Development of deep learning auto-encoder algorithms for predicting alcohol use in Korean adolescents based on cross-sectional data

Alcohol is a highly addictive substance, presenting significant global public health concerns, particularly among adolescents. Previous studies have been limited by traditional research methods, making it challenging to encompass diverse risk factors and automate screening or prediction of adolescents' alcohol use. This study aimed to develop prediction algorithms for adolescent alcohol use in South Korea using machine learning (ML) and deep learning (DL) models, and to identify important features. The study utilized a combination of DL (i.e., Auto-encoder) and ML (i.e., Logistic regression, Ridge, LASSO, Elasticnet, Decision tree, Random forest, AdaBoost, and XGBoost) algorithms to develop the prediction models. It involves 41,239 Korean adolescents and 46 socio-ecological input variables based on cross-sectional data. The analysis revealed that the prediction algorithms had AUC scores ranging from 0.6325 to 0.7214. The feature importance analysis indicates that variables within the domains of sociodemographic characteristics, physical and mental health, behavioral problems, family factors, school factors, and social factors all play significant roles. The developed algorithms enable automatic and early identification of adolescent alcohol use within public health practice settings. By leveraging a comprehensive array of input variables, these methods surpass the limitations of traditional regression approaches, offering novel insights into the critical risk factors associated with alcohol use among Korean adolescents, thereby facilitating early and targeted prevention efforts.