To Estimate or to Forecast? Lessons From a Comparative Analysis of Four Bayesian Fitting Methods Based on...

Using pharmacokinetic (PK) models and Bayesian methods in dosing software facilitates the analysis of individual PK data and precision dosing. Several Bayesian methods are available for computing Bayesian posterior distributions using nonparametric population models. The objective of this study was to compare the performance of the maximum a posteriori (MAP) model, multiple model (MM), interacting MM (IMM), and novel hybrid MM(HMM) in estimating past concentrations and predicting future concentrations during therapy. Amikacin and vancomycin PK data were analyzed in older hospitalized patients using 2 strategies. First, the entire data set of each patient was fitted using each of the 4 methods implemented in BestDose software. Then, the 4 methods were used in each therapeutic drug monitoring occasion to estimate the past concentrations available at this time and to predict the subsequent concentrations to be observed on the next occasion. The bias and precision of the model predictions were compared among the methods. A total of 406 amikacin concentrations from 96 patients and 718 vancomycin concentrations from 133 patients were available for analysis. Overall, significant differences were observed in the predictive performance of the 4 Bayesian methods. The IMM method showed the best fit to past concentration data of amikacin and vancomycin, whereas the MM method was the least precise. However, MM best predicted the future concentrations of amikacin. The MAP and HMM methods showed a similar predictive performance and seemed to be more appropriate for the prediction of future vancomycin concentrations than the other models were. The richness of the prior distribution may explain the discrepancies between the results of the 2 drugs. Although further research with other drugs and models is necessary to confirm our findings, these results challenge the widely accepted assumption in PK modeling that a better data fit indicates better forecasting of future observations.