MODELLING OF LONGITUDINAL DIGITAL HEALTH DATA TO UNDERSTAND UNDERLYING PHENOTYPES

Abstract

In this thesis, efforts have been put into understanding mental health by taking a quantitative approach to mobile health data. The focus has been on achieving two objectives: 1) novel application of cutting-edge statistical learning methods to longitudinal health data; 2) development of a Bayesian approach to model-based clustering of time series of categorical variables. In Chapter 1, we introduce topics in mental health, which is a relevant issue in modern healthcare. In Chapter 2, we provide brief introductions to few of the methods and other tools which have been used throughout the thesis. Chapters 3, 4 and 5 can be read independently as they are written in academic paper format, and each of these chapters includes a specific abstract at the beginning of the paper. In Chapter 3, we perform residual analysis and clustering of mood-pain trajectories on the basis of transitions taken from a longitudinal study where we discovered four digital phenotypes for the behaviour of moving from one mood-pain state to another. In Chapter 4, we perform Bayesian inference on the same data considered before in the previous chapter. We assume the data to be distributed multinomially and taking Dirichlet distribution as a conjugate prior, we use Hamiltonian Monte Carlo method to sample estimates of the model parameters. In doing so, we also address the problem of label-switching in the mixture model, and ultimately build a distribution over transition matrices. In Chapter 5, we consider all the self-reported symptoms, not just mood and pain, and implement dimensionality reduction to investigate the relationships amongst those. Final two chapters include further prospects of the thesis, and conclusion.

Date of Award	31 Dec 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Mark Muldoon (Supervisor), Mark Lunt (Supervisor) & Thomas House (Supervisor)