AI-Based Technology Developed for Detecting Online Suicide Risk

Professor Paek Jongwoo's Team at Kyung Hee University Hospital
"Expected to Help Prevent Suicide by Creating a Safer Digital Environment and Enabling Early Identification of High-Risk Groups"

Kyung Hee University Hospital announced on the 25th that a research team led by Professor Paek Jongwoo of the Department of Psychiatry has developed technology that uses artificial intelligence (AI) to detect harmful online suicide content at an early stage. The study was conducted jointly with Professor Park Jinyoung of Sungkyunkwan University, Dr. Park Sungjun of the Korea Suicide Prevention Association, and Professor Cho Kyunghyun of New York University.

Professor Baek Jongwoo and his research team at Kyung Hee University Hospital developed a technology to detect harmful online suicide-related content at an early stage. Illustration depicting a person feeling depressed. AI-generated image.

The research team analyzed 43,244 posts from social media and online communities to build benchmark data that was directly reviewed by psychiatrists and psychologists, and based on this, developed an AI detection system.

The developed system automatically classifies suicide-related posts into five levels according to risk: illegal, harmful, potentially harmful, harmless, and non-suicidal. It is designed to detect even censorship-evading expressions such as slang, metaphors, and abbreviations, thereby supplementing the limitations of conventional human-centered monitoring.

In performance evaluation, when the GPT-4 model was applied, the system achieved an accuracy rate of 66.46% for detecting illegal content and 77.09% for detecting harmful content, confirming its potential for real-world application.

Professor Paek stated, "The AI-based suicide harmful-content detection system will contribute to creating a safer digital environment and to the early identification of high-risk groups," adding, "I hope it will serve as a foundation for proactive and cost-effective policy responses and for building a national suicide prevention infrastructure." The research findings were published in the February issue of the international journal JMIR Medical Informatics.