From data streams to mental health predictions:Improving the use of passive measures from digital devices

Poor mental health is a global concern, with the World Health Organization reporting that one in eight people suffer from a mental disorder. Identification and treatment are hampered by limited access to care and inadequate health insurance coverage. Digital technologies, such as smartphones, offer promising tools for improving mental health through continuous monitoring and timely intervention. These devices can collect rich data on various factors, such as social context and behavior, through active (e.g., questionnaires) and passive (e.g., GPS tracking) methods. Researchers often aim to use this passively collected data to predict mental health outcomes. Despite its potential, passive data collection is still evolving, and current predictive accuracy remains low to moderate. The overall goal of this thesis is therefore to optimize the use of passive measures from digital devices for predicting mental health outcomes.The first part of this thesis focuses on improving the accuracy of predicting mental health outcomes. Results show that combining passive and active data collection methods outperforms passive measures alone, but predictive performance remains low to moderate. Advanced machine learning models also show only moderate success in predicting variability in depressive symptoms. The second part of this thesis focuses on improving the transparency and reproducibility of studies using passive measures. It highlights key challenges researchers face and provides guidance for researchers working with passive measures, for example by proposing a preregistration template. Preregistration involves publicly outlining the study plan before the research begins, which can increase transparency and prevent bias.