Impact of Factors Associated with Short-Term Transplant Outcomes

Outline of ThesisIn Chapter 2 we discuss the relevant background information relating to both the clinical and methodological sides of the project. We begin by describing end stage renal disease, the transplantation process, types of kidney donors, and the relevance of the treatment withdrawal to death (also referred to as the agonal) phase. Chapter 2 then proceeds by providing the methodological background, beginning by making the clear distinction between explanatory and predictive modelling. Relevant statistical learning theory is then covered, which is particularly relevant for benchmarking machine learning methods in Chapter 3. Chapter 2 concludes with a discussion relating to missing data in longitudinal studies and its relation to the joint modelling framework, where the joint model is introduced.In Chapter 3 we introduce various machine learning methods (adaptive boosting, extreme gradient boosting, random forests and conditional random forests). A simulation study is then performed to assess the ability of these methods to rank the importance of predictor variables based on various importance metrics when faced with data complications present in the motivating dataset (such as multicollinearity, variables with many categories and a hierarchical structure). The NHSBT dataset is then analysed using the methods introduced, and their performance is benchmarked against random intercept logistic regression. Finally, we propose the use of various methods of visualisation that can be used when multiple imputation is performed.The multivariate Bayesian joint model (MBJM) is introduced in chapter 4 and extensions that are relevant in our application are described. In particular, we discuss the handling of multiple longitudinal covariates, allowing a more elaborate parametrisation of the longitudinal covariates in the joint model (allowing the hazard to depend on functions of the current biomarker value, such as the current gradient) and also relaxing the assumption of a constant association between the (possible function of the) biomarker value and the risk of event over time. We formally describe how the joint model can be used to perform dynamic prediction and explain measures of discrimination and calibration that account for the dynamic nature of the problem at hand. Finally, we conduct the analysis and apply the discussed methods to the novel dataset to predict DCD donor event times in the treatment withdrawal period.In Chapter 5 a two-stage approach for deriving summaries of characteristics of the physiological variables in the treatment withdrawal period that are in turn used as covariates in another regression model to predict recipient transplant outcome is investigated. A simulation study is conducted to study how inferential properties of this approach compare to those of the alternative two-stage approach using a linear mixed effects model (LMEM) instead of the joint model. We also investigate how both of these approaches compare to simply using functions of the observed physiological trajectories, which we expect to suffer from bias resulting from measurement error. The novel dataset is thenformally analysed and the model derived to improve our understanding of how charac- teristics of the treatment withdrawal period relate to recipient outcome is interpreted.This thesis concludes with a final discussion in Chapter 6, where final conclusions are drawn, assumptions and limitations are discussed and potential scope for future work is outlined