The article explores the impact of model-misspecification on dose recommendations and clinical outcomes from personalised dosing.
Model Based Personalised Dosing (MBPD) requires a population pharmacokinetic (PK) or pharmacodynamic model to determine the optimal dose of medication for an individual. Often several models are published, and the decision of which model is implemented in MBPD may have a large impact on its clinical utility. As quoted by Box, “all models are wrong, the practical question is how wrong can they be and still be useful”. Voriconzole, a triazole antifungal and the example used in this manuscript, currently has nine population PK models published. To assess the impact of model-misspecification on MBPD, five structurally mis-specified models for voriconazole were developed. Intensive plasma concentrations were simulated for 100 virtual subjects. The dose adjustments required to reach a target exposure were determined by using the empirical Bayes estimates of the PK parameters under each of the mis-specified models. The predicted plasma concentrations and the probability of clinical outcomes, upon following the dose recommendations, were determined. Models that did not contain non-linear clearance performed poorly, with a median dose recommendation 155–310 mg higher than appropriate doses, when only one plasma concentration was available. Removal of body weight and CYP2C9 genotype as covariates had no clinically significant impact on outcomes. In summary, the removal of important structural components, such as non-linear clearance in the case of voriconazole, had a large impact on the clinical utility of MBPD. The removal of patient covariates, even highly influential covariates such as CYP2C9 genotype for voriconazole, had no clinical impact.