External Evaluation of Vancomycin Population Pharmacokinetic Models at Two Clinical Centers

Background: Numerous vancomycin population pharmacokinetic models in neonates have been published; however, their predictive performances remain unknown. This study aims to evaluate their external predictability and explore the factors that might affect model performance. Methods: Published population pharmacokinetic models in neonates were identified from the literature and evaluated using datasets from two clinical centers, including 171 neonates with a total of 319 measurements of vancomycin levels. Predictive performance was assessed by prediction- and simulation-based diagnostics and Bayesian forecasting. Furthermore, the effect of model structure and a number of identified covariates was also investigated. Results: Eighteen published pharmacokinetic models of vancomycin were identified after a systematic literature search. Using prediction-based diagnostics, no model had a median prediction error of ≤ ± 15%, a median absolute prediction error of ≤30%, and a percentage of prediction error that fell within ±30% of >50%. A simulation-based visual predictive check of most models showed there were large deviations between observations and simulations. After Bayesian forecasting with one or two prior observations, the predicted performance improved significantly. Weight, age, and serum creatinine were identified as the most important covariates. Moreover, employing a maturation model based on weight and age as well as nonlinear model to incorporate serum creatinine level significantly improved predictive performance. Conclusion: The predictability of the pharmacokinetic models for vancomycin is closely related to the approach used for modeling covariates. Bayesian forecasting can significantly improve the predictive performance of models.