Chihiro Hasegawa (1), Mari Shiomi (1), Hiroyuki Yoshitsugu (1)
MSD (Merck Sharp and Dohme), Japan
Introduction: Identifiability of models is an important component to be considered when estimating parameters. Structural identifiability is concerned with the uniqueness of the model parameters. For example, the bioavailability (F) cannot be distinguished only with concentration-time data after extravascular dosing. Such parameters are called unidentifiable and need to be fixed during model development. Shivva, et al. explored the property of an informal, information theory-based approach for identifiability analysis using a parent-metabolite modelling as one of the case examples where the volume of distribution for the central compartment of the metabolite (named V3) is unidentifiable [1]. One of the findings was that even when the fixed-effect parameter (THETA) was unidentifiable, its corresponding between-subject variability (OMEGA) was identifiable. This is what is often considered; e.g., the between-subject variability (BSV) for F is estimated (as a relative F) even only with concentration-time data after extravascular dosing. It is however still not clear if the estimated value is still reasonable in a real setting.
Objectives: Examine the accuracy and precision of BSV estimates for unidentifiable parameters through stochastic simulation-estimation approach in a real setting. We use a parent-metabolite modelling as a case example.
Methods: As a typical design of first-in-human phase 1 studies, intensively sampled profiles were simulated for 5 different dose levels (1, 3, 10, 30 and 100 mg) of 6 subjects each as single dose (over 72h). Two-compartment models were used for both parent and metabolite, while no first pass effect was considered (i.e., 1st order absorption only for the parent) for simplification. Two elimination pathways were considered for the parent, one converting to the metabolite and the other (e.g., urinary excretion), thus the volume V3 is unidentifiable in this case due to the lack of mass balance information. Parameter values were adopted from Duffull, et al. [2]. All BSVs were set at 30% (except the parameters related to peripheral compartments where no BSV was set), residual errors for both parent and metabolite at 50%. NONMEM®[3] 7.4 with FOCE-I was used for parameter estimation. The accuracy and precision of estimates were assessed using the relative estimation error (REE = (Est-True)/True). The ETA shrinkage for the BSV of V3 was also explored to assess the ability of exploring covariates on the parameter.
Results: In the stochastic simulation-estimation process, more than 95% of the replications converged successfully. The median of REE (reflecting the accuracy) for the BSV parameter of V3 was approximately 50%, while those for other parameters less than 50%. The variance of REE (reflecting the precision) was higher for the BSV parameter of V3 (inter quarter range of REE from 0 to 100%) compared to others. The median of ETA shrinkage for the BSV of V3 was the highest among parameters (>30%).
Conclusions: These findings suggest that while the BSV for unidentifiable parameters can be estimated, caution should be made with the appropriateness of the estimated BSV value and the possibly biased relationship between the corresponding posthoc ETA and covariates.
References:
[1] Shivva V et al. CPT:PSP (2013) 2, e49.
[2] Duffull SB et al. Eur J Pharm Sci (2000) 10, 285-94.
[3] Beal SL et al. 1989-2011. NONMEM Users Guides. Icon Development Solutions, Ellicott City, Maryland, USA.
Reference: PAGE () Abstr 9353 [www.page-meeting.org/?abstract=9353]
Poster: Methodology - Covariate/Variability Models