Uncertainty in model-informed precision dosing: predicting what you can and what you cannot predict

Thesis

Recent advances in Big Data, Machine Learning (ML), and Artificial Intelligence (AI) have affected almost all phases of the drug development life cycle, promising to reduce its extremely high attrition rates. However, drug development remains challenging, especially for those drugs displaying narrow therapeutic windows and large treatment response variability. For such drugs, dosing regimens often need to be individualised in order to balance the trade-off between their efficacy and their toxicity. ML & AI methods, among other modelling approaches, are increasingly used to support this dosing regimen individualisation – an approach known as model-informed precision dosing (MIPD).

In this thesis, we study aspects of the responsible use of MIPD in clinical practice, maximising its positive outcomes, while mitigating its risks. In this context, we particularly focus on the estimation of predictive uncertainty. We work in a Bayesian inference framework, where predictive uncertainty is accredited to the lack of information, resulting in structural and parametric uncertainty of models.

We begin our study by reviewing popular MIPD approaches, including pharmacokinetic & pharmacodynamic (PKPD) modelling as well as ML & AI approaches, such as regression and reinforcement learning. We then develop a simulation framework for clinical MIPD trials, enabling the development and evaluation of targeted MIPD approaches, while avoiding the astronomical costs of real clinical trials and guaranteeing the safety of patients throughout the development. Using this simulation framework, we find that MIPD approaches have different strengths and limitations, which demonstrates the importance of tailoring the MIPD approach to the use case. Overall, we find that for typical applications, where the availability of clinical data are limited, PKPD models provide the most responsible approach to MIPD – notably, PKPD modelling is not an ML or an AI approach.

We develop a new inference approach for population PKPD models, which we call ‘filter inference’. Filter inference reduces the computational costs of nonlinear mixed effects (NLME) inference, making it applicable to datasets with measurements from thousands of patients, such as those obtained during post-market monitoring. The use of post-marketing data has the potential to improve estimates of inter-individual variability, which would improve uncertainty estimates during MIPD. We also study the strengths and limitations of information criteria for the estimation of structural uncertainty of PKPD models, identifying future directions for research in responsible MIPD.

As part of this thesis research, an open source Python package has been developed (https://chi.readthedocs.io), facilitating the responsible use of MIPD.

Authors: Augustin, D

• Alternative title: Uncertainty in MIPD
• DOI: 10.5287/ora-7e8mmwwa1
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: NLME inference; machine learning; treatment response optimisation; treatment response individualisation; experimental design; PKPD modelling; Markov Chain Monte Carlo; filter inference; model-informed precision dosing; mechanistic modelling; uncertainty estimation; Bayesian inference; nonlinear mixed effects inference; entropic pressure; artificial intelligence; dosing regimen optimisation; numerical optimisation; clinical trial simulation; predictive uncertainty; MIPD; information criteria; parametric uncertainty; MCMC; structural uncertainty
• Subjects: Artificial intelligence; Bayesian inference; Model-informed precision dosing; Mechanistic modelling