Research
I am interested in a Bayesian technique called Variational Bayes. This is a faster alternative to the usual Markov Chain Monte Carlo (MCMC) methods used but is sometimes less accurate. My work involves deriving variational algorithms to use in a variety of health related applications, usually revolving around longitudinal data or risk prediction.
I'm interested in supervising PhD students in areas of Bayesian computation, longitudinal modelling or risk prediction.
I contributed to the analysis of COVID data for a number of projects, including analysis of in-hospital patients in the ISARIC cohort, and the assessment of SMART testing policies in collaboration with the CIPHA platform.
I previously completed a UKRI Innovation Fellowship funded by the MRC, investigating variational approximations in longitudinal data analysis. My research involves approximate computationally efficient methods of estimating mixed models for multivariate longitudinal data. I'm also working on jointly modelling longitudinal and survival data. Most of this work is focused on clinical risk prediction methods, and covers applications involving diabetes, epilepsy and liver cancer.
My postdoc research involved analysing longitudinal data and developing methods to correctly account for the correlation between biomarkers of various types. My main work is in methods of discriminant analysis to classify patients into disease risk groups. This was part of the MRC funded DiALog project.
Prior to this I completed my Ph.D in the Department of Mathematical Sciences. My PhD work was in the area of nonparametric regression and involved applying monotonicity constraints to regression estimates in psychometric studies.
Variational Approximations
Increasingly large amounts of data are being collected in clinical studies, with many variables being repeatedly measured over time. An ideal way to analyse this multivariate longitudinal data is through multivariate models. However, the complexity of these models and the computational burden in estimating the models mean that such methods are limited to relatively small datasets with reasonably small numbers of variables. Variational Bayes offers a potentially fast and efficient way of estimating large multivariate mixed models. Part of the cost is a potential loss in accuracy of the model parameter estimates. My research involves deriving variational approximations for multivariate mixed models and joint models for longitudinal and survival data. Part of the research involves assessing the accuracy of the model estimates and comparing them to existing methods.
Longitudinal Discriminant Analysis
I'm interested in ways of correctly modelling the correlation between various longitudinal profiles. In particular I am investigating how to account for correlation between markers of different types (eg continuous, binary and count data). We use multivariate generalised linear mixed models to achieve this. I'm interested in ways in which we can distinguish between groups of patients based on their characteristics. Our aim is to be able to allocate new patients into disease risk groups based on clinical biomarkers collected over time.
Nonparametric statistics
If the assumption that data follows a known distribution is hard to justify, nonparametric methods can be used to calculate regression models. I am interested in local polynomial kernel regression, and in particular adding constraints to regression estimates to enforce monotonicity for example.
Research grants
Data Accelerator Project (System P & AMR-X)
NHS CHESHIRE AND MERSEYSIDE ICB (UK)
December 2023 - October 2027
REliability of HRDD as a biomarker in Painful diabetic nEuropathy - a validation study (REPEL)
PROCTER & GAMBLE (USA)
July 2023 - December 2024
AMOUNT: A mixed methods investigation of the individual, sociocultural, and societal factors that underlie the recent increase in substance use among young people to inform policy
DEPARTMENT OF HEALTH & SOCIAL CARE (UK)
November 2021 - May 2023
Covid-SMART Release & Return
DEPARTMENT OF HEALTH & SOCIAL CARE (UK)
February 2022 - July 2022
Can existing risk models accurately predict cardiovascular disease in people with psoriasis and psoriatic arthritis?
THE PSORIASIS AND PSORIATIC ARTHRITIS ALLIANCE (UK)
September 2020 - June 2023
Variational Approximation Approaches for Efficient Clinical Predictions: NPIF Fellowship for David Hughes
MEDICAL RESEARCH COUNCIL
November 2017 - August 2021
Research collaborations
Matt Wand
Variational Approximations for Multivariate Mixed Models
University of Technology, Sydney, Australia
Deriving mean field variational bayes approximations for complex logitudinal and survival models.
Steven Zhao
Analysing the effects of smoking on axial spondyloarthritis in both cross-sectional and longitudinal analysis.
Dr Arnost Komarek
Charles University in Prague
Methods of Longitudinal discriminant analysis and corresponding R code in the package mixAK.
Dr Simon Maher
Analysis of Quadrupole Mass Spectrometers and their performance