(PS2-A8) Examining the Quality of Future-oriented Thoughts Among Suicidal Adolescents Using Natural Language Processing

Suicidal thoughts and behaviors increase drastically during adolescence and are a prevalent concern (Cha et al., 2018; Nock et al., 2013). Current research is attempting to better predict suicidal ideation (SI) (e.g., Franklin et al., 2017; Ribeiro et al., 2016). One risk factor for SI is deficits in episodic future thinking, such that suicidal individuals display poor future thinking abilities (e.g., Cha et al., 2022; Williams et al., 1996). One way to assess the associations between future thinking and SI is through the Autobiographical Interview (AI; Levine et al., 2002). To assess the association between SI and future thinking, descriptions of future events are transcribed and manually coded following the AI coding procedures to yield an overall count of future event details. Since these procedures are completed manually, they can be highly time consuming. Addressing this limitation of the AI, van Genugten & Schacter (2022) introduced an automated scoring process using Natural Language Processing (NLP) to automatically score the AI as an equivalent index of episodic future thinking.

This current study aims to use the NLP-based machine learning model to predict SI. More specifically, we aim to 1) compare the number of NLP-generated future event details between suicidal adolescents (i.e., past year history of SI) and non-suicidal adolescents (i.e., no history of SI/attempt), and 2) use NLP-generated future event detail counts to predict follow-up SI 3- and 6- months later. We hypothesize that previously suicidal adolescents will display fewer future event details at baseline compared to non-suicidal adolescents, and that fewer future event details would predict SI in the future. Additionally, we hypothesize that details that are not relevant to the imagined future event are not predictive of SI.

Results based on a preliminary round of automated coding suggest that, contrary to the hypothesis, no statistically significant difference between internal event detail count for SI (M = 244.16, SD = 63.23) and non-SI (M = 288.22, SD = 76.37) groups was detected, t = 1.43, p = 0.16. There was a small effect size (Cohen’s d = 0.23). However, we observed a significant difference between external event detail count for SI (M = 84.27, SD = 58.20) and non-SI (M = 108.45, SD = 65.33) groups, t = -2.45, p = 0.02. There was a small effect size (Cohen’s d = -0.39). Additionally, and once again contrary to Hypotheses, average internal detail count did not significantly predict SI at 3-months, β = 0.05, p = 0.57, yet average external detail count significantly predicted SI at 3-months, β = -0.22, p = 0.02. Similarly, average internal detail count did not significantly predict SI at 6-months, β = 0.08, p = 0.40, yet average external detail count significantly predicted SI at 6-months, β = -0.26, p = 0.01. Of note, though not significant, the internal detail counts were in the right direction predicting for SI at both 3- and 6-months follow-ups.

Assessing future thinking, more specifically event details, has been a new direction to understand risk factors associated with suicidal thoughts and behaviors. This on-going study provides a space to further investigate the association between future thinking and SI.