Monitoring depressive symptoms using social media data

Abstract

Social media data contains rich information about one's emotions and daily life experiences. In the recent decade, researchers have found links between people's behavior on social media platforms and their mental health status. However, little effort has been spent on mapping social media behaviors to the psychological processes underlying the psychopathological symptoms. Identifying these links may allow researchers to observe the trajectory of the illness through social media behaviors.

The psychological processes examined in this thesis include affective patterns, distorted cognitive thinking and topics relevant to mental health status. In the first part of the thesis, we conducted two studies to explore methods to extract affective patterns from social media text. We demonstrated that mood fluctuations and mood transitions extracted from social media text reflect an individual’s depressive symptom level. In another study, we demonstrated that the affect from content not written by social media users themselves, such as quotes and lyrics, also reflects depressive symptoms, but the implications from these are different from content written by the users themselves.

In the second part of the thesis, we identified distorted thinking from social media text. We found that these thinking patterns have a higher association with users' self-reported depressive symptom levels than affect extracted from users' text. In the last part of the thesis, we manually compiled topic dictionaries related to suicidal ideations according to the psychopathology literature. We found that users' suicidal risk levels can be estimated by using these topics. The estimation can be improved by combining these topics with results from a language model.

The data-driven empirical studies in this thesis demonstrated that we can characterize the social media signals in a way that impacts our understanding of mental disorder symptoms. We blended data-driven methods such as machine learning, natural language processing and data science with theoretical insights from psychology.