Individualized reference values for NT-proBNP in chronic heart failure: a Bayesian method to enhance biomarker-based patient monitoring
A Hecksteden, R Mazzolari, I Kindermann, T Kessler, C WernerAbstract
Background
N-terminal pro-B-type natriuretic peptide (NT-proBNP) is integral in the diagnosis of chronic heart failure (CHF) and an important predictor of morbidity and mortality in CHF. By contrast, NT-proBNP-guided monitoring and treatment of CHF patients remains imprecise. A plausible reason for this discrepancy is the considerable inter-individual variation in NT-proBNP, which leads to wide reference ranges and increased uncertainty in the interpretation of individual measurements. Individualized reference ranges could be based on the patient’s own previous NT-proBNP values, however this requires many measurements. Bayesian methods enable the stepwise individualization of the uniform (“prior”) reference with few values from the specific patient. The resulting synergy between group-based information and individual data has been shown to improve diagnostic accuracy for other biomarkers.
Purpose
The present work aims to investigate the potential benefit of individualizing the interpretation of NT-proBNP values in CHF patients using a novel statistical approach.
Methods
Analyses are based on retrospective, longitudinal data from a university hospital heart failure unit. Clinical status (compensated vs. acutely decompensated) is used as criterion for establishing error rates of NT-proBNP-based classifications. A closed formula Bayesian updating strategy is used for individualizing the group-based distribution underlying the uniform reference range. A sample-specific group-based distribution serves as comparator to avoid overestimating the benefit of individualization. Beyond taking the upper bound of the reference state (compensated) as cut-off value, a Bayesian classifier is constructed, which considers the distributions of NT-proBNP for both clinical states. Adjusting for glomerular filtration rate (GFR) has been proposed to account for renal NT-proBNP excretion. Therefore, GFR-adjusted NT-proBNP values are used as additional comparator to represent predictor-based individualization. The area under the receiver operating characteristic curve (ROC-AUC) is used to compare individualized and uniform classifiers (DeLong’s test, significance level α<0.05). Variability in measured NT-proBNP concentrations is analyzed by mixed effects modeling. All analyses are carried out with log-transformed NT-proBNP values.
Results
689 datapoints (compensated: 444, acutely decompensated: 245) from 50 CHF patients are included. Inter-individual differences account for 55% of total variance in NT-proBNP concentrations from compensated time points (figure 1). The ROC curve for the individualized Bayesian classifier is consistently above the uniform comparator (ROC-AUC 0.953 vs. 0.896, p=0.006, figure 2). Adjusting NT-proBNP for GFR does not improve performance (ROC-AUC 0.848).
Conclusion
Individualizing reference values using Bayesian methods could improve the accuracy of NT-proBNP-guided monitoring and clinical decision-making in CHF patients.