Data synthesis over a network of multiple treatment comparisons for joint longitudinal and event-time outcomes
Related longitudinal and time-to-event data is often collected from multiple heterogeneous data sources (such as studies, hospitals or centres) to answer clinical questions. Using such data, it is of interest to simultaneously assess all available treatment options, rather than perform multiple pairwise comparisons. However such analyses can be complicated if each data source examines only a subset of available treatment options. My project will develop methodology linking the areas of joint modelling for longitudinal and time-to-event data and network meta-analyses to allow analysis of such scenarios. The methodology will be implemented in freely available software, which will be streamlined through use of computer science techniques such as Sequential Monte Carlo Samplers. Use of such methods will allow fast fitting of joint models of complex networks of multi-source data, allowing estimations to be efficiently updated based on new information. Analyses of a range of clinical areas will be conducted, including datasets concerning HIV, hypertension, and intensive care unit patients

Development and evaluation of machine learning approaches for genetic-based disease prediction in population biobanks
Genetic association studies aim to discover DNA variants associated with complex binary and continuous traits in order to better understand the causal mechanisms of these conditions and enable the development of new treatments and interventions. One of the early aims of genetic association studies was to use associated variants to enable personalised medicine based on an individual’s genetics. However, complex traits are typically influenced by many genetic variants all contributing a small effect on the trait, which limits the utility of individual variants in risk prediction. Genetic risk scores, which aggregate effects from many associated genetic variants, have been demonstrated to have some utility in risk prediction, but these methods assume only additive effects, can include only common variants, and are chosen using a crude p-value threshold following association testing.

Quantification of uncertainty within systems pharmacology to optimise personalised therapy
The current one-dose-fits-all viewpoint, dosing for the 'average' human, and empirical derivations of dosing regimens are clearly not sufficient to effectively treat a diverse population. Personalised medicine, with the determination of treatment plans influenced by genetic factors and historical cases of toxicity, represents a step towards improving clinical outcomes. However, when an individual is deemed not suitable for treatment with the standard medicine/dose based on pharmacogenetics, an alternative must be found which is likely to either be costlier, less effective or potentially toxic. This scenario can be improved by having a greater understanding of both the mechanisms of toxicity in subpopulations and a proper quantification of risk and efficacy. Mathematical modelling can help to improve the understanding of the mechanisms of toxicity, simulate different dosing scenarios and propose optimal treatment. Statistical techniques can be applied to define and quantify uncertainties within model output and define risk when translating to optimised treatment strategies incorporating different genetic and non-genetic factors associated with variability in drug-induced adverse reactions. The integration of applied mathematics and statistics in this project comprises a potentially exciting new field bringing together two quantitative modelling approaches to better quantify uncertainty and inform drug development, testing and therapy.

Approximation Approaches for Efficient Clinical Predictions
High-dimensional data is now routinely collected in many settings, due to the number of different variables being measured on a subject, the number of times a variable is measured and the number of individuals in a study. This data can be used for screening patients to determine their risk of disease, or to classify patients into risk groups.
In the machine learning literature variational Bayes approximation methods have been shown to give fast and accurate estimates of model parameters in a variety of settings.
My proposal is to derive a mean-field variational Bayes (MFVB) approach to estimating joint models for multiple longitudinal biomarkers of different types and time-to-event data. Making it practicable to fit such models in a reasonable time frame, will allow personalised risk-prediction and diagnostics to be performed in real-time.
This project will be at the interface of statistics and computer science with significant statistical, methodological, and computational components.