From Thoughts to Actions: Predicting Suicidal Ideation and Attempts in Youth Using Sex-Based Machine Learning Models

Background
Suicidal ideation (SI) and suicide attempts (SA) are urgent global public health concerns, particularly among adolescents. Male youth are often underrepresented in suicide prevention efforts due to gender-based stigma and lower help-seeking behaviors. While previous research has primarily examined individual-level predictors, limited studies have explored how psychosocial risks and protective behaviors differ by gender. Grounded in the Interpersonal Theory of Suicide, which posits that thwarted belongingness and perceived burdensomeness contribute to suicidality, this study aimed to identify gender-specific pathways of suicide risk and protection. We focused on how key behavioral and social factors—cyberbullying, screen time, physical activity (PA), and sleep—shape distinct experiences for male and female youth. The goal was to build sex-specific predictive models for SI and SA and generate evidence-based insights for targeted suicide prevention strategies.

Methods
We analyzed 2021 U.S. Youth Risk Behavior Survey data, including responses from 3,760 high school students. The primary outcomes were past-year SI and SA. Predictor variables included cyberbullying, screen time, PA, sleep, sexual minority status, and demographic characteristics. Separate models were developed for male and female youth using the XGBoost algorithm. Three nested models were trained per group, incrementally adding predictors: Model 1 used demographics; Model 2 added cyberbullying and screen time; Model 3 added PA and sleep. Model performance was assessed using test accuracy, F1-score, and area under the curve (AUC), with hyperparameter tuning via Optuna. Variable importance was interpreted using SHAP (Shapley Additive Explanations) values.

Results
In terms of predictive performance, the female-specific Model 2—which included cyberbullying and screen time—achieved the highest accuracy for suicidal ideation, with an AUC of 0.817. For male youth, the most effective model for predicting suicide attempts was Model 3, which incorporated all predictors and yielded an F1 score of 0.34 and an AUC of 0.75. Cyberbullying (SHAP: female = 0.38, male = 0.29) and sexual minority status (SHAP: female = 0.37, male = 0.28) consistently emerged as strong predictors across groups. These results align with the Interpersonal Theory of Suicide, suggesting that social exclusion and identity-related stressors may intensify feelings of thwarted belongingness and perceived burdensomeness, thereby increasing suicide risk. Protective factors varied by gender. For males, physical activity (PA, SHAP = 0.35) was the strongest protective factor against suicide attempts, potentially mitigating psychological burdens by reinforcing a sense of control and self-efficacy. For females, sleep (SHAP = 0.44) exerted greater influence, which may reflect the role of emotional regulation and physiological recovery in maintaining social and emotional functioning.

Conclusion
This study underscores the need for gender-specific modeling in youth suicide prevention. While cyberbullying was a consistent predictor for both sexes, PA and sleep revealed protective effects by gender. These findings support tailored intervention strategies—such as PA promotion for boys and sleep-focused mental health programs for girls. By aligning machine learning-driven prediction with adolescents’ lived experiences, this research advances the 2026 SSWR theme of “Leading for Transformative Change,” offering practical and policy-relevant insights for reducing suicide risk in diverse youth populations.