Bayesian forecasting of serum vancomycin concentrations in neonates and infants

Abstract

A dynamic pharmacokinetic model for i.v. vancomycin administration was developed and tested in 47 neonates and infants. Twenty-nine patients (Group 1), having two or more concentrations, were used to estimate population parameters by nonlinear least-squares analysis. Multiple stepwise linear regression techniques showed that estimated creatinine clearance, Clcr, and postnatal age were significant demographic factors related to vancomycin clearance (CL). No strong associations were found for the apparent volume of distribution. A one-compartment model was constructed using the associations of CLcr and postnatal age with vancomycin CL. Eighteen patients (Group 2), receiving 35 courses of vancomycin therapy, with both initial and subsequent sets of peak and trough concentrations, were used to test the predictive performance of the model with and without the use of Bayesian forecasting. Using only population-based parameters, the respective mean error (ME) (bias) and mean absolute error (MAE) (precision) for predicting subsequent peak concentrations were -1.20 and 3.89 mg/L and for trough concentrations, 0.83 and 2.23 mg/L, respectively. For the Bayesian method, these values were, respectively, 0.45 and 4.13 mg/L for peak concentrations and 1.55 and 2.40 mg/L for trough concentrations. When predicted concentrations occurred within 30 days of feedback concentrations, the Bayesian method tended to be slightly less biased and more precise than the population-based parameters. The opposite was true > 30 days of the initial set of feedback concentrations. The use of population-specific pharmacokinetic parameters and Bayesian forecasting should allow accurate dosage regimen design as well as minimize the need for monitoring serum vancomycin concentrations in neonates and young infants.