Modelling pre-dose concentrations in steady-state data. The importance of accounting for between-occasion variability and poor adherence.
Paolo Denti (1), Helen McIlleron (1)
Division of Clinical Pharmacology, University of Cape Town
Objectives: In large Phase III/IV studies, data are often obtained from outpatients, who are under direct observation only on the day of their PK sampling. This leaves a large amount of uncertainty about the time of prior doses, since the information provided by patients may be imprecise or unreliable. Moreover, the PK of many drugs is subject to diurnal variation, meal-dependent absorption, and BOV. Thus, there is considerable variability in drug concentrations observed in the pre-dose samples. This variability needs to be properly accounted for in the model, even if the underlying causes cannot be measured because they are not related to the observed PK profile.
Methods: Using parameter values based on a published model of nevirapine [1], concentrations from a once-daily dosing regimen were simulated for 250 patients, with a rich sampling schedule (8 samples during 12 hours after dose), including a measurement 30 minutes before dose. Commonly encountered scenarios causing "unexplained" variability in the pre-dose concentrations were explored in the dataset: poor or no adherence to treatment, discrepancy between the actual and the reported dosing times, and large BOV in the PK. Different approaches to model such data were tested: inclusion or exclusion of the pre-dose samples, use of the assumption C0=C24 , introduction of BOV in bioavailability (BOVBIO), or baseline estimation using methods similar to those explained by Dansirikul et al. [2]. Results from 500 reiterations of stochastic simulation and estimation were analysed at population and individual level.
Results: In our simulations, when C0 is assumed to equal C24, either discarding the pre-dose samples or including them without properly accounting for the additional variability they are subject to, led to inflated estimates for the BSV of ka and CL/F (+29% for exclusion, +40% for inclusion respectively) and the RUV additive component (+60% or +173% bias). Introduction of BOVBIO did not solve these issues. The estimate of BOVBIO was also spuriously inflated, because it accounted for reduced adherence. The baseline estimation methods performed best, especially B2 and B4, reducing the bias of the TV and BSV parameters within 11%. At individual level, all baseline approaches greatly improved the fit for subjects whose pre-dose concentration was significantly different from the expected C24 value. In particular, more accurate estimates of individual clearance were obtained.
Conclusions: Variability in pre-dose concentrations should be appropriately modelled to avoid overestimation of RUV, and BSV of the PK parameters. Using a baseline estimation method can satisfactorily overcome this problem and greatly improve the model fit for poorly adherent or "anomalous" subjects. Further investigation will be necessary to compare the proposed methods and fathom the effect of factors such as the size of measurement error and the sampling schedule.
References:
[1] de Maat, M. M., Huitema, A. D., Mulder, J. W., Meenhorst, P. L., van Gorp, E. C., Beijnen, J. H., et al. (2002). Population pharmacokinetics of nevirapine in an unselected cohort of HIV-1-infected individuals. British journal of clinical pharmacology, 54(4), 378-85.
[2] Dansirikul, C., Silber, H. E., & Karlsson, M. O. (2008). Approaches to handling pharmacodynamic baseline responses. Journal of pharmacokinetics and pharmacodynamics, 35(3), 269-83.
