Bridging Online Behavior and Clinical Insight: A Longitudinal LLM-based Study of Suicidality on YouTube Reveals Novel Digital Markers

TL;DR

The paper investigates how YouTube language reflects suicidality by leveraging a novel longitudinal dataset of 181 suicide-attempt channels and 138 controls. It deploys three complementary frameworks—bottom-up LLM-based topic modeling, top-down clinical narrative assessment, and a hybrid expert-curated approach—analyzed via longitudinal GLMMs around reference events. Key findings show two robust pre/post markers (Mental Health Struggles and YouTube Engagement) and distinct motivational patterns before versus during attempts, illustrating how digital traces can align with clinical concepts while also revealing platform-specific dynamics. The work provides a publicly available dataset and AI-driven pipelines that bridge online behavior with clinical insight, offering a foundation for future ethically governed digital suicidality research.

Abstract

Suicide remains a leading cause of death in Western countries. As social media becomes central to daily life, digital footprints offer valuable insight into suicidal behavior. Focusing on individuals who attempted suicide while uploading videos to their channels, we investigate: How do linguistic patterns on YouTube reflect suicidal behavior, and how do these patterns align with or differ from expert knowledge? We examined linguistic changes around suicide attempts and compared individuals who attempted suicide while actively uploading to their channel with three control groups: those with prior attempts, those experiencing major life events, and matched individuals from the broader cohort. Applying complementary bottom-up, hybrid, and expert-driven approaches, we analyzed a novel longitudinal dataset of 181 suicide-attempt channels and 134 controls. In the bottom-up analysis, LLM-based topic-modeling identified 166 topics; five were linked to suicide attempts, two also showed attempt-related temporal changes (Mental Health Struggles, $OR = 1.74$; YouTube Engagement, $OR = 1.67$; $p < .01$). In the hybrid approach, clinical experts reviewed LLM-derived topics and flagged 19 as suicide-related. However, none showed significant effects beyond those identified bottom-up. YouTube Engagement, a platform-specific indicator, was not flagged, underscoring the value of bottom-up discovery. A top-down psychological assessment of suicide narratives revealed differing motivations: individuals describing prior attempts aimed to help others ($β=-1.69$, $p<.01$), whereas those attempted during the uploading period emphasized personal recovery ($β=1.08$, $p<.01$). By integrating these approaches, we offer a nuanced understanding of suicidality, bridging digital behavior and clinical insights.

Figures (20)

• Figure 1: An illustration of our research framework, showing the methodological approaches used in this study. Each circle represents a different approach, with the specific methods listed inside. The computational approach includes Language Models (LM) based
• Figure 2: The research groups are categorized into treatment - Attempted (During) - and three control groups - Attempted (Before), Control (Matches) and Control (Major Life Event). Both Attempted groups refer to those who attempted suicide during or before their channel upload period, while the other groups refer to control-based matches from the general population, or to those who experienced a major life event during their upload period (and didn't attempt suicide as far as we know). The sample size for each group is indicated below their respective categories. See data collection pipeline in Figure \ref{['fig:high_level_method_pipeline']}.
• Figure 3: The methodological pipeline consists of three approaches: 1. Computational Bottom-Up - Factors were generated using an LM-based topic modeling algorithm, and topic relevance was evaluated using a bottom-up approach; 2. Domain-Expert Top-Down - Factors were predefined by a clinical expert and assessed solely based on the suicide attempt narratives, and a 3. Hybrid Approach - A set of 166 LM-based topics was reduced to 19 based on their relevance to suicide as determined by a clinical expert. The bottom-up and hybrid approaches were also followed by a two-way mixed effect ANOVA examining both temporal and group effects, while the top-down approach was limited to group differences.
• Figure 4: YouTube Engagement and Mental Health Struggles: The figures present topics identified as significant in the bottom-up analysis (§\ref{['results_buttom_up_comp']}): YouTube Engagement (left) and Mental Health Struggles (right). The top row shows average topic values across four groups—Attempted (During), Attempted (Before), Control (Major Life Event), and Control (Personal)—in Pre- and Post-Event periods. TF-IDF word clouds are displayed alongside representative transcript excerpts, with word size reflecting importance. Below, the second row presents longitudinal trends over normalized time, aligned to each group's reference event (e.g., suicide attempt). Asterisks ($*$) denote statistically significant differences between Attempted (During) and other groups at each time point (t-test; see Appendix \ref{['appendix_temporal_testing']} for temporal testing procedures).
• Figure A.1: The data collection pipeline includes searching for relevant YouTubers (e.g., YouTubers who attempted suicide) (\ref{['inclusion_exclusion_appendix']}), a two-step human evaluation (\ref{['human_eval_protocol_full']}), involving non-expert assessment (personal channel validation and extracting demographic and event-related information), and psychological assessment (attempt intent validation and suicide-related factor extraction); both steps validating and then analyzing the narrative stories describing a suicide attempt or a major life event. The final steps involved control-based matching algorithm, which involved finding "similar" YouTube channels to the treatment group (Figure \ref{['fig:matching_algorithm']}). The last step included performing data processing (including filtering uploads and automatic transcription) on the 4 research groups (\ref{['data_preprocessing_appendix']}).