Abstract

This study aimed to develop machine learning-based classification models for suicidal behaviors (suicidal ideation, planning, and attempt) among young adults in their 20s and to provide empirical evidence for establishing tailored prevention strategies by exploring major risk factors at each stage. To this end, data from 828 young adults in their 20s were analyzed, incorporating variables from diverse domains such as sociodemographic factors, health-related variables, substance abuse, sleep, trauma, self-harm, interpersonal relationships, and psychological variables. A total of ten machine learning algorithms-including Logistic Regression, LASSO, Ridge, Elastic Net, Random Forest, GRRF, XGB, SVM, Naive Bayes, and KNN-were utilized to construct the classification models, and their validity was verified through ten repetitions of 10-fold cross-validation. The performance evaluation of the predictive models revealed that LASSO showed the best performance for classifying suicidal ideation, while Elastic Net demonstrated superior performance for classifying suicidal planning and attempts. From the exploration of major variables, 20 core risk factors were identified across four categories: suicide contagion-related factors, non-suicidal self-injury related factors, psychological factors, and sociodemographic and medical history factors. This study is significant in that it established precise machine learning-based classification models by subdividing suicidal behavior stages and provided empirical evidence for young adults in their 20s. In particular, by confirming that the optimal algorithms and core risk factors may vary according to the stage of suicidal behavior, these findings are expected to contribute to the construction of stage-specific, tailored suicide prevention strategies and crisis intervention systems for the young adult population.