Symptoms and Their Temporal Distribution: An Interpretable AI Approach for Depression Detection on Social Media1

Depression is a common mental disorder involving a depressed mood or the loss of pleasure for long periods, which can have serious societal and financial impacts. Social media-based depression detection is an effective method for facilitating early intervention to mitigate these impacts. Because decisions about whether to intervene based on such detection can have significant consequences for individuals, the reliability and interpretability of the detection process are essential. Although a few studies have explained such decisions based on the importance of linguistic or demographic features, these explanations do not directly relate to depression diagnosis criteria that are based on symptoms. To fill this gap, we developed a “focused temporal prototype network” (FTPNet) to detect depression and provide interpretations based on depressive symptoms as well as their temporal distribution. Extensive evaluations using large-scale datasets show that FTPNet outperformed comprehensive benchmark methods with an F1-score of 0.864. Our results also reveal fine-grained and emerging manifestations of depressive symptoms, such as “sharing admiration for a different life,” that are unnoted in traditional depression surveys like the Patient Health Questionnaire-9 (PHQ-9). We further conducted a user study to demonstrate improved interpretability over the benchmark. This study contributes to the information systems literature by introducing an interpretable approach to depression detection that models the temporal distribution of depressive symptoms. In practice, multiple stakeholders, such as social media platforms and volunteers, could apply our approach to identify depressed users and deliver targeted assistance.