Study population

Patients with a stable form of either heart failure with reduced ejection fraction (HFrEF) (LVEF < 40%) or heart failure with mid-range ejection fraction (HFmrEF) (LVEF 40%–49%) were eligible for inclusion in this study. The cohort included patients who were followed up and treated in outpatient clinics of cardiology departments (where authors of this study work) for stable chronic heart failure. All three cardiology departments provide specialised care for heart failure patients. In their medical history, all patients included in the study had an attack of heart failure with elevated natriuretic peptides and a reaction to heart failure treatment. Echocardiography was performed for all patients by experienced physicians working in echocardiography laboratories of cardiology departments. Other structural and/or functional abnormalities related to the patients’ heart failure were found, including LVEF values ranging between 40% and 49%: left ventricular hypertrophy (an interventricular septum ≥ 11 mm or a left ventricular mass index ≥ 115 g/m² for the men and ≥ 95 g/m² for the women), left atrial enlargement (a left atrial volume index > 34 ml/m²) and/or diastolic dysfunction (E/e’ ≥ 13 and mean e’ ˂ 9 cm/s). A current NT-proBNP level < 125 pg/mL was not among the exclusion criteria because all the patients had increased NT-proBNP levels in their medical history. The exclusion criteria were the following: not signing the informed consent, signs and symptoms of acute decompensation from heart failure, and conditions other than heart failure that would likely hinder the patients’ mid-term prognosis (e.g., advanced cancer, severe dementia, etc.). The final decision about the diagnosis of chronic heart failure and the enrolment of a specific patient in the study was done by a cardiologist experienced in care for heart failure patients.

Laboratory methods

The patient plasma NT-ProBNP levels were analysed using the Cobas E411 NT-proBNP electrochemiluminescence immunoassay Kit (Elecsys proBNP II, Roche Diagnostics, Indianapolis, IN, USA). The limit of blank (LoB) was 3 pg/mL, the limit of detection (LoD) was 5 pg/mL, the measuring range was 5–35,000 pg/mL, the functional sensitivity (the lowest analyte concentration that can be reproducibly measured with an intermediate precision CV of 20%) was 50 pg/mL, and the cut-off value was 125 pg/mL.

The plasma hs-cTnI levels were analysed using the Atellica^® IM High sensitivity Troponin I assay (Atellica^® IM TnIH, Siemens Healthcare Diagnostics Inc., Tarrytown, NY, USA), run on an Atellica^® IM Analyser. The LoB of the Atellica^® IM TnIH assay was 0.50 ng/L. The observed LoD ranged from 1.13 to 1.53 ng/L across three reagent lots and two matrices (serum and lithium heparin plasma). The limit of quantitation (LoQ) of the Atellica^® IM TnIH assay was 2.50 ng/l. The Atellica^® IM TnIH assay provides results from 2.50 to 25,000.00 ng/L. The lower end of the measuring interval is defined by the LoQ. The Atellica^® IM TnIH assay is a 3-site sandwich immunoassay using direct chemiluminometric technology. The solid phase reagent is based on magnetic latex particles conjugated with streptavidin with two bound biotinylated capture monoclonal antibodies, each recognising a unique cTnI epitope. The 99th percentile upper reference limit for healthy individuals was 34 ng/L for women and 53 ng/l for men. The total imprecision (CV) at the 99th percentile value of 45.20 pg/mL (ng/L) was below 10%.

Statistical methods

Standard descriptive statistics were applied in the analysis: the continuous variables were described as the mean ± SD and the median (5th percentile; 95th percentile), whereas the categorical variables were characterised by absolute and relative frequencies. The statistical significance of the differences among the groups of patients was analysed using the Mann-Whitney U test for continuous variables and the Fisher’s exact test for categorical variables. The contribution of the hs-cTnI biomarker to the NT-proBNP levels and clinical model was evaluated according to previously published recommendations [7, 8]. A logistic regression was adopted for the identification of predictors and development of multivariable models for scoring systems and biomarkers. The models were evaluated using a flexible calibration curve [9], C statistics, and a reclassification analysis of the model results. The analysis was completed using SPSS 24.0.0.1 (IBM Corporation, 2016) and R 3.5.1, with the PredictABEL and rms package.

Relationship of hs-cTnI to clinical and laboratory parameters

The hs-cTnI levels differed significantly among the groups of patients categorized by age, LVEF, and renal function (expressed by the eGFR). In contrast, there were no significant differences found among the groups classified by sex, BMI, the ischaemic aetiology of the patients’ heart failure, hypertension, atrial fibrillation, diabetes mellitus, chronic obstructive pulmonary disease, or lower extremity peripheral artery disease (S3 Table). According to the Spearman’s rank correlation coefficient, the hs-cTnI levels were most strongly linked to the NT-proBNP levels (r = 0.439; p < 0.001) and renal function (the correlation coefficients for creatinine, urea, and eGFR were r = 0.245, 0.241, and -0.240, respectively, with p < 0.001 for all three parameters). The link was somewhat weaker when LVEF (r = -0.208), heart rate (r = 0.198), and age (r = 0.186) were considered (p < 0.001 for all three parameters) (S4 Table).

The prognostic value of hs-cTnI

Based on the ROC analysis (Table 2), we determined that 17 ng/L is the optimum cut-off value for the hs-cTnI level for predicting the monitored primary endpoint. The area under the curve (AUC) in this model was 0.658 (p < 0.001) (S1A Fig). The ROC analysis of hs-cTnI as a continuous parameter did not demonstrate any statistical significance. Fig 1 shows the distribution of the hs-cTnI levels and the primary endpoint occurrence in the individual categories, according to the increasing levels of hs-cTnI. Patients with hs-cTnI levels < 17 ng/L are at a low risk of endpoint occurrence, whereas those with hs-cTnI levels ≥ 17 ng/L are at a high risk. It should be noted, however, that a further increase in troponin levels does not translate into a significant increase in the risk of primary endpoint occurrence. Patients with hs-cTnI levels < 17 ng/L accounted for 46.4% of all the patients. The monitored primary endpoint did not occur in any of the patients with hs-cTnI levels lower than the LoB (< 3 ng/L), but these patients only accounted for 2.9% of the total patients.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. hs-cTnI distribution in patients with chronic heart failure and the associated 2-year occurrence of the combined endpoints (death, HTX or LVAD).

https://doi.org/10.1371/journal.pone.0255271.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. NT-proBNP [ng/l] and hs-cTnI [ng/l] as predictors of the primary endpoint in the logistic regression models (i.e., the two-year prognosis in terms of all-cause mortality, heart transplantation and left ventricular assist device [LVAD] implantation).

https://doi.org/10.1371/journal.pone.0255271.t002

Similarly, based on the ROC analysis, we determined that 17 ng/L is also the optimum cut-off value for the hs-cTnI level for predicting the secondary endpoint. The area under the curve (AUC) in this model was 0.647 (p < 0.001) (S1B Fig).

Multivariable prognostic models for primary endpoint

The AUC value for the prediction of the primary endpoint based on the ROC analysis of the NT-proBNP levels was 0.749 (p < 0.001) for the categorical model. The predictive value of the NT-proBNP plus hs-cTnI model was 0.767 (p < 0.001), according to the AUC (Table 2).

Based on previously published prognostic models, we developed a multivariable model by selecting comorbidities and clinical and laboratory parameters that have been shown to influence the prognosis of chronic heart failure patients [3, 4, 10]. Using a univariable logistic regression, we identified parameters that were significantly related to the monitored primary endpoint (Table 3). These parameters, together with the NT-proBNP and hs-cTnI levels, were used as inputs for the multivariable logistic regression. The significance of the individual variables selected for the multivariable model is shown in S1 Fig. The final multivariable model, which included NT-proBNP levels, NYHA (> 2), urea, and hs-cTnI levels (≥ 17 ng/l), was created using a backward stepwise algorithm (Table 4). The excellent discriminatory power of the model to distinguish between patients with good and poor prognoses, assessed by the ROC analysis and expressed by the AUC, was 0.823 (p < 0.001) (S5 Table). Adding other parameters to this model did not improve its discriminatory power. After performing the 10-fold cross-validation to validate the results, the average AUC of 0.804 (in the range 0.689–0.892) was obtained. The calibration of the model was assessed by a flexible calibration curve, which confirmed that the predictive model is well calibrated (S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Patient characteristics as predictors of the primary endpoint in the univariable logistic regression models (i.e., the two-year prognosis in terms of all-cause mortality, heart transplantation and left ventricular assist device [LVAD] implantation).

https://doi.org/10.1371/journal.pone.0255271.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. The multivariable logistic regression model using a backward stepwise algorithm for the selection of independent predictors of the primary endpoint (i.e., the two-year prognosis in terms of all-cause mortality, heart transplantation and left ventricular assist device [LVAD] implantation).

https://doi.org/10.1371/journal.pone.0255271.t004

Our previously published multivariable model for predicting the monitored two-year endpoint (overall mortality, HTX, or LVAD implantation) is based solely on clinical parameters and NT-proBNP levels [4]. This model includes the following independent parameters: older age, advanced heart failure (NYHA III+IV), anaemia, hyponatraemia, hyperuricaemia, and a higher dose of furosemide (> 40 mg daily). According to the ROC statistics and expressed by the AUC, the predictive power of this model, applied to the population of 520 patients included in this analysis, was 0.777 (0.722; 0.832) (p < 0.001).

A reclassification analysis confirmed the improvement of the predictive power of the model that combined the NT-proBNP and hs-cTnI levels, urea, and NYHA, as opposed to the model based solely on the NT-proBNP level and clinical parameters: the category-free net reclassification improvement was 0.473 (0.224; 0.723) (p < 0.001) and the integrated discrimination improvement was 0.054 (0.014; 0.094) (p = 0.008).

The final model was visualised using a nomogram (Fig 2), which makes it possible to establish the risk for each patient individually, based on their NT-proBNP and hs-cTnI levels, urea, and NYHA.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The nomogram of the proposed risk score.

https://doi.org/10.1371/journal.pone.0255271.g002

Multivariable prognostic model for secondary endpoint

To create a multivariable model for secondary endpoint prediction, we proceeded similarly as in the primary endpoint analysis. Using a univariable logistic regression, we identified parameters that were significantly related to the occurrence of the secondary endpoint (age, systolic blood pressure, diastolic blood pressure, left ventricular ejection fraction, hypertension, diabetes mellitus, chronic obstructive pulmonary disease, NYHA classification, anaemia, natrium, urea, uric acid, estimated glomerular filtration rate, use of ACE inhibitors/angiotensin II receptor blockers, beta-blockers and higher dose of furosemide; the data are not presented). These parameters, together with NT-proBNP and hs-cTnI levels, were used as inputs for the multivariable logistic regression. The final multivariable model was created using a backward stepwise algorithm that included NYHA (> 2), NT-proBNP and urea levels. The level of hs-cTnI was not found to be an independent predictor of the secondary endpoint (S6 Table). The discriminatory power of the model to distinguish between patients with good and poor prognosis, assessed by the ROC analysis and expressed by the AUC, was 0.802 (p < 0.001).

Comparison of HFrEF and HFmrEF patient subpopulations

In our opinion, it is interesting to highlight some differences between the HFrEF and HFmrEF patient subpopulations. Men outnumber women in both groups; however, the proportion of women is higher in the HFmrEF group than in the HFrEF group (30% and 16.2%, respectively). Furthermore, HFmrEF patients are older, have a higher systolic blood pressure, and their heart failure is of ischaemic aetiology much more frequently. Overall, patients in the HFmrEF group are less frequently classified as NYHA III-IV, have lower median levels of hs-cTnI, NT-proBNP and uric acid, and are less frequently treated with furosemide ≥ 40 mg/day and with mineralocorticoid receptor antagonists (S7 Table). According to Kaplan-Meier curves, the two-year survival without the monitored endpoint is 90.7 (85.1–96.3) in the HRmrEF group, and 82.7 (78.7–86.7) in the HRrEF group. According to the ROC analysis of our multivariable model, the area under the curve (AUC) is 0.860 (N = 120) for HRmrEF and 0.813 (N = 400) for HRrEF. The cut-off values for hs-TnI for HRrEF and HRmrEF patient subpopulations were determined to be 17 ng/l (N = 400) and ideally 14 ng/l (N = 120), respectively; but even 17 ng/l is a very good cut-off value for the HRmrEF group. In view of the fact that the HRmrEF group involves a markedly lower number of patients, we consider a single cut-off value to be adequate, namely 17 ng/l.

Clinical benefit of the model and its interpretation

Our model makes it possible to identify the highest-risk patients in the population of HFrEF/HFmrEF patients. This subgroup of patients should be followed up in specialised centres for chronic heart failure patients and, in accordance with current guidelines, re-examinations, and possibly intensification of pharmacological treatments (increasing the dose of diuretics/ACE inhibitors/sartans/beta-blockers; replacing ACE inhibitors/sartans with sacubitril/valsartan; adding spironolactone/eplerenone, ivabradine, SGLT2 inhibitors or digoxin) and/or non-pharmacological treatments (revascularisation, ICD/CRT implantation, valvular heart disease surgery), might be considered. As a last resort, LVAD implantation or putting the patient on a waiting list for heart transplantation may also need to be considered [2].

From a practical standpoint, it is reasonable to find the boundary beyond which patients are at risk. During the two-year follow-up, the primary endpoint occurred in 14% of the patients in our cohort. In the group of patients with advanced heart failure, for whom the benefit of LVAD implantation was not yet definite (INTERMACS profile 6–7), the two-year occurrence of death/orthotopic heart transplantation/LVAD implantation was 42%. In the group of patients with an INTERMACS profile of 4–5, for whom the benefit of LVAD implantation (as compared with the optimum pharmacotherapy) had been demonstrated, the two-year occurrence of the combined endpoint was 53% [11]. In contrast, the two-year overall mortality rate of patients with advanced heart failure who had undergone the implantation of a centrifugal-flow pump (HeartMate 3) was 11.9%, and the survival at two years, free of disabling stroke or reoperation to replace or remove a malfunctioning device, was 79.5% [12].

With the use of the suggested predictive model, patients can be divided into four groups: low-risk patients (up to 10%, with a score of up to 120 points), low-to-moderate-risk patients (11%–30%, with a score between 121 and 180 points), moderate-to-high-risk (30%–45%, with a score between 181 and 208 points), and high-risk patients (45%, with a score of 209 or more points). Patients with extremely low hs-cTnI levels (< 3 ng/L) have a very good prognosis.

Intensification of pharmacological/non-pharmacological treatments might be considered for the group of low-to-moderate-risk patients, whereas moderate-to-high-risk and high-risk patients should be followed up in centres specialised in heart failure.

Comparison with other models

The main advantages of our model are as follows: (1) the use of a combined endpoint, where LVAD implantation and heart transplantation are monitored together with mortality and (2) the use of routinely available biomarkers. We have demonstrated the benefit of adding the hs-cTnI level to the NT-proBNP level and clinical parameters. The two-year mortality prediction was published in the Seattle Heart Failure Model, with an AUC of 0.792 [13]. The NT-proBNP level was also used in a recently published model for a two-year prediction of a primary composite endpoint (cardiovascular death or hospitalisation for heart failure) (AUC 0.71), cardiovascular death (AUC 0.71), and all-cause death (0.70) [5]. The use of biomarkers, particularly natriuretic peptides, should be a standard procedure to establish the degree of risk in heart failure patients. The NT-proBNP level was the most significant prognostic factor, not only in our study, but also in an extensive cohort of patients in the PARADIGM-HF study [5]. Our results are in accordance with a recently published, and rather extensive, meta-analysis that evaluated the prognostic value of hs-cTnT for individual patients with chronic heart failure. An hs-cTnT level ≥ 18 ng/L, added to clinical parameters and the NT-proBNP level, was an independent predictor for a more than two-year prognosis of monitored endpoints (all-cause mortality, cardiovascular mortality, and cardiovascular hospitalisation) [6].

Our study showed that including an acute heart failure hospitalization as the component of the combined endpoint leads to a diminished predictive power of increased hs-cTnI value. Similar findings were demonstrated by previous analyses of risk models in patients with heart failure published by Rahimi et al. The discriminatory ability of the models for prediction of death appeared to be higher than that for prediction of death or acute heart failure hospitalization or prediction of hospitalizations alone [3].

Limitations

Our study has several limitations. First, we only described a population of patients with mid-range and reduced ejection fractions; any patient with a preserved ejection fraction was excluded from the study. Due to the limited number of patients, we did not attempt to establish whether men and women might have different cut-off values for predicting a poor prognosis. Our model was not validated in an external cohort of patients. Additionally, our results respond to hs-cTnI measurements in chronic heart failure patients in a stable phase in outpatient clinics. And finally, population sampling might be influenced by a selection bias. This group of heart failure patients has been followed up in university centres; therefore, the patient cohort very probably was not a representative sample of heart failure patients with EF < 50% from across the Czech Republic. A study from the Netherlands, based on an extensive registry of heart failure patients, which was intended to described the real-world population with heart failure [14], revealed that about 77% of all patients had EF < 50%. Their mean age was 72 years (as compared to 65 years in our cohort) and men accounted for 64% only (as compared to 80,6% in our cohort).