Venelin Mitov (1), Daniel Lill (1), Mohammed Cherkaoui-Rbati (2), Anne Kümmel (1)
1. IntiQuan GmbH, Basel Switzerland; 2. Medicines for Malaria Venture, Geneva, Switzerland
Objectives: Simulation of models is a key element of model diagnostics as well as a popular method for model-based prospective decision making. In the context of NLME models, sampling of the model parameters before simulation is a key prerequisite ensuring that the simulation is consistent with the estimated parameter uncertainty, individual random variability, covariate effects and residual error of the model [1]. Despite its importance, to date there is neither a commonly accepted NLME parameter sampling tool, nor a specification for how NLME parameter sampling should be done. Instead, the parameter sampling is done in a rather adhoc way, differing between modeler and modeling project; the sampling procedures are scarcely reported and hard or impossible to reproduce. We present a unifying software approach for NLME parameter sampling.
Methods: Our proposed solution is based upon the following components:
- General Parameter Format (GPF): a new Excel file format providing a clear description of the NLME model estimation results and all additional information needed for sampling the parameters. Importantly, this format has been designed to be both programmable as well as readable by humans.
- A set of exporter R functions implementing the conversion of model estimation results from existing tools (e.g. NONMEM, MONOLIX) to GPF.
- A GPF constructor used to import a manually written or automatically generated GPF excel file, or to call an appropriate exporter function, based on its input.
- A set of sampler and calculator R functions implementing all steps from sampling population parameter values from the estimate’s uncertainty distribution to obtaining individual parameter values for a user-specified dataset of individual covariates.
- A set of validator R functions checking the consistency of the parameter samples with the GPF file.
Results: We present the above approach based on real life PKPD modeling examples. In particular, we illustrate how to create a GPF file (i) manually, based on a model published in the literature [2], and (ii) automatically, based on a NONMEM NLME project. For a well behaved NLME example, we show that our implementation provides parameter samples such that the sampled distributions are in line with the distributions defined. Using other examples, we highlight an important caveat concerning the sampling of positive-definite variance covariance matrix parameters. As the uncertainty of the elements of such matrix-parameters usually is reported as a multivariate normal distribution, there is an inherent risk of sampling invalid parameter sets, namely, negative-definite matrices. We present our current implemented approach to tackle this problem. Finally, we show how the validator functions can be used to detect and notify the modeler about mismatches between the sampled and specified parameter distributions.
Conclusions: GPF is a powerful yet human readable format for specifying NLME model parameters enabling easy sharing of NLME parameter estimation results independent of the used modeling platform. The implementation of a generic and transparent parameter sampling tool presents an important step towards improving the quality and reproducibility of NLME model simulations.
References:
[1] Kümmel, Anne & Bonate, Peter & Dingemanse, Jasper & Krause, Andreas. (2018). Confidence and Prediction Intervals for Pharmacometric Models. CPT: Pharmacometrics & Systems Pharmacology. 7. 10.1002/psp4.12286.
[2] Kleijn, Huub J et al. “Population pharmacokinetic-pharmacodynamic analysis for sugammadex-mediated reversal of rocuronium-induced neuromuscular blockade.” British journal of clinical pharmacology vol. 72,3 (2011): 415-33. doi:10.1111/j.1365-2125.2011.04000.x
Reference: PAGE 29 (2021) Abstr 9676 [www.page-meeting.org/?abstract=9676]
Poster: Methodology - Covariate/Variability Models