From May 2006 to July 2010, ambulatory patients treated at a multidisciplinary HF clinic were consecutively included in the study in an outpatient setting. Referral inclusion criteria and blood sample collection have been described elsewhere (11, 14). In summary, patients were referred to the HF clinic by cardiology or internal medicine departments and, to a lesser extent, from the emergency or other hospital departments. The principal referral criterion was HF according to the European Society of Cardiology guidelines irrespective of etiology, at least one hospitalization for HF, or a reduced left ventricular ejection fraction (LVEF). NT-proBNP, hs-TnT, and ST2 were analyzed from the same blood sample stored at −80°, without previous freeze–thaw cycles. All samples were obtained between 09:00 a.m. and 12:00 p.m.

All participants provided written informed consent, and the local ethics committee approved the study. All study procedures were in accord with the ethical standards outlined in the Helsinki Declaration of 1975, as revised in 1983.

All patients were followed at regular predefined intervals, with additional visits as required in case of decompensation. The regular visitation schedule included a minimum of quarterly visits with nurses, biannual visits with physicians, and elective visits with geriatricians, psychiatrists, and rehabilitation physicians (11, 14). Patients who did not attend the regular visits were contacted by telephone.

The total number of unplanned hospitalizations [all-cause, cardiovascular (CV)-, and HF-related] were selected as the primary endpoints. CV admissions were considered as those occurring due to acute coronary syndrome, arrhythmias, stroke, or other CV causes such as rupture of an aneurysm, peripheral ischemia, or aortic dissection. Hospitalizations were identified from clinic records of patients in the HF unit and hospital wards and from the electronic Catalan history record. Fatal events were identified from the clinical records from the HF unit, hospital wards, emergency room, and general practitioners, and by contacting the patient’s relatives. Furthermore, data were verified from the databases of the Catalan and Spanish Health Systems. Six patients lost during follow-up were adequately censored.

NT-proBNP levels were determined using an immuno-electrochemiluminescence method (Elecsys^®, Roche Diagnostics, Switzerland). This assay has <0.001% cross-reactivity with bioactive BNP, and in the constituent studies in this report, the assay had inter-run coefficients of variation ranging from 0.9 to 5.5%.

Troponin levels were measured by electrochemiluminescence immunoassay using the hs-TnT assay on the Modular Analytics E 170 (Roche Diagnostics). The hs-TnT assay has an analytic range from 3 to 10,000 ng/L. At the 99th percentile value of 13 ng/L, the coefficient of variation was 9%.

ST2 was measured from plasma samples using a high-sensitivity sandwich monoclonal immunoassay (Presage^® ST2 assay, Critical Diagnostics, San Diego, CA, USA). The ST2 assay had a within-run coefficient of <2.5%, a total coefficient of variation of 4%, and a limit of detection of 1.31 ng/mL.

Continuous variables were expressed as mean ± 1 SD or median [interquartile range (IQR)] when appropriate. Normal distribution was assessed by normal Q–Q plots. Discrete variables were summarized as percentages. Clinical characteristics were compared among patients with none, one, two, or three or more HF-related hospitalizations and also among biomarker strata. P values for trend for continuous variables and for each HF etiology were calculated using Spearman correlation. Categorical variables were compared using chi-square test (linear by linear).

Biomarker strata relied on categorical use of the three biomarkers using widely accepted cut-points, namely, 1000 ng/L for NT-proBNP, 14 ng/L for hs-TnT, and 35 ng/mL for ST2. Four strata, each indicative of the number of biomarkers elevated (0, 1, 2, and 3) were defined. Crude incidence rates (expressed as number of readmissions per 100 person-years) were calculated for each biomarker strata and for each readmission endpoint (all-cause, CV-, and HF-related). Also event-free curves from Cox regression analysis were plotted for the first and second CV-related hospital admission relative to biomarker strata.

The independent association between biomarkers and recurrent hospitalizations was assessed through a multivariable NBreg and expressed as IRR. The category “0” of the variable (none of the biomarkers elevated) was used as reference.

In addition to the three studied biomarkers, the covariates included in the clinical models were as follows for each endpoint: (1) all-cause recurrent hospitalizations – age, NYHA class, time since HF diagnosis, diabetes mellitus, peripheral artery disease, chronic obstructive pulmonary disease, glomerular filtration rate, serum sodium, hemoglobin, and treatment with beta-blockers; (2) CV-recurrent hospitalizations – age, NYHA class, time since HF diagnosis, diabetes mellitus, peripheral artery disease, glomerular filtration rate, serum sodium, treatment with beta-blockers, ACEI/ARB, and loop diuretics; and (3) HF-related recurrent hospitalizations – age, gender, NYHA class, time since HF diagnosis, diabetes mellitus, glomerular filtration rate, sodium, treatment with beta-blockers, and loop diuretics. The decision to include a covariate in each model was mainly based on backward stepwise selection with Akaike information criterion (AIC) as stopping criterion (P = 0.16) to achieve a parsimonious model and prevent model overfitting. During this process, the linearity assumption for continuous variables was simultaneously tested, and transformed if appropriate, with fractional polynomials (15).

The McFadden pseudo-R² and measures of goodness-of-fit such as the AIC and the Bayesian information criterion (BIC) were used to compare the performance of each multi-biomarker model vs. the clinical model without biomarkers. Given any two estimated models, the model with the lower BIC and AIC scores was preferred. No statistical tests compare different BIC or AIC estimations, and lower values indicate a better model. These comparisons were made, in turn, for each recurrent hospitalization.

A two-sided P value of <0.05 was considered to be statistically significant for all analyses. All analyses were performed using Stata 14.1 (StataCorp. 2015, Stata Statistical Software: Release 14.1 College Station, TX, USA: StataCorp LP) and SPSS 15 (SPSS Inc., Chicago, IL, USA).

From May 2006 to July 2010, a total of 891 consecutive HF patients were included in this analysis. Table 1 shows the baseline characteristics of the study cohort relative to the occurrence of none, one, two, and three or more HF-related hospitalizations. The mean age of the total cohort was 66.4 ± 12.4 years, 71.6% were male, 52.7% showed ischemic etiology, and 87.5% had LVEF below 50%. Most of them were in NYHA class II (65.5%). The median concentrations (IQR) for NT-proBNP, hs-TnT, and ST2 were 1237 ng/mL (481–2798), 22.4 ng/mL (10.5–40.2), and 38.1 ng/mL (30.8–50.9), respectively. The number of HF hospitalizations tended to increase with age, HF duration, worse NYHA functional class, the presence of diabetes mellitus, renal dysfunction, and anemia, and higher biomarker concentrations.

Table 2 summarizes patient characteristics relative to the multi-biomarker strata of none elevated or one, two, or three biomarkers above the cut-point. As the number of biomarkers above the cut-point increased, patients tended to be older, with shorter HF duration, worse NYHA functional class, and worse biochemistry parameters (estimated glomerular filtration rate, sodium, hemoglobin) (Table 2).

At a mean follow-up of 4.2 ± 2.1 years, 399 (44.8%) deaths, 1623 all-cause hospitalizations in 498 patients (55.9%), 710 CV-related hospitalizations in 331 patients (37.2%), and 444 HF-related hospitalizations in 214 patients (24.1%) were registered. The distribution of patients with one, two, three, four, and five or more all-cause, CV-, and HF-related hospitalizations is depicted in Figure S1 in Supplementary Material. Crude incidences for all-cause, CV-, and HF-related recurrent hospitalizations were 36, 19, and 12 per 100 person-years, respectively.

For patients with the multi-biomarker panel above the cut-point (hs-TnT > 14 ng/L, NT-proBNP > 1000 ng/L, and ST2 > 35 ng/mL), the crude incidences of all-cause, CV-, and HF-related recurrent hospitalizations were significantly higher (all P < 0.001; Table 3). Relative to the number of elevated biomarkers, a stepwise increase in the incidences of all-cause and specific causes of recurrent hospitalizations was found when moving from none to three biomarkers elevated (all P < 0.001; Figure 1). Indeed, the number of HF-related hospitalizations per 100 person-years significantly rose from 1 to 31 from none to three biomarkers elevated (P < 0.001).

Figure 2A shows event-free curves for first CV-related admission. Hazard ratios (HRs) based on the number of biomarkers elevated (taking 0 as reference) were as follows: one elevated biomarker, HR 2.57 (95% CI 1.52–4.34), P < 0.001; two elevated biomarkers, HR 4.54 (95% CI 4.74–7.54), P < 0.001; and three elevated biomarkers, HR 7.29 (95% CI 4.4–12.1), P < 0.001. Figure 2B shows event-free curves for the second CV-related admission, considering from discharge date of the first CV-related hospitalization. HRs based on the number of biomarkers elevated (taking 0 as reference) were as follows: one elevated biomarker, HR 3.42 (95% CI 1.03–11.4), P < 0.05; two elevated biomarkers, HR 5.58 (95% CI 1.75–17.9), P = 0.004; and three elevated biomarkers, HR 5.16 (95% CI 1.62–16.4), P = 0.006.

After multivariable adjustment, gradients of risk remained significant for all endpoints, with the strongest association for HF-related hospitalizations. Figure 3 shows the corresponding IRRs (95% CIs) for 1, 2, or 3 biomarkers elevated compared to none for all-cause, CV-, and HF-related recurrent hospitalizations.

For all-cause, CV-, and HF-related recurrent hospitalizations, the McFadden R² as well as the goodness-of-fit measures AIC and BIC supported the value of incorporating the multi-biomarker panel with NT-proBNP, hs-TnT, and ST2 into a clinical predictive model (Table 4).

Finally, in a sensitivity analysis, we tested the interaction between a multimarker risk score (TnThs + NTproBNP + ST-2) and eGFR < 60 mL/min/1.73 m² under the same multivariate scenario. This analysis showed a non-significant interaction among multimarker score and eGFR, revealing, a non-differential prognostic effect of this proposed score across renal function strata (omnibus P-value for all-cause and CV-readmission were 0.381 and 0.112, respectively). Importantly, the multimarker risk score was strongly associated with the risk of HF readmissions in patients with and without renal dysfunction; however, the magnitude of the association was greater in patient without renal dysfunction.