You're Making Me Depressed: Leveraging Texts from Contact Subsets to Predict Depression

Depression, a prevalent and debilitating mental illness, is frequently undiagnosed. Diagnosis is an important step towards treatment. Currently screening tools, such as the Patient Health Questionnaire-9 (PHQ-9), require patient input. Many studies have used a variety of data types and features to predict depression scores for individuals. In this study, we focus on the predictive ability of a single under-utilized modality indicative of the impact of social interactions: received text messages. Our approach encompasses creating subsets of influential contacts for each participant and engineering features from the text messages of those contact subsets. Overall, our study demonstrates that received text communications are a promising modality when predicting depression scores. Specifically, we found that the F1 score of Gaussian Naive Bayes models leveraging just the text messages from a subset of top contacts performed statistically significantly better by 13.2 percent than the models leveraging text messages from all contacts.