From statistical significance to clinical relevance: A simple algorithm to integrate brain natriuretic peptide and the Seattle Heart Failure Model for risk stratification in heart failure

Abstract

Background: Heart failure (HF) guidelines recommend brain natriuretic peptide (BNP) and multivariable risk scores, such as the Seattle Heart Failure Model (SHFM), to predict risk in HF with reduced ejection fraction (HFrEF). A practical way to integrate information from these 2 prognostic tools is lacking. We sought to establish a SHFM+BNP risk-stratification algorithm.

Methods: The retrospective derivation cohort included consecutive patients with HFrEF at the Mayo Clinic. One-year outcome (death, transplantation or ventricular assist device) was assessed. The SHFM+BNP algorithm was derived by stratifying patients within SHFM-predicted risk categories (≤2.5%, 2.6% to ≤10%, >10%) according to BNP above or below 700 pg/ml and comparing SHFM-predicted and observed event rates within each SHFM+BNP category. The algorithm was validated in a prospective, multicenter HFrEF registry (Penn HF Study).

Results: Derivation (n = 441; 1-year event rate 17%) and validation (n = 1,513; 1-year event rate 12%) cohorts differed with the former being older and more likely ischemic with worse symptoms, lower EF, worse renal function and higher BNP and SHFM scores. In both cohorts, across the 3 SHFM-predicted risk strata, a BNP >700 pg/ml consistently identified patients with approximately 3-fold the risk that the SHFM would have otherwise estimated, regardless of stage of HF, intensity and duration of HF therapy and comorbidities. Conversely, the SHFM was appropriately calibrated in patients with a BNP <700 pg/ml.

Conclusion: The simple SHFM+BNP algorithm displays stable performance across diverse HFrEF cohorts and may enhance risk stratification to enable appropriate decision-making regarding HF therapeutic or palliative strategies.

Keywords: Seattle Heart Failure Model; biomarkers; heart failure; natriuretic peptides; prognosis; risk stratification.

Figure 1

Receiver operating characteristic (ROC) curves demonstrating the prediction accuracy of BNP for the composite outcome of death, cardiac transplantation or ventricular assist device implantation at one year in the derivation and validation cohorts. Circles indicate (1–specificity, sensitivity) for the BNP cut-point of 700 pg/mL.

Figure 2

Calibration histograms in each of derivation (D) and validation (V) cohorts. Patients are stratified within the pre-specified categories of clinically relevant one-year risk [low (≤2.5%), intermediate (2.6–≤10%), and high risk (>10%)] according to SHFM -predicted probabilities. One-year predicted SHFM mortality risk in each of the derivation cohort (grey bar) and validation cohort (pink bar). Observed composite one-year outcome of mortality, cardiac transplantation or ventricular assist device implantation in each of the derivation cohort (black bar) and validation cohort (red bar).

Figure 3

Calibration histograms in each of derivation (D) and validation (V) cohorts are shown with one-year predicted SHFM mortality risk in each of the derivation cohort (grey bar) and validation cohort (pink bar). Observed composite one-year outcome of mortality, cardiac transplantation or ventricular assist device implantation in each of the derivation cohort (black bar) and validation cohort (red bar). Patients are stratified within the pre-specified categories of clinically relevant one-year risk [low (≤2.5%), intermediate (2.5–≤10%), and high risk (>10%)] according to SHFM-predicted probabilities and subsequently according to BNP levels above (↑) or below (↓) the optimal partition-value (700 pg/ml) based on ROC-analysis in the derivation cohort.

Figure 4

Hazard ratios for the composite outcome of death, cardiac transplantation or ventricular assist device implantation between BNP ≥ 700 pg/mL and BNP < 700 pg/mL, stratified by categories of clinically relevant one-year risk (≤2.5%, 2.5–≤10%, and >10%) according to the SHFM. Sample sizes in each group are provided in Figure 3.

Figure 5

NT-proBNP sensitivity analysis within validation cohort. Calibration histograms shown: one-year observed mortality, cardiac transplantation or ventricular assist device implantation (red bar) and SHFM-predicted mortality risk (pink bar). Patients are stratified within the pre-specified categories of clinically relevant one-year risk [low (≤2.5%), intermediate (2.5–≤10%), and high risk (>10%)] according to SHFM -predicted probabilities and subsequently according to NT-proBNP levels above (↑) or below (↓) the optimal partition-value (NT-proBNP > 1100) based on ROC-analysis in the validation cohort.