A Psychosocial Approach to Predicting Substance Use Disorder (SUD) Among Adolescents

Substance Use Disorder (SUD) affects about 5% of adolescents and can lead to many personal and societal problems. Risk factors such as peer pressure, permissive parenting, and impulsiveness make SUD more probable, whereas protective factors like community engagement alleviate this risk. No factor, however, is a sole determinant of SUD. The objective of this research is to build an ensemble learning framework to comprehensively predict adolescents at risk for SUD, considering the interplay between different factors. A data-driven model of 34 factors reflecting multiple dimensions of an adolescent's sphere of life, many of which comprise an adolescent's social network, is built from over 100 questions in the National Survey on Drug Use and Health. These factors are split into two groups; Proximal includes characteristics that are individual-centric, whereas Distal includes environmental influences. A labeled data set is curated by pooling the observations from the 2016 and 2017 editions of the survey. Two ensemble classifiers are trained based on the labeled data set, while applying the SMOTE algorithm to consider class imbalance. Both classifiers can distinguish between adolescents with and without SUDs exceptionally accurately, with Area Under the ROC curve over 0.90, outperforming multivariate logistic regression, a commonly used model in public health studies. Obesity combined with being approached with drugs poses the highest risk from over 1000 interactions. It is possible that the legalization of marijuana may exacerbate this problem. Based on these findings, we may infer that SUD among adolescents may not be exclusively attributed to natural tendencies or environmental influences but arises from their confluence.