Population Pharmacokinetic Model Evaluation with a Small Real-World Dataset Versus a Large Virtual Dataset: Does Sample Size Affect Decision-Making?

Background and objective: In a recent simulation-based study, we found that sample size had minimal influence on the external evaluation of population pharmacokinetic (PK) models. However, the applicability of these findings to clinical data remains unexplored. This study aims to validate our previous simulation-based results using real-world clinical data.

Methods: Data from a prospective clinical study in the > 75-year-old population admitted to the McGill University Health Center (MUHC) receiving piperacillin/tazobactam were collected. A virtual population of 1000 patients representative of the characteristics of MUHC patients was also simulated. A population PK model was externally evaluated both using the small clinical dataset and a larger simulated dataset. The predictive performance of the model was assessed using bias, imprecision, goodness-of-fit plots (GOF), and prediction-corrected visual predictive checks (pcVPC). The distribution of prediction errors between the clinical and simulated datasets was compared using the Wilcoxon rank-sum test.

Results: Data from 13 patients undergoing piperacillin/tazobactam therapy were collected. The Ishihara et al. model showed low bias (2.4% population, 0.5% individual) and imprecision (23.8% and 3.2%) and was therefore chosen for Monte Carlo simulation of the virtual population. The Hemmersbach-Miller et al. model showed bias values of - 37.8% (population) and - 21.4% (individual), with imprecision values of 43.2% (population) and 31.3% (individual) for the clinical dataset. For the simulated population, bias values were - 28.4% (population) and - 13.9% (individual), with imprecision values of 40.2% (population) and 18.1% (individual). No significant difference was observed between the prediction error distributions of the clinical and simulated datasets. Both GOF plots and pcVPCs showed similar model misspecification across the clinical and simulated datasets.

Conclusions: This study confirms that small clinical datasets may be used to externally evaluate population PK models.