Predicting changes in depression using person-generated health data

Depression is a leading cause of disability, impacting the lives of an increasingly large number of individuals worldwide. Despite a range of treatment options, a substantial fraction of individuals experiencing depressive symptoms do not seek or receive treatment. Person-generated health data (PGHD), including self-reported data and consumer-grade wearable technology, can be used to observe individual behaviour and improve outcomes in depression. In this thesis, we propose PSYCHE-D (Prediction of SeveritY CHange - Depression), a two-phase classication model that predicts longitudinal change in depression severity using sparse depression severity labels, self-reported survey responses and consumer wearable data, trained on a large and diverse cohort comprising more than 10,000 samples from more than 4,000 individual participants. In the first phase, the model enhances the density of depression severity labels by generating intermediate monthly labels. The generated monthly labels, combined with the existing labels, self-reported survey responses and consumer-grade wearable data are used as inputs for the second phase, where PSYCHE-D predicts whether an individual experiences an increase in depression severity over a 3-month period, achieving a sensitivity of 55.4% and a specicity of 65.3%. We demonstrate a promising PGHD-based approach that can be used as a foundation for a low-burden consumer-facing system that could minimize barriers to depression diagnosis and treatment, and eventually improve outcomes in depression.