved ejection fraction (HFpEF) represents a clinical syndrome characterized by signs, symptoms, and structural changes of the heart in the presence of normal or near-normal ejection fraction (EF >50%) [1, 2]. The HFpEF subtype constitutes more than half of the total heart failure population [1, 2], with an annual rising rate of 1% [3]. Its complex and poorly defined etiology and pathophysiology significantly challenge therapy for patients with HFpEF [4]. Indeed, most of the clinical studies conducted to identify specific therapies for this subtype of heart failure have been unsuccessful and have failed to achieve their primary outcome [4‒7]. Furthermore, no significant improvement in morbidity and mortality for these patients has been reported so far [4‒7]. These findings lead to a recognition of HFpEF patients as a diverse phenotype population [8], attributing its heterogeneity to the various phenotypes within the HFpEF subtype [5].

In recent years, a new discipline named phenomics [9, 10] has been successfully introduced into clinical practice [11, 12] to recognize and divide subgroups of patients within unique cohorts of individuals, according to their qualitative and quantitative characteristics: environmental, genetic, lifestyle, clinical or imaging data, or pharmacological response. Phenomics represents a systematic collection and analysis of multidimensional data, visualized by heat maps and analyzed by clustering algorithms [9, 10]. The methods applied (machine learning) allow the refinement and the specification of the particular phenotype, learning relationships between the objects [13], and suggest a personalized approach to patients and phenotype-specific treatment [5]. Most recently, using machine learning, phenotype mapping (“phenomapping”) of the HFpEF population revealed subpopulations of patients who are at greater risk of disease progression and defined long-term adverse outcomes [5, 11, 14‒19].

Using this knowledge, we hypothesized the existence of distinct biomarker phenotypes (for biomarkers beyond routine biochemistry) within the unique cohort of HFpEF patients, demonstrating different clinical, laboratory, echocardiographic, and electrocardiographic parameters and therapies. We sought to identify these phenotypes, using unsupervised machine learning approaches applied to plasma concentration of cardiac remodeling biomarkers: soluble suppression of tumorigenicity 2 (sST2), galectin-3 (Gal-3), growth differentiation factor (GDF)-15, syndecan (Syn)-1, BNP, and cystatin C. We also aimed to understand them as new phenotypes, based on differences in their clinical, biochemical, echocardiographic, and electrocardiographic variables and pharmacological therapy.

This cross-sectional, single-center study was conducted at the Institute for Treatment and Rehabilitation Niska Banja, Serbia. 75 patients, previously diagnosed with HFpEF, regardless of etiology, who had been admitted to the institute for the purpose of rehabilitation were studied. Diagnosis of HFpEF was established according to 2016 ESC Guidelines [1] and required the following: the presence of signs and symptoms of heart failure, left ventricular EF (LVEF) ≥50%, BNP plasma concentration over 35 pg/mL and at least one additional criteria, relevant structural heart disease (LV hypertrophy and/or left atrial enlargement), and/or diastolic dysfunction [1]. Key structural alterations were considered as a left atrial volume index >34 mL/m² or a left ventricular mass index ≥ 115 g/m² for males and ≥95 g/m² for females, while diastolic abnormality was regarded as E/A ratio <0.75 or ≥1.5.

All patients with HFpEF were uniformly sampled in a compensated HF status and received recommended pharmacological therapy. Patients whose EF initially was <50% and was later improved were excluded from the study. HFpEF patients with comorbidities such as malignancies, neurological disorders, end-stage renal failure, liver cirrhosis, and systemic or infectious diseases were also excluded from the study. Medical history, physical, laboratory, echocardiography, and 24-h electrocardiography monitoring were recorded at baseline. Thirty-five volunteers who were aged and gender-matched with the clinical group were included as controls. All participants in the study gave informed consent prior to inclusion, and the study was approved by the two institutional Ethics Eommittees: the Faculty of Medicine, Nis, University Nis (Number 12-10580-2/3) and the Institute for Treatment and Rehabilitation Niska Banja (Number 03-4185/1).

The selected HFpEF patients (75) were denoted as a clinical group that was divided into two phenotypes (subgroups), according to the hierarchical clustering based on their biomarker profile. In fact, we applied hierarchical clustering to the concentrations of six biomarkers (sST2, Gal-3, GDF-15, Syn-1, BNP, and cystatin C), and the heat map of this analysis is presented as Figure 1. The first phenotype of HFpEF patients included those who had statistically higher concentrations of measured biomarkers, and this subgroup was classified as the “remodeled phenotype.” The second subgroup consisted of HFpEF patients who had significantly lower levels of biomarkers and they were identified as the “non-remodeled phenotype.” Finally, the differences between these phenotypes were analyzed.

All participants underwent two-dimensional echocardiography using a commercially available system (ACUSON Sequoia 256, New York) and were analyzed according to the current guidelines [20]. LVEF and LV volumes were assessed using the biplane method, and the dimensions of the LV, left atrium, and LVM were provided by M-mode imaging. Doppler‐derived mitral valve flow velocity waves (E-wave, A-wave, E/A ratio) were determined, where the E/A ratio as the ratio of the early (E) to late (A) ventricular filling velocities was regarded as a parameter of diastolic dysfunction; E/A ratio <0.75 or ≥1.5. Right heart structure and function were assessed by the dimensions of the right ventricle, systolic pulmonary artery pressure, and the tricuspid annular plane systolic excursion in an apical 4-chamber view. The maximum systolic excursion of the lateral tricuspid annulus was measured by M-mode, with a tricuspid annular plane systolic excursion of <17 mm indicating right ventricle dysfunction.

The 24 h electrocardiography monitoring was done using the Del Mar Medical Systems (Tennessee, USA). The electrodes were attached to the participants within 24 h of hospital admission, and after 24 h of wear, we performed manual verification of the following: atrial fibrillation (AF), mean heart rate (max and min), bradycardia and intermediate pauses, atrioventricular and sinoatrial blocks, supraventricular and ventricular arrhythmias, ST-depression, and heart rate variability. We considered intermediate pauses where the prolongation of the R-R interval was >2 s or >2.5 s.

Blood sampling was performed on admission, immediately after informed consent forms were signed, and all standard biochemical measurements were obtained using the Sysmex XS 1000 apparatus (Europe GmbH). Plasma samples were immediately stored and frozen at -80°C until all the samples were collected. Standard biochemical measurements and biomarkers were all analyzed from the same plasma sample. Plasma concentrations of evaluated biomarkers were determined using a colorimetric quantitative (sandwich) enzyme-linked immunoassay technique, suitable for cell culture supernatant, human plasma, and serum, according to the manufacturer’s instructions for each of them. Plasma concentrations of human sST2, galectin-3, and GDF-15 were determined using the Quantikine kits (R&D Systems, Inc. Minneapolis, MN, USA). Assay precision was measured by percent error or coefficient of variation (%CV). The human ST2/IL-33R kit had sensitivity of 13.5 pg/mL (assay range: 31.3–2,000 pg/mL) and %CV for intra-assay of 4.4 and inter-assay of 5.4. The Human Galectin-3 Kit had a sensitivity of 0.085 ng/mL (assay range: 0.3–10 ng/mL) and %CV for intra-assay of 3.7 and inter-assay of 6.2, while the human GDF-15 kit had sensitivity of 0.947 pg/mL (assay range: 23.4–1,500 pg/mL) and %CV for intra-assay of 6.7 and inter-assay of 8.4. Syndecan-1 plasma concentration was determined using a Human Syndecan-1 ELISA Kit (Abcam, Cambridge, UK). The test sensitivity is 4.94 ng/mL (range: 8 ng/mL–256 ng/mL), %CV: intra-assay is 6.2, and inter-assay is 10.2. Human BNP plasma concentration was determined using a human BNP ELISA Kit (Abcam, Cambridge, UK); the sensitivity of this kit was 14 pg/mL (range: 14 pg/mL–1,000 pg/mL), reported precision (%CV): intra-assay is <10 and inter-assay is < 12.

The data are presented as mean ± standard deviation or as frequency and percentages. Agglomerative hierarchical clustering (AHC), an unsupervised learning tool, was adopted to classify patients based on their plasma levels of evaluated biomarkers, sST2, Gal-3, GDF-15, Syn-1, BNP, and cystatin C. The AHC used as the possible number of clusters was not known to us a priori and was performed using the Ward method, squared Euclidian distance as appropriate for continuous data. Based on AHC, patients were allocated to two phenotypes. Differences in demographic, clinical, biochemical, echocardiographic, and 24-h ECG monitoring parameters between clusters were tested with the χ² test, Fisher’s test, t-test, and Mann-Whitney test. The level of significance was set at p < 0.05. All statistical analyses were performed using R software, version 3.0.3 (R Foundation for Statistical Computing, Vienna, Austria) [21].

Our study sample included 75 HFpEF patients for the final analysis, who were separated into two phenotypes according to the plasma concentrations of biomarkers (Table 1). The levels of sST2, Gal-3, GDF-15, Syn-1, and cystatin C were significantly different between phenotypes (p < 0.001 for all biomarkers) and were significantly higher in the “remodeled phenotype.” However, the concentration of BNP was higher in the first subgroup, but not statistically significant, so we did not use it for subsequent statistical analyses. Their clinical variables, laboratory, echocardiographic, and 24-h ECG parameters were compared. The “remodeled phenotype” was characterized by older age (65.88 ± 11.22) patients, predominantly belonging to NYHA class association I (75%) and II (25%). Concerning HF etiology, 75% of them had previous cardiomyopathy, 62.5% had a valvular disease, and 52.9% had coronary artery disease (CAD), whereas 70.6% of them reported previous acute myocardial infarction and, 52.9% were classified as STEMI and 17.6% as NSTEMI. Hypertension was documented in 87.5%, obesity in 37.5%, hypercholesterolemia in 100%, and diabetes mellitus in 37.5% of patients within the “remodeled phenotype.” The “non-remodeled phenotype” was composed predominantly of women (87.5%) who were younger compared to the first phenotype, 62.47 ± 8.33 with relatively preserved NYHA classification I (88.2%) and II (11.8%). The etiology of heart failure was as follows: valvular heart disease 29.5%, cardiomyopathy 70.6%, and CAD 37.5%, followed by previous myocardial infarction in 25%, including equally represented STEMI (12.5%) and NSTEMI (12.5%). Regarding clinical history, 88.2% of these patients presented with hypertension, 100% had hypercholesterolemia, 41.2% had diabetes mellitus, and 76.5% were obese. All demographic and clinical data are presented in Table 2. Table 3 summarizes baseline laboratory data, Table 4 summarizes baseline echocardiographic parameters, and Table 5 summarizes parameters of the 24-h ECG monitoring across the phenotypes for the overall study group.

Patients with the “remodeled phenotype” presented with a significantly higher prevalence of myocardial infarction (p = 0.047), in particular, STEMI (p = 0.045) (Table 2), anemia: lower number of red blood cells (p = 0.037), hemoglobin concentration (p = 0.034) and hematocrit (p = 0.046) (Table 3), and higher prevalence of AF (p = 0.047) (Table 5). Considering echocardiographic parameters that imply the extent of left ventricular remodeling, the “remodeled phenotype” presented with significantly increased left ventricular mass index (p < 0.001), LVM (p = 0.001), and increased diameters of the interventricular septum (p = 0.027) and the posterior wall (p = 0.003), compared to the “non-remodeled phenotype” (Table 4). Besides the prevalence of AF (p = 0.047), statistical significance was obtained in the prevalence of the verified intermediate pauses duration over 2.0 s (p < 0.006) since 50% of “remodeled phenotype” patients had pauses with duration >2 s. Both phenotypes had similar rates of pharmacological therapy (Table 1), and interventional strategies such as pacemaker implantation and coronary artery graft bypass grafting (data not shown).

In the present study, we identified two distinct biomarker phenotypes within HFpEF patients using hierarchical clustering analysis applied to various biomarkers; these two phenotypes differed in clinical, echocardiographic, and laboratory characteristics. The phenotype with significantly higher concentrations of biomarkers of myocardial remodeling and cystatin C had a higher rate of previous myocardial infarction, particularly STEMI, AF, and anemia. The phenotypes also presented significantly different echocardiographic parameters, which refer to cardiac structure and electrocardiographic parameters. However, no significant differences in applied therapy or medical interventions were observed between the phenotypes.

HFpEF integrates cardiometabolic and pro-inflammatory backgrounds, combining them with old age and numerous comorbidities [22]. HFpEF patients predominantly have a history of arterial hypertension, CAD, AF, diabetes mellitus, obesity, and renal impairment [3‒5]. None of these comorbidities are exclusive to HFpEF; therefore, more precise pathophysiological underpinning to an accurate biochemical profile needs to be established in order to find the most beneficial therapy for the specific phenotype of HFpEF. Unsupervised statistical methods may need to be performed to define distinct subtypes of HFpEF, using a set of different variables [5, 11, 14‒19].

One of the first cluster analyses in chronic HF patients introduced four distinct phenotypes that differed in various clinical parameters and evaluated biomarkers (sST2, NT-proBNP, and Gal-3), concluding that each cluster responded differently to therapy [18]. This study raised the possibility that a more appropriate phenomapping tool than disease severity for the subclassification of patients may be their underlying pathophysiology as suggested in our model. Furthermore, the distinctive phenotypes in HFpEF have been shown to hold significant differences in event-free survival and can be identified using clinical characteristics routinely obtained during the patient management. This implies that several factors in combination identify the subgroup with a specific clinical outcome as it is not possible to calculate the hazard ratio for all combinations of clinical characteristics [14]. For instance, the combination of diabetes mellitus, anemia, obesity, and renal dysfunction increases the severity of chronic inflammation and endothelial dysfunction, leading to cardiac hypertrophy and HF progression. Additionally, it has been demonstrated that despite great disease diversity, HFpEF can be subdivided into three exclusive phenotypes, demonstrating different pathophysiologies, risk factors, therapies, and outcomes [5]. Some authors give priority to biochemical over clinical profiling of HFpEF patients, considering that the patients may phenotypically appear to be the same, but may have different outcomes depending on the therapy employed [19]. A potential advantage of biomarker profiling is that it indicates the biological processes at a specific time point and may be reclassified into individual biological responses and that choosing a therapy based on biomarker profiling, rather than clinical variables, shows improved outcomes [19]. Recent research documented results on phenotyping, based on laboratory parameters, combined with clinical and echo variables, deriving six distinct phenogroups, which is in accordance with our results (the prevalence of CAD, AF, anemia, and cardiac structure and function within the obtained phenogroups) [11]. Most of these phenotypes (including AF and anemia) were predictors of the composite endpoint in the HFpEF population [23]. This similar result to ours showed that the most severe forms of HFpEF had the highest plasma sST2 concentrations [11], giving support to this protein in the course [11, 24] and phenotypization of HFpEF [11]. Recent research [15] also revealed three distinct phenotypes within HFpEF that, similarly to our results, differed in the prevalence of ischemic heart disease, AF, anemia, and some echocardiographic parameters, further suggesting that these variables may be included in criteria for the HFpEF categorization [11]. Moreover, two recent studies on clinical HFpEF phenomapping also concluded three mutually exclusive phenogroups [16, 17]. A phenogroup which presented with the highest prevalence of LV abnormalities and comorbidities had a significantly higher risk of all adverse clinical events [16], whereas a phenogroup that demonstrated more functional impairment, LV hypertrophy, higher rate of comorbidities, biomarkers of inflammation and remodeling, and response to spironolactone therapy [17] had the highest risk of adverse cardiovascular (CV) events.

All the aforementioned studies revealed that a total cohort of HFpEF patients may be subdivided and refined into particular phenotypes, based on clinical and/or biochemical data, presented with different risks of clinical endpoints. The results of our study contribute to that body of knowledge and point to a few important implications. First, biomarker profiling allowed us to form two clinically meaningful subgroups, where one phenotype demonstrated the ischemic etiology of the disease, resembling HF with a reduced EF. These results extend the importance of CAD management in HFpEF. As 35–60% of HFpEF patients present with CAD [5, 11, 15, 25], treatment should not be neglected, especially when a diagnosis of CAD could not be established by noninvasive measures, even in the presence of lesion, in one-third of HFpEF patients [26]. When present in HFpEF patients, CAD correlates with a rapid decline in LVEF, thus with mortality [5, 25].

It may be hypothesized that biomarkers from one pathophysiological domain, such as cardiac remodeling, rather than that of myocardial stretch (BNP) reflect a proper pathophysiological mechanism in the course of the disease, which makes this subclassification potentially valid for proper management of HFpEF patients. Plasma BNP concentration did not differ between the clusters in our model, even though its elevation is essential to establish a diagnosis of HFpEF [1, 2]. Accordingly, our results do not support BNP as a useful biomarker for phenotypization in HFpEF. This may be due to a small number of enrolled patients, or to intra-individual variations of BNP in cardiac patients, even in healthy individuals [27, 28]. Secretion of BNP is precisely regulated by specific pathophysiologic mechanisms, and very small fluctuations in neuro-hormonal and stress-related cytokines lead to variations in BNP levels [27‒30]. This knowledge provides a proof of concept for BNP to be an index of activation of the neuroendocrine system, rather than a marker of myocardial dysfunction [27]. For the most accurate analyses, it may prove valuable to include a cohort of healthy individuals as a control group, for the initial determination of plasma BNP levels [31], together with the concentration of biomarkers from different pathophysiological domains (sST2, Gal-3, GDF-15, Syn-1, and cystatin C). This approach may aid in establishing their relationship for the calculation of appropriate values, and in that manner, BNP may be taken into consideration for phenomapping analyses.

The most rational and useful clinical classification encounters four clinical phenotypes of HFpEF, aging, obesity, pulmonary hypertension, and CAD [32], making routine patient care somewhat easier. Considering that comorbidities (hypertension, obesity, and diabetes mellitus) occur more frequently in the elderly, for optimal phenotypization, additional biomarker data or parameters on cardiac structure and function are needed as suggested by our results. AF is also an expected comorbidity in HFpEF, especially in older hypertensive patients, with a prevalence of between 44 and 51% [14], which is consistent with our findings (62%), being a predictor of a poor outcome [1, 28]. Based on our results, the prevalence of AF was also significantly different between phenotypes. This emphasizes its role in HFpEF progression, suggesting that restoring sinus rhythm should be included in a timely manner within HFpEF treatment [5]. Our findings also revealed that patients in the first cluster had a statistically higher rate of intermediate pauses >2 s, whereas it is demonstrated that intermediate pauses of 2–3 s in length are associated with an increased risk of different adverse CV events [33].

We acknowledge a few important limitations of the study. Machine learning (including clustering) analyses usually require large data, preferably obtained from multiple centers; the small number of enrolled patients selected from a single center and the small number of phenotypic variables used for phenotype-based clustering most likely affect the validity of generalized results and should be regarded only as preliminary. Furthermore, the cross-sectional design of the study disables the possibility of drawing conclusions based on the follow-up of disease outcomes; therefore, these results must be validated in further prospective-designed studies. The lack of validation in an external cohort also limits the interpretation of findings. Still, a control group of healthy individuals presenting with BNP, and other evaluated biomarkers, values within the normal range, may allow different results in the clustering analysis and may be an additional task to be clarified. Moreover, according to the initial design of our study, we did not include the high-sensitivity cardiac troponins for phenomapping, but only biomarkers beyond routine biochemical analyses. Accordingly, it requires mentioning that recent evidence suggests their determination for the stratification of CV risk in the general population, being responsive to preventive pharmacological or lifestyle interventions, change in parallel to risk modifications [34, 35], and for the risk prediction of all-cause and CV mortality, CV events, and HF hospitalization [36]. Cardiac troponins, when added to well-established prognosticators, offer incremental risk prediction, affecting patient treatments and likely reverting the initial myocardial remodeling. Notwithstanding the limited experimental information, it is possible that phenomapping, including cardiac troponins and biomarkers analyzed in this study, may be beneficial for stratification of CV risk and strategies for prevention. This assumption requires further scrutiny, and thus the preliminary results presented here should be considered, mainly exploratory to create a better knowledge base and a framework for subsequent investigations. Finally, the presence of previous myocardial infarction was the only clinical parameter related to the CV system significantly different between the two clusters. Moreover, our analyses did not include the timing of myocardial infarction, and this may have affected the concentrations of evaluated biomarkers.

Despite these substantial limitations, we are optimistic that these results may lead to further research that will indicate new criteria for better HFpEF phenomapping based on selected biomarkers and underlying pathophysiologic mechanisms. Regardless of the small number of participants, we were still able to identify two meaningful phenotypes and draw relevant conclusions and are of the opinion that these could be applied to other populations within the same cohort. Machine learning algorithms and multi-marker panels offer engaging new perspectives for future research [8].

The present study demonstrated that using machine learning-based clustering applied to the plasma concentration of cardiac remodeling biomarkers (sST2, Gal-3, GDF-15, syndecan-1), BNP, and cystatin C, within the cohort of HFpEF, two distinct phenotypes may be identified, significantly different in some clinical characteristics and cardiac structure and function.

These results suggest that distinct pathogenetic mechanisms underlie each of the resulting phenotypes and this may be used for the most optimal pathophysiological subclassification of HFpEF patients. We anticipate that these efforts will enable new approaches to HFpEF phenomapping, enabling better prevention, treatment, and outcome.

The research was conducted ethically in accordance with the Declaration of Helsinki and was approved by the two institutional Ethics Committees: the Faculty of Medicine, Nis, University Nis (Number 12-10580-2/3) and the Institute for Treatment and Rehabilitation Niska Banja (Number 03-4185/1).

The authors have no conflicts of interest to declare.

This work was supported by a grant by the Faculty of Medicine, University of Nis, INT-MFN-46, 2020-2023, by a grant from the Ministry of Science and Technological Development, Project Number 451-03-47/2023-01/200113, and by the Project of the Serbian Academy of Science O-06.17.

Valentina Mitic and Dijana Stojanovic conceived the concept of the study, contributed to the design of the research, and drafted the manuscript. Valentina Mitic, Dejan Petrovic, and Marina Deljanin Ilic were involved in clinical work and data collection. All authors of the study analyzed, discussed, and interpreted results and searched the literature. Dijana Stojanovic, Aleksandra Ignjatovic, and Jelena Milenkovic completed statistical analyses. All authors edited and approved the final version of the manuscript critically for important intellectual content.

Raw data will be made available upon reasonable request.