Patients were consecutively enrolled from 2002 to 2010 at 2 multidisciplinary HF units from the Division of Cardiology at the Donostia University Hospital (San Sebastián, Spain) and the Division of Cardiology at the Hospital Universitari Germans Trias i Pujol (Badalona, Spain). The patient population consisted of 686 patients (mean age 66.6 years, range 21-96 years, 32.7% women) and the principal inclusion criterion was HF according to the European Society of Cardiology guidelines irrespective of etiology, together with at least 1 HF hospitalization and/or a reduced left ventricular ejection fraction.16,17 The median time from HF diagnosis until inclusion in the study was 8.0 [interquartile range: 2.0-45.7] months.

Blood samples were taken from the antecubital vein at the time recruitment, kept in ice for 1 to 2hours maximum before processing, aliquoted and kept at−80°C until measured. Additionally, samples were batch-analyzed to minimize differences due to sample processing and storage.

The main study outcome was death from cardiovascular causes and the second study outcome was all-cause mortality. For additional information see the supplementary data.

Serum IGF-1, IGFBP2, the N-terminal pro-B-type natriuretic peptide (NT-proBNP) and high-sensitivity troponin T (hs-TnT) were measured using ELISA assays. Serum creatinine levels were analyzed using the CREA method and estimated glomerular filtration rate (eGFR) was calculated using the Chronic Kidney Disease Epidemiology Collaboration 2009 creatinine equation (CKD-EPI).18 The presence of CKD was established according to decreased eGFR (ie, <60mL/min/1.73 m2).19 For additional information see the supplementary data.

The analyses were performed in all patients and in patients classified according to the absence or presence of CKD with decreased eGFR. Normality was demonstrated by the Shapiro-Wilks or Kolmogorov-Smirnov tests. Nonnormally distributed variables were examined after logarithmic transformation. Multivariable linear regression models were performed, with adjustment for covariables significant in univariable analyses. Optimal cutoff values for predicting the outcomes of interest were determined by performing receiver operating characteristic (ROC) curve analysis followed by calculation of the Younden's J statistic.

The cumulative incidence of all-cause death was estimated by the Kaplan-Meier method; unadjusted differences were assessed with log-rank tests. Univariable and multivariable Cox or competing risk regression models (Fine-Gray) with adjustment for noncardiovascular death were used to calculate hazard (HR) or subhazard (SHR) ratios and corresponding 95% confidence intervals for the risk of all-cause death or cardiovascular death, respectively. Patients without outcome were censored at the date of their last follow-up. To determine whether the association of IGFBP2 and IGF-1 with the outcomes of interest differed by the presence of CKD with decreased eGFR, quantitative interaction analyses were performed by Cox or Competitive risk regression analyses in models including IGFBP2 or IGF-1 as continuous variables, CKD with decreased eGFR (yes/no) and their respective interaction terms. In addition, qualitative interaction analyses were performed by Cox or competing risk regression analyses in models including IGFBP-1 or IGF-1 (≤ or >cutoff as determined in ROC analyses), CKD with decreased eGFR (yes/no), and their respective interaction terms.

The baseline characteristics considered as covariables were identified as significant in univariable competing risk regression and Cox analyses for cardiovascular or all-cause death, respectively (tables 1 and 2 of the supplementary data), followed by a backward stepwise selection with minimization of the Akaike information criterion and the P value set at .15 for elimination. In patients with nondecreased eGFR, the covariables selected were included in 3 different models due to the low number of cardiovascular death outcomes. In CKD patients with decreased eGFR, a first model was selected according to the previous procedure and a second model was also considered including clinically relevant covariables significant in univariable analyses.

The additional value of the combination of biomarkers for risk prediction of the outcomes of interest was assessed with Harrell's c-statistics and the integrated discrimination (IDI) and the continuous net reclassification (NRI) indexes. Harrell's C estimates were calculated using the Stata package “somersd”. The variances for the NRI and IDI estimates were calculated using bootstrapping (1000 resamples).

Values are expressed as mean±SD or median [interquartile range], and categorical variables as numbers and percentages. Robust standard errors were calculated by using the stata option vce (cluster center), considering the clustering effect of the enrollment of patients in 2 different multidisplicinary units of HF. Statistical significance was set as a 2-sided P value of .05. The statistical analyses were performed by using SPSS (15.0 version) and STATA (12.1 version) software.

For additional information see the Statistical analysis section of the supplementary data.

The patient population consisted of 686 patients, (mean age 66.6 years, range 21-96 years, 32.7% women). Table 1 shows the clinical and echocardiographic characteristics of all HF patients and of patients classified in 2 groups: patients with CKD as defined by a decreased eGFR (ie, <60mL/min/1.73 m2) (n=290) and patients with nondecreased eGFR (ie, ≥ 60mL/min/1.73 m2) (n=396). Compared with patients with nondecreased eGFR, CKD patients were older, had lower levels of serum sodium, hemoglobin and reduced left ventricular ejection fraction, and showed increased levels of serum potassium, NT-proBNP and a higher rate of hospitalizations in the previous 3 months, ischemic heart disease, hypertension, diabetes mellitus, hypercholesterolemia, peripheral vascular disease and diuretic treatment (table 1).

Serum IGF-1 levels were similar in the 2 groups, but serum IGFBP2 was higher in CKD patients with decreased eGFR than in those with nondecreased eGFR (table 1).

In all patients, IGFBP2 was inversely correlated with eGFR (figure 1A) and directly correlated with NT-proBNP (figure 1B). Multiple regression analyses in table 2 showed that eGFR and NT-proBNP were predictors of the increase in IGFBP2 independently of age, body mass index, serum sodium, hemoglobin, New York Heart Association (NYHA) class, HF duration, hospitalization in the previous 3 months, the presence of ischemic heart disease, hypertension, diabetes mellitus, peripheral vascular disease, chronic obstructive pulmonary disease and diuretic treatment. Interestingly, among the previously mentioned variables, eGFR and NT-proBNP explained the highest percentages of the IGFBP2 variance (table 2).

Figure 1.

Linear regression diagrams between the estimated glomerular filtration rate (eGFR) and insulin-like growth factor binding protein 2 (IGFBP2) (log2) (y=10.3-0.02x) (A) and between the N-terminal pro–B-type natriuretic peptide (NT-proBNP) (log2) and IGFBP2 (log2) (y=0.32x-5.62) (B) in all heart failure patients. Dotted lines are the 95% prediction limits.

(0.2MB).

On the other hand, circulating IGF-1 levels showed associations with age, female sex, NYHA class, hospitalization in the previous 3 months and treatment with beta-blockers and digoxin (table 3 of the supplementary data). However, this protein did not show univariable associations with either eGFR or NT-proBNP (table 3 of the supplementary data).

During a median follow-up of 3.5 (min-max range: 0.1-6) years, cardiovascular death occurred in 98 patients (14.3%) and all-cause death occurred in 177 patients (25.8%). In patients with nondecreased eGFR, cardiovascular death and all-cause death occurred in 36 (9.1%) and 83 (21.0%) patients, respectively. In the CKD group, cardiovascular death and all-cause death occurred in 62 (21.4%) and 94 (32.4%) patients, respectively. Longitudinal analyses showed that patients with CKD (and decreased eGFR) had a higher risk of cardiovascular death and all-cause death than patients with nondecreased eGFR (figure 1 of the supplementary data).

Multivariable competing risk and Cox regression analyses showed that whereas increased IGFBP2 and decreased IGF-1 were associated with all-cause death independently of a baseline model including age, body mass index, NYHA class, eGFR, NT-proBNP (log2), hemoglobin, chronic obstructive pulmonary disease, diabetes mellitus and peripheral vasculopathy, only IGFBP2 was associated with cardiovascular death independently of age, NYHA class, eGFR, NT-proBNP (log2), hemoglobin, chronic obstructive pulmonary disease, ischemic cardiomyopathy, and diuretic treatment (table 3). However, no significant prognostic improvement was observed for IGFBP2 in addition to any of the baseline models considered in the discrimination and IDI reclassification analyses. Nonetheless, IGFBP2 improved discrimination for both outcomes according to NRI values (P ≤ .05, table 3).

Cutoff values to predict cardiovascular and all-cause death for IGFBP2 and IGF-1 were calculated by ROC analyses in all patients (table 4 of the supplementary data). With these thresholds, 237 patients (34.5%) were classified as the high IGFBP2 subgroup and 316 patients (46.1%) were classified as the low IGF-1 subgroup. Interaction analyses were performed to assess whether the associations among IGF-1 or IGFBP2 with the outcomes of interest could be influenced by the presence of CKD (with decreased eGFR), considering the levels of these proteins as continuous or categorical (> or <cutoff as determined in ROC analyses) variables in quantitative and qualitative analyses, respectively. In adjusted analyses controlling for the previously mentioned baseline models (selected for each outcome of interest in the whole population), a significant interaction was observed for the influence of CKD on the association between IGFBP2 and cardiovascular death in quantitative (P=.038) and qualitative (P <.001) analyses (chi-square=7.5, df=2, omnibus P=.023 for the overall interaction effect in the qualitative analysis). In fact, whereas in patients with nondecreased eGFR, high levels of IGFBP2 (above the cutoff=686 ng/mL) were associated with a subhazard ratio of 1.97 (95% confidence interval [95%CI], 1.58-2.47; P <.001) for the risk of cardiovascular death (figure 2A of the supplementary data), in CKD patients, high levels of IGFBP2 were associated with a subhazard ratio of 6.48 (95%CI, 3.21-13.0; P <.001) (figure 2B of the supplementary data). In addition, figure 2C of the supplementary data shows that patients with both high levels of IGFBP2 and an eGFR <30mL/min/1.73 m2 had the highest rate of cardiovascular death (40.6%). Regarding the association between IGFBP2 and all-cause death, interactions were not observed in either qualitative (P=.78) or quantitative (P=.63) analyses.

No significant interactions were found for the influence of CKD on the association between IGF-1 and the outcomes of interest in either quantitative (cardiovascular death: P=.55; all-cause death: P=.99) or qualitative (cardiovascular death: P=.72; all-cause death: P=.87) analyses.

Given that CKD status can influence the potential prognostic value of IGFBP2, longitudinal analyses to evaluate the association of IGFBP2 with cardiovascular death (as the outcome for which an interaction has been identified) have been performed separately in patients with nondecreased eGFR and in CKD patients (with decreased eGFR).

ROC analysis showed that the AUC for IGFBP2 to predict cardiovascular death in patients with nondecreased eGFR was 0.674 (95%CI: 0.585-0.763, P=.001) and the best cutoff value was 564 ng/mL with a sensitivity of 61.1% and a specificity of 69.7%. With this threshold, 131 patients (33.1%) were included in the high IGFBP2 subgroup.

Adjusted analyses showed independent associations of the continuous and the categorical IGFBP2 variable (below or above 564 ng/mL) with cardiovascular death in patients with nondecreased eGFR (figure 2A-C, table 5 of the supplementary data). Nonetheless, no significant prognostic improvement was observed by adding the continuous or categorical IGFBP2 variables to any of the baseline models considered in the discrimination and IDI reclassification analyses, although significant NRI values were found for the categorical IGFBP2 variable (table 5 of the supplementary data).

Figure 2.

Subhazard ratios (solid black lines) and 95% confidence intervals (dotted blue lines) for the association between baseline insulin-like growth factor binding protein 2 (IGFBP2) and cardiovascular (CV) death occurrence in heart failure patients with nondecreased estimated glomerular filtration rate (eGFR ≥ 60mL/min/1.73 m2) after adjustment by age and sex (A), serum sodium and NYHA class (B) and ischemic heart disease and N-terminal pro–B-type natriuretic peptide (NT-proBNP) (log2) (C), and in patients with HF and chronic kidney disease (CKD) with decreased eGFR (< 60mL/min/1.73 m2) after adjustment by age, NYHA, ischemic cardiomyopathy, eGFR and NT-proBNP (log2) (D), and after adjustment by hemoglobin, diabetes mellitus and treatment with angiotensin converting enzyme inhibitor or angiotensin II type 1 receptor blockers (E). In each model, the baseline IGFBP2 levels were modeled with restricted cubic splines with 3 knots in the competing risk regression models (Fine-Gray) where the competing event was noncardiovascular death. The reference level was set at the 25th percentile in the whole population (324 ng/mL, dotted vertical red line). All graphs were truncated at the 75th percentile (831 ng/mL). The horizontal red line indicates SHR=1.

(0.25MB).

ROC analysis showed that the AUC for IGFBP2 to predict cardiovascular death in CKD patients was 0.730 (95%CI, 0.667-0.794; P <.001) and that the best cutoff value was 726 ng/mL with a sensitivity of 83.9% and a specificity of 59.3%. With this threshold, 145 patients (50.0%) were included in the high IGFBP2 subgroup.

Adjusted analyses showed associations of the continuous and the categorical IGFBP2 variable (below or above 726 ng/mL) with cardiovascular death in CKD patients after adjustment for covariables both in model 1 (age, NYHA, ischemic cardiomyopathy, eGFR and NT-proBNP) and in model 2 (hemoglobin, diabetes mellitus, treatment with ACEIs/ARBs) (figure 2C, D, table 4). In addition, Harrell's c-statistics indicated a significant discrimination improvement in the prediction of cardiovascular death after the addition of the continuous and the categorical IGFBP2 variables on top of the baseline models (table 4). Moreover, the IDI and continuous NRI indexes indicated that the addition of IGFBP2 to the baseline models improved their ability to predict cardiovascular death in CKD patients (table 4).