AI Analyzes Nearly 450,000 Teens to Spot Suicide Risk Factors

In an effort to identify young people in psychological crisis before tragedy strikes, researchers have turned to computational tools to analyze nationwide youth health patterns. By evaluating survey responses from nearly 450,000 South Korean teenagers, a team of scientists explored whether machine learning can effectively detect which adolescents face the highest vulnerability to suicide attempts. Published in Frontiers in Psychiatry, the study reveals how automated models paired with transparent analytics can help untangle the complex personal and social forces behind adolescent mental health emergencies.

Mapping Adolescent Crisis With Big Data

Researchers Eun Sun So and Ji‐Young Yeo analyzed pooled records from 448,798 students who participated in the Korea Youth Risk Behavior Web-Based Survey between 2017 and 2024. The study evaluated several machine learning models—including logistic regression, random forest, and XGBoost—to classify whether a student had attempted suicide.

Across the board, the models demonstrated moderate predictive accuracy. XGBoost attained the highest overall balance between precision and recall, while the random forest model captured the greatest proportion of at-risk teens under screening-oriented conditions. According to the authors, these findings show that algorithmic tools hold promise for high-level population screening, functioning as an early warning layer rather than a standalone clinical diagnostic device.

Opening the AI ‘Black Box’ on Suicide Risk

Because artificial intelligence can often operate as an opaque “black box,” the researchers employed explainable artificial intelligence using SHapley Additive exPlanations (SHAP). This technique quantifies exactly how much individual variables shift the algorithm’s predictions, shedding light on the primary drivers of adolescent crisis.

The SHAP analysis revealed that profound feelings of hopelessness emerged as the single strongest contributor to predicted suicide attempts, followed closely by experiences of school violence and severe perceived stress. The models also identified several secondary influences, including dissatisfaction with sleep, poor self-rated health, household economic hardship, and specific behavioral warning indicators, demonstrating how psychological distress intertwines with environmental adversity.

The Screening Dilemma: Catching Cases vs. False Alarms

To serve broad public health objectives, the researchers calibrated the models to prioritize sensitivity, casting a wide net to avoid missing adolescents in danger. While this adjustment successfully boosted case detection, positive predictive values remained low—a statistical reality common when screening for events that, despite their severity, occur at relatively low base rates across the general population.

Because a lower positive predictive value inevitably produces false alarms, the researchers emphasize that such algorithms cannot serve as definitive diagnostic tests. Instead, they operate best as a broad safety net designed to prompt non-invasive wellness check-ins, supportive dialogue, and timely guidance from school counselors.

A Stepping Stone for Youth Public Health

By providing clear rationales behind algorithmic predictions, explainable AI equips educators and public health officials with actionable insights. Rather than relying on guesswork, authorities can observe how specific social stressors amplify emotional distress across a massive population.

The researchers suggest that suicide prevention initiatives will likely achieve the greatest impact by addressing the leading risk factors identified in the data. Focusing public health resources on school bullying prevention, stress-reduction strategies, and emotional support networks offers a clear path toward protecting vulnerable adolescents before a crisis escalates.

---