Predictors of Suicidal Ideation Among South Korean Adolescents: A Machine Learning Approach

Adolescents with suicidal thoughts report greater maladjustment in physical, psychological, and behavioral domains, develop maladaptive social relationships, and often experience social exclusion. Accurate and early detection of adolescents with signs of suicidal ideation is imperative for intervention. However, issues of multicollinearity and power when examining multitudinous predictors, as well as overfitting of data may emerge when using regression-based analysis. This research examines predictors of adolescent suicidal ideation across physical, psychological, behavioral, academic, and social domains by applying machine learning techniques.

A merged dataset of 7,319 second-year middle school adolescents in 2011 (n = 2,351), 2014 (n = 2,378), and 2019 (n = 2,590) were analyzed (Korea Children and Youth Panel Survey). Suicidal ideation was used as the dependent variable and 23 variables were examined as predictors of suicidal ideation: year, sex, physical (weekdays/weekend sleep hours, health), psychological (concentration, aggression, somatic symptoms, social withdrawal, depression), behavioral (status/violent/property/sexual/cyber delinquency), academic (academic commitment/achievement), and social (nurturing, peer/teacher relationship, youth program/club/fandom involvement) factors. Year was coded as ordinal scale (1 = 2011, 2 = 2014, 3 = 2019) and others were all coded as binary scales with 0 and 1. With 10-fold cross validation, the most efficient classification prediction model among Random forest, Logistic regression, Decision tree, and as well as the most important predictors of suicidal ideation were identified.

Random forest (AUC = 0.889~0.891) and Logistic regression (AUC = 0.894~0.897) were identified as a “good” prediction model, and Decision tree (AUC = 0.830~0.833) was identified as a “fair” prediction model. First, Random forest showed that psychological (depression, social withdrawal, somatic symptoms), physical (weekend sleep hours), and social factors (peer relationship, club involvement) as the most important predictors. Depression and social withdrawal stood out as the most important predictors of suicidal ideation, followed by somatic symptoms, weekend sleep hours, peer relationship, and club activity involvement. Second, Logistic regression pointed to 10 statistically significant predictors of suicide ideation: year (OR = 0.75, p < .05), weekdays (OR = 0.47, p < .001) and weekend sleep hours (OR = 0.36, p < .001), somatic symptoms (OR = 2.04, p < .001), social withdrawal (OR = 4.31, p < .001), depression (OR = 18.73, p < .001), negative nurturing (OR = 1.37, p < .01), positive peer relationship (OR = 0.57, p < .001), status deviance (OR = 0.45 , p < .001), violent deviance (OR = 0.48, p < .001). Lastly, Decision tree showed that adolescents who experience depression, social withdrawal, and somatic symptoms have a higher likelihood of reporting suicidal ideation than counterparts by 89%.

To identify adolescents with high risk of suicide, great attention should be drawn towards internalization symptoms such as depression and social withdrawal. Additionally, ensuring consistent and sufficient amount of sleep and promoting the sense of belonging and social support through positive friendships and community activities will help lower the suicide risk of adolescents.