Machine learning–based prediction of adolescent smoking behavior using psychological predictors

This study aimed to examine how psychological factors are associated with adolescent smoking and evaluate the use of machine learning for prediction. Data were drawn from the 20th Korea Youth Risk Behavior Survey (2024), including 54,653 adolescents. The outcome variable was lifetime smoking experience, and the predictors consisted of multiple psychological factors. Five classification methods—logistic regression, naïve Bayes, decision tree, boosting, and random forest—were applied and compared. Among them, the random forest model showed the highest performance among the evaluated models using a holdout validation approach. The variable importance analysis indicated that loneliness and feelings of sadness or despair were the strongest associated factors, followed by subjective stress, irritation, and worry. Overall, the findings suggest that adolescent smoking is closely associated with underlying emotional vulnerability. This study highlights the need for prevention strategies that address these psychological aspects. It also suggests that machine learning can serve as a useful tool in public health for identifying and understanding psychological risk factors.