Analyzing Natural Language for Real-Time Risk Identification and Communication

Depression, anxiety, suicidality, and intimate partner violence are leading addressable causes of mortality and morbidity. With early identi?cation, there are effective interventions to address these kinds of psychosocial risks. However, obstacles like stigma and limited access to qualified providers mean typical screening procedures do not catch every occurrence. As patient-generated text becomes more common, natural language processing techniques can estimate semantic content like affect in writing, opening another route to risk detection. This thesis identi?es methods to supplement traditional risk detection procedures like self report measures and discussion with clinicians with computational analysis of reflective written language. I apply natural language processing techniques to extract information like emotional or thematic content from people's writing, then customize and train models to estimate writers' psychosocial risks. Two of the studies additionally present methods for elicitation of original data for this purpose, with both datasets coming from pregnant and postpartum respondents. Computational risk detection models could expand the scope of mental health screening beyond doctors' offices and formalized measures to anywhere a person in need of help decides to reach out. In addition, identifying those language features which best indicate elevated risk could extend practitioners' understanding of the ways that that risk can appear, potentially providing or reinforcing heuristics for clinicians to screen for risk during normal appointments.