Effect of dataset characteristics on estimation method performance: a TMDD model example
Richard Dimelow, Chiara Zecchin, Stefano Zamuner, Monica Simeoni
GlaxoSmithKline
Objectives: Expectation maximisation (EM) algorithms included in NONMEM from version 7.0 have shown increased numerical stability and reduced parameter bias in comparison to the traditional gradient based algorithms, FO and FOCE, in a selection of models [1, 2]. The aim of this work is to compare the performance of the FOCE and importance sampling EM methods in relation to different dataset characteristics. The target mediated drug disposition (TMDD) model [3] was chosen as the data descriptor.
Methods: As a reference, the model application by Ng and colleagues [4] has been selected. Phase 1-like datasets were simulated in different scenarios derived by the combination of low (15%) and high (30%) proportional residual variability, rich and sparse (50% of rich scheme) PK and PD sampling, and dose levels from a partial (2 dose levels) or full (3 dose levels) dose range. The variance of the inter-subject variability of the log-normally distributed parameters was estimated from the data. Argument settings for the importance sampling method were selected according to [2]. The comparison between estimation methods has been based on the following metrics: estimated versus true parameter values, standard error estimates of each parameter, number of iterations to convergence, and run time.
Results: The IMPMAP algorithm (importance sampling algorithm assisted by mode a posterior) was selected over the IMP algorithm for model parameter estimation, due to a more stable route to convergence. Both the FOCE and IMPMAP (with ISAMPLE=1000) algorithms successfully converged, and estimated well the fixed effect and residual model parameters, in all tested datasets. Run times were comparable between the FOCE and IMPMAP methods.
Conclusions: IMPMAP showed superior convergence stability over IMP when fitting a TMDD model to a PK/PD dataset. FOCE and IMPMAP performed equally well in the tested noise levels and sampling schemes.
References:
[1] Johansson, A. M., et al. Evaluation of bias, precision, robustness, and runtime for estimation methods in NONMEM 7. JPKPD (2014) 41:223-238
[2] Sahota, T and Johnson B, Efficient argument settings for NONMEM 7 expectation maximisation methods. A. M. PAGE 2015
[3] Mager DE, Jusko WJ. General pharmacokinetic model for drugs exhibiting target-mediated drug disposition. J Pharmacokinet Pharmacodyn. 2001 Dec;28(6):507-32.
[4] Ng CM et al. Pharmacokinetics/pharmacodynamics of nondepleting anti-CD4 monoclonal antibody (TRX1) in healthy human volunteers. Pharm Res. 2006 Jan;23(1):95-103
