Detecting Early and Implicit Suicidal Ideation via Longitudinal and Information Environment Signals on Social Media

A team of researchers from the University of Illinois Urbana-Champaign and other institutions has published a novel AI framework for detecting early and implicit suicidal ideation (SI) on social media platforms. The paper, titled 'Detecting Early and Implicit Suicidal Ideation via Longitudinal and Information Environment Signals on Social Media,' addresses the critical challenge of identifying distress signals that users do not disclose explicitly, instead surfacing through everyday posts and peer interactions.

The technical approach frames the problem as a forward-looking prediction task. The framework models a user's 'information environment' by combining two key data streams: the user's own longitudinal posting history and the discourse of their socially proximal peers. Researchers used a composite network centrality measure to identify a user's top neighbors, then temporally aligned the user's and neighbors' interactions. These multi-layered signals were integrated into a fine-tuned DeBERTa-v3 model, a powerful transformer architecture known for its language understanding capabilities.

In a controlled study involving 1,000 Reddit users (500 case and 500 control), this novel approach demonstrated a significant 10% average improvement in early and implicit SI detection over all other baseline methods. The findings underscore that peer interactions and community context provide valuable predictive signals beyond an individual's direct posts. This work, accepted for the 18th ACM Conference on Web Science (WebSci 2026), carries major implications for designing more sensitive, proactive online safety and mental health support systems that can capture indirect and masked expressions of risk.