External Cross-validation of Two Ciprofloxacin Population Pharmacokinetic Models in Patients in Intensive Care

Abstract

Background: The population pharmacokinetic (popPK) variability of ciprofloxacin in patients in intensive care units (ICUs) is unclear. Two popPK models of ciprofloxacin in those in the ICU were externally cross-validated to determine if a published popPK model can be applied for model-informed precision dosing or if a new popPK model needs to be developed. The predictive performance of the 2 popPK models was evaluated.

Methods: Data were collected from patients in the ICU at Amsterdam University Medical Center (AUMC), and a popPK model for ciprofloxacin was developed using nonlinear mixed-effects modeling. The data and the published pharmacokinetic model from the ICU of the Radboud University Medical Center (RUMC) were used for cross-validation. The RUMC dataset was used to externally validate the AUMC model and vice versa. The predictive performance of the models was evaluated by comparing the population-predicted and corresponding observed concentrations in the dataset. The primary endpoints were bias and precision, calculated as the mean percentage error (MPE) and normalized root mean squared error (NRMSE), respectively. Visual predictive checks (VPCs) and Bland-Altman plots visualized predictive performance.

Results: The AUMC dataset consisted of 159 concentration-time data points from 32 patients, and the RUMC dataset consisted of 531 samples from 39 patients. A 2-compartment linear model with modification of diet in renal disease as a covariate for ciprofloxacin clearance most accurately fit both study populations. The final AUMC model predicted the RUMC population data with an MPE of -3.87% (95% CI, -7.56 to -0.185) and an NRMSE of 44.05% (95% CI, 39.48-48.19). The final RUMC model predicted the AUMC population data with a nonsignificant MPE of -31.29% (95% CI, -73.56 to -10.98) and an NRMSE of 64.02% (95% CI, 48.61-76.38). pcVPC indicated acceptable predictive performance because the observed data fell within the 95% prediction CIs; the AUMC model overestimated the variability. The Bland-Altman plots confirmed that both models were imprecise, overrepresenting large negative relative errors.

Conclusions: Neither ciprofloxacin popPK model accurately predicted external data, and the AUMC model exhibited bias. The prior RUMC model is unsuitable for the AUMC ICU population and vice versa. We recommend either adapting an existing popPK model from literature or creating a new popPK model specifically tailored to the ICU population.

Keywords: ICU; ciprofloxacin; dosing; external validation; population pharmacokinetics.