Comparison of factors distinguishing suicide attempters among suicidal ideators and factors associated with suicidal ideation and attempts using machine learning models in South Korean adolescents

Most adolescents with suicidal ideation (SI) do not attempt suicide; fewer than one-third report a suicide attempt. Yet, characteristics distinguishing suicide attempters among adolescents with (SI) remain underexplored. This study applied four machine learning (ML) algorithms, namely regularized logistic regression (LR), decision tree, random forest, and extreme gradient boosting, to identify key variables for suicide attempt (SA) among adolescents with SI (SA-SI) and for SA and SI in the total sample (SA-All, SI-All). We analyzed nationally representative survey data from 54,948 South Korean adolescents aged 12–18 years. Model performance was moderate for SA-SI (AUC 0.705), excellent for SA-All (AUC 0.944), and good for SI-All (AUC 0.874). LR showed the best performance among the four ML models. Positive predictive values were at least 1.6-fold, 7.9-fold, and 3.1-fold higher than the raw prevalences (15.7 %, 1.9 %, and 10.7 %) in SA-SI, SA-All, and SI-All, respectively, yet appeared modest due to class imbalance, highlighting the fundamental challenges of detecting rare events. SA models incorporated mental, sociodemographic, economic, health status, and health behavior factors, whereas SI models were driven solely by mental health variables. Despite the limitations inherent in a cross-sectional design, these findings imply that emotional vulnerability underlies SI, while progression to SA reflects the interplay of psychological, behavioral, and contextual factors. A comprehensive approach addressing multiple domains may enhance identification and prevention strategies for adolescents at risk of suicidal behavior.