Modelling choices in pharmacokinetic and pharmacodynamic models

Thesis

Pharmacokinetic and pharmacodynamic (PKPD) models can be used to predict the benefits or toxicity from drugs, such as the neutropaenia, anaemia and thrombocytopaenia caused by chemo-therapeutic drugs. In this thesis, I will be exploring the methodology used to make modelling decisions in these PKPD models, and apply them to a commonly used model by Friberg et al. I will use these methods to compare different ways to model the noise observed in the data and to build a population-based model. Using Bayesian methods to determine model parameters, profile likelihoods to ensure identifiability, and Widely Applicable Information Criteria (WAIC) for model selection, I found that only multiplicative or constant Gaussian noise were fully identifiable and constant Gaussian noise had the greatest out-of-sample predictive power. These methods were also utilised to build a mixed-effects population model and I found that a single mixed-effects parameter had better out-of-sample predictive power, and including further parameters affected convergence of the Monte Carlo samplers in Bayesian inference. I also developed a more computationally efficient method for acquiring Profile likelihoods to determine practical identifiability. I propose that the methodology outlined in this thesis for making modelling decisions should be standard for a general PKPD case. However, there may need to be further work done to ensure it is suitable for all modelling needs.

Authors: Rumney, R

• DOI: 10.5287/ora-wrm07yqrx
• Type of award: MSc by Research
• Level of award: Masters
• Awarding institution: University of Oxford
• Language: English
• Keywords: PKPD; Bayesian; profile likelihoods; parameter identifiability; parameter estimation
• Subjects: Biology--Mathematical models; Parameter estimation; Mathematics