Predictors of illicit substance abuse/dependence during young adulthood: A machine learning approach

Abstract

Prior studies have found an especially high prevalence of illicit substance use among adolescents and young adults in Brazil. The current study aimed to employ machine learning techniques to identify predictors of illicit substance abuse/dependence among a large community sample of young adults followed for 5 years. This prospective, population-based cohort study included a sample of young adults between the ages of 18–24 years from Pelotas, Brazil at baseline (T1). The Alcohol, Smoking and Substance Involvement Screening Test (ASSIST) was used to assess illicit substance abuse/dependence. A clinical interview was conducted to collect data on sociodemographic characteristics and psychopathology. Elastic net was used to generate a regularized linear model for the machine learning component of this study, which followed standard machine learning protocols. A total of 1560 young adults were assessed at T1, while 1244 were reassessed at the 5-year follow-up period (T2). The strongest predictors of illicit substance abuse/dependence at baseline (AUC of 0.83) were alcohol abuse/dependence, tobacco abuse/dependence, being in a current major depressive episode, history of a lifetime manic episode, current suicide risk, and male sex. The strongest predictors for illicit substance abuse/dependence at the 5-year follow-up (AUC: 0.79) were tobacco abuse/dependence at T1, history of a lifetime manic episode at T1, male sex, alcohol abuse/dependence at T1, and current suicide risk at T1. Our findings indicate that machine learning techniques hold the potential to predict illicit substance abuse/dependence among young adults using sociodemographic/clinical characteristics, with relatively high accuracy.