Advancing Mental Health Research Using Data Science: Investigating Vulnerability to Depression with Language and Network Analysis

Abstract

Depression affects over 5% of the global population, yet treatments for depression are only effective in 30-50% of people. Changing the status quo requires an improved understanding of how depression manifests in real life. This thesis tackled methodological shortcomings in the literature to take the field beyond cross-sectional studies in order to develop new technology-based ways to gather rich and repeated data within individuals. We assessed language use patterns from social media and collected self-reported emotions through a series of ecological momentary assessment (EMA) studies. These methods were used to test predictions of depression as a complex and dynamic system allowing us to probe key aspects of network theory, language use, and co-morbidity. Language use on Twitter was shown to be only weakly predictive of depression and not suitable for individual predictions. By modelling mental health as a network, emotion network connectivity - based on EMA data ? was found to be primarily related to fluctuations in depression, rather than simply severity. Finally, we used longitudinal time-series data from Twitter as a proxy for EMA to measure longer term changes in network connectivity. Networks constructed from depression-relevant language were found to be more connected during depressive episodes. The studies presented in this thesis, therefore, evaluated depression as a complex system and provide new ways of understanding how depression manifests and changes over time.