Obese-Inflammatory Phenotypes in Heart Failure with Preserved Ejection Fraction

Abstract

BACKGROUND:

Comorbidity driven microvascular inflammation is posited as a unifying pathophysiologic mechanism for heart failure (HF) with preserved ejection fraction (HFpEF). Obesity is pro-inflammatory and common in HFpEF. We hypothesized that unique obesity-inflammation HFpEF phenotypes exist and are associated with differences in clinical features, fibrosis biomarkers and functional performance.

METHODS:

Patients (n=301) from three HFpEF clinical trials were studied. Unsupervised machine learning (hierarchical clustering, HC) with obese status and 13 inflammatory biomarkers as input variables was performed. Associations of clusters with HFpEF severity and fibrosis biomarkers (PIIINP, CITP, IGFBP-7, Gal-3) were assessed.

RESULTS:

HC revealed 3 phenotypes: Pan-Inflammatory (n=129; 64% obese), Non-Inflammatory (n=83; 55% obese) and Obese High CRP (n=89; 98% obese). The Pan-Inflammatory phenotype had more comorbidities and HF hospitalizations, higher left atrial volume, NT-proBNP and fibrosis biomarkers and lower glomerular filtration rate, peak VO₂, 6-min walk distance and active hours/day (p<0.05 for all). The Non-Inflammatory phenotype had most favorable values for all measures. The Obese High CRP phenotype resembled the Non-Inflammatory phenotype except for isolated elevation of CRP and lower functional performance. HC cluster assignment was independent of CRP genotype combinations that alter CRP levels, and more biologically plausible than other clustering approaches. Multiple traditional analytic techniques confirmed and extended the HC findings.

CONCLUSIONS:

Unique obesity-inflammation phenotypes exist in HFpEF and are associated with differences in comorbidity burden, HFpEF severity and fibrosis. These data support comorbidity driven microvascular inflammation as a pathophysiologic mechanism for many but not all HFpEF patients.

A novel unifying pathophysiologic heart failure (HF) with preserved ejection fraction (HFpEF) paradigm, the microvascular inflammation hypothesis (MIH)¹, posits that pro-inflammatory comorbidities result in low-grade systemic and coronary microvascular endothelial inflammation, myocardial inflammation, and subsequent microvascular dysfunction and cardiac fibrosis.^{1, 2} Microvascular inflammation produces oxidative stress, impairs nitric oxide (NO)-cGMP-protein kinase G (PKG) signaling and alters myofilament protein phosphorylation producing diastolic dysfunction independent of fibrosis.¹

Obesity is a potent HFpEF risk factor.³ Elevation of some inflammatory biomarkers is more common in obese persons and inflammation mediates obesity associated metabolic derangements.⁴ However, obesity may lead to exercise intolerance and HFpEF via other (non-inflammatory) mechanisms.³ Furthermore, differences in inflammation related to other comorbidities or environmental and hemodynamic stressors variably associated with obese status may result in phenotypic variability even amongst obese HFpEF patients.⁵

Shah et al used unsupervised machine leaning to identify three prognostically-unique phenotypes among patients with HFpEF,⁶ highlighting the potential value of machine learning to characterize HFpEF heterogeneity. Here, we studied patients enrolled in three Heart Failure Clinical Research Network (HFN) HFpEF trials. Thirteen inflammatory mediators, adhesion molecules, chemokines and angiogenic biomarkers implicated in the MIH were measured. We utilized hierarchical cluster (HC) machine learning algorithms and extensive standard statistical methods to identify and characterize pathophysiologic phenotypes using inflammatory mediator biomarkers and obese status as input variables. We hypothesized that unique obesity-inflammation phenotypes exist within HFpEF, are identified by different patterns of obese status and inflammatory mediators and are associated with differences in variables reflective of the upstream drivers and down-stream biological sequela of inflammation.

METHODS:

The data that support the findings of this study are available from the corresponding author upon reasonable and justified request.

Study Population

All patients in three HFN HFpEF clinical trials (RELAX, NEAT, INDIE) ^7-9 with adequate biorepository samples (n=301) comprised the study group. Entry criteria (Supplementary Table 1) and collected baseline phenotypic data (Supplemental Table 2) for the trials are provided. All data were de-identified and supplied by the Data Coordinating Center (Duke Clinical Research Institute, Duke University). The HFN Biomarker Committee and the Duke University institutional review board approved the study.

Inflammatory, pro-fibrotic and other biomarkers

The HFN Core Biomarker Laboratory (University of Vermont) performed all assays. Inflammatory mediators were measured using multi-plexed immunoassay panels (Meso Scale Discovery, Rockville, MD).

Input variable inflammatory mediator assays:

Inflammatory mediators associated with comorbidities common in HFpEF were assessed (interferon-gamma (INFγ), interleukin-1 beta (IL-1β), interleukin-6 (IL-6), tumor necrosis factor (TNF), interleukin-8 (IL-8), C-reactive protein (CRP) and serum amyloid A (SAA)).^10-14 Adhesion molecules (vascular cell adhesion molecule-1 (VCAM-1) and intercellular adhesion molecule-1 (ICAM-1)) and transforming growth factor beta-1 (TGF-β) were measured as they have been implicated in the MIH.¹ Monocyte chemoattractant protein-1 (MCP-1), a chemokine implicated in inflammation in atherosclerosis, aging and HF,^{15, 16} interleukin 16 (IL-16), a cytokine implicated in a study of human and rodent HFpEF,¹⁷ interleukin-10 (IL-10), an anti-inflammatory cytokine with evidence of protective effects in HFrEF¹⁸ and hypertension,¹² and vascular endothelial growth factor-A (VEGF-A), an angiogenic mediator whose down-regulation is believed to contribute to microvascular rarefaction in HFrEF,¹⁹ were measured.

Biomarkers compared across obesity-inflammation phenotypes:

Profibrotic biomarkers including procollagen III N-terminal peptide (PIIINTP), C-telopeptide for type 1 collagen (CITP), Galectin-3 (GAL-3) and insulin-like growth factor binding protein-7 (IGFBP-7), neurohumoral biomarkers (NT-proBNP, aldosterone and endothelin) and renal function biomarkers (creatinine and estimated glomerular filtration rate, eGFR) were measured as previously described.^{7, 8, 20-22}

CRP genotypes:

CRP levels are heritable and variants in the CRP gene are associated with differences in CRP levels but not cardiovascular risk.^{14, 23, 24} As previously reported,^{25, 26} the HFN Genomics Core laboratory (Université de Montréal Beaulieu-Saucier Pharmacogenomics Centre at the Montreal Heart Institue) has genotyped consenting patients using custom candidate gene and broad-based assays including the Illumina HumanOmni2.5-8 BeadChip and Illumina HumanExome v1.0 Beadchip (RELAX trial) and the Illumina Infinium Omni2-5Exome-8v1-3 BeadChip (INDIE and NEAT trials).

PyGenClean versions 1.2.5/1.8.3 and PLINK version 1.07 were used for the quality checks and genetic data cleanup process for RELAX and separately for INDIE and NEAT.²⁷ Three CRP polymorphisms (rs1205, rs1130864, rs3093077; 100% genotype completion rates) associated with plasma CRP levels were used to generate the nine most common genotype combinations and subsequently, these were compressed into three groups ranked according to incrementally higher CRP levels as previously described²⁴.

Statistical analysis

We performed HC analysis in JMP (14.1.0). Input variables included obese (BMI≥30) status and the 13 inflammatory mediators. Details regarding all analytic methods, the rationale for selecting HC clustering and clustering methods are provided in Supplemental Methods. We performed several sensitivity analyses. First, we averaged the inflammatory mediators for each patient and clustered by tertiles of the averaged value. Second, patients were clustered by obese status. Third, we provide results for 2 and 4 clusters within the HC algorithm. Fourth, we performed a multivariate normal mixture clustering model in R (mclust version 5.3) using Bayesian Information Criteria to select the optimal multivariate normal mixture model and cluster number.^{28, 29} We also confirmed the optimal cluster number using the NbClust program in R³⁰ via the majority rule of 23 separate indices of cluster cohesiveness and separation. A single measure of cluster cohesiveness and separateness (average silhouette width) was assessed for HC and Model based approaches. Fifth, multivariable standard least squares regression models using markers of fibrosis and clinical severity as dependent variables and inflammatory markers/obese status as independent variables were constructed. Lastly, multivariable logistic regression models were constructed using cluster as the dependent variable and inflammatory markers/obesity as independent variables.

Association between variables (CRP and IL-6) was assessed with standard least square regression and potential differences in associations between groups (≥ 3 clusters or genotype combinations) assessed using dummy variables.

RESULTS:

Patient Characteristics:

Overall, subjects had a median age of 69, were predominantly white (91%) and had advanced HFpEF with 65% satisfying clinical trial entry criteria associated with poor prognosis in HFpEF.^{31, 32} Participants had multiple comorbidities including hypertension (85%), hyperlipidemia (73%), obesity (median BMI 33.5) diabetes (38%), chronic obstructive lung disease (16%), previous smoking (58%), and anemia (33%) (Supplemental Table 3).

Obesity-Inflammation Clusters:

By HC, three clusters were identified (Figure 1). All inflammatory mediators except IL-10 and VEGF varied significantly across cluster. The largest (n=129) cluster, referred to as the Pan-Inflammatory phenotype, displayed the numerically highest levels of most inflammatory mediators (Figure 1B and Figure 2). Absolute levels are in Supplemental Table 4. The smallest cluster (n=83), termed the Non-Inflammatory phenotype, displayed the numerically lowest levels of most inflammatory markers. The remaining cluster (n=89), referred to as the Obese High CRP phenotype, had the numerically highest obesity prevalence. This cluster had CRP and SAA levels similar to the Pan-Inflammatory cluster but lower IL-6 levels. Other inflammatory mediators were in general similar to the Non-Inflammatory phenotype.

Figure 1:

(A) Constellation plot of study HFpEF cohort. The lengths of the lines in the constellation plot represent the distance between and within the clusters. The axis scaling, orientation of points, and angles of the lines are arbitrary and the length values are meaningful only with respect to each other. Colored nodes represent individual subjects. Red = Pan-Inflammatory cluster; Blue = Non-Inflammatory cluster; Green = Obese High CRP cluster (B) Phenotype Heatmap of study HFpEF cohort. Columns represent inflammatory markers, and rows represent individual subjects. Red = increased value of an inflammatory mediator; Blue = decreased value of an inflammatory mediator.

Figure 2:

Inflammatory mediators organized by cluster. Bar graph depicts the average standardized value for each inflammatory mediator ± SEM. Red = Pan-Inflammatory cluster; Blue = Non-Inflammatory cluster; Green = Obese High CRP Cluster.

As IL-6 is an important stimulus for hepatic CRP production, the relationship between CRP and IL-6 was examined overall (Supplemental Figure 1 A) and by cluster. CRP levels increased with increasing IL-6 (Model R² = 0.20, p<0.001). Cluster increased the Model R² from 0.20 to 0.32 (Supplemental Figure 1 B). The Obese High CRP group had the highest CRP levels for any given IL-6 level. CRP genotype combination was also associated with higher CRP levels for any IL-6 level but adjusting for genotype, cluster remained highly predictive of CRP levels (Supplemental Table 5). These data indicate that the cluster allocation in the machine learning algorithm was driven in part by differences in the relationship between IL-6 and CRP and independent of CRP genotype.

Clinical Characteristics according to Obesity-Inflammation Clusters:

There were statistically significant differences in age, sex and BMI but not race, blood pressure or heart rate across clusters (Table 1). While the total number of comorbidities varied by cluster, only anemia prevalence varied significantly across clusters. The Pan-Inflammatory cluster displayed numerically highest prevalence’s of male sex and anemia and number of comorbidities. The Obese High CRP cluster had lowest age and percentage of males and the highest BMI. The Non-Inflammatory cluster had the numerically lowest BMI. The cluster distribution did vary marginally (p=0.04) by trial but not to a clinically relevant degree (Supplemental Figure 2).

Table 1:

Clinical Characteristics of Study HFpEF Cohort

Clinical Characteristic	Pan-Inflammatory (n=129)	Non-Inflammatory (n=83)	Obese-High CRP (n=89)	p-value
Age, years	69 (65-78)	70 (63-76)	66 (59-72)	0.001
Male	55%	46%	37%	0.031
White	91%	90%	91%	0.23
BMI	32.9 (27.9-38.9)	30.7 (27.9-35.1)	37 (33.5-43.1)	<0.001
SBP, mmHg	124 (114-134)	127 (119-141)	128 (119-142)	0.44
Heart Rate, BPM	68 (62-78)	68 (60-76)	71 (63-83)	0.07
Atrial Fibrillation (n=300)	53%	43%	44%	0.21
COPD	19%	12%	16%	0.43
Diabetes	43%	29%	39%	0.10
Hyperlipidemia	76%	72%	69%	0.48
Hypertension	88%	83%	83%	0.56
Smoking (n=299)	63%	54%	53%	0.32
Anemia	41%	23%	31%	0.020
Sum Comorbidities (n=298)	4 ± 1	3 ± 1	3 ± 1	0.003

Data is presented as median (Interquartile range), percent, or mean ± standard deviation. Abbreviations: BMI body mass index; COPD, chronic obstructive pulmonary disease; SBP, systolic blood pressure.

Obesity-Inflammation Clusters: HF Severity and Fibrosis Biomarkers

Pan-Inflammatory Phenotype

Number of HF hospitalizations in the 12 months prior to enrollment, prevalence of loop diuretic use, edema and jugular vein distention, left atrial volume index (LAVI), NT-proBNP levels, endothelin levels and profibrotic biomarker levels (PIIINTP, CITP, GAL-3, IGFBP-7), eGFR and functional performance as assessed by peak VO₂, 6-minute walk distance and hours active per day all varied significantly by cluster (Figure 3 and Supplementary Table 6). Numerically, the Pan-Inflammatory Phenotype had more adverse values for all variables, particularly as compared to the Non-Inflammatory cluster. The Minnesota Living with HF Score, other relevant medication use, ejection fraction, e’, E/e’, left ventricular mass index, estimated right ventricular systolic pressure and aldosterone did not vary significantly by cluster.

Figure 3:

HFpEF severity and fibrosis makers organized by cluster. Bar graph depicts the standardized average value or frequency (bar graph insert) for each severity and fibrosis marker ± SEM. Red = Pan-Inflammatory cluster; Blue = Non-Inflammatory cluster; Green = Obese High CRP Cluster.

Non-Inflammatory and Obese High CRP Phenotypes

Relative to the Pan-Inflammatory cluster, the Non-Inflammatory and Obese High CRP clusters resembled each other with numerically lower values for HF hospitalizations, jugular venous pressure (JVP) elevation, LAVI, NT-proBNP and fibrosis biomarkers and with numerically better renal function (Figure 3 and Supplemental Table 6). However, some important differences between these two clusters were observed. The Non-Inflammatory cluster had the numerically highest peak VO₂, 6-minute walk distance and hours active per day. In contrast, the Obese High CRP cluster had poorer functional performance, with values similar to the Pan-Inflammatory cluster. Additionally, the Obese High CRP cluster had numerically higher frequencies of loop diuretic use and edema, similar to the Pan-Inflammatory cluster.

Number of Clusters Sensitivity Analysis

Specifying two clusters combined the Non-Inflammatory and Obese High CRP clusters together without changing the membership of the Pan-Inflammatory Cluster (Supplemental Figure 3). In the two cluster model, all inflammatory mediators except VEGF were still higher (Supplemental Figure 4) and variables reflective of HF severity, fibrosis and functional performance limitation were still more severe in the Pan-Inflammatory cluster (Supplemental Figure 5).

Use of four clusters split the Pan-Inflammatory cluster into two clusters (Pan-Inflammatory 1, n=94 and Pan-Inflammatory 2, n=35) while membership of the other clusters did not change (Supplemental Figure 3). The Pan-inflammatory 2 cluster was characterized by a higher rate of obesity (Supplemental Figure 1C) and by higher CRP and IL-6 (Supplemental Figures 1D and 6). The 4 cluster model better explained the relationship between CRP and IL 6 (higher R² value (0.40) than the 3 cluster model (0.32), Supplemental Table 4). This was not explained by differences in CRP genotype (Supplemental Table 4). Variables reflective of HF severity, fibrosis and functional performance limitation were generally less favorable in both the Pan-Inflammatory clusters. LAVI and NT-proBNP were less elevated in the smaller Pan-Inflammatory 2 cluster (Supplemental Figure 7) and likely reflect the higher body size with its implications for indexing LAV and for natriuretic peptide levels³³. This analysis supports three Clusters as the most parsimonious model in our data set.

Clustering by Inflammatory Mediator Values

We clustered patients by tertiles of the numerical average of all 13 inflammatory mediators for each patient. Accordingly, each individual inflammatory mediator value increased across the three tertiles (p<0.001 for all, data not shown). The number of comorbidities and HF hospitalizations, frequency of edema, JVP elevation and loop diuretic use, LAVI, NTproBNP and fibrosis biomarkers (except PIIINP), severity of renal dysfunction and severity of functional impairment (except 6-minute walk distance (6MWD), p=0.061) increased significantly (p<0.05) across the tertiles (Figure 4). The highest tertile was comprised primarily (80%) of members of the Pan-Inflammatory cluster while the lowest tertile was composed primarily (67%) of members of the Non-Inflammatory cluster (Supplemental Figure 8). These data confirm the association between inflammation as assessed by our inflammatory panel and HFpEF severity, fibrosis and functional limitation but did not discriminate the Obese High CRP group with its unique inflammatory, fibrosis, HF severity and functional performance profile.

Figure 4:

HFpEF severity and fibrosis makers organized by low, middle (mid), and high tertiles of averaged standardized inflammatory mediator values for each subject. Bar graph depicts the average value or frequency (bar graph insert) for each severity and fibrosis marker ± SEM.

Clustering by Obesity

Clustering by obese status, HFpEF patients with obesity had higher CRP, SAA and IL-6 levels, lower VCAM-1 and ICAM-1 levels, and no significant differences in other inflammatory mediators (Figure 5). Despite lower LAVI and NTproBNP, obese patients had higher frequencies of edema and loop diuretic use and worse peak VO₂ and 6-minute walk distance (Figure 6). However, other HFpEF severity indices and fibrotic biomarkers did not differ by obese status alone. The obese cluster was comprised of roughly equal numbers of members from all 3 HC groups (Supplemental Figure 9). These data illustrate that obese status did not identify patients with consistent activation of inflammatory mediators and fibrosis biomarkers and did not identify a group with uniformly adverse HF severity profile, as seen with unsupervised clustering by inflammatory mediators.

Figure 5:

Inflammatory mediators according to obese status. Bar graph depicts the average standardized value for each inflammatory mediator ± SEM.

Figure 6:

HFpEF severity and fibrosis makers organized by obese status. Bar graph depicts the standardized average value or frequency (bar graph insert) for each severity and fibrosis marker ± SEM.

Model based Clustering:

By Bayesian information criteria, mixture models with 2 clusters was deemed optimal consistently across multiple multivariate covariance structure assumptions. Furthermore, the majority rule convention in NbClust concurred with 2 clusters. The silhouette width for this solution was 0.17 compared to 0.12, 0.06 and 0.07 for the HC 2, 3, and 4 cluster solutions respectively. Despite the higher silhouette width, only 6 of the inflammatory mediators varied across the 2 clusters in the model based solution. This is in contrast to the 3 Cluster HC solution where 11 of the 13 inflammatory mediators varied significantly. More importantly, while all 15 of the “upstream drivers of and down-stream biological sequela of inflammation” variables varied significantly across clusters with the 3 Cluster HC solution (Figure 3), none of these varied between the 2 clusters in the model based solution (all p>0.05; median ANOVA p was 0.47, IQR 0.28-0.79). Thus, in this data set, the HC algorithm identified more biologically plausible clusters.

Regression Analysis:

Adjusting for all inflammatory mediators, obesity was associated with lower LAVI, NTproBNP and 6-minute walk distance and higher prevalence of edema, consistent with indexing to weight (LAVI) and the well-known impact of obesity on NTproBNP levels, volume status and exercise tolerance in HFpEF (Table 2).³ Of inflammatory mediators, VCAM-1, IL-6, TNFα, IL-8 and IL-16 each had independent and biologically plausible (higher levels with more adverse status) associations with several (between four and nine associations per mediator) of the 15 adverse clinical variables and pro-fibrotic biomarkers examined. The remaining inflammatory mediators had no (CRP and IL-10) or counter-intuitive (lower values with more adverse status) associations of borderline statistical significance with outcome variables, suggesting that they were not physiologically important contributors to HFpEF pathophysiology.

Table 2.

Regression analysis predicting outcome variables or cluster membership

	Obese	VCAM-1	IL 6	TNFα	IL 8	IL 16	ICAM-1	INFγ	VEGF	IL 1β	MCP-1	SAA	CRP	IL 10	Model R2	Model p
Loop Diuretic Use		1.8	5.6					1.6*							0.11	<0.001
Edema	3.5	4.7	1.7												0.13	<0.001
JVP Elevation		4.1								1.9					0.06	<0.001
Sum CM				5.3				1.9*							0.07	<0.001
HF Hsp					1.6							1.7			0.03	0.01
LAVI	2.9*	2.7	1.6		1.5		1.8*								0.19	<0.001
NTproBNP	3.0*	4.5		1.7		3.1		1.5*							0.22	<0.001
CITP			2.6	9.4	1.6			1.6*	1.6*		1.4*				0.38	<0.001
PIIINP		6.3			3.4					1.4*					0.23	<0.001
IGFBP7		4.9		4.5	1.5			1.5*			1.5*				0.38	<0.001
GAL3				3.7		1.7									0.15	<0.001
eGFR		3.1*		8.3*		6.2*	4.7		1.8						0.29	<0.001
Peak VO2			3.0*			1.9*									0.08	<0.001
6MWD	1.4*	1.4*	2.1*				2.4		1.8			2.1*			0.18	<0.001
HAPD			2.8*					1.9							0.10	0.001
Models predicting cluster membership
Pan-Inflam	4.5*	2.2	8.6	1.9	8.9	3.8		3.2	1.5*	5.8	1.6				0.74	<0.001
Non-Inflam	2.2*		5.9*		4.8*		4.3*		4.7	1.6*	1.4*	8.6*	3.3*	1.8*	0.71	<0.001
Obese CRP	10.6			4.1*		4.3		4.1*		1.8*		2.6	3.1		0.44	<0.001

^(*)LogWorth values (−log10(p-value)) for the effect of final variables on the dependent variable (1st Column) are shown. Inverse associations are indicated with an asterisk. A LogWorth value that exceeds 2 is significant at the 0.01 level. Abbreviations: CITP, C-terminal telopeptide of collagen I; COPD, chronic obstructive pulmonary disease; CM, comorbidities; CRP. C-reactive protein eGFR, estimated glomerular filtration rate; GAL3, galectin-3; HF, heart failure; Hsp, hospitalizations; HAPD, hours active per day; IGFBP7, insulin like growth factor binding protein 7; JVP, jugular venous pressure; LAVI, left atrial volume index; Non-Inflam, Non-Inflammatory Cluster; NNT-proBNP, N-terminal pro b-type natriuretic peptide; PIIINP, N-terminal propeptide of procollagen III; Pan-Inflam; Pan-Inflammatory Cluster; RVSP, right ventricular systolic pressure; 6MWD, six minute walk distance; SBP, systolic blood pressure; VO₂, oxygen consumption.

In logistic regression analysis, Pan-Inflammatory cluster assignment was driven by obese status (lower prevalence of obesity) and those inflammatory mediators most associated with adverse outcome variable profiles in multivariable regression (higher levels of VCAM-1, IL-6, TNFα, IL-8 and IL-16) but also by higher values of INFγ, IL1β and MCP 1 and lower levels of VEGF. These additional mediators likely contribute to the association of Pan-Inflammatory cluster assignment with more adverse outcome variables and suggest that INFγ, IL1β, and MCP 1 may also contribute to pathophysiology in some HFpEF patients.

The lower VEGF levels in the Pan-Inflammatory cluster and the higher VEGF levels with lower levels of inflammatory mediators in the Non-Inflammatory cluster supports the hypothesized link between inflammation and microvascular rarefaction.²

DISCUSSION:

We used unsupervised machine learning to determine if unique obese-inflammatory phenotypes exist in HFpEF patients. We identified 3 obese-inflammatory HFpEF phenotypes. Patients with a “Pan-Inflammatory” phenotype exhibited the highest circulating levels of inflammatory mediators, more comorbidity, more HF hospitalizations, higher LAVI and NT-proBNP, worse renal function, highest levels of fibrotic biomarkers and the lowest functional capacity. A “Non-Inflammatory” phenotype had lowest levels of inflammatory mediators and most favorable levels of all other variables. An “Obese High CRP” phenotype resembled the Non-Inflammatory phenotype except for isolated elevation in CRP and impaired functional capacity.

Standard statistical methods of clustering by tertiles of the averaged inflammatory mediator values confirmed the associations between elevated inflammatory mediators and HFpEF severity, pro-fibrotic biomarkers and functional limitation. Clustering by obese status suggested that obesity is not universally associated with inflammation, clinical HFpEF severity indices or a pro-fibrotic state. Cluster membership was independent of CRP genotype variants known to affect CRP levels. The HC machine learning algorithm identified more biologically plausible clusters than other machine learning methods and was driven in part by the relationship between IL-6 and CRP. Standard multivariable regression models confirmed the association of some inflammatory mediators with some indices reflective of clinical HFpEF severity or pro-fibrotic state while cluster was associated with all indices reflective of clinical HFpEF severity or pro-fibrotic state. These data support comorbidity driven microvascular inflammation as a pathophysiologic mechanism for some but not all HFpEF patients and demonstrate complementary roles of machine learning and standard analytic techniques in studying HFpEF pathophysiology.

Inflammation in HFpEF

The MIH proposes that comorbidities and environmental or hemodynamic stressors result in low-grade systemic microvascular endothelial inflammation through circulating signaling molecules, which trigger endothelial cells to produce reactive oxygen species (ROS) and express adhesion molecules such as VCAM-1 and ICAM-1. These ROS in turn reduce the NO content which lowers soluble guanylate cyclase activity, alters myofilament protein phosphorylation and promotes myocyte diastolic dysfunction. Adhesion molecules promote inflammatory cell microvascular adhesion and myocardial migration, inflammation, release of TGF-β and cardiac fibrosis in some patients.¹

Limited studies have reported on a few “upstream” inflammatory markers implicated in the MIH in patients with or at risk for HFPEF such as CRP.³⁴ In a cross-sectional study, HFpEF patients were found to have increased IL-6, IL-8, and MCP-1 when compared to asymptomatic hypertensive patients.³⁵ In a large combined population cohort study, accounting for clinical characteristics, NT-proBNP and urinary albumin to creatinine ratio were predictive of future HFpEF while CRP, Gal-3, soluble ST2, and fibrinogen were not.³⁶

Studying one or two candidate inflammatory mediators associated with any single comorbidity may not be reflective of the cumulative impact of different factors that cause the “comorbidity-driven systemic inflammation” in the MIH. One proteomics study suggested the biological processes associated with HFpEF include cell adhesion and leukocyte migration, inflammatory response, integrin mediated signaling pathways and extracellular matrix organization.³⁷ However, no previous study assessed inflammatory phenotypic variation within HFpEF. In the current study, we identified three obese-inflammatory HFpEF phenotypes that were uniquely associated with differences in HFpEF severity, pro-fibrotic state and functional capacity.

Inflammation Biomarkers in HFpEF: Potential for Precision Medicine

To date, several multicenter randomized controlled trials targeting the impairment in NO-cGMP signaling postulated to result from inflammation in HFpEF patients have not shown benefit.^{7-9, 38} Our data suggest that phenotyping is needed to identify HFpEF sub-groups with inflammation that would be positioned to respond favorably to interventions based on the MIH. Data extrapolated from patients with HFrEF and coronary artery disease provide further support for this concept.

In a post-hoc analysis of a neutral trial of statins in HF with reduced EF (HFrEF), patients with high CRP benefited from statins; suggesting that HFrEF patients with inflammation derived a benefit from anti-inflammatory effects of statins.³⁹ CRP is produced by the liver in response to IL-1 activation via IL-6. In patients with previous MI and elevated hs-CRP, canakinumab (an IL-1β inhibitor) decreased adverse vascular⁴⁰ and HF events.⁴¹ However, only 55% of canakinumab treated patients achieved an on-therapy CRP level of < 2 mg/L and the benefit of canakinumab on vascular or HF events was confined to this subgroup.^{41, 42} These data suggest that not all CRP elevations necessarily reflect IL-1/IL-6 activation. The Obese High CRP phenotype is of particular relevance here because in this cluster, CRP elevation was out of proportion to IL-6, not related to the CRP genotypes we measured and not associated with other inflammatory mediators, fibrosis or other adverse clinical features except exercise intolerance. The exercise intolerance in some obese patients may be mediated by non-inflammatory mechanisms³ (see below). In contrast, in the four cluster model sensitivity analysis, a subset of the Pan-Inflammatory cluster (Pan-Inflammatory Cluster 2) had the highest CRP and IL-6 levels and worse inflammatory mediators, fibrosis and adverse clinical features. Thus, while the Obese High CRP and Pan-Inflammatory 2 cluster were both nearly all obese and had high CRP, they had very different biological features with differing IL-6 levels. Collectively, these data suggest that not all CRP elevation is related to IL-1/IL-6 mediated inflammation, that better biomarkers of inflammation are needed to guide therapy and that the machine learning algorithm used here was driven in large part by differences in the relationship between IL-6 and CRP, a feature than was associated with differences in HFpEF severity.

The Obese High CRP phenotype also illustrates the potential for additional pathophysiologic phenotypic variability beyond inflammation which can affect diastolic function. Obokata et al demonstrated that obese HFpEF patients had worse exercise performance and greater estimated plasma volume, epicardial fat, total heart volume and exercise induced elevation in filling pressures due to pericardial constraint than non-obese HFpEF patients.³ This is consistent with our study, where the Obese High CRP phenotype had more edema and loop diuretic use and lower functional performance than the Non-Inflammatory phenotype, despite similar inflammatory mediators, HF severity indices and fibrosis biomarkers.

Unsupervised cluster analysis in HFpEF

Machine learning has great potential to advance the diagnosis, risk stratification and therapy of cardiovascular diseases. Important potential limitations of machine learning include the concept of “un-observability” wherein the input variables must contain the critical information needed to characterize a disease process from a pathophysiologic, prognostic or therapeutic perspective; the persistence of causal inference limitations and the ability to act on the data stemming from the analysis.⁴³

Shah et al used unsupervised hierarchical and penalized model-based cluster analyses in a large (n≈400) HFpEF cohort to define three prognostically distinct groups.⁶ The input variables for this analysis were 47 clinical variables, basic laboratory, ECG and echocardiographic parameters. This novel study introduced the concept of unsupervised machine learning to define HFpEF phenotypes. However, the clustering input variables did not provide the opportunity to explore pathophysiologic phenotypes and the cluster assignment did not add greatly to prognostic information provided by an existing risk score defined using similar input variables and traditional statistical methods.⁶ As emphasized by the authors, the input variables included known prognostic variables and it is unclear whether the resulting clusters represented unique phenotypes or different stages of HFpEF. In our cohort, despite better cohesiveness and separation, the model based cluster method did not identify clusters which differed in the upstream mediators and downstream sequela of inflammation. The HC method did and was consistent with traditional statistical methods where tertiles of the averaged inflammatory mediators were associated with more severe upstream mediators and downstream sequela of inflammation. Others used latent class analysis of simple clinical risk factors and found clusters that were associated with prognosis and differences in some inflammatory and pro-fibrotic biomarkers.^{44, 45}

To our knowledge, our study is the first to use unsupervised machine learning to probe unique pathophysiologic inflammation phenotypes in HFpEF. We used input variables postulated to reflect the diverse, chronic low grade inflammation resulting from pro-inflammatory comorbidities addressing the “un-observability” limitation inherent in using only clinical data to assess potential HFpEF pathophysiologic phenotypes with machine learning. Our results are still subject to “causal inference limitations” as we cannot prove that the associations between the obesity-inflammation clusters and HF severity, fibrosis and functional performance are causal. Further studies of inflammatory mediator profiles across the natural history (ie, AHA/ACC stages) of HFpEF would be helpful in this regard as inflammation severity should increase across the stages of HFpEF according to the MIH.

Traditional statistics vs machine learning to understand HFpEF pathophysiology

It remains unclear whether unsupervised machine learning can reveal pathophysiologic insights that traditional analytic techniques cannot. Here, cluster was more strongly associated with a larger number of adverse clinical variables and pro-fibrotic biomarkers than tertiles of an average of inflammatory mediator levels, clustering by obese status or multivariable regression analysis models. We probed the mechanism of the superior performance of cluster using standard regression analysis and found that cluster assignment was influenced not only by the five mediators identified as predictive by multivariable regression analysis but also by multiple other mediators and their interplay with obese status. We conclude that the machine learning algorithm provided additional pathophysiologic information than standard regression analysis alone but that post-hoc traditional statistical analyses may complement the machine learning algorithm and assist in better understanding the machine learning output.

Limitations

The aim of this study was not to identify a new HFpEF clinical classification, but rather to use novel and standard analytic techniques to determine if unique inflammation patterns exist in HFpEF and are associated with clinical severity or pro-fibrotic state. A validation cohort was not analyzed due to our study’s limited size. Future studies of these and other inflammatory mediators using novel and standard analytic techniques in different stages of HFpEF are needed to further define variations in inflammation in HFpEF. Our clinical profiling did not include imaging for fibrosis and not all variables were collected in all three trials. There is biological relatedness in our cluster variables (inflammatory mediators) which were chosen based on previously described associations with HFpEF or its risk factors. The relatedness and overlap amongst the cluster variables may contribute to the unimpressive silhouette values.³⁰

CONCLUSIONS:

Using unsupervised machine learning, we identified three unique obesity-inflammation HFpEF phenotypes with significant differences in comorbidity burden, HFpEF severity, cardiac fibrosis biomarkers and functional performance. Traditional statistical techniques confirmed and extended the machine learning approach and together, these data support comorbidity driven microvascular inflammation as a pathophysiologic mechanism for many but not all HFpEF patients.

What is new?

• A “pan-inflammatory” phenotype exhibited the highest circulating levels of inflammatory mediators and fibrosis biomarkers, lowest angiogenic biomarker (VEGF) levels and worst HFpEF severity variables.

• A “non-inflammatory” phenotype had the lowest levels of inflammatory mediators, lower fibrosis biomarkers, higher VEGF levels and most favorable levels of HF severity variables.

• An “obese high CRP” phenotype resembled the Non-Inflammatory phenotype except for an isolated elevation in CRP and poorer functional status.

• Our findings show that inflammation is present in some but not all HFpEF patients and is associated with impaired microvascular status, a pro-fibrotic state and worse clinical status.

What are the clinical implications?

• A precision approach is needed to identify those patients with inflammation related HFpEF for targeted anti-inflammatory therapy.

Non-standard Abbreviations and Acronyms

6MWD

6-minute walk distance

CITP

C-telopeptide for type I collagen

CRP

C-reactive protein

GAL

Galectin

HC

Hierarchical Cluster

HF

Heart failure

HFN

Heart Failure Network

HFpEF

Heart failure with preserved ejection fraction

IGFBP

Insulin-like growth factor binding protein

JVP

Jugular venous pressure

LAVI

Left atrial volume index

MCP

Monocyte chemoattractant protein

MIH

Microvascular inflammatory hypothesis

NT-proBNP

N-terminal-pro B-type natriuretic peptide

PIIINP

Procollagen III N-terminal peptide

PKG

Protein Kinase G

SAA

Serum Amyloid A