Predictive accuracy of population meropenem models in critically ill pneumonia patients

Population pharmacokinetic (PK) models may not generalize across critically ill populations. We evaluated the transportability of published parametric and nonparametric meropenem models as Bayesian priors. Seven published meropenem models were implemented in Monolix 2024R1. Nonparametric models were converted to log-normal priors from reported medians and CV%. An independent cohort was used for evaluation. A priori and a posteriori predictions were summarized across eligible observations and separately for observations obtained during continuous renal replacement therapy (CRRT) and in the absence of CRRT. Predictive performance was assessed using relative mean prediction error (rMPE) and relative median absolute prediction error (rMdAPE). An rMPE within ±20% and an rMdAPE ≤30% were defined as acceptable. Seventeen patients, including six receiving CRRT, contributed 74 plasma concentrations. In the pooled cohort, only the Rohani and Shekar models met predefined performance thresholds at both population and individual levels (a priori rMdAPE 24.6%-26.4%; a posteriori rMdAPE 6.7%-11.1%). In CRRT patients, no model achieved acceptable a priori performance; after Bayesian updating, all CRRT-specified models met accuracy and precision criteria, with the nonparametric models (Reed and Rohani) demonstrating low imprecision. In non-CRRT patients, several models showed population-level bias, but all met predefined accuracy and precision thresholds after Bayesian updating. Among evaluated models, Rohani, Reed, and Shekar demonstrated the most consistent predictive performance across the CRRT and non-CRRT subgroups. Nonparametric model summaries can be translated into parametric priors, but model selection remains critical when applying Bayesian dosing in critically ill patients.