Identifying Predictors of Adolescent Suicide Attempts: A Comparative Machine Learning Study

Suicide attempts among adolescents are an urgent public health crisis in the United States. Traditional statistical methods have limited success in predicting such attempts, as they often overlook complex interactions among diverse risk factors. This study applies four machine learning models —LASSO Logistic Regression, Classification and Regression Trees, Random Forest, and eXtreme Gradient Boosting— to predict suicide attempts using data from the 2017 Youth Risk Behavior Surveillance System. Variables included school safety, substance use, sexual and physical violence, health behaviors, media use, cognitive factors, and demographic characteristics. The primary outcome was adolescents who attempted suicide within the previous year. Sensitivity was used as the primary metric to select the best-performing model, prioritizing the detection of true positive cases.LASSO Logistic Regression demonstrated the highest sensitivity (88.9%) and area under the ROC curve (AUC) of 94.6%, outperforming the tree-based models in true positive detection. The analysis revealed that suicidal ideation, suicide planning, substance use, and feelings of insecurity were the most influential predictors across models. These findings suggest that ML models can improve adolescent suicide risk prediction by incorporating broader behavioral and environmental variables than traditional methods. Enhanced sensitivity in ML models offers a promising tool for developing effective prevention and intervention strategies tailored to high-risk youth populations.