This retrospective study evaluated the diagnostic value of B-type natriuretic peptide (BNP) for acute heart failure (AHF) among patients presenting with acute dyspnea and developed prediction models that integrate BNP with routine clinical variables. A key aim was to determine whether BNP retains diagnostic accuracy across different levels of renal function assessed by estimated glomerular filtration rate (eGFR).
Patients included were those admitted to the Emergency Department of Peking University First Hospital between January 2019 and June 2023 whose primary or initial symptom was acute dyspnea. A total of 605 patients were enrolled; 236 (39.0%) were diagnosed with AHF and 369 (61.0%) were non-AHF.
Demographic information, clinical history, and laboratory parameters were collected as candidate predictors. Renal function was categorized by eGFR into stages (the manuscript reports analysis in eGFR ≥60 and eGFR <60 ml·min⁻¹·(1.73 m²)⁻¹ subgroups for BNP performance assessment).
Thirty-four candidate variables were subjected to least absolute shrinkage and selection operator (LASSO) regression to identify variables associated with AHF diagnosis. LASSO selection yielded seven variables retained for further analysis and model building: history of diuretic use, diabetes mellitus, prior cardiac dysfunction, persistent atrial fibrillation, acute myocardial infarction, eGFR stage, and BNP level.
Single-variable receiver operating characteristic (ROC) analysis of the selected predictors identified BNP as the single most discriminatory variable for AHF. Reported performance for BNP was: area under the curve (AUC) 0.949 (95% CI 0.931–0.967), optimal cutoff 488 ng/L, sensitivity 80.9%, specificity 96.7%, and Youden index 0.777.
When stratified by renal function, BNP AUCs were 0.944 (95% CI 0.902–0.987) for patients with eGFR ≥60 ml·min⁻¹·(1.73 m²)⁻¹ and 0.938 (95% CI 0.913–0.962) for those with eGFR <60 ml·min⁻¹·(1.73 m²)⁻¹. The difference between these subgroup AUCs was not statistically significant (P = 0.793), indicating BNP maintained high diagnostic value regardless of the eGFR subgroup reported.
Using the seven variables selected by LASSO, five predictive models were developed for AHF diagnosis: logistic regression, random forest, AdaBoost, Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost). Model performance was evaluated using ROC AUC, calibration (including Hosmer–Lemeshow testing and calibration curves), and decision curve analysis (DCA). The DeLong test was used for comparing differences in AUC across subgroups.
Reported AUCs for the five models were:
Among these, the logistic regression model demonstrated the best overall discrimination. Reported operating characteristics for the logistic model were sensitivity 87.7%, specificity 91.6%, and Youden index 0.793. Hosmer–Lemeshow testing, calibration curves, and decision curve analysis indicated acceptable calibration and potential clinical utility for the logistic regression model in the study cohort.
Shapley additive explanation (SHAP) analysis was applied to assess model interpretability and variable importance. BNP was identified as the most influential predictor across the model(s), with the highest mean absolute SHAP value reported (4.278), supporting the dominant role of BNP in the predictive framework.
In this single-center retrospective cohort of emergency patients with acute dyspnea, BNP showed high diagnostic accuracy for AHF, and its performance did not differ significantly between patients with preserved versus reduced eGFR as defined in the study. A logistic regression model combining BNP with routinely available clinical indicators (history of diuretic use, diabetes, prior cardiac dysfunction, persistent atrial fibrillation, acute myocardial infarction, and eGFR stage) yielded the best diagnostic performance, good calibration, and indicated clinical utility by DCA.
The authors conclude that BNP retains diagnostic value for AHF across renal function levels and that a BNP-centered logistic model may serve as a practical adjunct to emergency diagnosis of AHF. Details on external validation, prospective testing, or implementation were not reported in the abstract.