Association of Biomarker Clusters with Cardiac Phenotypes and Mortality in Patients with HIV Infection

Abstract

BACKGROUND

Although individual cardiac biomarkers are associated with heart failure risk and all-cause mortality in HIV-infected individuals, their combined utility for prediction has not been well studied.

METHODS AND RESULTS

Unsupervised k-means cluster analysis was performed blinded to the study outcomes in 332 HIV-infected persons on 8 biomarkers: ST2, NT-proBNP, hsCRP, GDF-15, Cystatin C, IL-6, D-dimer, and Troponin. We evaluated cross-sectional associations of each cluster with diastolic dysfunction (DD), pulmonary hypertension (PH, defined as echocardiographic PASP ≥35mmHg), and longitudinal associations with all-cause mortality. The biomarker-derived clusters partitioned subjects into 3 groups. Cluster 3 (N=103) had higher levels of CRP, IL-6, and D-dimer (“inflammatory phenotype”). Cluster 2 (N=86) displayed elevated levels of ST2, NT-proBNP, and GDF-15 (“cardiac phenotype”). Cluster 1 (N=143) had lower levels of both phenotype-associated biomarkers. After multivariable adjustment for traditional and HIV-related risk factors, cluster 3 was associated with a 51% increased risk of DD (95%CI: 1.12-2.02) and cluster 2 was associated with a 67% increased risk of PH (95%CI: 1.04-2.68), relative to cluster 1. Over a median 6.9 years of follow-up, 48 deaths occurred. Cluster 3 was independently associated with a 3.3-fold higher risk of mortality relative to cluster 1 (95%CI: 1.3-8.1), and cluster 2 had a 3.1-fold increased risk (95%CI: 1.1-8.4), even after controlling for DD, PH, left ventricular mass, and ejection fraction.

CONCLUSIONS

Serum biomarkers can be used to classify HIV-infected individuals into separate clusters for differentiating cardiopulmonary structural and functional abnormalities, and can predict mortality independent of these structural and functional measures.

INTRODUCTION

Highly active antiretroviral therapy has dramatically reduced morbidity and mortality in persons with HIV infection, and yet HIV disease remains associated with early onset of cardiovascular disease (CVD). CVD is a leading cause of death among HIV-infected persons, and similar to the uninfected population, heart failure (HF) is a growing source of morbidity. The reported prevalence of diastolic dysfunction ranges from 26-50% in contemporary HIV-infected cohorts¹. Although symptomatic pulmonary hypertension is rare in HIV infection, our group and others^{2, 3} have reported elevated pulmonary artery systolic pressures in HIV-infected persons. Diastolic dysfunction (DD) is a risk factor for progression to HF and is also a significant cause of pulmonary hypertension (PH)⁴. Both DD and PH are strongly associated with all-cause mortality among individuals without HIV^{5, 6}.

We recently reported² that biomarkers of inflammation, thrombosis, apoptosis, and myocardial injury are elevated in HIV-infected persons relative to uninfected controls. In addition, we found that several of these markers were individually associated with both cardiac dysfunction and all-cause mortality, independent of traditional CVD risk factors and HIV-related factors. However, HF is a heterogeneous condition, especially in the setting of HIV infection⁷, where the epidemiology has changed over time with the use of effective combination antiretroviral therapy, and the pathophysiology appears to be multifactorial.

The heterogeneity of the HF syndrome suggests that multiple cardiac biomarkers reflecting different pathophysiologic pathways and consequences may be needed to represent the spectrum of the disease⁸. How such biomarkers inter-relate⁹, and how to integrate them in risk prediction and patient management remains unclear. In other settings such as cancer and rhinosinusitis, cluster analysis has been used to identify disease subtypes and to predict treatment outcomes^{10, 11}. Classifying patients beyond traditional risk factors using a biomarker-generated “phenotype” could improve diagnosis and determination of prognosis and could inform therapeutic strategies¹².

In this study we utilized an unsupervised cluster approach (i.e., blinded to the outcome variable) to categorize participants into separate subgroups based on the results of multiple serum biomarkers to obtain maximum prognostic utility. We hypothesized that in combination, a complementary, parsimonious set of biomarkers would identify separate clusters of risk for diastolic dysfunction (DD) and pulmonary hypertension (PH), while improving mortality risk prediction beyond that ascertained by traditional CVD risk factors and HIV-related risk factors.

METHODS

Study Population

Individuals with HIV infection were consecutively enrolled between September 2004 and March 2011 from the Study of the Consequences of the Protease Inhibitor Era (SCOPE), a large clinic-based cohort at San Francisco General Hospital. All participants of SCOPE were documented to be HIV-infected by either medical records, letter of diagnosis, or HIV-antibody testing. The only inclusion criterion for this analysis was HIV infection, and there were no exclusion criteria. Individuals were not preselected on the basis of cardiovascular risk factors, symptoms, or anti-retroviral drug regimens. The present analysis includes: 1) 71 untreated patients, defined as no ART in the preceding 6 months; 2) 83 treated patients with detectable viremia, as defined as >24 weeks of ART with the most recent or previous HIV RNA level >75 copies/ml; and 3) 178 treated patients who achieved full viral suppression, as defined as >24 weeks of ART with 2 most recent HIV RNA levels <75 copies/ml. The study was approved by the University of California, San Francisco Committee on Human Research, and all study participants provided written informed consent prior to study enrollment. The data, analytic methods, and study materials will not be made available to other researchers for purposes of reproducing the results or replicating the procedure.

Serum Biomarkers

Biomarkers measured in this study at baseline included: ST2 (fibrosis); GDF-15 (apoptosis); NT-proBNP (myocyte stretch); cTnI (myocardial injury); hsCRP and IL-6 (inflammation); Cystatin C (renal dysfunction); and D-dimer (thrombosis).^13-18 GDF-15, Cystatin C, and IL-6 were measured using a Quantikine Human Immunoassay (R&D systems, Minneapolis, MN). ST2 was measured using the Presage Assay (Critical Diagnostics, San Diego, CA). CRP was measured using the CardioPhase High Sensitivity C-Reactive Protein Immunoassay (Siemens Medical Solutions Diagnostics, Tarrytown, NY). Troponin was measured using the Advia Centaur TnI-Ultra Assay (Siemens Medical Solutions Diagnostics, Tarrytown, NY). NT-proBNP was measured by the Roche E Modular assay (Roche Diagnostics Corporation, Indianapolis, IN). D-dimer was measured using the Zymutest D-Dimer ELISA (Aniara, West Chester, Ohio). Details regarding sensitivity, range, and coefficients of variation have been published previously².

Echocardiography

As described previously¹⁹, a 2D transthoracic echocardiogram was performed on each participant within 6 months of enrollment by a sonographer blinded to participant’s HIV status. The presence of diastolic dysfunction (DD) was determined using the guidelines from the American Society of Echocardiography.²⁰ Left ventricular (LV) end-diastolic and end-systolic volumes (LVEDV and LVESV respectively) and LV ejection fraction (LVEF) were assessed using the modified Simpson’s rule and indexed to body surface area. Pulmonary artery systolic pressure (PASP) was quantified by using the modified Bernoulli equation to obtain the calculated pressure gradient (based on peak tricuspid regurgitation jet velocity) and then added to the mean right atrial pressure, which was estimated from the diameter of the inferior vena cava, and degree of inspiratory collapse.²¹ All calculations and interpretations were performed off-line by 2 cardiologists who were blinded to participants’ HIV infection and clinical status.

Outcomes

All patients were followed longitudinally after enrollment as part of the SCOPE study. The main echocardiographic outcomes (measured at baseline only) were DD (defined as ≥ grade 1), pulmonary hypertension (PH, classified as PASP >35mmHg)²¹ and LV hypertrophy (LVH, defined as LV mass ≥110 or ≥125 g/m² in females and males, respectively). DD was classified as grade 1 (n=128), grade 2 (n=5), grade 3 (n=3), grade 5 (n=2), or none (n=166), and was missing in 28 participants. Systolic dysfunction, defined as a LVEF of <50%, was not considered a primary outcome due to the low prevalence in the HIV population²² and in our cohort (14 out of 332 participants). PASP measures were available in all but 2 participants. We did not assess for presence of atrial fibrillation, but previous studies suggest that AF remains uncommon (<3% incidence) in the setting of HIV infection²³. For all-cause mortality, participants were followed through January 2015 or until the time of death as determined by the National Death Index.

Construction of clusters

The goal of our clustering procedure was to simplify the data from 8 distinct serum biomarkers to partition subjects into a small number of groups based on the totality of biomarker information. This grouping was based solely on the aggregate biomarker data and did not utilize clinical characteristics or subsequent outcomes. We first examined unadjusted Spearman correlations between markers. We used k-means clustering (chosen for its relative efficiency) to perform unsupervised clustering of biomarkers²⁴. K-means is considered a “flat” clustering algorithm, because it generates clusters without any prespecified ordering or structure, unlike hierarchical clustering methods. Groups of participants having similar biomarker patterns can be identified as clusters.

Because the biomarkers were right-skewed, we log-transformed each marker to normalize its distribution. We also standardized each biomarker to the same scale (mean=0, SD=1) so that biomarkers with larger variances would not have a greater influence on cluster assignment. We used k-means clustering to partition subjects into separate clusters, using the SAS FASTCLUS procedure to identify outliers and reduce their effect on cluster centers. We identified a parsimonious set of biomarkers using the six most distinguishing biomarkers to construct clusters. Markers used in cluster construction were those with the highest ratio of between- to within-group variance in biomarker levels. We used canonical correlation analysis to construct integrated biomarker scores, which represent weighted sums of the 6 biomarkers (standardized to the same scale). Biomarkers with larger coefficients contribute more weight to the score.

In addition to using the k-means algorithm to generate clusters, we compared other clustering algorithms including hierarchical (using Ward’s method), Single, Complete, Average, McQuitty, Median, and Centroid (Supplemental Figure 1)²⁵. We also performed internal validation to assess the quality and stability of clusters generated using k-means and other algorithms (using the clValid package in R, Supplemental Figures 2-3). Additional details and results are summarized in the Supplemental File.

Statistical analysis

We compared baseline clinical and demographic characteristics across clusters using chi-square and Kruskal-Wallis tests for categorical and continuous variables, respectively. Multinomial logistic regression was used to identify factors associated with cluster membership, using cluster 1 as the reference group. Relative risk regression (using a modified Poisson approach) was used to examine associations of clusters with DD, PH, and LVH, and Cox proportional hazards regression was used to examine associations with all-cause mortality. We assessed the proportional hazards assumption for mortality by testing weighted Schoenfeld residuals. Covariates from the baseline exam included demographics (age, gender, race/ethnicity), CVD risk factors (smoking, hypertension, diabetes, BMI, HDL, LDL, triglycerides), and HIV-related risk factors (CD4 count, HIVRNA, hepatitis C, history of opportunistic infection and HAART use). Additionally, we included baseline echocardiographic measures as covariates in models of mortality. We calculated risk ratios for each outcome using cluster 1 as the reference category.

We assessed model performance using discrimination, calibration, Nagelkerke’s R² (for overall performance), and the category-less IDI (integrated discrimination improvement, for reclassification)²⁶. For survival models, Harrell’s c was used to assess discrimination²⁷. Results are summarized in Results are summarized in the Supplemental File (Supplemental Figure 4).

Statistical analyses for comparisons of clinical data and associations with outcomes were conducted using SAS version 9.4 (SAS Institute Inc., Cary, NC, USA).

RESULTS

Construction of Biomarker Clusters

In unadjusted analysis, the serum biomarkers were only modestly inter-correlated (all correlation coefficients <0.4). A heat map and dendogram showing biomarker relationships found separate “cardiac” and “inflammatory” domains (Figure 1). The strongest correlation was between CRP and IL-6 (r=0.35), and the weakest correlations were between Troponin and the other biomarkers (all r≤0.11). Clusters were derived by an unsupervised cluster approach using the 6 most distinguishing biomarkers (ST2, NT-proBNP, GDF-15, CRP, IL-6, and D-dimer), blinded to study outcomes. The two excluded markers (troponin and cystatin C) contributed very little to cluster membership (Supplemental Figure 5). Cluster membership was similar when all 8 markers versus the 6 most distinguishing were used.

Figure 1.

Correlations of biomarkers among HIV-infected participants. Spearman correlation coefficients are shown for intracorrelations of biomarkers. Heat map with overlaid dendrogram depicts separation of “cardiac” and “inflammatory” biomarkers.

Increased cardiac markers (ST2, NT-proBNP, and GDF-15) were most able to distinguish cluster 2 from cluster 1, while increased inflammatory markers (CRP, IL-6, and D-Dimer) best distinguished cluster 3 from cluster 1 (Supplemental Figure 5 and Table 1). In addition, both clusters 2 and 3 had worse eGFR by cystatin C compared with cluster 1, while cluster 2 also had worse eGFR by creatinine compared with clusters 1 and 3. Rates of detectable troponin were similar across the clusters (27% overall).

Weighted biomarker scores generated using canonical correlation analysis showed that the first score “I” was driven by inflammatory markers (CRP, IL-6, and D-Dimer) and the second score “C” by heart failure markers (ST2, NT-proBNP, and GDF-15). A comparison of scores “I” and “C” showed good differentiation between the three k-means derived clusters (Figure 2): cluster 1 was characterized by low levels of all markers, cluster 2 had elevated cardiac markers (score C) but not inflammatory markers, and cluster 3 had elevated inflammatory markers (score I), with some having increases in cardiac markers.

Figure 2.

Association of integrated biomarker scores with cluster membership. Note: scatterplot shows separation of clusters by integrated biomarker scores

Standardized canonical coefficients were used to calculate integrated biomarker scores as follows:

Score I = −0.08xST2 − 0.40xBNP − 0.27xGDF15 + 0.81xCRP + 0.61xIL6 + 0.71xDdimer

Score C = +0.53xST2 + 0.59xBNP + 0.80xGDF15 + 0.05xCRP + 0.03xIL6 + 0.12xDdimer.

Each biomarker is log-transformed and standardized to a mean of zero and SD of one.

Cohort characteristics

Demographic and baseline clinical characteristics stratified by biomarker-derived cluster are shown in Table 1. The overall median age was 49 years, 19% were female, 32% were African-American, 56% were Caucasian, and 12% were of other race or ethnicity. In multinomial logistic regression analysis, factors independently associated with cluster 2 (“Cardiac”) membership included Caucasian race, hypertension, lower LDL, lower nadir CD4 count, and HCV coinfection (Table 2). Factors independently associated with cluster 3 (“Inflammatory”) membership included older age, female gender, higher BMI, and lower nadir CD4 count.

Table 1.

Baseline and demographic characteristics, stratified by biomarker-derived phenotype^*

Parameter	Cluster 1	Cluster 2 “Cardiac Phenotype”	Cluster 3 “Inflammatory Phenotype”	P-value
	N=143	N=86	N=103
Demographic and clinical characteristics:
Age (y)	47 (41-54)	49 (43-54)	50 (43-54)	0.096
Female	17 (12%)	20 (23%)	25 (24%)	0.022
Race/ethnicity
African American	46 (32%)	21 (24%)	38 (37%)	0.21
Caucasian	76 (53%)	57 (66%)	53 (51%)
Other	21 (15%)	8 (9%)	12 (12%)
History of CAD	0	8 (9%)	5 (5%)	0.0016
Current smoking	48 (34%)	27 (31%)	42 (41%)	0.18
Diabetes mellitus	8 (6%)	11 (13%)	8 (8%)	0.15
Hypertension	36 (25%)	42 (49%)	34 (33%)	0.0012
History of HF	0	4 (5%)	3 (3%)	0.046
Lipid lowering therapy	37 (26%)	27 (31%)	33 (32%)	0.50
LDL (mg/dL)	104 (82-132)	92 (72-115)	95 (74-131)	0.15
HDL (mg/dL)	47 (39-58)	46 (37-54)	48 (38-55)	0.51
TG (mg/dL)	149 (93-249)	178 (106-301)	148 (95-245)	0.41
T Chol (mg/dL)	188 (155-215)	176 (156-204)	182 (150-216)	0.46
BMI (kg/m2)	26 (23-29)	25 (23-29)	25 (23-30)	0.82
Duration of HIV infection (y)	14 (8-18)	14 (10-17)	15 (11-19)	0.13
HAART use	102 (71%)	74 (86%)	89 (86%)	0.0036
NRTI use	112 (78%)	78 (91%)	90 (87%)	0.026
NNRTI use	69 (48%)	41 (48%)	47 (46%)	0.92
PI use	91 (64%)	70 (81%)	81 (79%)	0.0039
Current abacavir use	33 (23%)	33 (38%)	33 (32%)	0.042
Current tenofovir use	63 (44%)	37 (44%)	57 (55%)	0.16
Current CD4 count (cells/mm3)	489 (318-740)	397 (270-732)	450 (260-578)	0.066
Nadir CD4 count (cells/mm3)	200 (66-390)	98 (33-228)	144 (33-260)	0.0003
Plasma HIV RNA < 75 copies/mL	87 (61%)	50 (58%)	61 (60%)	0.68
Hepatitis C	27 (19%)	35 (41%)	25 (24%)	0.0012
Biomarker levels:
ST2 (ng/mL)	25 (20-30)	36 (29-45)	32 (24-40)	<.0001
NT-proBNP (pg/mL)	26 (14-50)	128 (46-293)	37 (22-99)	<.0001
GDF-15 (pg/mL)	649 (458-994)	2415 (1225-5482)	1087 (633-1700)	<.0001
hsCRP (mg/L)	1.3 (0.5-2.8)	1.9 (0.7-3.6)	7.3 (4.6-12.1)	<.0001
IL-6 (pg/mL)	1.9 (1.0-3.5)	2.3 (1.1-4.6)	6.0 (4.2-10.4)	<.0001
D-dimer (ng/mL)	188 (120-262)	205 (139-345)	411 (297-654)	<.0001
Cystatin C (mg/L)	0.7 (0.6-0.8)	0.9 (0.7-1.1)	0.8 (0.7-1.0)	<.0001
eGFRcys (mL/min/1.73m2)	115 (106-124)	98 (66-113)	99 (78-117)	<.0001
eGFRcr (mL/min/1.73m2)	92 (74-103)	82 (64-100)	92 (74-104)	0.038
Detectable Troponin	34 (25%)	21 (27%)	30 (30%)	0.68
Echocardiographic measures:
EF (%)	62 (58-66)	60 (55-65)	61 (56-66)	0.16
Systolic Dysfunction	4 (3%)	8 (10%)	2 (2%)	0.025
Diastolic Dysfunction†
DD Stage 2+	2 (2%)	4 (5%)	4 (4%)	0.026
DD Stage 1	42 (33%)	36 (44%)	50 (52%)
No DD	82 (65%)	41 (51%)	43 (44%)
PASP (mmHg)	29 (24-34)	35 (28-39)	31 (25-37)	0.0002
Pulmonary Hypertension	26 (18%)	33 (39%)	31 (30%)	0.0023
LV Mass Index (g/m2)	86 (70-124)	101 (79-153)	92 (72-137)	0.045
LV Hypertrophy	31 (24%)	32 (41%)	31 (32%)	0.046

Data are presented as Median (IQR) or numbers (percent).

^*List of biomarkers used to derive phenotypes: ST2, NT-proBNP, GDF-15, hsCRP, IL-6, D-dimer

^†DD stage 2+ includes 5 participants with stage 2, 3 participants with stage 3, and 2 participants with stage 5.

Table 2.

Multinomial logistic regression analysis of factors associated with cluster membership

Parameter	Cluster 2 vs. Cluster 1	Cluster 3 vs. Cluster 1
	OR (95%CI)	OR (95%CI)
Age (per decade)	1.39 (0.96, 2.01) p=0.082	1.48 (1.06, 2.08) p=0.023
Female vs. Male	2.34 (0.99, 5.49) p=0.052	2.21 (1.04, 4.72) p=0.040
Black vs. Caucasian	0.31 (0.14, 0.67) p=0.0030	0.90 (0.48, 1.70) p=0.75
Other vs. Caucasian	0.66 (0.25, 1.75) p=0.40	0.95 (0.41, 2.18) p=0.90
BMI (kg/m2)	0.99 (0.93, 1.06) p=0.81	1.06 (1.00, 1.11) p=0.039
Hypertension	4.20 (2.11, 8.36) p<.0001	1.15 (0.61, 2.17) p=0.67
LDL (per 10 mg/dL)	0.89 (0.82, 0.98) p=0.015	0.94 (0.85, 1.04) p=0.24
Nadir CD4 (per doubling)	0.77 (0.65, 0.90) p=0.0010	0.81 (0.71, 0.94) p=0.0040
HCV infection	3.91 (1.95, 7.81) p=0.0001	1.30 (0.67, 2.54) p=0.44

Abbreviations: OR = odds ratio, CI = confidence interval, BMI = body mass index, LDL = low density lipoprotein, HCV = hepatitis C.

Associations of biomarker clusters with echocardiographic measures

Associations of biomarker-derived clusters with echocardiographic measures are also shown in Table 1. Levels of LVEF were similar across the clusters (median 61% overall), and rates of systolic dysfunction were somewhat higher in cluster 2 but low overall (5%). DD was most prevalent in cluster 3 and intermediate in cluster 2. In contrast, PH and LV hypertrophy were most prevalent in cluster 2 and intermediate in cluster 3.

In demographic adjusted analysis, cluster 3 was associated with a 59% increased risk (p=0.0018) of DD relative to cluster 1 (Table 3). Cluster 2 had a 36% increased risk of DD relative to cluster 1, but the association did not reach statistical significance (p=0.064). After adjustment for traditional and HIV-related factors, cluster 3 remained independently associated with a 55% increased risk of DD (p=0.0047), while the association for cluster 2 weakened substantially. Similarly, cluster 3 was associated with a 51% increased risk of DD relative to cluster 1 (p=0.0060) in a parsimonious model which retained only statistically significant covariates.

Table 3.

Association of biomarker-derived phenotype^* with echo-derived outcomes and all-cause mortality

Outcome		1. Demographic adjusted	2. Traditional RF and HIV-related RF	3. Model 2 + Echo parameters	4. Parsimonious model
Diastolic Dysfunction	Event Rate	PR (95% CI)	PR (95% CI)		PR (95% CI)
Cluster 1	44/126 (35%)	reference	reference		reference
Cluster 2	40/81 (49%)	1.36 (0.98, 1.89), p=0.064	1.13 (0.80, 1.60), p=0.48		1.25 (0.90, 1.74), p=0.19
Cluster 3	54/97 (56%)	1.59 (1.19, 2.12), p=0.0018	1.55 (1.14, 2.10), p=0.0047		1.51 (1.12, 2.02), p=0.0060
Pulmonary HTN	Event Rate	PR (95% CI)	PR (95% CI)		PR (95% CI)
Cluster 1	26/143 (18%)	reference	reference		reference
Cluster 2	33/85 (39%)	2.00 (1.28, 3.12), p=0.0024	1.52 (0.94, 2.47), p=0.089		1.67 (1.04, 2.68), p=0.034
Cluster 3	31/102 (30%)	1.50 (0.95, 2.36), p=0.082	1.47 (0.91, 2.36), p=0.11		1.48 (0.94, 2.33), p=0.088
LV Hypertrophy	Event Rate	PR (95% CI)	PR (95% CI)		PR (95% CI)
Cluster 1	31/128 (24%)	reference	reference		reference
Cluster 2	32/79 (41%)	1.43 (0.98, 2.09), p=0.064	1.32 (0.90, 1.93), p=0.15		1.37 (0.95, 1.98), p=0.090
Cluster 3	31/97 (32%)	1.10 (0.75, 1.60), p=0.63	1.09 (0.74, 1.62), p=0.65		1.12 (0.76, 1.65), p=0.57
All-cause mortality	5 year event rate (95% CI)	HR (95% CI)	HR (95% CI)	HR (95% CI)	HR (95% CI)
Cluster 1	3.5 (1.7, 7.0)	reference	reference	reference	reference
Cluster 2	14.7 (9.1, 23.6)	3.92 (1.66, 9.22), p=0.0018	3.17 (1.19, 8.47), p=0.021	3.07 (1.13, 8.39), p=0.028	2.79 (1.12, 6.93), p=0.027
Cluster 3	16.0 (10.6, 24.0)	4.20 (1.87, 9.44), p=0.0005	3.50 (1.47, 8.35), p=0.0048	3.29 (1.33, 8.12), p=0.0097	3.32 (1.41, 7.81), p=0.0060

Model 1 controls for age, gender, and race/ethnicity. Model 2 controls for model 1 + traditional CVD (smoking, BMI, DM, HTN, HDL, TG, LDL) and HIV-related risk factors (HAART use, CD4 count, HIVRNA, HCV, and OI). Model 3 controls for model 2 + echo parameters (PASP, LV Mass, DD, EF). Model 4 controls for age, gender, race, DM, HDL, OI, HCV, HIVRNA, CD4 count, and PASP.

^*List of biomarkers used to derive phenotypes: ST2, NT-proBNP, GDF-15, hsCRP, IL-6, D-dimer. Abbreviations: RF = risk factor, HTN = hypertension, LV = left ventricular, PR = prevalence ratio, HR = hazard ratio, CI = confidence interval. ** five year event rates were calculated as: # events/person-years of follow-up × 5 years

Cluster 2 was associated with a 2-fold higher risk of PH in demographic adjusted analysis (p=0.0024) relative to cluster 1, while cluster 3 had a 50% increased risk (p=0.082). After adjustment for traditional and HIV-related factors, cluster 2 was associated with a 52% increased risk of PH (p=0.089). In a parsimonious model which retained only statistically significant covariates, cluster 2 was associated with a 67% increased risk of PH (p=0.034).

Cluster 2 was associated with a 43% increased risk of LVH in demographic adjusted analysis (p=0.064). However, the association weakened after multivariable adjustments.

Associations of biomarker clusters with all-cause mortality

We next examined associations of cluster membership with all-cause mortality. A total of 48 deaths occurred over a median 6.9 years of follow-up. Rates of all-cause mortality were lowest in cluster 1 and highest in clusters 2 and 3 (Figure 3). In demographic adjusted analysis, both clusters 2 and 3 were associated with a 4-fold higher risk of mortality (Table 3). After adjustment for traditional risk factors, HIV-related factors, and echocardiographic measures, clusters 2 and 3 each remained independently associated with a 3-fold higher risk of mortality, relative to cluster 1 (both p<0.05).

Figure 3.

Cumulative mortality by biomarker-derived phenotype*. * List of biomarkers used to derive phenotypes: ST2, NT-proBNP, GDF-15, hsCRP, IL-6, D-dimer

As an alternative, we tested the two integrated biomarker scores in place of the clusters. When adjusted simultaneously, we found that the first score (inflammatory) showed a statistically significant association only with diastolic dysfunction (HR=1.10 per 1 SD increase, p=0.014). By contrast, the second score (cardiac) had statistically significant associations with all outcomes, independent of traditional and HIV-related risk factors. Specifically, each 1 SD increase in the cardiac score was associated with diastolic dysfunction (PR=1.10, p=0.035), PHTN (PR=1.19, p=0.0009), LVH (PR=1.15, p=0.0016), and all-cause mortality (HR=1.43, p=0.0009).

Because our earlier work found that ST2 was independently associated with both mortality and echocardiography findings, we performed sensitivity analyses to determine whether our cluster variable was independent of ST2. When we tested ST2 and the cluster variable simultaneously in multivariable adjusted models, we found that ST2 was no longer statistically significant for DD, PH, or mortality (data not shown). By contrast, cluster 3 remained independently associated with DD (PR=1.4, p=0.029) and mortality (HR=2.5, p=0.045) and cluster 2 remained independently associated with PH (PR=1.7, p=0.032). Additional models found that cluster 3 was independent of GDF-15 and cluster 2 was independent of D-dimer for mortality (HR=2.8, p=0.028 and HR=2.6, p=0.040, respectively).

Because a large percentage of our cohort had detectable viremia at baseline, we also assessed interactions of cluster membership with viremia for each outcome. Tests for cluster by HIVRNA interaction were weak and did not reach statistical significance for DD (p=0.63), PH (p=0.62), or mortality (p=0.71).

Alternative cluster results

We also performed sensitivity analyses in which we varied the number of clusters and compared our method (k-means) to an alternative, hierarchical clustering procedure (Ward’s method) for prediction of all-cause mortality. Models adjusting for traditional CVD risk factors, HIV-related factors, and echocardiographic measures are shown in Table 4. When the number of clusters was chosen to be two, we found a statistically significant difference between clusters for mortality risk for both k-means (HR=2.78, p=0.0035) and Ward’s method (HR=2.77, p=0.0015). When the number of clusters was chosen to be three, we found that clusters 2 and 3 were both associated with increased mortality risk for k-means, while only cluster 3 reached statistical significance for Ward’s method (HR=3.17, p=0.013). When the number of clusters was chosen to be four, we found that only cluster 4 had a significantly increased mortality risk (k-means: HR=5.38, p=0.0046; Ward: HR=5.46, p=0.027). By contrast, clusters 2 and 3 showed 2-fold and 2.5-fold increased mortality risks respectively for k-means, but neither association reached statistical significance (p=0.24 and p=0.13). Similarly, Ward’s method showed that the intermediate clusters did not reach statistical significance (HR=1.6 for cluster 2, p=0.54; HR=2.6 for cluster 3, p=0.24).

Table 4.

Associations of cluster membership with all-cause mortality, by procedure and number of clusters

Clustering Procedure	Number of clusters	Cluster 1 (reference)	Cluster 2 HR (95%CI)	Cluster 3 HR (95%CI)	Cluster 4 HR (95%CI)
K-means	2	–	2.78 (1.40, 5.51), p=0.0035
		N=187	N=145
	3	–	2.79 (1.12, 6.93), p=0.027	3.32 (1.41, 7.81), p=0.0060
		N=143	N=86	N=103
	4	–	2.05 (0.62, 6.78), p=0.24	2.49 (0.76, 8.18), p=0.13	5.38 (1.68, 17.24), p=0.0046
		N=93	N=98	N=83	N=58
Ward (Hierarchical)	2	–	2.77 (1.48, 5.21), p=0.0015
		N=247	N=85
	3	–	1.22 (0.47, 3.19), p=0.69	3.17 (1.27, 7.90), p=0.013
		N=120	N=127	N=85
	4	–	1.64 (0.33, 8.11), p=0.54	2.57 (0.54, 12.33), p=0.24	5.46 (1.21, 24.57), p=0.027
		N=54	N=120	N=73	N=85

Models are adjusted for demographics, traditional CVD, HIV-related, and echo, as listed in Table 3.

DISCUSSION

In this cohort of 332 HIV-infected men and women, we found that serum biomarker-derived clusters can be used to partition subjects into different “phenotypes” for differentiating cardiopulmonary structural and functional abnormalities, and for predicting mortality risk. We used six distinct biomarkers to classify participants into separate clusters, without incorporating any clinical information. Cluster 2 (“cardiac phenotype”) had an increased risk of both PH and mortality, while those classified into cluster 3 (“inflammatory phenotype”) had an increased risk of both DD and mortality, independent of standard CVD risk factors and HIV-related factors. Both clusters 2 and 3 predicted mortality independent of echocardiographic findings (PH and DD).

Clusters 2 and 3 were both characterized by lower nadir CD4 cell counts consistent with having more advanced HIV disease. In addition, cluster 2 participants had a higher prevalence of hypertension, lower LDL, and HCV coinfection, while cluster 3 participants were characterized by older age, female gender, and higher BMI. Our cluster variable was independently associated with DD, PH, and all-cause mortality, even when controlling for traditional and CVD risk factors.

We previously reported that ST2 and GDF-15 are associated with both cardiovascular dysfunction and mortality risk, even after controlling for traditional CVD risk factors and HIV-related factors². In this analysis, we found that our cluster categories had strong, independent associations with diastolic dysfunction, pulmonary hypertension, and all-cause mortality. While individual biomarkers also showed statistically significant associations with cardiac dysfunction and mortality, these markers were not simultaneously predictive, making it difficult to reconcile the findings of multiple markers.

We examined associations of both clusters and integrated biomarker scores with our study outcomes. While the continuous biomarker scores had more statistical power than the clusters (which are categorical and therefore discard information), a disadvantage is that they are less interpretable and harder to communicate clinically. Clusters are a useful first step in phenotyping a population. Our goal in this study was not to maximize prediction but rather to create clinically meaningful categories that could be used to differentiate structural and functional abnormalities. An advantage of our cluster-based analysis is that it allows partitioning of subjects into discrete categories of risk, based solely on biomarker levels. In addition, our analysis identified risk factors for these “cardiac” and “inflammatory” phenotypes which are readily available to the clinician. While clusters have value in understanding the heterogeneity in a population, translation to the clinical setting can be facilitated using integrated biomarker scores. Such scores could be applied to an individual patient to quantify their risk, giving a more flexible assessment since the degree of heart failure risk exists on a continuum.

To our knowledge, this is the first study to use biomarker-based cluster analysis to classify cardiopulmonary abnormalities and to predict mortality in the setting of HIV infection. A recent study of HF patients used phenotype data (67 variables including ECG, echocardiography, clinical, and laboratory measures) to cluster patients into 3 distinct risk categories, and the authors found strong associations with clinical outcomes¹². By contrast, our study used biomarkers alone to define the clusters, so that we could distinguish and evaluate the biomarkers’ ability to stratify patients into distinct phenotype and risk categories. Although the clusters’ prediction ability would certainly improve if they were comprised of biomarkers combined with clinical risk factors, our goal was to evaluate serum biomarkers in isolation to measure their incremental contribution, as well as to understand the pathophysiologic pathways that mediate risk.

One potential advantage of using biomarkers from just one sample type (i.e., serum) is that the selected panel can be used to develop a multiplex assay, which may offer advantages of efficiency over individual ELISAs and could be performed using a single methodology²⁸. Multiplex assays could be used to facilitate clinical applicability, providing a simple, clinically relevant message to clinicians and patients about the risk of disease. Cluster phenotypes could also be derived from larger discovery based platforms. A recent study of a cohort with stable CHD²⁹ derived a 9 protein prognostic score from an initial panel of 1130 proteins. This score had associations with a composite cardiovascular outcome that were independent of traditional Framingham variables. Additional work is needed to study these proteins in the setting of HIV infection.

Our results illustrate the strengths and limitations of the use of serum biomarkers for the classification of cardiac structural abnormalities and prediction of mortality in the setting of HIV infection. Our novel cluster method allowed us to select a relevant subset of 6 markers from 8 candidates, without knowledge of study outcomes. A strength of this approach is that clusters can define useful groups of patients, and can mitigate the problems of multicollinearity³⁰ that may arise with the inclusion of multiple correlated measures in a multivariable regression model. A limitation of this approach is that the optimal number of clusters and optimal procedure can be difficult to determine. While increasing the number of clusters might allow the detection of more heterogeneity of heart failure, this comes at a cost of generalizability due to the smaller resulting sample sizes. We found that increasing the number of clusters resulted in smaller, less informative clusters and did not improve model fit materially, while using only two clusters ignored the distinction between the cardiac and inflammatory phenotypes. An advantage of our chosen procedure (k-means) is that it does not impose a hierarchical structure on the clusters, so that subjects can be partitioned based on unspecified, distinct patterns (e.g., elevated in one domain and normal in another).

Although our results have not yet been validated in an external cohort, our work serves as a prototype for future cardiac biomarker studies. Precedent for this approach includes use of gene expression data to identify cancer subtypes¹⁰, and use of prognostic biomarkers to identify patients who can benefit from a treatment¹¹. We envision that clinicians will be able to use such a biomarker panel to determine both the stage and type of HF risk (cardiac vs. inflammatory) of an individual HIV-infected patient for HF and early mortality. Although our biomarker clusters were derived using cross-sectional data, they may represent a serial progression of risk in which inflammation occurs first, leading to DD, which in turn leads to cardiac biomarker elevation and PH^{31, 32}. In addition, this biomarker panel could be repeated to provide updated information on a patient’s cardiac health.

Our study includes several limitations. Our 8 biomarkers were measured at a single time point, so we cannot rule out the possibility that DD and PH were a cause rather than a consequence of biomarker elevations. We did not use a high sensitivity troponin assay, and only one-third of our participants had detectable troponin levels. We did not have adjudicated HF outcomes, and our diagnosis of PH by echocardiogram lacks specificity compared with hemodynamic studies, and cannot discrimination between the etiologies of PH³³. Our analysis found that GDF-15 and sST2 clustered with NT-proBNP, and not with the inflammatory markers. While both GDF-15 and sST2 have been strongly associated with cardiovascular disease in multiple studies and have shown utility in heart failure staging³⁴, GDF-15 is also elevated in a variety of malignancies such as colorectal cancer³⁵, while sST2 is elevated in allergic and inflammatory disorders³⁶. However, our biomarker clusters were associated with PH, DD, and mortality even after controlling for immune status (CD4 count and HIVRNA level), suggesting that GDF-15 and sST2 have utility in staging heart failure in the setting of HIV infection. Although NT-proBNP is considered the gold standard biomarker for heart failure, it does not capture all dimensions of heart failure. These results suggest that markers that identify other aspects of heart function/damage may be needed to inform prognosis and diagnosis. Additional work is needed to understand the role of potentially cardiotoxic antiretroviral drugs, such as abacavir, and to validate our findings in other cohorts. Forty percent of our participants had detectable viremia; however, we found that cluster associations were similar in those with both detectable and undetectable viremia. Finally, there may have been incomplete or inadequate control for factors that may confound or mediate the association of elevations in cardiac biomarkers with HF and mortality.

CONCLUSIONS

In summary, we have shown that unsupervised cluster analysis (i.e., uninformed by the outcome variable) of cardiac biomarkers can identify distinct categories of risk and thereby help differentiate cardiopulmonary structural and functional abnormalities as well as mortality risk. Further studies of HIV-infected persons are needed to validate these results. A study of low-dose methotrexate is currently being conducted to determine whether inflammatory marker reduction can improve endothelial function in HIV-infected adults (NCT01949116). Additional studies of other anti-inflammatory therapies may also be needed to identify safe and effective methods of reducing inflammation in HIV infected persons. A broader array of candidate biomarkers may improve discrimination potential. In the future, the use of cardiac biomarker panels could help inform the diagnosis and staging of HF and may be used to identify patients who are at risk of drug toxicity.

CLINICAL PERSPECTIVE.

What is new?

• The analyses in this article build on our earlier finding that elevated levels of ST2 and GDF-15 in HIV-infected persons are associated with cardiac dysfunction and mortality. Despite effective antiretroviral therapy, HIV-infected persons have an increased risk of heart failure. However, heart failure is a heterogeneous condition, and the pathophysiology in HIV infection appears to be multifactorial.

• Our current study finds that patients can be partitioned into separate phenotypes of “inflammatory” and “cardiac” risk, based solely on the measurements of 6 serum biomarkers.

What are the clinical implications?

• Patients with a cardiac phenotype appear to have an increased risk of both pulmonary hypertension and mortality. Patients with an inflammatory phenotype have an increased risk of both diastolic dysfunction and mortality. Both phenotypes are associated with a 3-fold higher risk of all-cause mortality

• These associations are independent of traditional CVD risk factors, HIV-related factors, and echocardiographic findings.

• Serum biomarker derived phenotypes can help inform the diagnosis and staging of heart failure, and can improve prediction of mortality in HIV-infected patients.