Linguistic Social Media Markers of Suicide Risk in Youth Admitted to the Emergency Department

Adolescent suicide rates have been steadily increasing in recent years. Pediatric emergency departments (ED) often serve as a primary site of contact for adolescents with acute suicidal thoughts and behaviors. Risk of suicide death is heightened in the period directly after ED discharge, with roughly 5-10% of youth making an attempt within 90 days. These statistics highlight the need to move beyond static risk factors toward dynamic, person-specific models of suicide prediction and prevention.

Leading theories of suicide, including diathesis-stress models, emphasize the role of interpersonal stress and suicide-related cognitive distortions in triggering suicidal thoughts and behaviors (STB). Further, with many adolescent interactions now occurring on social media, detecting specific online exposures may be a promising avenue for understanding STB risk. Bypassing limitations of self-report measures of social media experiences used in most prior work, recently developed machine learning tools developed by our team are capable of reliably analyzing linguistic features from social media, including online bullying, social support, and suicide-related language (e.g.kill myself, kms, wanna die), by extracting activity logs from platforms like Instagram and TikTok. Although these linguistic measures have been linked to depression and suicidal ideation in adults, no research has yet explored whether they longitudinally predict short-term changes in STB risk among youth.

Directly addressing these gaps, the current study will enroll adolescents (N=80; ages 13-17) admitted to the pediatric ED for suicide concerns. At baseline, participants will complete clinical assessments and upload social media data (e.g., Instagram, TikTok), including all public (e.g., posts) and private (e.g., direct message) communications sent and received in the 6 months prior to the ED visit. Using protocols validated in our work, participants will complete weekly experience sampling for 3 months to assess suicidal ideation. At the 3-month follow-up assessments, youth will provide social media data and complete clinical assessments. We will build on recent advances in natural language processing (NLP) to best capture interpersonal stressors (i.e., online bullying, social support) and suicide-related cognitive distortions (i.e., suicide language use) on social media using transformer-based models. To test a diathesis-stress model, we will evaluate whether interpersonal stressors and suicide-related cognitive distortions predict suicidal ideation, and critically, whether their interaction confers heightened short-term risk for suicidal events (e.g., suicide attempts, ED visits, or hospitalizations). By identifying person-specific linguistic signatures of short-term youth suicide risk, this project will advance precision psychiatry approaches that enable digital health tools and social media platforms to recognize individual risk factors and deliver timely, personalized interventions during periods of crisis. This work directly addresses a critical public health priority by informing dynamic, individualized strategies for suicide prevention and by providing training in data-driven approaches to youth suicide research.