N-terminal pro b-type natriuretic peptide (NT-pro-BNP) –based score can predict in-hospital mortality in patients with heart failure

Serum N-terminal pro b-type natriuretic peptide (NT-pro-BNP) testing is recommended in the patients with heart failure (HF). We hypothesized that NT-pro-BNP, in combination with other clinical factors in terms of a novel NT-pro BNP-based score, may provide even better predictive power for in-hospital mortality among patients with HF. A retrospective study enrolled adult patients with hospitalization-requiring HF who fulfilled the predefined criteria during the period from January 2011 to December 2013. We proposed a novel scoring system consisting of several independent predictors including NT-pro-BNP for predicting in-hospital mortality, and then compared the prognosis-predictive power of the novel NT-pro BNP-based score with other prognosis-predictive scores. A total of 269 patients were enrolled in the current study. Factors such as “serum NT-pro-BNP level above 8100 mg/dl,” “age above 79 years,” “without taking angiotensin converting enzyme inhibitors/angiotensin receptor blocker,” “without taking beta-blocker,” “without taking loop diuretics,” “with mechanical ventilator support,” “with non-invasive ventilator support,” “with vasopressors use,” and “experience of cardio-pulmonary resuscitation” were found as independent predictors. A novel NT-pro BNP-based score composed of these risk factors was proposed with excellent predictability for in-hospital mortality. The proposed novel NT-pro BNP-based score was extremely effective in predicting in-hospital mortality in HF patients.


Result
A total of 1276 patients were screened during the study period. During the selection procedure, 990 patients were excluded due to a variety of factors, including: a lack of final diagnosis of HF at discharge, an age younger than 18, presence of severe chronic pulmonary diseases, the presence of decompensated hepatic diseases with ascites, and renal failure requiring renal replacement therapy. A total of 17 patients were additionally excluded due to a lack of echocardiography examinations. Finally, 269 patients (mean age, 74.5 ± 13.6 years; female, 53.9%) were included in this study. The mean and median levels of serum NT-pro BNP in all patients are 11530.9 ± 12062.0 pg/ml and 6494.0 pg/ml, respectively. Also, 139 patients (51.7%) had been admitted to the intensive care unit (ICU), and 48 patients (18.2%) received mechanical ventilator support. Overall, the mean length of total hospital stay was 12.1 ± 11.5 days, and 29 patients (10.8%) expired during the hospitalization (Table 1).
When setting α as 0.05 and odds ratio (OR) as 4.48, obtained from simple logistic regression evaluating the association between in-hospital mortality and NT-pro-BNP levels categorized by its best cut point by generalized additive models (GAM), the calculated power of logistic regression model in our study using the G-Power 26 reached 1.00.

NT-pro-BNP levels and in-hospital mortality.
Comparing the patients without in-hospital mortality, those who expired during the hospitalization had significantly higher serum NT-pro-BNP levels (19829.0 ± 13542.9 pg/ml versus 10528.3 ± 11501.9 pg/ml in mean levels, and 15942.0 pg/ml versus 6013.0 pg/ml in median levels) (z value, − 3.80; p < 0.001). The GAM revealed that the probability of in-hospital mortality initiated to elevate since NT-pro-BNP level of 0 pg/ml, and the best cut-off point of NT-pro-BNP level was 8100 pg/ml ( Supplementary Fig. 1).
NT-pro BNP-based score and in-hospital mortality. By using Pearson's correlation analysis, 17 variables were significantly correlated with in-hospital mortality in our HF population, including age, New York Heart Association Functional Classification (NYHA Fc) of the heart, laboratory results upon initial hospitalization (serum sodium and serum NT-pro-BNP), ICU admission, mechanical ventilator and noninvasive positive pressure ventilator (NIPPV) support, experience of cardio-pulmonary resuscitation (CPR), vasopressor administration, comorbid diseases of hypertension and chronic kidney disease, infection during hospitalization, as well as oral medications during hospitalization (angiotensin converting enzyme inhibitors (ACEI), angiotensin receptor blocker (ARB), beta-blocker, Aldo. blocker, loop diuretics and digoxin) ( Table 2).
Furthermore, by using the plot displaying the association between the "NT-pro-BNP"-based score and in-hospital mortality, we categorized the NT-pro BNP-based score into low (< 8 points), medium (≧ 8 to < 12 points) and high (≧ 12 points) risk groups. The predicted in-hospital mortality of the low, medium and high-risk groups were 0.5%, 21.7%, and 88.2%, respectively (Fig. 2). While applying the scores back to the current study population, the actual in-hospital mortality rates in low, medium and high-risk groups were 0.5%, 22.1%, and 86.9%. The actual in-hospital mortality and that predicted by the NT-pro BNP-based score was therefore significantly correlated. (Pearson's correlation coefficient, 1.00; p = 0.005). (Fig. 3) The proposed model was demonstrated with an adequate calibration by using the Hosmer and Lemeshow test. Comparison between HF with reduced ejection fraction (HFpEF) and HF with preserved ejection fraction (HFrEF). Since HFrEF and HFpEF are considered two distinctly dissimilar entities in terms of therapy and survival 14 , we made an additional comparison in demographics, etiology, management and in-hospital mortality between the two patient groups.

Discussion
Previously, several scores 4-9 for predicting prognoses of HF patients had been proposed but none of them took serum NT-pro-BNP level as a risk factor. The current study is the first one to propose a novel NT-pro BNP-based score for in-hospital mortality in hospitalization-requiring HF patients. In our single-centered HF cohort, the score exhibited an excellent predictive power which was even better than HF revised score 9 and OPTIMIZE-HF score 4 . This scoring formula was composed of 9 clinical predictors, which were weighted by logistic regression method. These predictors included "serum NT-pro-BNP level > 8100 mg/dl, " "age > 79 years, " "without ACEI/ ARB, " "without beta-blocker, " "without loop diuretics, " "with mechanical ventilator support, " "with NIPPV support," "with vasopressor," and "experience of CPR." The predictors in current NT-pro BNP-based scores were distinct from those in previous prognosis-predicting models in hospitalized HF patients 5,9,4,8 . There are several explanations for the differences of the current NT-pro BNP-based score from others.
Scientific RepoRts | 6:29590 | DOI: 10.1038/srep29590 Database (AHEAD) registry 7 and HF revised score 9 , the NT-pro-BNP level proved more difficult in being taken into consideration because its data was available in only about half of the participants in these studies.
Since the serum NT-pro BNP test has been recommended as a diagnostic tool for HF 20 , the population of HF patients diagnosed with the serum NT-pro BNP criteria in the current study were probably different from the HF populations diagnosed without the serum NT-pro BNP criteria in these studies 5,4,8,7,9 . Besides, NT-pro-BNP has also been proven as a reliable prognosis-predictive factor in HF patients [22][23][24][25]27 . Thus, any risk predictive scoring system that does not evaluate this major risk factor may result in an altered final predictive model and subsequent inaccuracy for outcome predicting.
Previously, several variables such as heart rate > 140 bpm, creatinine clearance rate < 60 ml/min/1.73m2, and serum sodium < 130 mEq/L were exhibited as predictors for mortality. However in the current study, serum sodium < 130 mEq/L (adjusted OR = 3.74; p = 0.005), heart rate > 140 bpm (adjusted OR = 6.25; p = 0.008), and creatinine clearance rate < 60 ml/min/1.73m2 (adjusted OR = 2.90; p = 0.045) were indeed disclosed as independent predictors before other variables, including serum NT-pro BPN level, that were inserted into the multivariate analysis. Conversely, the predictive abilities of the three variables were overtaken and replaced after placing a number of more powerful predictors, including the NT-pro-BNP level, into the multivariate comparison in the later steps (Supplementary Table 2).
In addition, previous studies tended to assess basic characteristics or clinical variables at initial hospitalization -rather than the procedure or pharmacologic intervention during the hospitalization -as potential risk factors 4,5,8,9 . In fact, the procedure itself or intervention such as medications, non-invasive or invasive ventilation support, and administration of vasopressor or experience of CPR also play important roles influencing in-hospital mortality among hospitalized HF patients 7,10,28 . According to the American College of Cardiology (ACC)/AHA guideline of HF pharmacological treatment, ACEI/ARB, and beta-blocker are suggested for decreasing mortality and improving symptoms in HF patients 20 . The OPTIMIZE-HF study 4 demonstrated a lower risk of in-hospital mortality in patients taking ACEIs or beta-blockers at the time of admission, and similar results in acute HF patients were also found by the clinical quality improvement network (CQIN) investigators 29 . Aside from ACEI, ARBs are recommended in patients with systolic dysfunctional HF accompanied by current or prior symptoms who are ACEI intolerant by the ACC/AHA, to reduce morbidity and mortality 20 . Consistent with previous studies, the current study demonstrated the protective role of ACEI/ARB or beta-blocker impacting in-hospital mortality. The patients who took beta-blockers or ACEI/ARB during hospitalization had a lower risk of in-hospital mortality.
Theoretically, diuretics can increase urinary sodium excretion and decrease physical signs of fluid retention in patients with HF, resulting in improved symptoms and exercise tolerance. But the direct effects of diuretics on morbidity and mortality were not known previously 20,30 . Nevertheless, the current study also showed the protective effect of diuretics regarding in-hospital mortality in the HF population.     Non-invasive or invasive ventilation support, administration of vasopressor, and experience of CPR were considered as prognostic endpoints or dependent variables in several studies 10,31 . Spinar et al. 7 discovered that mechanical ventilation support (OR = 15.30; p < 0.05) and use of vasopressors (OR = 5.90; p < 0.05) were The AUC of NT-pro-BNP based score was higher than that of the serum NT-pro-BNP level, HF revised score 9 and OPTIMIZE-HF score 4 (p < 0.0001), while the AUC of serum NT-pro-BNP level was higher than the HF revised score 9 and OPTIMIZE-HF score 4 (p < 0.0001).

Figure 2. Association between the NT-pro BNP-based score and the probability of in-hospital mortality.
Note: The NT-pro BNP-based score has a range of 0-20 points. According to the trend of increasing with the predicted in-hospital mortality, the NT-pro BNP-based score was categorized into low (< 8 points), medium (≧ 8 to < 12 points) and high risk (≧ 12 points), which represented the probability of in-hospital mortality of 0.52%, 21.70%, and 88.20%, respectively. Scientific RepoRts | 6:29590 | DOI: 10.1038/srep29590 associated with higher in-hospital mortality. Sadhu et al. 28 disclosed that in-hospital mortality rate increased to as high as 75% after CPR for in-hospital cardiac arrest among hospitalized patients with HF, with a majority of patients dying within the first post-CPR day. Taking the results of the current study and previous investigations together, the use of non-invasive or invasive ventilation, administration of vasopressor and CPR should be considered as adverse prognostic factors and included into the risk predictive scores of increasing predictability.
It is worth mentioning that left ventricular ejection flow (LVEF) did not exhibit its predictive ability for in-hospital mortality in our study, whether in the entire HF population or subgroup with HFrEF or HFpEF. Actually, while LVEF was considered as a predictor of mortality in many studies, there were still other studies that showed dissimilar results 8,9 . The population in current study consisted of hospitalized HF patients who might have an acute component of HF rather than chronic stable HF. It is probable that the LVEF determined by echocardiography during the hospitalization could reflect an acutely deteriorated heart function rather than a chronic stable status; thus, it could not predict prognosis. Similarly, the HFrEF or HFpEF defined by the LVEF and evaluated during hospitalization may also not truly reflect the chronic stable heart function of the patients. This may further explain why the pharmacologic intervention, such as the use of ACEI/ARB or beta-blocker and even in-hospital mortality, was not statistically different between the two groups with HFrEF and HFpEF.

Limitations.
Several limitations of the current study should be addressed. First, the retrospective nature of the study was subjected to bias. Second, the current study enrolled HF patients who probably had acute components and fulfilled the age-specific serum NT-pro BNP cut-points. Those (such as obese patients) with clinical HF but lower than expected serum NT-pro-BNP levels would be excluded from this study. The results from the current study may not be applied to all HF patients, and the predictability of NT-pro-BNP was probably inflating. Third, the investigation was mainly limited to exploring the relationship between clinical prognosis and one serum NT-pro-BNP level at initial admission. The serial changes of serum NT-pro-BNP levels after management during hospitalization were not taken into consideration. Fourth, this is a single-center study without an external validation, which is its biggest limitation. Although we have done some cross-validation using the Hosmer and Lemeshow test to demonstrate the goodness-of-fit statistic of the HF revised score and OPTIMIZE-HF risk score in our study cohort, the finding that the proposed NT-pro BNP-based score has better predictability for in-hospital mortality than the serum NT-pro-BNP level, HF revised score and OPTIMIZE-HF score in our study cohort could not become a claim applying to other HF cohorts. Further multicentered, prospective research is warranted to confirm the predictive value of the proposed NT-pro BNP-based scores in all HF patients.

Conclusion
In the current single-centered cohort, serum NT-pro BNP testing is not only a good diagnostic tool but also an adequate prognostic marker for hospitalized HF patients. The current study proposed a novel NT-pro BNP-based score with excellent predictability for in-hospital mortality in HF patients.

Method
Study population. This retrospective study was conducted during the period from January 1, 2011, to December 31, 2013, in a regional teaching hospital in the eastern part of Taiwan. All adult hospitalized patients were eligible for this study once they met the following criteria: (1) had serum NT-pro BNP drawn within 24 hours upon hospitalization, which had reached the predefined age-specific cut-points; (2) had undergone echocardiography during the hospitalization; and (3) had final diagnosis of HF with International Classification of Diseases (ICD)-9 code of 428, 428.0, 428.1, and 428.9 at their discharge. The age-specific serum NT-pro BNP cut-points which had been proven to identify acute HF patients were as follows: (1) > 1800 pg/ml in patients over 75 years; (2) > 900 pg/ml in patients aged 50-75 years; and (3) > 450 pg/ml in patients below 50 years [32][33][34][35][36][37][38] . The exclusion criteria included patients less than 18 years of age, along with those who had severe chronic pulmonary diseases (i.e. forced expiratory volume in one second < 1 liter in pulmonary function), decompensated hepatic diseases with ascites, and/or renal failure requiring renal replacement therapy. For those with more than one hospitalization, which was eligible with the inclusion and exclusion criteria during the study period, only the first hospitalization was included into the current study.
Information gathered from patients' medical charts included baseline demographic data, comorbid diseases, Charlson combine scores, etiologies of HF, and NYHA Fc of the heart, along with other clinical variables including vital signs and laboratory results upon initial hospitalization, oral medications during hospitalization, LVEF from echocardiography examinations, chest roentgenogram, and atrial fibrillation (Af) from electrocardiogram, length of hospital stays, and in-hospital mortality. HFrEF and HFpEF were defined as LVEF < 45% and ≧ 45%, respectively 14 .
In those admitted to ICU, additional information such as the implementation of a mechanical ventilator or NIPPV support, the experience of CPR, as well as the length of a mechanical ventilator or NIPPV support, ICU stays, and vasopressors support was also documented. The main statistical analyses were listed below: (1) All variables were put into Pearson's correlation test to evaluate the correlation between in-hospital mortality and these variables. (2) The variables which exhibited significant correlation with in-hospital mortality in Pearson's correlation test were put into the logistic regression model, using the conditional forward stepwise procedure for multivariate analysis to investigate their regression coefficient, OR, and p-value. The elimination criterion for the multivariate analysis was set at p > 0.05. The continuous variables would be transformed into categorical variables in advance by using their best cut points for measuring the probability of death by GAM. In addition, a bootstrap simulation (* 2000), which was proposed for estimating sampling distributions and associated statistics for regression variables in multivariate models 39 , was used to verify the internal validity of this analysis. Resampling procedures were performed and applied to an appropriate joint distribution to estimate covariance matrices, make bias corrections, and construct CIs. (3) An NT-pro BNP-based scoring formula was created using the regression coefficient identified in the multivariate modeling. The scores of the individual predictors were composed of the arithmetic sum of b coefficients derived from logistic regression analysis, including all independent predictors after each numerical rounding. The calibration of the model was evaluated by the Hosmer and Lemeshow Goodness-of-Fit test and the plot of the change in deviance 40 . (4) ROC curve with AUC, as well as sensitivity, specificity, PPV, NPV, and accuracy were used to exam the predictive ability for in-hospital mortality of both the serum NT-pro-BNP level and the NT-pro BNP-based score.
Additionally, the two well-known prognostic scores, namely, HF revised score 9 and OPTIMIZE-HF risk score 4 , which were proven with good fits (goodness-of-fit statistic with 8 degrees of freedom of 9.436 (p = 0.307) for HF revised score and of 10.84 (p = 0.211) for OPTIMIZE-HF risk score) in our study cohort by the Hosmer and Lemeshow test, were applied to the current study population for predictive power comparisons using the method of Hanley and McNeil 41 . In addition, the ROC curve with AUC has used to exam the predictive ability for in-hospital mortality of both HFrEF and HFpEF groups. Finally, plots were drawn to compare the association between the risk of in-hospital mortality and the NT-pro BNP-based scores. In all statistical analyses, a two-sided p ≦ 0.05 was considered statistically significant.