Shijun Wang, Yaqing Zhang, Andrew C. Hooker, Mats O. Karlsson
Department of pharmaceutical bioscience, Uppsala University
Objectives: It is recognized that nonlinear mixed effect (NLME) models are a more appropriate representation of data from multiple subjects than naïve pooled data (NPD) models. However, given that NPD models have shorter runtimes and are more robust, there may exist an advantage to use them for model building decisions when NLME models are too run-time intensive or when high robustness is sought, as in automatic model building algorithms. However, the use of NPD models as proxies for NLME models depend on the similarity one can expect in the relative goodness-of-fit of contenting models. Therefore, this study aims to explore how similar model selection is between NLME and NPD models.
Methods: NLME and NPD were compared for 13 previously developed population pharmacokinetic (PPK) models based on real data. Each developed model was structurally divided into 4 parts, which contained the following components: oral absorption delay, absorption rate, distribution and elimination. For the 13 original models, 42 test models were generated by iteratively changing one of the components. The test models and original models were fit to the corresponding real data using both NLME and NPD methods, followed by the calculation of the difference of objective function value (OFV) and Akaike information criteria (AIC) values between each test model and its related original model for NLME and NPD separately (the OFV or AIC of test models minus original models). Model selection criteria was then performed using the likelihood ratio test (5% significance level) as well as choosing the lowest AIC value of the compared models. In a second step, simulation studies were performed to test the sensitivity of the NLME and NPD methods to identifying different model structures. A “full” PPK model with 2-compartment distribution kinetics, non-linear elimination and a transit-compartment first-order absorption model was used to simulate relatively densely sampled data. By varying parameters in the “full” model, characteristics of the model could be emphasized or hidden. For example, by varying the Vmax and Km in the Mechaelis-Menten elimination kinetics, one could mimic first order or zero order elimination, given the dose amount. Similarly, by varying the inter-compartmental clearance Q, one could hide or emphasize the 2nd elimination phase of the “full” model. Relevant reduced models (one-compartment, linear elimination, no transit compartment, zero order absorption) as well as the “full” model were then fit to the simulated data and the ability of the NPD and NLME methods to detect the true “full” model were compared using OFV and AIC values as in the real data examples.
Results: In the comparison of real data, the range of difference in OFV were -101 ~ 2775 and -52 ~ 2332 for NLME and NPD respectively and the model selection of the two methods was consistent for 38 out of 42 test models. For AIC, the model selection was consistent between NLME and NPD for 36 out 42 test models. In 3 out of the 4 differences between NLME and NPD in the OFV comparison and in 4 out of the 6 differences in the AIC comparison, the differences occurred when the test model and original model differed in the structure of the distribution model, which indicated that differences might exist in the selection of the distribution model when using NLME or NPD. For simulation data sets, the average difference in OFV when comparing a 2-compartment (“full”) model and a 1-compartment alternative were -224 (-310 ~ 0.06) and -62 (-152 ~ 45) for NLME and NPD respectively, showing the relatively lower power for NPD to identify a more complex distributional model.
Conclusions:NPD can act as an aid for NLME in PPK model building in terms of model structure selection. Nevertheless, the selection of the distribution model by NPD needs extra attention.
Reference: PAGE 27 (2018) Abstr 8732 [www.page-meeting.org/?abstract=8732]
Poster: Methodology - New Modelling Approaches