Machine learning for real-world data from digital mental health

Abstract

Mental healthcare demands are substantial worldwide, overwhelming an understaffed mental healthcare workforce. At the same time, recovery from mental health problems, and its interaction with treatment and person characteristics is still not well enough understood, which has negative consequences for people suffering from mental health problems. Real-world mental health data is data relating to a person`s mental health status, factors which may influence this status, and, potentially, treatment delivery, that is routinely collected, often using digital technology. With the advancing digitisation of daily life, and increasing availability of such data, real-world data have received increasing attention from mental health researchers. Researchers are particularly interested in the ecologically valid, detailed trajectories of mental health and its interaction with treatment for mental illness that these data promise to yield. Due to the volume and dimensionality that these data often have, researchers often call on machine learning methodology to process them. Such real-world data also accrue from the use of digital mental health interventions, which are researched as alternatives and adjuncts to face-to-face mental healthcare, depending on the severity of the mental illness. Real-world studies using data from such digital interventions are only beginning to be explored as a possibility to generate evidence about digital intervention use and users. In this thesis, I use real-world data from a digital substance dependence intervention presently available in UK addiction treatment services within three separate studies dealing with different aspects of digital intervention outcomes research. Specifically, I worked with questionnaire, and intervention module completion data from registrants with this intervention. The first study that I present in this thesis explores initially available data in order to identify associations between digital intervention outcomes, and person and intervention characteristics. In my second study, I evaluate the feasibility of behavioural engagement prediction. My third study comments on the feasibility of explainable outcome prediction modelling with machine learning models using mental health data more generally, which are often characterised by high feature set multicollinearity. Data from the digital substance dependence intervention serve as example data in this study. Documenting the lessons learned conducting these studies, I hope to contribute to characterising the potential and challenge of using real-world data from digital mental health interventions for research.

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	David Wong (Supervisor), Caroline Jay (Supervisor) & Iliada Eleftheriou (Supervisor)