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

Area of Science:

• Pharmacokinetics and Pharmacodynamics
• Computational Biology and Bioinformatics
• Clinical Pharmacology

Background:

• Pharmacokinetic (PK) models and Bayesian methods in dosing software are crucial for analyzing individual PK data and enabling precision dosing.
• Several Bayesian methods exist for computing posterior distributions using nonparametric population models, but their comparative performance in clinical settings requires further investigation.

Purpose of the Study:

• To compare the performance of four Bayesian methods—maximum a posteriori (MAP), multiple model (MM), interacting MM (IMM), and hybrid MM (HMM)—in estimating past and predicting future drug concentrations.
• To evaluate these methods using amikacin and vancomycin PK data from older hospitalized patients.

Main Methods:

• PK data from 96 patients (406 amikacin concentrations) and 133 patients (718 vancomycin concentrations) were analyzed using BestDose software.
• Two strategies were employed: fitting the entire dataset and using each therapeutic drug monitoring occasion to estimate past and predict future concentrations.
• Bias and precision of model predictions were compared across the four Bayesian methods.

Main Results:

• Significant differences in predictive performance were observed among the four Bayesian methods.
• The IMM method demonstrated the best fit for past amikacin and vancomycin concentrations, while MM was the least precise.
• MM excelled in predicting future amikacin concentrations, whereas MAP and HMM showed similar performance and were more suitable for predicting future vancomycin concentrations.

Conclusions:

• The study challenges the assumption that a better fit to past data necessarily leads to better forecasting of future drug concentrations.
• The richness of the prior distribution may influence the discrepancies observed between amikacin and vancomycin results.
• Further research with diverse drugs and models is needed to validate these findings and refine Bayesian methods for precision dosing.