This work investigates how longitudinal behavioral signals on social media reflect suicidal behavior and how these signals align with, or diverge from, established clinical knowledge. Using a novel, clinically validated dataset of 181 YouTube channels belonging to individuals who attempted suicide and 134 matched controls, we analyze linguistic and engagement patterns before and after suicide attempts. We integrate three complementary approaches: a bottom-up LLM-based topic modeling framework (BERTopic with transformer embeddings and density-based clustering), a top-down clinical psychological assessment of suicide narratives, and a hybrid expert review of model-derived topics. Our analysis reveals both clinically grounded markers (e.g., mental health struggles) and previously underexplored, platform-specific digital markers (e.g., reduced YouTube engagement prior to attempts) that were not identified by expert-driven methods alone. By combining longitudinal modeling, mixed-effects statistical analysis, and expert validation, this work demonstrates the value of data-driven discovery for uncovering early behavioral indicators of suicide risk. The findings have practical implications for understanding suicidality in digital environments and for informing responsible, interpretable AI tools that complement clinical expertise rather than replace it.
Ilanit Sobol is a data scientist and applied researcher specializing in machine learning, natural language processing, and large language models, with experience spanning both corporate and startup environments. She holds an M.Sc. in Data Science from the Technion, under the supervision of prof. Roi Reichart, where her research focused on longitudinal, LLM-based analysis of suicidality on YouTube, integrating computational modeling with clinical psychological expertise. Ilanit has worked on end-to-end ML systems across healthcare, robotics, and document intelligence domains, developing and deploying models for anomaly detection, clinical NLP, and LLM-powered applications in production settings. Her work emphasizes statistical rigor, longitudinal analysis, and interpretable, responsible AI. In parallel, she is actively involved in the data science community through mentoring, organizing research workshops, and leading data-for-good initiatives.
Keynote session: Hadas Grossmon Ella
Break
Lightning talks session
Roundtable closing
Talk by Hila Paz
Talk by Dr. Moran Mizrahi
Closing remarks
End
