Introduction
Mental illness is a leading cause of disability worldwide. It is estimated that nearly 300 million people suffer from de-pression (World Health Organization, 2001). Reports on lifetime prevalence show high variance, with 3% reported in Japan to 17% in the US. In North America, the probabil-ity of having a major depressive episode within a one year period of time is 3–5% for males and 8–10% for females (Andrade et al., 2003).
However, global provisions and services for identifying, supporting, and treating mental illness of this nature have been considered as insufficient (Detels, 2009). Although 87% of the world’s governments offer some primary care health services to tackle mental illness, 30% do not have programs, and 28% have no budget specifically identified for mental health (Detels, 2009). In fact, there is no reliablelaboratory test for diagnosing most forms of mental illness; typically, the diagnosis is based on the patient’s self-reported experiences, behaviors reported by relatives or friends, and a mental status examination.
In the context of all of these challenges, we examine the potential of social media as a tool in detecting and predict-ing affective disorders in individuals. We focus on a com-mon mental illness: Major Depressive Disorder or MDD1. MDD is characterized by episodes of all-encompassing low mood accompanied by low self-esteem, and loss of in-terest or pleasure in normally enjoyable activities. It is also well-established that people suffering from MDD tend to focus their attention on unhappy and unflattering infor-mation, to interpret ambiguous information negatively, and to harbor pervasively pessimistic beliefs (Kessler et al., 2003; Rude et al., 2004).
People are increasingly using social media platforms, such as Twitter and Facebook, to share their thoughts and opinions with their contacts. Postings on these sites are made in a naturalistic setting and in the course of daily ac-tivities and happenings. As such, social media provides a means for capturing behavioral attributes that are relevant to an individual’s thinking, mood, communication, activi-ties, and socialization. The emotion and language used in social media postings may indicate feelings of worthless-ness, guilt, helplessness, and self-hatred that characterize major depression. Additionally, depression sufferers often withdraw from social situations and activities. Such chang-es in activity might be salient with changes in activity on social media. Also, social media might reflect changing so-cial ties. We pursue the hypothesis that changes in lan-guage, activity, and social ties may be used jointly to con-struct statistical models to detect and even predict MDD in a fine-grained manner, including ways that can comple-ment and extend traditional approaches to diagnosis.
Our main contributions in this paper are as follows:
(1) We use crowdsourcing to collect (gold standard) as-sessments from several hundred Twitter users who report that they have been diagnosed with clinical MDD, using the CES-D2 (Center for Epidemiologic Studies Depression Scale) screening test.(2) Based on the identified cohort, we introduce several measures and use them to quantify an individual’s social media behavior for a year in advance of their reported on-set of depression. These include measures of: user en-gagement and emotion, egocentric social graph, linguistic style, depressive language use, and mentions of antidepres-sant medications.
(3) We compare the behaviors of the depressed user class, and the standard user class through these measures. Our findings indicate, for instance, that individuals with de-pression show lowered social activity, greater negative emotion, high self-attentional focus, increased relational and medicinal concerns, and heightened expression of reli-gious thoughts. Further, despite having smaller egonet-works, people in the depressed class appear to belong to tightly clustered close-knit networks, and are typically highly embedded with the contacts in their egonetwork.
(4) We leverage the multiple types of signals obtained thus to build an MDD classifier, that can predict, ahead of MDD onset time, whether an individual is vulnerable to depression. Our models show promise in predicting out-comes with an accuracy of 70% and precision of 0.74.
We believe that this research can enable new mecha-nisms to identify at-risk individuals, variables related to the exacerbation of major depression, and can frame directions on guiding valuable interventions.