Identifying Signs of Depression on Twitter

Abstract

Affecting 350 million people worldwide, depression is a source of serious costs to both personal and public well-being. However, most individuals with depression do not receive proper treatment. One possible way to address this issue is through automatic screening of individuals for depression. In this study we explore the possibility of automatically detecting individuals with depression through their behavior on Twitter. We build off of previous studies in the following ways. First, we identify a group of depressed and non-depressed users through Twitter’s API. Next, we explore the effect of additional features—topic-based features, removing retweets, and standardization of tweets—on the predictive ability of the logistic regression and SVM models to differentiate between depressed and non-depressed users. Using data collected at two different time points, we test the robustness of our model over time and over different incidence rates. Finally, we explore the possible implementations of a screening tool based on our work. Our findings demonstrate, first, the feasibility of discriminating depressed from non-depressed users through Twitter’s API, and second, the relative robustness of our model over time, although our additional features do not make a large difference in predictive power. However, our classifier’s performance significantly decreases when the incidence rate of the testing dataset is decreased to a more realistic level of 7.6%. This finding indicates that the incidence rate of depression in the training and testing datasets is an important additional factor to consider in future studies. Overall, although much work remains to be done before such a tool could be implemented, our work provides additional evidence that it may indeed be feasible to identify individuals struggling with depression through their behavior on Twitter.