Student Mental Health Screening with Text Message Metadata

Passive mental health screening can facilitate earlier detection and treatment of mental illnesses. Text log distribution features have previously proven promising for mental illness screening, providing detailed information regarding communication patterns while maintaining user privacy. In this research, we extract reply latency, consecutive message, and conversation ratio distribution features from a college student population for the first time, using them in machine learning models to screen for social anxiety and depression. We also compare screening ability of the text log distribution features to a set of more basic conversational features. Tree ensembles achieved a balanced accuracy of 0.81 (sensitivity = 0.92, specificity = 0.68) for depression screening with consecutive message distribution feature set and a balanced accuracy of 0.78 (sensitivity = 0.69, specificity = 0.86) for social anxiety screening with a basic conversational feature set. Our results indicate that text message logs can be a valuable modality for student mental health screening.