Jonathan Kertawidjaja
Research Mentor(s): Aparna Ananthasubramaniam
Mentor Department:
Authors: Aparna Ananthasubramaniam, Jonathan Kertawidjaja, Silas Falde
Session: Session 4 (1:00pm – 1:50pm)
Presentation Type: Poster 120
Abstract
Suicide ranks among the leading causes of death for individuals aged 10 to 24 in the United States with over 5,800 youth suicides in 2022. Social media engagement has emerged as a potential factor contributing to youth suicide, as it is correlated with increased rates of anxiety and depression: two primary risk factors for suicide. Despite the wealth of information in databases like the Centers for Disease Control’s National Violent Death Reporting System (NVDRS), many existing variables in these narrative datasets inadequately address the role of online spaces in youth suicide death. With recent advancements in large language models (LLMs), there is significant potential for progress in the automation and accuracy of theme detection in narrative datasets. Applying LLMs to analyze NVDRS data can enrich our understanding of how online activities impact youth suicide and improve suicide prevention strategies. Leveraging zero-shot learning with LLMs, 1,279 narratives referencing online engagement by the decedent were identified with precision of 95.8, recall 94.5, and F1 95.1. Using a mixed-methods approach based on Durkheim’s theory of suicide and the Integrated Motivational-Volitional model, we conducted thematic analysis of 4,000 NVDRS narratives. From this analysis, we proposed 14 new variables derived from analyzing non-normative behaviors in online spaces that may influence youth suicide (for instance: sharing vulnerable content online, interpersonal conflicts that begin or progress online, producing explicit content, etc.). LLMs showed potential in classifying each variable, but they exhibited poor precision with certain variables like Private Sharing and Conflict. Overall, findings highlight LLMs’ utility in scalable qualitative analysis, prompting timely interventions and deepening insight into social media’s influence on youth suicide mortality. We recommend incorporating these newly identified variables into the NVDRS and improving social media documentation by law enforcement to enhance data accuracy related to suicide-related events.
