Biomarkers of Myocardial Stress and Fibrosis as Predictors of Mode of Death in Patients with Chronic Heart Failure

Study Population

Details of the design, rationale, and primary results of the Heart failure and a controlled trial investigating outcomes of exercise training (HF-ACTION) study have been published elsewhere (9,10). Briefly, HF-ACTION (clinicaltrials.gov, NCT00047437) was a randomized clinical trial evaluating the effect of exercise training vs. usual care on long-term morbidity and mortality in patients with chronic heart failure due to left ventricular systolic dysfunction (NYHA class II–IV, left ventricular ejection fraction (LVEF) < 35%). The primary endpoint of HF-ACTION was a composite of all-cause mortality and all-cause hospitalization over a median follow up of 2.5 years. HF-ACTION was approved by local Institutional Review Boards, and all enrolled patients provided written informed consent.

Adjudication of Mode of Death

An independent clinical events committee that comprised of 11 cardiologists blinded to treatment allocation reviewed and adjudicated all suspected events by consensus. Definitions were as follows:

Biomarker Measurements

A sub-set of patients enrolled in the HF-ACTION study who agreed to participate in the biomarker substudy underwent plasma collection of at baseline. Baseline blood samples were obtained on the same day as baseline exercise testing but were obtained prior to exercise. Samples were collected via peripheral vein into EDTA containing tubes, and then centrifuged immediately and stored at −70°C prior to analysis at core laboratories. Plasma NT-proBNP was measured by a standard electrochemiluminesence immunoassay (Elecsys proBNP, Roche Diagnostics, Indianapolis, IN). The inter-assay and intra-assay coefficients of variation were 2.9 and 6.1%, respectively. Galectin-3 was assessed using enzyme linked immunosorbent assays (BGM Galectin-3, BG Medicine, Waltham, MA). The inter-assay and intra-assay variances were 5.6 and 8.6%, respectively. ST2 was measured from banked plasma samples via a highly sensitive sandwich monoclonal immunoassay (Presage ST2 assay, Critical Diagnostics, New York, NY). The inter-assay and intra-assay coefficients of variation were less than 2.5% and 4.0%, respectively. The core laboratories were blinded to all clinical data.

Statistical Analysis

Baseline patient characteristics were described using medians and intra-quartile ranges or proportions, first in the entire cohort and then stratified by median level of each biomarker. For continuous variables, differences were compared using the ANOVA and Kruskal-Wallis tests. Categorical variables were compared using the chi-square and exact tests. The primary outcome variables of interests were death from pump failure and sudden cardiac death; other modes of death were treated as competing risks in the analysis. ST2 and NT-proBNP were log-base₂ transformed to approximate normality. After assessment of various methods for handling galectin-3, including log transformation, the best model fit was provided by handling galectin-3 as a continuous variable with truncation at >20 ng/mL to avoid violation of the linearity assumption. To describe the timing and frequency of pump failure and sudden cardiac death and their relationship to the 3 biomarkers, cumulative incidence curves stratified by median level of each biomarker were plotted for each outcome. Using Gray’s K-sample test for comparing the cumulative incidence, P-values were derived that tested the null hypothesis of equality between the median-stratified sub-distributions (11).

To determine whether each biomarker was associated with pump failure or sudden cardiac death, and whether the association differed (in magnitude or direction) between the two outcomes, a series of cause-specific Cox proportional hazard models was fit. Cause-specific Kaplan-Meier curves were created in order to select the appropriate cause-specific modeling strategy; because if the curves crossed, an indicator for mode of death was included as a stratification variable in the cause-specific Cox regression model using the data augmentation method described by Lunn and McNeil (12). The association between each biomarker and outcome of interest was characterized by a hazard ratio (HR) and the corresponding 95% confidence interval (CI). To determine if each biomarker had a differential effect on mode of death, the interaction between biomarker and mode of death was tested in the cause-specific Cox regression model described above and the differential association P-value was reported. Models included adjustment first for a comprehensive set of predictors (BMI, sex, dosage of loop diuretic, LVEF, Canadian Cardiovascular Society angina classification, creatinine, and resting ECG ventricular conduction abnormalities) that had been identified as independently contributing to risk prediction in the final adjusted model in the overall HF-ACTION cohort for all-cause mortality, and then additionally for the other two biomarkers (13).

The prognostic utility of each biomarker was examined using two metrics. First, Therneau’s survival concordance index was provided to help assess each model’s discriminative ability. This metric is useful in determining a model’s ability to assign higher risk to individuals that experienced an event. Second, the categorical net reclassification improvement (NRI) was derived to help assess the added predictive ability of these novel biomarkers. The “benchmark” model used to construct the NRI tables was the HF-ACTION adjustment model covariates and NT-proBNP. Patients were categorized as low/high risk based on the Kaplan-Meier failure rate at 4 years. ST2 and galectin-3 levels were added to the benchmark model individually and simultaneously in order to compare the reclassification of risk. Next, we estimated the expected number of cases and controls at 4-years and tabulated the NRI among cases and controls separately. All analyses were performed with SAS 9.3 (SAS Institute Incorporated, Cary, NC) and R 3.0.0 (R Development Core Team). A P-value ≤ 0.05 was considered statistically significant for all analyses. The authors had full access to and take full responsibility for the integrity of the data.

Baseline characteristics

Baseline plasma samples on NT-proBNP, galectin-3, and ST2 were available for 813 patients, who were broadly similar to the HF-ACTION cohort as a whole (N=2331, data not shown). Baseline characteristics for the overall cohort and stratified by median biomarker levels are shown in Table 1. Median age of the study cohort was 59 years; 64% were white, and 71% were male. Median biomarker levels were 852 pg/mL (NT-proBNP), 13.9 ng/mL (galectin-3), and 23.8 pg/dL (ST2). Patients with elevated biomarker levels were older, had a higher prevalence of co-morbidities and NYHA function class, and impaired functional capacity on exercise testing.

Biomarkers and Risk of Mode of Death

During a median follow-up time of 32 months, there were 155 deaths. 49 were classified as due to pump failure, 42 as due to sudden cardiac death, 36 from other cardiac causes (CVA, cardiac procedure-related, fatal myocardial infarction, or unknown), and 28 from non-cardiac causes. Figures 1 shows the cumulative probability of events according to mode of death. Figure 2 shows the cumulative incidences of non-cardiovascular, pump failure, and sudden cardiac deaths, stratified by median biomarker levels. As shown, the incidence of non-cardiovascular death was similar regardless of biomarker levels, whereas it differed significantly in case of cardiovascular causes of death. Patients with higher than median levels of NT-proBNP, galectin-3, and ST2 were more likely to die of pump failure (P<0.001, all). Sudden cardiac death occurred more commonly in patients with higher than median levels of NT-proBNP (P=0.05) and galectin-3 (P<0.01), while the association in the case of ST2 was of borderline significance (P=0.06).

To explore the differential association between each biomarker and mode of death, we first looked at the unadjusted effect, then adjusted for known clinical predictors of adverse outcomes, and lastly for the two remaining biomarkers (Table 2). After adjustment for clinical predictors, all three biomarkers remained significantly associated with both pump failure and sudden cardiac death. After additional adjustment for other biomarkers (ST+galectin-3), NT-proBNP remained predictive of both pump failure and sudden cardiac death. In case of galectin-3, the associations with either mode of death were attenuated, but remained significant, after adjustment for clinical predictors and NT-proBNP, whereas ST2 only retained a significant association with risk of pump failure.

There was a differential association between pump failure and sudden cardiac death for all three biomarkers in the unadjusted models (Table 2). In all cases, increases in the biomarker had a stronger association with pump failure than with sudden cardiac death (P≤0.04). After adjustment for clinical risk factors, there was no further evidence of a significant differential association for NT-proBNP and galectin-3, whereas ST2 still had borderline evidence for a stronger association with pump failure (P=0.06).

Due to potential confounding that may results from patients with ICD therapies, we performed an additional sensitivity analysis by using ICD shocks as an SCD equivalent (Supplementary Table 1), and noted trends similar to those reported in Table 2. We also noted statistical evidence of a differentially stronger association of all three biomarkers with risk of pump failure.

Incremental Changes in Prediction of Mode of Death Using Biomarkers

To explore the clinical implications of our findings, we calculated changes in C statistics for prediction of pump failure and sudden cardiac death with incremental addition of biomarkers to the clinical model (Table 3). For pump failure, the C statistic for the clinical model was 0.82 (95% CI: 0.73–0.91), and similar to that for univariate biomarkers: 0.76 (95% CI: 0.67–0.84) for galectin-3, 0.79 (95% CI: 0.70–0.88) for ST2, and 0.83 (95% CI: 0.74–0.91) for NT-proBNP. Addition of NT-proBNP led to an increase in the C statistic (0.82→0.87); similarly subsequent additions of ST2 and galectin-3 increased the C statistic (0.87→0.89). The ability to discriminate among individuals that experienced a pump failure death was better than that for sudden cardiac death using the clinical model: 0.68 (95% CI: 0.58–0.78)(Table 3). Addition of NT-proBNP to the clinical model improved the C statistic (0.68→0.73), and subsequent additions of ST2 and galectin-3 increased it to 0.75.

Due to potential insensitivities in detection of clinically important risk differences using AUC methods, we quantified the NRI for pump failure and sudden cardiac death. We calculated whether addition to the biomarkers to a risk model with clinic predictors and NT-proBNP would lead to more appropriate classification of risk for each mode of death. Patients were classified into risk categories (low/high) of pump failure and sudden cardiac death based on a dichotomy of risk from a cause-specific model that included clinical covariates and NT-proBNP levels. The Kaplan-Meier failure rate at 4 years was chosen as the cut-off for high/low risk. For pump failure, this value was 10%, for sudden cardiac death this value was 8%. For the endpoint of pump failure (Figure 3; Supplementary Table 2), addition of ST2, galectin-3, or ST2 + galectin-3 to the clinical model + NT-proBNP did not result in a significant amount of net reclassification. In the case of sudden cardiac death (Figure 4; Supplementary Table 3), the addition of ST2 + galectin-3 led to a significant degree of appropriate reclassification [NRI: 0.11 (95% CI: 0.02 – 0.20]. These findings suggest that, after consideration of clinical risk factors and NT-proBNP, addition of novel biomarkers galectin-3 and ST2 may lead to improvements in risk classification of sudden cardiac death.