Mental health rating scales are used to diagnose and assess risk. They are often retrospective (e.g., ask about symptoms in the past), rely on an individual accurately interpreting the wording of items, and can miss key risk factors if not included in the questionnaire. When humans describe their symptoms using their own natural language, they can potentially describe many more risk factors in an ecological way. The challenge is using psychometric properties to reliably measure symptoms from text as has been done with rating scales.
I will present a series of methods for measuring psychological constructs from text including semi-automated ways of generating a lexicon for a given construct as well as few-shot learning and semantic-similarity approaches that use embeddings to overcome lexicons' need for an exact match. I report how well these methods measure 38 known risk factors for suicidal thoughts and behaviors, which highlight the need for explainable and fair assessments. I validate these measurements of risk factors by using them to predict the severity of a crisis counseling session using a dataset of over 100k sessions. I also test how well each method is able to quantify one of 12 types of crisis (e.g., suicide, sexual assault, anxiety, eating disorders) as annotated by counselors.
In a second study, I apply these methods to longitudinal data, an ecological momentary assessment study (N=106), and assess their reliability. Suicidal individuals responded to rating scales about their mental health and also described the specific content of their suicidal ideation daily for 22 days on average. Our results allow us to provide an answer to a fundamental question: what do people actually think about when they think about suicide? And we demonstrate how capturing the dynamics of specific thoughts over time can help assess risk.