Admission, Discharge, or Change in B-Type Natriuretic Peptide and Long-Term OutcomesClinical Perspective
Data From Organized Program to Initiate Lifesaving Treatment in Hospitalized Patients With Heart Failure (OPTIMIZE-HF) Linked to Medicare Claims
Background—B-type natriuretic peptide (BNP) has been associated with short- and long-term postdischarge prognosis among hospitalized patients with heart failure. It is unknown if admission, discharge, or change from admission to discharge BNP measure is the most important predictor of long-term outcomes.
Methods and Results—We linked patients ≥65 years of age from hospitals in Organized Program to Initiate Lifesaving Treatment in Hospitalized Patients with Heart Failure (OPTIMIZE-HF) to Medicare claims. Among patients with recorded admission and discharge BNP, we compared Cox models predicting 1-year mortality and/or rehospitalization, including clinical variables and clinical variables plus BNP. We calculated the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) for the best-fit model for each outcome versus the model with clinical variables alone. Among 7039 patients in 220 hospitals, median (25th, 75th) admission and discharge BNP were 832 pg/mL (451, 1660) and 534 pg/mL (281, 1111). Observed 1-year mortality and 1-year mortality or rehospitalization rates were 35.2% and 79.4%. The discharge BNP model had the best performance and was the most important characteristic for predicting 1-year mortality (hazard ratio for log transformation, 1.34; 95% confidence interval, 1.28 to 1.40) and 1-year death or rehospitalization (hazard ratio, 1.15; 95% confidence interval, 1.12 to 1.18). Compared with a clinical variables only model, the discharge BNP model improved risk reclassification and discrimination in predicting each outcome (1-year mortality: NRI, 5.5%, P<0.0001; IDI, 0.023, P<0.0001; 1-year mortality or rehospitalization: NRI, 4.2%, P<0.0001; IDI, 0.010, P<0.0001).
Conclusions—Discharge BNP best predicts 1-year mortality and/or rehospitalization among older patients hospitalized with heart failure. Discharge BNP plus clinical variables modestly improves risk classification and model discrimination for long-term outcomes.
Natriuretic peptides such as B-type natriuretic peptide (BNP) and N-terminal pro–B-type natriuretic peptide (NT-proBNP) have become valuable biomarkers to confirm the diagnosis of heart failure (HF).1–3 Further work has demonstrated that BNP and NT-proBNP levels are predictive of outcomes among ambulatory and hospitalized HF patients.4–7 For more than a decade, studies have examined the importance of timing for natriuretic peptide measurements among hospitalized HF patients to determine the optimal period for predicting long-term outcomes; however, the results have been mixed because of small sample sizes, study designs, or being limited to single centers.8–16
Clinical Perspective on p 636
Given substantial rates of postdischarge morbidity and mortality among patients hospitalized with HF, additional measures to better risk-stratify or guide therapeutic decisions among this population are increasingly important. The ability to stratify patients into high-risk categories can guide enrollment in disease management programs or early intensive follow-up, potentially costly interventions which may be cost-effective for those at highest risk.17–20 Although natriuretic peptide measurements are common in clinical practice, it remains unclear whether admission, discharge, or the ratio of discharge/admission natriuretic peptide levels best predict postdischarge outcome and whether these levels provide prognostic information incremental to that provided by standard clinical and laboratory variables.
The objective of the present study was to use a large real-world registry linked to Medicare claims to determine which measure of BNP—admission, discharge, or the change from admission to discharge—best predicts 1-year mortality and 1-mortality or rehospitalization. Moreover, we sought to determine whether the model with the measure of BNP that performed best would reclassify HF patients appropriately among risk strata compared with a model with only clinical variables.
We used clinical data from the OPTIMIZE-HF (Organized Program to Initiate Lifesaving Treatment in Hospitalized Patients with Heart Failure) registry.21 We obtained research-identifiable Medicare claims data including Medicare part A claims and corresponding denominator files. The OPTIMIZE-HF registry includes clinical and laboratory data for patients with HF admitted to 259 participating hospitals from January 1, 2003, through December 31, 2004. Patients eligible for participation were those with a primary admission for worsening HF or increasing HF during an admission resulting in a primary discharge diagnosis of HF. Representative hospitals varied in size and geographic location. Medicare beneficiaries in OPTIMIZE-HF had similar characteristics and outcomes as the general fee-for-service Medicare population.22
Medicare part A claims data included institutional claims for facility costs; beneficiary, physician, and hospital identifiers; admission and discharge dates; and International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) diagnosis and procedure codes. The denominator files included date of birth, sex, race, date of death, and information about program eligibility and enrollment.
Patients included in the analysis were ≥65 years of age and enrolled in OPTIMIZE-HF and Medicare fee-for-service so we could link their OPTIMIZE registry record to a Medicare claim. Because OPTIMIZE-HF and Medicare records do not contain direct patient identifiers, we used a previously described method22 to link records, based on indirect patient identifiers including sex, admission data, discharge data, and date of birth (all used in combination). The index admission for each patient was defined as the earliest admission during the specified study period. We included patients who survived to index hospitalization discharge. Eligible patients were enrolled in Medicare fee-for-service for at least 12 months before the index admission.
Clinical data including demographics, medical history, laboratory findings, and physical examination findings were obtained from the OPTIMIZE-HF registry.
We were interested in 3 exposure variables: admission value of BNP, the ratio of discharge to admission BNP, and discharge value of BNP. Admission BNP was drawn within the first 24 hours of admission (generally the first blood draw). The discharge BNP was that drawn closest to discharge within the last 48 hours of hospitalization. Our analysis only included patients with documented admission and discharge values of BNP. BNP, as opposed to NT-proBNP, was used for this study since at the time the OPTIMIZE-HF registry was enrolling patients, BNP was the primary natriuretic peptide assay utilized and the one for which we had the least missing data. BNP testing was performed locally at each site using their standard clinical assay. Neither site investigators nor patient providers were blinded to the results.
We assessed outcomes for patients at 1 year after discharge. The 2 primary outcomes of interest were 1-year all-cause mortality and the combined outcome of 1-year mortality or rehospitalization. Mortality information was derived from Centers for Medicare and Medicaid Services denominator files. Further, we calculated time to first rehospitalization within the 1-year time period after the index discharge date, excluding transfers to or from another hospital or rehabilitation readmissions.
Baseline variables are presented as frequencies (%) for categorical variables and medians (25th, 75th percentiles) for continuous variables. We constructed Cox proportional hazards regression models for 1-year mortality and 1-year mortality or rehospitalization for each of the 3 exposure variables of interest (a total of 6 models). Stepwise selection was used to select covariates from a candidate list: Entry and exit criteria were 0.05. The list of potential variables was based on significant components of models published in prior OPTIMIZE analyses. The candidate variable list included: admission systolic blood pressure, age, weight, reactive airway disease, admission sodium, admission serum creatinine, liver disease, lower extremity edema, statin prescribed at discharge, β-blocker at discharge, discharge serum creatinine, discharge systolic blood pressure, admission hemoglobin, troponin, angiotensin-converting enzyme inhibitor at discharge, chronic obstructive pulmonary disease (COPD), angiotensin receptor blocker at discharge, coronary angiography performed during admission, nitrates on admission, mechanical ventilation during hospitalization, digoxin on admission, diuretic on admission, prior cerebrovascular accidence/transient ischemic attack, site, and implantable cardioverter-defibrillator placed during hospitalization. We eliminated troponin from our candidate variable list because the values were a mixture of troponin T and troponin I (varying assays) with significant missing data. However, we did use a dichotomous value of troponin (normal or abnormal) in a sensitivity analysis. We eliminated lower extremity edema from our candidate list due to high rates of missing data. Clinical and demographic covariates used for adjustment in the multivariable models included age, race, sex, history of coronary artery disease, history of COPD, history of cerebrovascular accident/transient ischemic attack, history of diabetes, history of hyperlipidemia, left ventricular dysfunction (ejection fraction <40%), discharge serum creatinine, admission systolic blood pressure, discharge systolic blood pressure, admission weight, admission sodium, history of arrhythmias, and admission hemoglobin. These variables were independent clinical predictors in at least 1 of the outcomes of interest. For each continuous variable, linearity was assessed and an appropriate transformation applied, if necessary. Because of nonlinearity, BNP variables were logarithmically transformed. On the basis of knowledge that BNP values tend to be lower for those without left ventricular systolic dysfunction, we tested the interaction term between BNP and left ventricular systolic dysfunction. For each outcome of interest, models with different measures of BNP (admission, discharge, ratio of discharge/admission) were compared using the Akaike information criterion (AIC) to compare goodness-of-fit among the models and to select the best model.23
To test the ability of a model including BNP to correctly reclassify risk for each outcome of interest over a model with clinical variables only, we constructed reclassification tables with low-, intermediate-, and high-risk strata comparing predicted outcomes and observed outcomes, a method that has been previously described.24 These groups were determined based on tertile grouping of prognostic estimates for each outcome. Moreover, for each outcome, we analyzed the proportion of patients in the clinical model who were appropriately reclassified utilizing the model with discharge BNP. We calculated a net reclassification improvement value and integrated discrimination improvement value.25 A probability value of 0.05 was used to determine statistical significance for all results. Both R and SAS version 9.2 (SAS Institute, Inc, Cary, NC) were employed for these analyses.
As a sensitivity analysis, we tested for an interaction term between the variable left ventricular (LV) systolic dysfunction and log discharge BNP in our final mortality model. As a second sensitivity analysis, we added a dichotomous troponin variable to the model to assess its impact on model discrimination.
Among 23 931 HF patients, 7610 (32%) had no BNP recorded at admission or discharge, 9122 (38%) had an admission value only, 160 (1%) had a discharge value only, and 7039 (29%) had an admission and discharge value.
Baseline characteristics are reported in Table 1. Compared with the cohort with admission and discharge BNP values, the cohorts with admission BNP only, discharge BNP only, or neither were similar with respect to demographics; however, there was a significantly higher proportion of black patients in the cohort that had neither admission nor discharge BNP recorded versus those with both admission and discharge BNP (16.0% versus 6.9%; P<0.001). Baseline comorbidities, blood pressure, and laboratory values were similar across all cohorts. Although there were small statistically significant differences, they were not clinically meaningful.
Among the entire cohort, the 7039 patients with an admission and discharge BNP were included in the analysis. The median (25th, 75th) admission and discharge BNP was 832 pg/mL (451, 1660) and 534 pg/mL (281, 1111), respectively. The median age was 80 years (74, 86), 3073 (44%) were male, and 467 (7%) were black. More than half of the patients had a history of coronary artery disease, a third had a history of COPD, and more than 70% had a history of hypertension. The median left ventricular ejection fraction was 42% (28, 58). The majority of patients (67%) were taking a diuretic on admission and almost 50% had been hospitalized within the previous 6 months.
The observed 1-year mortality rate was 35.2%. Among 3 Cox proportional hazards models constructed to predict 1-year mortality (Table 2), the model with discharge BNP (log-transformed) versus admission BNP or ratio of discharge/admission BNP had the best performance characteristics based on model AIC, χ2 associated with the BNP measure (168.3), and c-index (0.693).
The adjusted 1-year mortality model including discharge BNP is shown in Table 3. The logarithmically transformed value of discharge BNP was the strongest predictor of 1-year mortality (hazard ratio [HR], 1.34; 95% confidence interval [CI], 1.28, 1.40). Other significant associations with mortality included age (per 10 years) (HR, 1.41; 95% CI, 1.33, 1.50), admission systolic blood pressure (per 10 mm Hg) (HR, 0.91; 95% CI, 0.89, 0.92), and history of COPD (HR, 1.49; 95% CI, 1.37, 1.63). Increasing weight, hemoglobin, sodium, and a personal history of hyperlipidemia were protective.
Table 4 shows a reclassification table comparing predicted and observed 1-year mortality for our clinical model with and without discharge BNP based on 3 risk strata: low risk (≤25% mortality), intermediate risk (26% to 34% mortality), and high risk (≥35% mortality). A total of 297 (13.7%), 824 (50.6%), and 423 (14.7%) patients in the low-, intermediate-, and high-risk strata, based on clinical variables were reclassified after adding discharge BNP to the clinical model. Among those in the intermediate risk strata based on clinical variables, the addition of discharge BNP to the model allowed the appropriate up-reclassification of 166 (32.1%) patients who died within 1 year and down-reclassification of 338 (30.4%) patients who did not die within 1 year. The net reclassification improvement (NRI) using the clinical model plus discharge BNP, which takes into account all appropriate and inappropriate reclassification, was 5.5% (P<0.0001) versus the clinical model alone, indicating a net improvement in risk-classification by adding discharge BNP to the model. Compared with those who survived, patients who died were 5.5% more likely to move up rather than down a risk category by adding discharge BNP to the clinical model. The integrated discrimination improvement (IDI) was 0.023.
One-Year Mortality or Rehospitalization
The observed 1-year mortality or rehospitalization rate was 79.4%. Among the 3 Cox proportional hazards models we constructed to predict 1-year death or rehospitalization, the model with log-transformed discharge BNP, versus admission BNP or the ratio of discharge/admission BNP, performed best based on the model AIC, χ2 value associated with BNP (83.4), and c-index (0.606) (Table 2).
Table 5 shows the final adjusted 1-year mortality or rehospitalization model. Discharge BNP, log transformed, was the second strongest predictor of 1-year death or rehospitalization (HR, 1.15; 95% CI, 1.12, 1.18). A history of COPD was the strongest predictor of the combined outcome (HR, 1.41; 95% CI, 1.33, 1.51). Other significant components included increasing discharge creatinine values and admission systolic blood pressure. Similar to the 1-year mortality model, increasing weight, hemoglobin, sodium, and a personal history of hyperlipidemia were protective.
Table 6 shows a reclassification table comparing predicted and observed 1-year mortality or rehospitalization for our clinical model with and without discharge BNP, based on 3 risk strata: low risk (≤70% mortality/rehospitalization), intermediate risk (71% to 84% mortality/rehospitalization), and high risk (≥85% mortality or rehospitalization). Risk reclassification occurred in 297 (17.6%), 765 (25.1%), and 267 (15.3%) patients in the low-, intermediate-, and high-risk strata when discharge BNP was added to the model with clinical variables alone. Among those in the intermediate-risk strata based on clinical variables, the addition of discharge BNP to the model allowed the appropriate up-reclassification of 309 (14.8%) patients who died within 1 year and down-reclassification of 151 (15.8%) patients who did not die within 1 year. The NRI using the clinical model plus discharge BNP was 4.2% (P<0.001). The IDI was 0.010 (P<0.0001).
Relationship Between Discharge BNP Values and Clinical Outcomes
The Figure and shows the relationship between values of discharge BNP and the hazard of 1- year mortality and 1-year mortality or rehospitalization. The relationship is curvilinear. The risk of adverse clinical outcomes increases with increasing levels of BNP. However, the incremental increase in risk per unit change in BNP is greatest at the lower levels and tends to be less at the higher levels.
We tested the interaction term between left ventricular systolic dysfunction and log discharge BNP in our final mortality model to determine if reduced versus preserved ejection fraction was an effect modifier. There was no significant interaction (P for interaction=0.36).
Troponin, dichotomized as normal or abnormal, was significantly associated with each outcome, but the effect on model discrimination was negligible (mortality: change in c-statistic 0.003; mortality/rehospitalization: change in c-statistic 0.001).
This study is the largest to examine the relationship of admission or discharge BNP measurements and long-term clinical outcomes as well as the first to explore reclassification of risk using this biomarker. Although any measure of BNP during a hospitalization is highly predictive of outcomes, discharge BNP is the most informative for long-term postdischarge outcomes. Discharge BNP provides better discrimination for 1-year mortality and the composite of 1-year mortality or rehospitalization when compared with clinical variables alone, admission BNP, or the ratio of discharge/admission BNP. Furthermore, addition of discharge BNP to models of long-term outcomes improves the ability to appropriately stratify patients among low-, medium-, and high-risk categories defined by standard clinical and laboratory variables.
Prior studies examining the ideal time for measuring BNP to predict long-term outcomes after discharge have been limited for several reasons. Most studies have been from single centers or had small sample sizes limiting the evaluation of BNP in conjunction with other important clinical characteristics. Other studies have had varied follow-up time and heterogeneous endpoints. Among the studies to examine admission and discharge BNP values, several have also found that predischarge BNP was the most important prognostic marker.8,12,14,26 For example, the largest single-center study by Logeart et al10 (n=114) revealed discharge BNP was the most important predictor of 6-month outcomes. However, given the small sample size, there was limited ability to adjust for other covariates. Similarly, O'Connor et al created a predictive modeling tool for 6-month postdischarge adverse outcomes among HF patients using variables collected in the Evaluation Study of Congestive Heart Failure and Pulmonary Artery Catheterization Effectiveness (ESCAPE) and found discharge BNP to be a strong independent predictor of outcomes. The generalizability of this study, however, was limited by inclusion of only patients with advanced HF enrolled in a clinical trial.26 Other studies have found that the percent reduction in natriuretic peptide level during an acute HF admission is the best predictor of outcome; these studies were also small with less than 200 patients, limiting their ability to control for confounding variables.9,13 Others have suggested that discharge BNP and percent change in BNP provide complementary prognostic information.15
Our analysis of more than 7000 patients from nationally distributed hospitals improves the ability to address which measure of BNP best predicts long-term outcomes. The combination of rich clinical data and long-term outcomes from Medicare claims provide a large sample size to appropriately adjust for other important clinical variables with a sufficient number of events. This approach provides a more robust assessment of the timing of BNP measurements during a hospitalization for predicting long-term outcomes.
In addition, this is the first analysis to explore the ability of BNP to reclassify risk over traditional clinical variables among HF inpatients with respect to long-term outcomes. Our results demonstrate that we appropriately reclassified patients who had an event into higher risk categories and likewise reclassified patients without events into lower risk categories. These findings may lead to integration of discharge BNP into important clinical decision tools or programs for risk stratification.
Although reclassification occurs in all 3 risk strata, the highest proportion of reclassified patients are in the intermediate-risk strata. This holds true for both outcomes of interest, 1-year mortality, and 1-year mortality or rehospitalization. This finding suggests that intermediate-risk patients are most likely to benefit from additional information to better stratify risk and thus determine if more intensive treatment would be beneficial. Yet, unlike other biomarkers such as c-reactive peptide for cardiovascular disease risk, there is not yet a specific therapy that could be differentially applied once patients are better risk stratified. Nevertheless, this model could be used for prospective evaluation of intensive postdischarge care or disease management.
Future decision models may include discharge BNP to identify a group of high-risk patients who may benefit from more intensive outpatient interventions. In particular, disease management programs, though heterogeneous in design, are potentially costly and have had mixed results in the literature.27,28 However, improving our ability to identify patients who are at high risk for death or rehospitalization might allow providers to refer those at highest risk to disease management programs or to varying intensities of disease management. Although this hypothesis would need to be studied prospectively, it provides a potential mechanism to improve care for HF patients. Further, patients at high risk could be scheduled for earlier outpatient physician follow-up appointments. Early physician follow-up has been demonstrated to be associated with decreased rates of 30-day readmission.20 Yet, not all patients see a physician within 1 or 2 weeks of hospital discharge. The ability to stratify patients into high-risk groups may provide a cohort who should be preferentially targeted for early physician follow-up.
As suggested in prior publications looking at the association between inpatient BNP levels and postdischarge outcomes, there is the potential to guide therapy and discharge timing based on natriuretic peptide levels. Although there have been several trials with mixed results of BNP-guided therapy in the outpatient setting, this has not been studied in the inpatient setting. Yet, patients admitted with acute decompensated HF are at higher risk for subsequent death or rehospitalization than their outpatient counterparts. The ability to guide treatment and discharge decisions based on natriuretic peptide levels in combination with clinical variables could potentially improve HF patient outcomes, a hypothesis that could be tested in prospective randomized clinical trials.
Given the retrospective design of our analysis, there is the potential for unmeasured or residual confounding. Patients enrolled in the OPTIMIZE-HF registry may be dissimilar to the general population, restricting our ability to generalize results of this study. Likewise, hospitals participating in OPTIMIZE-HF, a quality-improvement registry, may be unlike other hospitals, thus introducing bias into our analysis. The analysis was restricted to older patients enrolled in Medicare fee-for-service, hence the results may not be generalizable to younger HF patients. Because the analysis only included patients with both admission and discharge BNP levels recorded, they may be systematically different in terms of severity of illness or characteristics than other patients in the registry. The OPTIMIZE-HF registry studied real world variables and outcomes among patients admitted with acute decompensated HF. Therefore, we used site-reported BNP values. This is both an advantage and a limitation of the present analysis. Because we did not use core laboratory BNP values, there is likely variation in BNP assays utilized across sites that could produce varied results. Although this may introduce increased variability in measured BNP levels, the results are more consistent with real world clinical practice. Because site investigators and clinicians were not blinded to the BNP results, this may have affected treatment and outcome. Yet, no clear guidelines or evidence from prospective randomized trials currently exist that guide inpatient therapy based on BNP values. Therefore, it is unlikely that this lack of blinding would result in systematic bias. Prior studies have shown consistently higher natriuretic peptide levels among those with HF and LV systolic dysfunction versus those with a preserved ejection fraction.29 These differences could have potentially modified the effect of BNP on outcomes based on LV function. However, when we tested the interaction term between BNP and LV systolic dysfunction, we found no statistically significant interaction. Other potential HF biomarkers such as troponin, galectin-3, ST-2, and midregional pro-ANP were not studied. Finally, there may be variables such as blood urea nitrogen that might have improved model discrimination further. However blood urea nitrogen was not captured in this data set. Creatinine, however, tends to perform nearly as well.
Compared with admission or the admission/discharge ratio, discharge BNP best predicts 1-year mortality and 1-year mortality and rehospitalization among older patients hospitalized with HF. The addition of discharge BNP to clinical variables modestly improves risk classification and model discrimination for long-term outcomes. The model with clinical variables plus discharge BNP also appropriately reclassifies patients among tertiles of risk and improves discrimination compared with a model with clinical variables alone. Further research is needed to test the role that these prognostic models could play in improving postdischarge treatment and outcomes for HF patients.
Sources of Funding
OPTIMIZE-HF was funded by GlaxoSmithKline. This analysis was supported by the Duke Clinical Research Institute. Dr Kociol has received funding from an American Heart Association Pharmaceutical Roundtable outcomes training grant (0875142N); Dr Reyes's research was partially supported by National Institutes of Health grant T32HL079896.
Dr Fonarow received honoraria from GSK; Dr O'Connor received consulting and grant support from Roche Diagnostics; and Dr Felker received consulting and research for Roche Diagnostics.
- Received March 17, 2011.
- Accepted July 1, 2011.
- © 2011 American Heart Association, Inc.
- Maisel AS,
- Krishnaswamy P,
- Nowak RM,
- McCord J,
- Hollander JE,
- Duc P,
- Ornland T,
- Storrow AB,
- Abraham WT,
- Wu AH,
- Clopton P,
- Steg PG,
- Westheim A,
- Knudsen CW,
- Perez A,
- Kazanegra R,
- Herrmann HC,
- McCullough PA
- McCullough PA,
- Nowak RM,
- McCord J,
- Hollander JE,
- Herrmann HC,
- Steg PG,
- Duc P,
- Westheim A,
- Ornland T,
- Knudsen CW,
- Storrow AB,
- Abraham WT,
- Lamba S,
- Wu AH,
- Perez A,
- Clopton P,
- Krishnaswamy P,
- Kazanegra R,
- Maisel AS
- Gegenhuber A,
- Mueller T,
- Dieplinger B,
- Poelz W,
- Pacher R,
- Haltmayer M
- Bettencourt P,
- Azevedo A,
- Pimenta J,
- Frioes F,
- Ferreira S,
- Ferreira A
- Valle R,
- Prevaldi C,
- D'Eri A,
- Fontebasso A,
- Giovinazzo P,
- Noventa F,
- Barro S,
- Carbonieri E,
- Milani L,
- Aspromonte N
- Miller G,
- Randolph S,
- Forkner E,
- Smith B,
- Galbreath AD
- Cook N
- O'Connor CM,
- Hasselblad V,
- Mehta RH,
- Tasissa G,
- Califf RM,
- Fiuzat M,
- Rogers JG,
- Leier CV,
- Stevenson LW
B-type natriuretic peptide (BNP) is associated with short- and long-term prognosis among patients hospitalized with decompensated heart failure (HF). It is not known which measure of BNP (admission, discharge, or the change from admission to discharge) best predicts postdischarge outcomes. We analyzed data from the Organized Program to Initiate Lifesaving Treatment in Hospitalized Patients with Heart Failure (OPTIMIZE-HF) registry linked to Medicare claims. Our analysis included 7039 patients ≥65 years of age admitted to the hospital with a HF diagnosis, surviving to hospital discharge, with admission and discharge BNP levels recorded. Observed 1-year mortality and 1-year mortality or rehospitalization were 35.2% and 79.4%, respectively. Our analysis found that after adjustment for patient characteristics, the model containing discharge level of BNP performed best to predict 1-year mortality (c-index, 0.693) and 1-year mortality or rehospitalization (c-index, 0.606). Moreover, these models can be used to improve risk classification and model discrimination over models using only clinical variables (1-year mortality net reclassification index, 5.5%, P<0.0001; integrated discrimination improvement, 0.023, P<0.0001) and (1-year mortality or rehospitalization net reclassification index, 4.2%, P<0.0001; integrated discrimination improvement, 0.010, P<0.0001). These results suggest potential methods using BNP to improve risk stratification among HF patients at the time of hospital discharge and may be useful for identifying those who would benefit from higher intensity outpatient interventions. Further research is needed to determine if BNP may also be useful to guide treatment.