External evaluation of population pharmacokinetic models for vancomycin in neonates

Abstract

IntroductionNumerous vancomycin population pharmacokinetic (PK) models of neonates have been published. We aimed to comparatively evaluate a set of these models by quantifying their model-based and Bayesian concentration prediction performances using an external retrospective dataset, and estimate their attainment rates in predefined therapeutic target ranges.\n\nMethodsImplementations of 12 published PK models were added in the Bayesian dose optimisation tool, DosOpt. Model based concentration predictions informed by variable number of individual concentrations were evaluated using multiple error metrics. A simulation study assessed the probabilities of target attainment (PTA) in trough concentration target ranges 10-15 mg/L and 10-20 mg/L.\n\nResultsNormalized prediction distribution error analysis revealed external validation dataset discordances (global P < 0.05) with all population PK models. Inclusion of a single concentration improved both precision and accuracy. The model by Marques-Minana et al. (2010) attained 68% of predictions within 30% of true concentrations. Absolute percentage errors of most models were within 20-30%. Mean PTA with Zhao et al. (2013) was 40.4% [coefficient-of-variation (CV) 0.5%] and 62.9% (CV 0.4%) within 10-15 mg/L and 10-20 mg/L, respectively.\n\nConclusionPredictive performances varied widely between models. Population based predictions were discordant with external validation dataset but Bayesian modelling with individual concentrations improved both precision and accuracy. Current vancomycin PK models achieve relatively low attainment of commonly recommended therapeutic target ranges.