Improving Mental Health Screening with Predictive and Generative Modeling of Text Messages

Tlachac, ML

Etd

Improving Mental Health Screening with Predictive and Generative Modeling of Text Messages

Public

Screening for mental illnesses is vital, but traditional screening questionnaires are susceptible to conscious and unconscious bias. In my dissertation, I explore the mental illness screening capabilities of retrospectively harvested text messages. Leveraging lexical category features derived from the text message content of crowd-sourced participants, I trained traditional machine learning models and evaluated their ability to screen for depression and suicidal ideation. For sent texts, I discovered the most recent weeks of texts were more predictive than greater temporal quantities like the last year of texts. I further constructed lexicons with less formal language to improve the depression screening models. For received texts, I identified the 25 percent most prolific contacts as the subset with the messages most predictive of depression. To mitigate privacy concerns, I also explore depression screening potential of text reply latencies and time series of communications. I then collect a new dataset with a larger quantity of call and text logs labeled with depression and anxiety screening scores. Deep learning was more effective at screening for lower score cutoffs while machine learning was more effective at screening for higher score cutoffs. Lastly, I explore the depression screening potential of generated text content. I identify and adopt nine different conditional approaches for sequence generation. I then conduct a comparative evaluation of their ability to generate text messages from depressed and not depressed participants. The transformer-based classifiers proved better able to screen for depression with texts generated by the unconditioned models than the conditioned models, revealing future research opportunities.

Creator