| dc.description.abstract | There is increasing interest within different fields of medicine in developing and validating prediction models that can aid healthcare professionals in providing valuable insights into disease risks, diagnosis, prognosis and treatment response. Specifically, precision medicine models are designed to consider individual patient characteristics, biomarkers, and other pertinent data to estimate specific outcomes for each patient.
Various applications for prediction models in diabetes clinical practice have been deployed that improve patient care and management. These models are utilised as diagnostic and disease progression tools, tools to assess the risk of major diabetes complications in patients with diabetes and as glycaemic control prognostic tools. However, several challenges are encountered when developing prediction models for use in diabetes clinical practice. In this thesis, we investigate the utility of Bayesian modelling to help address some key challenges encountered when developing precision medicine models for diabetes, building upon and motivated by existing applications developed within the Exeter Diabetes group.
Firstly, we focus on developing and validating treatment selection models in observational data, which aim to estimate individualised treatment responses to competing second-line therapies for patients affected by Type 2 diabetes. We utilise Bayesian hierarchical models to help provide a powerful and flexible means of imputing missing data when developing such models. In our case, we combine a previously developed treatment selection model to help decide between SGLT2 and DPP4 inhibitors treatments, and we utilise non-parametric Dirichlet process mixture models as a way of modelling the complex joint distribution of the predictor variables, which can then be leveraged to impute any missing predictor variables. This modelling framework can help improve the fitted model's robustness but also, crucially, provides a way of producing predictions for new individuals even if their predictor information is incomplete, which is important in clinical practice. We also explore how such models can be used to inform further data collection.
We then explore recent advances in Bayesian non-parametric modelling aimed specifically at estimating treatment effects, so-called Bayesian Causal Forests, which provides a flexible and powerful means of estimating treatment effects whilst adjusting for clinical indication bias in observational data. We apply these ideas to develop a novel Type 2 diabetes treatment selection model that estimates individualised treatment response for two competing treatments, SGLT2 inhibitors and GLP1-receptor agonists. We also assess how targeting the treatment based on the optimal glycaemic benefit is associated with 12-month weight change, 6-month tolerability and 5-year risk of new-onset macro- and microvascular outcomes.
Finally, we focus on challenges when developing prediction models for rare diseases – Maturity-Onset Diabetes of the Young (MODY). In this case, suitable data sources that contain sufficient and complete information on population-representative patient characteristics to aid clinical diagnosis may be challenging to acquire. The low disease prevalence means that population-representative datasets often have small numbers of cases, making robust inference about risk factors difficult. Case-control studies are often used to identify disease risk factors for rare diseases, but these can only produce relative measures of risk in the form of odds ratios. In contrast, it is often desirable in clinical practice to produce absolute measures of disease risk in the general population. We investigate different approaches for adjusting disease risk prediction from a model developed in case-control data to a population-representative dataset. The approaches investigated are divided into three groups: approaches that only utilise a case-control dataset and adjust predictions to a known population prevalence; approaches that utilise a case-control dataset and an additional population representative dataset; and approaches that utilise additional data on informative biomarkers. The last group takes advantage of a Bayesian hierarchical model to build biologically plausible constraints into model predictions based on the prior information about the biomarkers. Although the model structure is built around these biomarkers, the model still allows for predictions to be made for patients whose biomarkers are yet to be measured.
Although the benefits highlighted in this thesis focus on problems encountered in diabetes, the implementation of these methods was done in general-purpose software and could be translated to other settings providing similar benefits. | en_GB |
