Introduction

The prevalence of heart failure (HF) continues to rise with over 26 million people suffering from the disease worldwide [1]. By the year 2030, the anticipated prevalence of HF in the United States (US) will increase by 46% to more than 8 million adults, and the total direct medical costs of congestive HF total over $60 billion [1]. Improving prognostic assessment of HF patients may help optimize clinical management and reduce hospitalizations and mortality.

Circulating natriuretic peptide (NP) levels are useful biomarkers for detecting the presence of HF and monitoring its severity [2]. Brain natriuretic peptide (BNP) and amino-terminal (NT)-proBNP provide compensatory effects such as diuresis and inhibition of the renin–angiotensin–aldosterone system in response to increased fluid status in patients with mild HF. However, patients with chronic HF often become insensitive to the effects of NPs, resulting in a wide range of observed plasma NP levels [3].

Although BNP and NT-proBNP are well-established in international guidelines for acute HF settings, their role in the chronic setting is unclear [2, 4]. Studies evaluating NPs in relation to chronic HF prognosis and associated outcomes vary significantly in their methods of NP reporting, as well as their overall effect estimates, making interpretation and application of NP results challenging [2]. As a result, the current American College of Cardiology (ACC)/American Heart Association (AHA) and European Society of Cardiology guidelines remain without clearly defined recommendations for the use of NP levels to predict the risk of HF hospitalizations or mortality [2, 4]. To clarify the utility of NPs in predicting mortality and hospitalizations in ambulatory patients with HF, we performed a systematic review and meta-analysis.

Methods

Data sources and searches

In October 2018, with the assistance of an experienced librarian, a systematic search of bibliographic databases was conducted using terms related to HF, mortality, admission, and prognosis (Supplemental Material 1). We searched Embase, MEDLINE, Cochrane Central Register for RCTs, Cochrane database for systematic reviews, and PubMed databases for citations published since 2005. The year 2005 was chosen as a threshold because the predictive utility of NPs may differ in remote versus contemporary cohorts of patients treated with guideline-recommended HF therapies. Ethics approval was not required for this study as it includes publicly available data.

Study selection and data abstraction

The review included ambulatory adult (≥ 18 years) patients with reduced or preserved ejection fraction HF that evaluated the association of BNP or NT-proBNP with mortality and/or hospitalization using multivariable Cox proportional hazard models and reported hazard ratios (HR) and their respective 95% confidence intervals (CI). Supplemental Material 2 provides further details of the study selection and data abstraction processes.

Risk of bias and certainty assessment

We used the six domain Quality in Prognostic Studies (QUIPS) instrument to assess the risk of bias of individual studies [5]. To evaluate certainty in the evidence we used the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach [6], Supplemental Material 2 provides further details of the QUIPS instrument and the GRADE approach.

Data synthesis and statistical analysis

When required, NP levels and associated HR and 95% CI were converted to units of 100 pg/mL. Using an established formula [7], the NT-proBNP values reported were converted to BNP values. A subgroup analysis was performed to determine if converted BNP values provided different effect estimates in comparison to directly reported BNP values, and when there was no difference, their association with outcomes was addressed in a single meta-analysis.

Separate meta-analyses were conducted for the outcomes of all-cause mortality, composite of all-cause mortality and HF hospitalization, and composite of CV mortality and CV hospitalization. Predictors measured continuously versus categorically and predictors evaluated in log units versus those in natural units were meta-analyzed separately. A pooled HR estimate was calculated for each meta-analysis using the inverse variance method with random-effect models, considering a two-sided p-value < 0.05 statistically significant. Heterogeneity was evaluated using the I2 statistic and visual assessment of forest plots. Pre-defined subgroup analyses were performed on low risk of bias studies only. Supplemental Material 2 presents details regarding subgroup analyses. This review was prepared following the PRISMA guidelines [8].

Calculation of absolute risks

Absolute risk estimates were calculated for patients with varying levels of BNP using the calculated pooled relative risk estimates evaluating the effect of BNP (per 100-pg/mL increase) on all-cause mortality across a range of risks [9]. Baseline risks were informed by hypothetical integer scores from the MAGGIC risk prediction model [10].

Patient and public involvement

Patients were not involved in the design and conduct of this review.

Results

Study selection and characteristics

After screening 83,416 citations, 67 eligible articles reporting on 76,178 ambulatory HF patients were included (Fig. 1). Supplemental Table 1 summarizes the individual study characteristics. Included studies had median and IQRs as follows: duration of follow-up period 3.1 years (2.0–4.6 years), age 68 years (61–72 years), proportion females 28% (22–35%), LVEF 34% (30–37%), NYHA class III–IV 35% (25–50%), and BNP level of 407 pg/mL (261–574 pg/mL). The median proportion of patients using beta-blocker was 77% (59–89%), using ACEi/ARB 86% (73–93%), with implantable cardiac defibrillator 12% (6–27%) and with cardiac resynchronization therapy was 9% (5–25%).

Fig. 1
figure 1

Study selection process (2005–2018)

Full size image

Risk of bias of individual studies

Of the 67 studies included, we considered 21 to have a high risk of bias (Supplemental Table 2) [11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31]. The most common reason was limitations in statistical analysis and reporting, particularly overfitting of multivariable models, building the multivariable model according to significant variables in univariable analysis, and only reporting on significant risk factors.

Association of NPs with outcomes

All-cause mortality

Sixty-five studies evaluated the association of BNP or NT-proBNP with all-cause mortality. Of these, 18 studies with 5,925 events observed evaluated BNP continuously per 100 pg/mL (10 requiring conversion from NT-proBNP), 44 studies with 16,162 events observed evaluated either BNP or NT-proBNP per 1-log unit, and 3 studies with 1,404 events observed evaluated NT-proBNP categorically for values > 3,500 pg/mL (equivalent to 1,000 pg/mL BNP).

A 100-pg/mL increase in BNP was associated with a 14% increase in hazard of all-cause mortality (HR 1.14, 95% CI 1.06–1.22; moderate certainty due to inconsistency) (Table 1, Fig. 2). High risk of bias studies underestimated the association between BNP and mortality (HR 1.00, 95% CI 1.00–1.01) when compared to low risk of bias studies (HR 1.14, 95% CI 1.06–1.22) (p-value for subgroup interaction < 0.001). There was no difference in effect estimates for converted BNP values compared to directly reported BNP values. Other subgroup analyses were not feasible due to insufficient number of low risk of bias studies.

Table 1 Summary of findings table for all prognostic factors meta-analyzed
Full size table
Fig. 2
figure 2

Meta-analysis of studies reporting the association of BNP per 100-pg/mL increase with all-cause mortality

Full size image

A 1-log unit increase in NP level was associated with a 51% increase in hazard of all-cause mortality (HR 1.51, 95% CI 1.41–1.61; moderate certainty due to suspected publication bias) (Table 1, Fig. 3). The magnitude of association between NPs and mortality decreased with increasing prevalence of diabetes. To illustrate this association, in a cohort with 10% prevalence of diabetes, the hazard of all-cause mortality would increase from 1.51 to 1.65, whereas if the prevalence of diabetes were to increase to 60%, the HR would decrease to 1.28. There was no difference between studies at high versus low risk of bias. Other prespecified subgroup analyses did not change the association between NPs and all-cause mortality.

Fig. 3
figure 3

Meta-analysis of studies reporting the association of NPs per 1-log unit increase with all-cause mortality

Full size image

NT-proBNP > 3,500 pg/mL versus ≤ 3,500 pg/mL (equivalent BNP value of 1,000 pg/mL) was associated with a 55% increase in hazard of all-cause mortality (HR 1.55, 95% CI 1.20–1.98; high certainty) (Table 1, Supplemental Fig. 1). Publication bias and the proposed subgroup analyses were not performed due to the low number of studies.

Composite all-cause mortality and HF hospitalizations

Seven studies with 2,758 events observed evaluated the association of BNP or NT-proBNP continuously per 1-log unit with the composite of all-cause mortality and HF hospitalizations. A 1-log unit increase in NP level was associated with a 48% increase in hazard of the composite outcome (HR 1.48, 95% CI 1.29–1.69; high certainty) (Table 1, Supplemental Fig. 2). Subgroup analyses did not explain any of the observed heterogeneity and publication bias was not evaluated due to the small number of studies.

Composite cardiovascular mortality and cardiovascular hospitalizations

Five studies with 658 events observed evaluated the association of BNP with the composite of CV mortality and CV hospitalizations. A 100-pg/mL increase in BNP was not significantly associated with the composite outcome (HR 1.06, 95% CI 0.98–1.15; moderate certainty due to imprecision) (Table 1, Supplemental Fig. 3). Publication bias and subgroup analyses were not performed for this outcome due to the low number of studies.

Absolute effect of prognostic value of BNP levels

We applied the pooled relative effect estimates of BNP per 100-pg/mL increase on all-cause mortality (Fig. 2) to illustrate the corresponding absolute increases in risk across low (100 pg/mL) and high (1,500 pg/mL) levels of BNP (Table 2). We calculated baseline risks using the MAGGIC Risk Calculator, which ranges from 0 to 50 and does not incorporate the effect of BNP. For a patient with the lowest possible MAGGIC score of 0, the baseline risk of 1-year all-cause mortality is 1.5%. The absolute effect of increasing BNP for a patient with a MAGGIC score of 0 ranges from 1 to 6% from the lowest to the highest BNP values. For a patient with a MAGGIC score of 40, the baseline risk of 1-year all-cause mortality is 52%. The absolute effect of increasing BNP for a patient with a MAGGIC score of 40 ranges from 39 to 96% from the lowest to the highest BNP values.

Table 2 Calculated absolute risk estimates for various MAGGIC risk scores. For a patient with the lowest possible MAGGIC score of 0, the baseline risk of 1-year all-cause mortality is 1.5%. The absolute effect of increasing BNP for a patient with a MAGGIC score of 0 ranges from 1 to 6% from the lowest to the highest BNP values. For a patient with a MAGGIC score of 40, the baseline risk of 1-year all-cause mortality is 52%. The absolute effect of increasing BNP for a patient with a MAGGIC score of 40 ranges from 39 to 96% from the lowest to the highest BNP values
Full size table

Discussion

Main study findings

Our meta-analyses highlight the prognostic value of both BNP and NT-proBNP in the setting of adult ambulatory patients with HF. Moderate to high certainty evidence shows a 100-pg/mL increase in a population with median BNP level of 407 pg/mL is associated with a 14% increased risk of all-cause mortality, that a 1-log-unit increase is associated with a 51% increased risk of all-cause mortality and 48% increased risk of the composite all-cause mortality or HF hospitalization. Moderate to high certainty evidence also shows that NT-proBNP levels > 3,500 pg/mL compared to ≤ 3,500 pg/mL are associated with a 55% increased risk of all-cause mortality.

The estimation of impact on absolute risks demonstrates how clinicians can use these results to evaluate prognosis in ambulatory patients with HF. For patients with high estimated baseline risk of 1-year all-cause mortality based on non-NP factors, small changes in NP can have a substantial impact on overall risk. In patients whose estimated risk is low, larger changes in BNP levels will translate into smaller changes in risk that may still be clinically meaningful. These gradients could likely have an important bearing on patient management.

Relation to previous work

Current guidelines recommend the use of BNP or NT-proBNP to establish prognosis in chronic HF [32]. The current AHA statement regarding the use of BNP for prognosis is informed by the 2005 systematic review by Doust et al. [33], which summarized only four studies and found a 35% (95% CI 22–49%) increase in the relative risk of all-cause mortality per 100-pg/mL increase. The authors did not assess the quality of evidence or estimate the impact on absolute risk and included studies reporting both unadjusted and adjusted effects, introducing the possibility of overestimation of the effect of BNP. Our meta-analysis included a larger number of high-quality studies, leading to a moderate to high level of certainty in the effect of NPs on adverse outcomes.

Recent work has demonstrated the prognostic value of adding NPs to existing risk prediction models for estimating the risk of mortality in patients with HF. The addition of BNP or NT-proBNP to the MAGGIC risk score resulted in significant improvements in predicting mortality in both the ambulatory HF setting [34] and among patients admitted to hospital with HF [35, 36]. Similar improvements in prognostic performance have been observed with the addition of BNP to the Seattle HF Model (SHFM) [37, 38], and NT-proBNP to both the Get With The Guidelines HF risk score and the Intermountain Risk Score [39].

Despite consistent evidence demonstrating improved prognostic performance of risk models that include NP levels, current online risk calculators, including the MAGGIC risk score and SHFM, have not been revised to incorporate NPs for use in the clinical setting. A new risk prediction model, derived from the Prospective Comparison of ARNI with ACEI to Determine Impact on Global Mortality and Morbidity in HF (PARADIGM-HF) trial population [40] and externally validated in 2 independent cohorts, now includes NT-proBNP and is available as an online calculator [41].

Our study provides further evidence on the prognostic value of BNP, which supports its routine assessment and incorporation in risk prediction models for evaluating prognosis among ambulatory patients with HF. Our review demonstrates that NPs should be considered in the context of other clinical risk factors, as changes in NP levels have a varying effect depending on a patient’s baseline risk (considering non-NP factors).

Limitations

Our search of the literature was conducted in October 2018; therefore, it may not include more recently published studies. Given the large scope of literature evaluated in this review, we believe that the studies included in the current analyses adequately inform the pooled effect estimates reported, and it is unlikely that any additional studies published would significantly alter the conclusions from our analysis. We also combined BNP and NT-proBNP effect estimates into single meta-analyses where appropriate. This was done by either converting continuous NT-proBNP per 100 pg/mL into BNP units, or by combining BNP and NT-proBNP when reported in log units. We did this to utilize as much available evidence to inform our meta-analyses and subgroup analysis of studies reporting converted BNP values and directly reported BNP values that did not demonstrate any difference in association with all-cause mortality.

Despite observing some heterogeneity in our analyses, subgroup analyses according to patient BMI, eGFR, and proportion of patients with AF, which are factors known to influence measures of BNP, were not feasible as these data were too infrequently reported. Additionally, we did not perform subgroup analysis according to proportion of patients on sacubitril/valsartan due to this data not being available in most studies. However, results from the PARADIGM-HF trial have shown consistent prognostic importance of NP values in patients with HF with reduced ejection fraction treated with sacubitril/valsartan. The authors demonstrated that, despite significant reduction, NP levels during follow-up were significantly associated with cardiovascular mortality or first hospitalization for HF in adjusted models.

One of the challenges in testing the association between NP levels and outcomes is related to its exponential increase. Assuming a linear association when in fact there is an exponential increase could lead to the overestimation of predicted risk for each unit change in predictor. In the case of BNP and NT-proBNP, logarithmic transformations are applied to allow for such analyses to be performed. However, these effect estimates cannot easily be applied into a clinical setting. An alternative is to report the effect of NP levels categorically; however, this reduces the granularity to assess their association with a given outcome. For example, NT-proBNP levels > 3,500 pg/mL versus ≤ 3,500 pg/mL were associated with a 55% increase in the risk of all-cause mortality. This creates uncertainty on how different the risk could be in a patient with a NT-proBNP level of 1,000 pg/mL in comparison to a patient with a level of 100 pg/mL.

Implications in clinical practice and research

The findings from our review support the use of either BNP or NT-proBNP when assessing prognosis and stratifying risk in patients with HF. Current guidelines, however, remain without any specific thresholds or instructions to use these biomarker levels for clinical management or for delineating patient risk. This can be attributed in part due to the heterogeneity in current RCTs that aim to discern clinical benefit for NP guided therapy compared to standard of care. Future trials may need to evaluate targeting BNP levels based on their impact on overall prognosis and not just simply a value.

Our results also support the incorporation of NP values into HF risk prediction models. Current risk models for predicting 1-year mortality in ambulatory HF patients do not typically include NP measures (HF Meta-Score [42], Seattle HF Model [43], MAGGIC score [10], 3C-HF [44]), and those that do (BCN Bio-HF [45]) only include them as optional variables. The addition of NPs to existing risk prediction models has demonstrated increased discrimination and calibration of these multivariable models to predict adverse HF outcomes [26, 38]. Incorporation of BNP or NT-proBNP into current and future risk scores may provide the most accurate interpretation of NPs for risk assessment in the context of known factors associated with HF outcomes and factors associated with increased levels of NPs (e.g., age, sex, LVEF, diabetes, obesity, and renal dysfunction).

Conclusion

With moderate to high certainty in the evidence, this systematic review and meta-analysis highlight the prognostic value of both BNP and NT-proBNP in the setting of adult ambulatory patients with HF. We also provide measures of absolute risk to aid physicians in applying these findings to the clinical context to better inform patient management, facilitate patient communication, and perform shared decision-making.