Measuring and Modeling Moment-by-Moment Emotional Deescalation in Suicide Crisis conversations

Research aims and hypotheses: Individuals in crisis experience emotional dysregulation involving intense negative emotion. In this project, we propose to leverage a unique dataset in combination with cutting-edge natural language processing (NLP) techniques to identify conversational strategies related to emotional deescalation over the course of suicide crisis conversations. Crisis Text Line (CTL) has taken over 11 million crisis conversations with texters, about 20% of whom have reported suicidal ideation. We plan to complete the following aims: 1) Develop models that predict emotional de-escalation over the course of a crisis conversation in suicidal individuals. We will develop machine learning models that predict whether crisis conversations are likely to result in successful de-escalation, based on emotional valence trajectories during early portions of conversations. The hypothesis is that early improvement in emotional valence will predict successful de-escalation. 2) Identify crisis counselor statements that predict emotional de-escalation in suicidal texters. Using unsupervised topic modeling techniques, we will extract clusters of topics from counselor messages related to reductions in negative emotional valence in suicidal texters. We will develop machine learning models to test the hypothesis that application of these crisis counselor effective statements results in de-escalation.

Sample: 17,271 individuals at imminent risk for suicide who completed the survey, they were ages 13 or younger (19%); 14-17 (38%); 18-24 (22%); 25-34 (11%); 35-44 (5%); 45-54 (3%); 55-64 (1%); and 65+ (0.2%). 66% identified as girls/women; 12% as boys/men; 22% reported any non-conforming gender identity. For race/ethnicity, they were White (68%); Black or African American (15%); Asian (6%); Native American, Native Alaskan, or Indigenous (5%); Native Hawaiian or Pacific Islander (1.2%); Middle Eastern, North African or Arab (1.3%); Latinx or Hispanic (17%). 39% were straight or heterosexual; 36% were bisexual or pansexual; 11% were gay or lesbian; 8% were questioning or unsure, 5% were asexual, and 2% were queer. To summarize, this sample is young, racially diverse, and includes more than 10,000 LGBTQ+ individuals, a population at particularly high risk for suicide.

Measures: The Crisis Text Line dataset includes message-level (conversation) data between counselors and texters, logs related to emergency services intervention, self-reported texter and counselor demographics, and counselor post-conversation survey data, including indications of issues discussed and the suicide risk assessment.

Procedures: There will be no participant recruitment; this is an analysis of existing anonymized data.

Potential impact: This project would provide foundational funding that could lead to multiple R01-level project grants. In the long term, this research may [1] help shift the field of crisis intervention so that more effective conversational techniques are deployed at higher rates, and [2] contribute to basic knowledge of emotional escalation and de-escalation in suicidal individuals.

Future directions: Beyond the planned analyses, future projects could include identifying training strategies that promote or hinder crisis counselor acquisition and deployment of effective deescalation techniques, examining how deescalation evolves longitudinally in repeat texters, identifying similarities and differences in emotional deescalation in vulnerable subpopulations, and extending these analyses to Crisis Text Line’s bilingual Spanish service.