A Multimarker Approach for Prediction of Heart Failure Incidence in the Community

Abstract

Background

Several biological pathways are activated in ventricular remodeling and in overt heart failure (HF). There are no data, however, on the incremental utility of a parsimonious set of biomarkers (reflecting pathways implicated in HF) for predicting HF risk in the community.

Methods and Results

We related a multi-biomarker panel to the incidence of a first HF event in 2754 Framingham Heart Study participants (mean age 58 years; 54% women), who were free of HF and underwent routine assays for 6 biomarkers (c-reactive protein, plasminogen activator inhibitor-1, homocysteine, aldosterone-to-renin ratio, b-type natriuretic peptide [BNP] and urinary albumin-to-creatinine ratio [UACR]). We estimated model c-statistic, calibration and net reclassification improvement (NRI) to assess the incremental predictive usefulness of biomarkers. We also related biomarkers to incidence of non-ischemic HF in participants without prevalent coronary heart disease.

On follow-up (mean 9.4 years), 95 first HF events occurred (54 in men). In multivariable-adjusted models, the biomarker panel was significantly related to HF risk (p=0.00005). Upon backwards elimination, BNP and UACR emerged as key biomarkers predicting HF risk: hazards ratio (HR; confidence interval [CI]) per standard deviation increment in log-marker were 1.52 (1.24-1.87) and 1.35 (1.11-1.66), respectively. BNP and UACR significantly improved the model c-statistic (CI) from 0.84 (0.80-0.88) in standard models to 0.86 (0.83-0.90), enhanced risk reclassification (NRI = 0.13; p=0.002), and were also independently associated with non-ischemic HF risk.

Conclusion

Using a multimarker strategy, we identified BNP and UACR as key risk factors for new-onset HF with incremental predictive utility over standard risk factors.

Introduction

Heart failure (HF) is associated with high morbidity and mortality, making its prevention a public health priority.¹ Identification of people who are at higher risk of developing HF is critical for targeting of prevention strategies. Investigators from the Framingham Heart Study previously described a HF “risk profile”² based on clinical, electrographic and X-ray features, but these clinical factors do not fully explain HF risk.³ Recently, numerous investigations have highlighted that several biological pathways are activated during left ventricular (LV) remodeling and HF evolution. Several reports focused on individual circulating and urinary biomarkers representing some of these key pathways,⁴ but few have assessed the incremental predictive utility of multiple biomarkers considered together.

We recently applied a multimarker strategy to identify key biomarkers associated with indices of LV remodeling⁵ and vascular stiffness.⁶ In the present investigation, we extend the multimarker strategy to overt HF by relating the panel of biomarkers to the incidence of a first HF event in a large community-based sample. The biomarker panel included: aldosterone-to-renin ratio⁵ (ARR; renin-angiotensin-aldosterone axis), c-reactive protein⁷ (CRP; inflammation), plasminogen activator inhibitor-1⁸ (PAI-I; fibrinolysis), b-type natriuretic peptide⁹ (BNP; natriuretic peptide system), homocysteine¹⁰ (oxidative stress) and the urine albumin-to-creatinine ratio¹¹ (UACR; endothelial function). We hypothesized that one or more of these circulating and urinary biomarkers will be associated with HF risk, and will incrementally predict HF incidence beyond established risk factors.

Methods

Study Sample

Details regarding the Framingham Offspring Study have been published previously.¹² In brief, 5124 individuals who were children (or spouses of children) of the original Framingham cohort participants were enrolled in 1971 into the Offspring Study, and these individuals have been examined approximately every 4 years. For the present investigation, we included attendees at the sixth examination cycle (1995-1998; n=3532), referred to as the baseline examination for our analysis. Of these, we excluded 737 participants because of non-available biomarker information, 10 participants for missing covariate information, and another 31 participants who had prevalent HF. Thus, 2754 participants (54% women) remained eligible for this investigation. All participants provided written informed consent and the Institutional Review Board of Boston University Medical Center approved the study protocol.

Measurement of Biomarkers

Biosamples were obtained on the morning of the baseline examination (when covariate information was also collected and follow-up started) after an overnight fast, usually between 8AM and 9AM, and frozen at -80° C without any freeze-thaw cycles until assays were performed. Plasma PAI-1 was determined using an ELISA test for PAI-1 antigen (TintElize PAI-1, Biopool, Ventura, CA).¹³ CRP was measured using the Dade-Behring BN100 nephelometer.¹⁴ Serum aldosterone was measured using a radioimmunoassay¹⁵ (Quest Diagnostics, Cambridge, MA) and plasma renin concentrations were measured by an immunochemiluminometric assay (Nichols assay, Quest Diagnostics).¹⁶ Plasma BNP was measured using a high-sensitivity immunoradiometric assay (Shionogi, Osaka, Japan).¹⁷ We used high-performance liquid chromatography with fluorometric detection to measure total plasma homocysteine.¹⁸ We assessed UACR on a morning urine specimen, using immunoturbidimetry (Tina-quant Albumin assay; Roche Diagnostics, Indianapolis, IN) to measure urine albumin,¹⁹ and the modified Jaffe method to measure urinary creatinine.²⁰

The following were the average interassay coefficients of variation for the biomarker measurements: PAI-1, 7.7%; CRP, 2.2%; renin, 2.0 (high concentrations) -10% (low concentrations); aldosterone, 4.0 (high concentrations) - 9.8% (low concentrations); BNP, 12.2%.; homocysteine 9%; urine albumin, 7.2% and urine creatinine 2.3%.

HF Assessment

Follow-up for the present investigation extended from the baseline examination through December 2007. An endpoints adjudication committee consisting of 3 physicians evaluated all suspected cardiovascular disease events (including HF) by reviewing Heart Study clinic charts, and hospitalization and physician office records, and ascertained the incidence of events according to criteria described previously.²¹ We used Framingham HF criteria²² (Supplementary Table 1) to adjudicate HF incidence. In secondary analyses, we refer to participants who developed HF without an interim myocardial infarction (MI) or unstable angina (also known as coronary insufficiency) as the “non-ischemic” HF group for simplicity.

Statistical Analyses

Biomarker values were natural logarithmically-transformed (to account for skewed distributions) and standardized within sex to account for sex-related differences in biomarker distributions. We modeled aldosterone and renin together as a ratio, ARR, because in our cohort such combined modeling has been most informative.²³ We calculated age- and sex-adjusted Spearman coefficients to evaluate correlations between biomarkers.

To evaluate the predictive utility of biomarkers (with regard to HF risk), we performed the following analyses. First, after confirming that the assumption of proportionality of hazards was met, we fitted Cox models²⁴ that incorporated age, sex, BMI, systolic blood pressure, hypertension treatment, diabetes, total cholesterol to high-density lipoprotein cholesterol ratio, smoking, prevalent MI and valvular heart disease. Next, we tested whether the set of 6 biomarkers (denoted m1-m6 below) was associated with HF risk using a 6 degree of freedom likelihood ratio test (LRT) of the null hypothesis H0: β_m1 = β _m2…= β _m6 = 0. The LRT chi-squared statistic was obtained by comparing likelihoods from two models: (i) with covariates only and (ii) with covariates plus 6 biomarkers. Third, after determining that the set of biomarkers improved the model, we conducted backward elimination (p for retention in model=0.05) to identify the biomarker(s) with the strongest association with HF incidence, forcing in the 10 clinical covariates in the model. We also used stepwise selection to confirm the final model.

HF risk portended by a biomarker in any given individual is a function of the concentration of the biomarker, and the relative risk associated with that concentration. Also, biomarkers carry different relative risks and may vary in concentrations independent of each other. Therefore, to assess the composite HF risk associated with several biomarkers in any individual, we constructed a weighted biomarker score (including only biomarkers that emerged in backward elimination) thus:

Biomarker score = (beta-coefficient estimate of biomarker₁ × sex-standardized log-biomarker₁ concentration) + (beta-coefficient estimate of biomarker₂ × sex-standardized log-biomarker₂ concentration)…. etc

We plotted the cumulative incidence of HF (accounting for competing risk of mortality) according to tertiles of the biomarker score; ascertained HF event proportions in each tertile, and evaluated if participants in the 2^nd and 3^rd tertiles had higher HF risk compared to those in the 1^st tertile in multivariable-adjusted models described above. We also repeated these analyses in a sub-sample with Framingham Risk Score predicted baseline 10-year CHD risk ≥ 10%. In addition, we used the top tertile of the biomarker score as a threshold to calculate the sensitivity, specificity, positive predictive value, negative predictive value, likelihood ratio of positive test and likelihood ratio of negative test to evaluate the performance characteristics of the score for possible HF risk screening.

To assess the utility of biomarkers associated with HF, we evaluated several metrics. First, we assessed model calibration (i.e. concordance of observed risk and that predicted by the model with biomarkers) by calculating the Hosmer-Lemeshow chi-squared statistic for Cox models. Second, we calculated c-statistic for the model with clinical covariates alone and compared it with that for the model with clinical covariates and biomarkers to estimate the increase in the c-statistic in the latter.^25,26 Third, we classified participants into 3 categories of absolute 12-year HF risk (the maximum follow-up duration), thus: low risk (<2%), intermediate risk (2-8%) and high risk (>8%). We evaluated if inclusion of biomarkers improved risk classification of participants by calculating “net reclassification improvement” (NRI).²⁷ The NRI is used to assess how well a new marker “reclassifies” people from one risk category to another (higher or lower). It is calculated as a sum of two separate components (one for individuals with events and the other for individuals without events) as follows:

NRI = [(events reclassified higher minus events reclassified lower)/events] + [(nonevents reclassified lower minus nonevents reclassified higher)/nonevents)].

We calculated NRI using an extension to survival analysis that employs Kaplan-Meier estimates of event probabilities at 12 years.²⁷ A large NRI indicates that the marker causes a large improvement in reclassification. Since there are no previously established categories for the absolute risk of HF, we defined these strata empirically based on the distribution of the risk estimates from the model with clinical covariates. Also, we implemented a 10-fold jackknife cross-validation approach to correct for over-optimism associated with validating the model on the same sample on which it was developed. The sample was split into 10 subsamples and predictions for each one tenth were obtained using models developed on the remaining nine tenths. These cross-validated probabilities were used to calculate jackknife c statistics.

In secondary analyses, we performed additional adjustments for prevalent and intervening ischemic events, and evaluated the relations of biomarkers to non-ischemic HF. We also addressed confounding related to changing values of covariates over time, underlying renal function (assessed as estimated glomerular filtration rate [eGFR]) and evaluated the utility of biomarkers in a subgroup with normal eGFR. We also tested for interactions with age, sex and BMI. Statistical methods for these analyses are described in Supplementary Information Section III.

All statistical analyses were performed using SAS software version 8.2 (SAS institute, Cary, NC) and a two-sided p-value of 0.05 denoted statistical significance. The authors had full access to and take full responsibility for the integrity of the data. All authors have read and agree to the manuscript as written.

Results

Baseline clinical and biochemical characteristics of our sample are displayed in Table 1. Of the 2754 participants in our sample, 118 had a history of MI or unstable angina at baseline. Approximately half the participants in our sample had a baseline Framingham Risk Score predicted 10-year CHD risk ≥10% and a fifth had predicted 10-year CHD risk ≥20%. Of note, women comprised approximately 70% of the group with 10-year CHD risk <10%. Approximately 80% of the HF events were concentrated in the subgroup of participants with Framingham Risk Score predicted 10-year CHD risk ≥10%. Age- and sex-adjusted correlations among the biomarkers are presented in Supplementary Table 2. The strongest positive and inverse correlations were between CRP and PAI-1 and PAI-1 and BNP respectively (Supplementary Table 2).

Table 1. Baseline Clinical and Biochemical Characteristics of Study Participants.

Variable	Men (N = 1278)	Women (N = 1476)
Clinical characteristics
Age, years	59 (10)	58 (10)
Body mass index, kg/m2	28.7 (4.4)	27.4 (5.8)
Systolic blood pressure, mmHg	130 (17)	127 (20)
Total/high-density lipoprotein cholesterol ratio	4.9 (2.0)	3.9 (1.4)
Hypertension treatment, %	30	25
Current smoking, %	14	16
Diabetes, %	12	9
Valvular heart disease, %	3	2
Prevalent myocardial infarction, %	7	1
Biochemical characteristics
C-reactive protein, mg/L	1.81 (0.90, 3.91)	2.40 (0.99, 5.63)
Plasminogen activator inhibitor– 1, ng/ml	25.6 (17.1, 36.0)	20.2 (12.1, 31.8)
Homocysteine, mmol/L	9.81 (8.26, 11.92)	8.30 (6.97, 10.13)
Aldosterone/renin ratio	0.65 (0.38, 1.14)	1.00 (0.55, 1.67)
B-type natriuretic peptide, pg/ml	6.10 (4.00, 15.9)	9.70 (4.00, 19.65)
Urine albumin/creatinine ratio mg/gm	4.88 (2.15, 10.93)	8.55 (3.57, 17.24)

Values are mean (standard deviation) or % for clinical characteristics and median (first quartile, third quartile) for biochemical parameters.

Over a mean follow-up of 9.4 years (maximum 12.8 years), 95 participants (41 women) developed HF. The panel of biomarkers was significantly related to HF risk (p=0.00005). Upon backwards elimination, BNP and UACR emerged as significant correlates; each standard deviation increase in log-BNP and log-UACR was associated with a 52% and 35% higher risk of developing HF, respectively (Table 2.A). When BNP and UACR were modeled together as a biomarker score, unadjusted HF proportions increased 10-fold across tertiles. Cumulative HF incidence curves by tertile of biomarker score are presented in Figure 1. Participants in the 2^nd and 3^rd tertile of biomarker score carried multivariable-adjusted HF hazards that were three-fold and four-fold higher respectively, compared to those with a biomarker score in the first tertile (Table 3). Results were similar when we repeated these analyses in a sub-sample with Framingham Risk Score predicted baseline 10-year CHD risk ≥ 10% (Supplementary Table 3). The sensitivity, specificity, positive predictive value, negative predictive value, likelihood ratio of positive test and likelihood ratio of negative test of the top tertile of biomarker score were 68%, 68%, 7%, 98%, 2.13 and 0.46 respectively. When these metrics were recalculated using a sample of individuals with at least 6 years of follow-up, the estimates were very similar.

Table 2. Relations of Biomarkers to HF risk.

	Adjusted HR* (95% CI)	p-value
Primary Analyses
A. Relations of Biomarkers to All HF (N = 2754)
Log-BNP	1.52 (1.24 – 1.87)	< 0.0001
Log-UACR	1.35 (1.11 – 1.66)	0.004
Secondary Analysis
B. Relations of Biomarkers to All HF (N = 2754), Adjusting for Interim MI†
Log-BNP	1.52 (1.24 – 1.86)	< 0.0001
Log-UACR	1.38 (1.13 – 1.69)	0.002
C. Relations of Biomarkers to Non-ischemic HF (N = 2636) ¶
Log-BNP	1.78 (1.37 – 2.31)	< 0.0001
Log-UACR	1.39 (1.06 – 1.82)	0.02

HR = hazards ratio; CI = confidence interval. BNP = b-type natriuretic peptide; UACR = urinary albumin-to-creatinine ratio.

^*Hazards ratio are per SD increment in log-biomarker, and are adjusted for age, sex, body mass index, systolic blood pressure, hypertension treatment, diabetes, current smoking, total/HDL cholesterol ratio, valvular heart disease and prevalent myocardial infarction.

^†Model adjusted for interim myocardial infarction in addition to age, sex, body mass index, systolic blood pressure, hypertension treatment, diabetes, current smoking, total/HDL cholesterol ratio, valvular heart disease and prevalent myocardial infarction.

^¶Analysis restricted to participants without prevalent myocardial infarction or unstable angina. Model adjusted for age, sex, body mass index, systolic blood pressure, hypertension treatment, diabetes, current smoking, total/HDL cholesterol ratio and valvular heart disease.

Figure 1.

Cumulative incidence of HF according to tertiles of the biomarker score incorporating BNP and UACR.

Table 3. Relations of Biomarker Score to HF Risk.

	Tertile 1	Tertile 2	Tertile 3
No. of events/no. at risk (%)*	6/918 (0.7)	24/918 (2.6)	65/918 (7.1)
HF risk†	Referent	2.9 (1.7 – 7.0)	4.2 (1.8 – 10.2)

^*Event proportions by tertile of biomarker score based on BNP and UACR

^†Multivariable-adjusted HF hazards (confidence intervals) compared to referent group. Model adjusted for age, sex, body mass index, systolic blood pressure, hypertension treatment, diabetes, current smoking, total/HDL cholesterol ratio, valvular heart disease and prevalent myocardial infarction.

Addition of biomarkers to the model with clinical characteristics improved the model c-statistic (CI) from 0.84 (0.80-0.88) to 0.86 (0.83-0.90; p=0.007 for improvement), and the model with biomarkers had excellent calibration (Hosmer-Lemeshow chi-square=9.45; p=0.40). The contributions of BNP and UACR to the improvement in c-statistic were of equal magnitude (0.01 each); c-statistics for models with clinical covariates alone, with covariates and BNP and with covariates and BNP and UACR were 0.84, 0.85 and 0.86 respectively. When participants who developed HF and those who did not were separately classified into risk categories based on clinical characteristics alone, the addition of biomarkers reclassified 13% of participants in the appropriate direction; i.e. those without HF on follow-up in the intermediate risk group were reclassified “downward”, and those with HF on follow-up in the intermediate risk group were reclassified “upward” (Table 4; NRI=0.13; p-value=0.002).

Table 4. Classification of Participants into HF risk groups Based on Multivariable Models With and Without Biomarkers.

A. Risk reclassification in participants without HF, N (row %)
	<2 %*	2–8 %*	> 8 %*	Total
<2 %†	1529 (93)	115 (7)	0 (0)	1644
2–8 %†	176 (23)	540 (71)	49 (6)	765
>8 %†	0 (0)	74 (30)	176 (70)	250
Total	1705	729	225	2659
B. Risk reclassification in participants with HF, N (row %)
	<2 %*	2–8 %*	> 8 %*	Total
<2 %†	26 (79)	7 (21)	0 (0)	33
2–8 %†	3 (8)	25 (69)	8 (22)	36
>8 %†	0 (0)	3 (12)	23 (88)	26
Total	29	35	31	95

Net Reclassification Improvement = [(7+8)-(3+3)]/95 + [(176+74) – (115+49)]/2659 = 0.13; p – value = 0.009

^*HF risk categories based on clinical covariates plus biomarker score

^†HF risk categories based on clinical covariates alone

In secondary analyses, relations of BNP and UACR to HF risk remained robust upon adjustment for interim MI (Table 2.B). Results of the jackknife cross-validation showed that the basic model not including the two biomarkers produced a jackknife c statistic of 0.838 (0.797-0.879), whereas addition of the two markers to the basic model produced a jackknife c statistic of 0.862 (0.825-0.899) yielding a jackknife over-optimism estimate of 0.024 (0.007-0.041, p=0.007), therefore suggesting only a modest degree of over-optimism. In the sub-sample of participants without baseline MI or unstable angina (N=2636), 57 developed non-ischemic HF on follow-up. In these analyses, each standard deviation increment in log-BNP and log-UACR was associated with a 78% (p<0.001) and 39% (p=0.02) higher HF risk, respectively (Table 2.C). When analyses evaluating the relations of biomarker score to HF risk were adjusted for all incident and prevalent ischemic events (MI, unstable angina and angina pectoris), the results were unchanged (Supplementary Table 4). Also, additional adjustment of the multivariable model with biomarker score for incident and prevalent unrecognized MI did not alter the results (Supplementary Table 5). In addition, results did not differ when our analyses were repeated in a sub-group of participants with normal LV systolic function (data not shown).

Relations of BNP and UACR to HF risk were not altered by additional adjustment for eGFR, and were similar in participants with eGFR ≥ 60 ml/min/1.73 m² (data not shown). Relations of biomarker score tertiles to HF risk were unchanged when clinical risk factors were modeled as time-dependent covariates (Supplementary Table 6). Lastly, relations of BNP and UACR to HF risk were not modified by age, sex or BMI (none of the interaction terms was statistically significant).

Discussion

Principal Findings

In our large community-based sample, we identified BNP and UACR (from a multimarker panel) as key predictors of HF risk, emphasizing the importance of natriuretic peptide system activation and endothelial dysfunction as markers of disease progression. BNP and UACR were also independently associated with non-ischemic HF, suggesting that relations of biomarkers to HF risk are not mediated solely by interim occurrence of ischemic events, and were significantly associated with HF risk in the subgroup with normal LV systolic function, suggesting that our findings were not driven by individuals with LV systolic dysfunction in our sample. Also, participants who were excluded due to intercurrent ischemic events as having “ischemic HF” may still have non-ischemic HF (i.e. the intervening event may not be causally related to HF occurrence). These two biomarkers improved HF risk prediction very modestly, as evidenced by improvements in model discrimination and risk reclassification. The biomarker score may have potential utility as a screening tool, a premise that would require additional studies; the high negative predictive value may be important to note in this context. We also demonstrate the robustness of the biomarkers in predicting HF risk in participants with high baseline CHD risk; however, we should be cautious in generalizing these findings to people with low baseline CHD risk (10-year risk <10%), and biomarker utility in the latter group needs further study.

Several previous investigations reported the relations of biomarkers from various biological domains to HF risk.^7,11,28-30 However, our investigation is novel in several respects. Whereas earlier studies evaluated biomarkers individually, we used a multi-marker strategy, which permitted a comparison of several biomarkers while limiting multiple statistical testing. Also, we assessed the potential incremental utility of biomarkers (and biomarker scores) for predicting HF risk (above and beyond standard clinical risk factors). An additional strength of our report is the demonstration that both BNP and UACR are associated with risk of developing non-ischemic HF in a large sub-group without history of MI or unstable angina, thereby avoiding potential confounding by pre-existing and interim ischemic events, which can activate several of the pathways represented by the biomarkers we investigated.

Although the exact mechanisms underlying the predictive value of BNP and UACR cannot be conclusively determined based on epidemiological data alone, the results from our investigation, if confirmed, can potentially be used in clinical settings for the purpose of evaluating HF risk and identifying specific high risk individuals. Analogous to the use of multivariable risk profiles in the assessment of CHD risk in people with dyslipidemia, with attendant determination of treatment targets, biomarker risk scores can be used to identify those who are at high risk for HF, and may potentially benefit from treatment of currently established HF risk factors (i.e. hypertension, diabetes, obesity etc) to targets that are lower than conventionally recommended in order to reduce HF incidence. However, this approach (using biomarker scores for risk stratification) has yet to be tested, and will need validation prior to use in clinical settings. Indeed, it may be argued that all patients at risk of developing HF warrant aggressive management of their risk factor burden.

Mechanisms Underlying the Relations of Biomarkers to HF Risk

BNP is a hormone with natriuretic, diuretic and vasodilatory properties that is released in response to increased LV filling pressures,³¹ and/or greater LV wall stress.³² BNP has been used in the diagnosis of clinical HF,³³ but is also elevated in people with asymptomatic LV dysfunction.³⁴ Thus, one explanation for relations of BNP to HF risk is that participants with elevated BNP concentrations are those with sub-clinical LV remodeling,³⁵ systolic dysfunction,³⁶ or diastolic dysfunction,³⁷ and therefore develop HF at higher rates. Another possibility is that relations of BNP to HF risk are mediated by its relations to incident ischemic cardiovascular events.⁹ BNP may be released in the setting of ischemic events other than MI, or with sub-clinical ischemia, and may therefore predict HF secondary to clinical or sub-clinical ischemia. In our investigation, a variety of adjustments for ischemic events (including adjustment for unrecognized MI) and exclusion of participants with MI and unstable angina did not alter the results. However, it is possible that BNP released in response to sub-clinical ischemia may be an explanation for the association between this biomarker and HF risk. Overall, it is likely that BNP is a marker for each of the mechanisms noted above.

Similarly, UACR is a risk marker for endothelial dysfunction,³⁸ target organ damage,³⁹ ischemic events,⁴⁰ and an atherogenic risk profile,⁴¹ which may explain why it is related to HF risk. As with BNP, it is likely that UACR is a cumulative measure for all these mechanisms for HF risk.

The other biomarkers in our panel (CRP, PAI-1, ARR and homocysteine) have been previously implicated in ventricular remodeling, alterations in LV function and HF incidence.^29,30,42-44 However, in our analysis, we did not observe an independent association between these biomarkers and HF risk. It is conceivable that the relations of these biomarkers to HF risk may be mediated through their relations to clinical risk factors we adjusted for in our model. Indeed, previous reports have described the associations between CRP and hypertension⁴⁵ and diabetes,⁴⁶ and between PAI-1 and hypertension⁴⁷ and metabolic syndrome.⁴⁸ Prior investigations also have noted the relations of ARR to hypertension²³ and both ARR and homocysteine to vascular stiffness.^6,49 Thus, our results do not imply that the pathways represented by these biomarkers do not contribute to HF risk.

Strengths and Limitations

Our study is strengthened by a large sample size, standardized measurements of biomarkers and clinical variables, rigorous definition of HF events and a conservative analysis strategy to minimize multiple testing. However, several limitations should be acknowledged. First, we tested only a small set of biomarkers that were available at a routine examination, and that have been previously implicated in LV remodeling and/or HF risk; markers of other biological pathways (or other biomarkers of domains we evaluated) that were not tested may be important in influencing HF risk. Second, differences in the analytical precision of the assays for these biomarkers may have influenced the results of analysis. Third, we lack information on quantitative LV ejection fraction, or measures of diastolic function, or measures of endothelial or vascular function at the baseline examination, and therefore could not adjust for these measures in our multivariable models. Fourth, reclassification metrics and performance characteristics of the biomarker score are susceptible to misclassification of participants (cases vs. non-cases) in whom follow-up information is not complete. Lastly, our sample comprised of middle-aged white individuals of European ancestry and our results may not be generalizable to other age groups or ethnicities.

Conclusions

Refining HF prediction is a fundamental step for preventing the condition. In our prospective investigation of a large community-based sample of middle-aged whites, we identified BNP and UACR as key biomarkers associated with HF risk. However, the incremental usefulness of these biomarkers over standard clinical factors (as assessed by c-statistics and NRI) is very modest. Additional investigations are therefore warranted, both to replicate our findings and to assess their applicability to routine clinical settings.

Clinical Summary.

Several biological pathways have been individually implicated in left ventricular (LV) remodeling and/or HF development, but it is unclear whether biomarkers reflecting these pathways aid in prediction and stratification of HF risk beyond standard risk factors. In a community-based sample, we prospectively related a panel of circulating biomarkers representing distinct biological pathways, viz. aldosterone-to-renin ratio (renin-angiotensin-aldosterone axis), c-reactive protein (inflammation), plasminogen activator inhibitor–1 (fibrinolysis), b-type natriuretic peptide (BNP; natriuretic peptide system), homocysteine (oxidative stress) and the urine albumin-to-creatinine ratio (UACR; endothelial function) to the risk of developing new-onset HF. We used a multi-marker approach that permitted a comparison of the biomarkers in relation to their contributions to HF risk while limiting multiple testing. We also related a biomarker score (based on biomarkers associated with HF) to HF risk. After adjustment for conventional HF risk factors, BNP and UACR emerged as key HF risk predictors. When BNP and UACR were modeled as a biomarker score, we observed a striking 10-fold increase in HF incidence across tertiles of biomarker score. Biomarkers provided incremental information over clinical risk factors for predicting HF as assessed by significant improvements in c-statistic and net reclassification. The predictive ability of biomarkers was maintained in sub-groups with normal LV function, renal function, and those without intervening ischemic events. Our findings are consistent with the notion that activation of the natriuretic peptide system and presence of endothelial dysfunction antedate and predict HF.