Understanding, Modeling and Leveraging Temporal Change in Mental Health
This event is free and open to the publicAdd to Google Calendar
Virtual: Zoom Password: laura
Abstract: Mental health conditions such as depression affect hundreds of millions of people, and the prevalence of these conditions has grown as the COVID-19 pandemic has swept across the world. The global mental health crisis, along with the proliferation of related text data online, has led to the emergence of a community that applies Natural Language Processing techniques to aid mental health treatment, detection, and monitoring.
In this dissertation, we focus on leveraging the fact that signals of human behavior unfold over time to build models of mental health. Then, we quantify temporal changes in the presentation of mental health symptoms to better understand communities and individuals with various conditions. Temporal change can be viewed as a challenge, making it harder to build robust models; however, we believe that it can be regarded as an opportunity for insight, as it can lead to enhanced understanding. In the right circumstances, temporal change can also be leveraged and incorporated during the modeling process to improve the robustness of the models.
We explore how the targeted selection of time periods for noisy training data allows us to build classifiers that better generalize to out-of-domain data. Using location-based data, we build representations that can be used to classify student’s mental health status. We then shift towards the goal of better understanding temporal change in mental health. We study how the COVID-19 pandemic led to major changes in the conversations and community structure on mental health forums, and propose a technique to account for seasonal and longitudinal variation to add to the robustness of our findings. We find that at an individual level, announcing a depression diagnosis online corresponds with a reduction in linguistic signals that are indicative of depression. Finally, we study how metrics connected to mental health vary over a one-year period, and how that variation is connected to seasonal depression and many linguistic patterns.
This thesis crosses disciplinary bounds, providing insights that can be built upon by computer scientists and social scientists in the pursuit of improving global mental health.