Mental health factors, especially persistent sadness and self-rated mental health, were the strongest predictors of adolescent suicidal ideation, planning, and attempts in interpretable XGBoost models, while lifestyle behaviors (sleep, physical activity, diet, screen time, substance use) contributed in stage-specific ways.
Key Findings
Results
XGBoost models predicted all three stages of adolescent suicidal behavior (ideation, planning, attempts) with satisfactory performance.
Data came from the 2021 Youth Risk Behavior Surveillance System (YRBS), a nationally representative U.S. high school survey.
After data cleaning and multiple imputation, 16,946 adolescents aged 14-18 years were analyzed.
Performance was assessed using AUC and related metrics; the abstract reports performance as 'satisfactory' without giving specific values.
Results
Persistent feelings of sadness and self-rated mental health were the strongest predictors across all stages of suicidal behaviors.
Feature importance was assessed with gain-based measures and SHAP analyses.
SHAP analyses indicated that greater sadness and poorer mental health were strongly associated with higher suicide risk.
Mental health factors were the most prominent predictors for ideation, planning, and attempts.
Results
Lifestyle behaviors contributed meaningfully to risk prediction, but their relative importance differed by suicidal behavior outcome.
Lifestyle factors included sleep duration, physical activity, diet, screen time, and substance use.
The associations were described as important and stage-specific.
Results
Adequate sleep and higher physical activity were associated with lower predicted suicide risk.
This direction of effect was identified through SHAP analyses.
Sleep duration also showed nonlinear and threshold effects in partial dependence analyses.
Results
Substance use (alcohol, cigarettes, marijuana) played a more important role in predicting suicide attempts than suicidal ideation or planning.
This finding comes from comparing feature contributions across the three outcome models.
It suggests substance use is more relevant at the most severe stage of the suicidal behavior spectrum.
Results
Partial dependence analyses revealed nonlinear and threshold effects for key predictors, especially sadness, mental health status, and sleep duration.
Partial dependence plots were used alongside SHAP to explore the nonlinear effects of key predictors.
These patterns would not be captured by simple linear models.
Conclusions
Interpretable machine learning can identify complex, nonlinear patterns associated with adolescent suicide risk that may guide targeted prevention.
The combination of XGBoost, gain-based importance, SHAP, and partial dependence plots allowed examination of direction, strength, and nonlinearity of effects.
The authors state the findings offer insights to guide targeted, evidence-based prevention efforts.
What This Means
This study used a type of machine learning to look at which factors best predict suicidal thoughts, plans, and attempts among U.S. high school students. It analyzed 2021 survey data from nearly 17,000 adolescents aged 14-18, and built separate models for each stage of suicidal behavior. All three models predicted outcomes reasonably well.
The strongest predictors at every stage were mental health factors, specifically persistent feelings of sadness and how teens rated their own mental health. Lifestyle habits also mattered: getting enough sleep and being physically active were linked to lower predicted risk, while substance use (alcohol, cigarettes, marijuana) was especially important for predicting suicide attempts compared with thoughts or plans. The effects were not always smooth or linear; some factors, such as sadness and sleep duration, seemed to have thresholds beyond which risk changed.
This research suggests that prevention efforts may benefit from attending to both emotional well-being and everyday behaviors, and from tailoring approaches to the stage of suicidal behavior. Because the data come from a one-time survey and show associations, the findings cannot show that any factor causes suicidal behavior.
Ding Y, Lei Y. (2026). Investigating the role of mental health and lifestyle behaviours in the risk of adolescent suicidal behaviours through machine learning.. PloS one. https://doi.org/10.1371/journal.pone.0358758