Distinguishing a drug use disorder from drug use in a high-risk sample of youth: A random forest classification and explanatory analysis.

Although many adolescents and young adults experiment with drugs, a subset may develop a drug use disorder (DUD). Few studies have used machine learning to identify risk and protective factors associated with DUDs, and to the best of our knowledge, none have been conducted with a high-risk youth sample or incorporated explanatory analyses to understand how such factors influence model predictions. We sought to address these gaps by estimating a random forest to identify salient risk and protective factors that differentiate youth with a DUD from their peers who use drugs but do not meet diagnostic criteria for a DUD. We extend traditional supervised learning methods by conducting explanatory model analyses to unpack what the model learned from the data, allowing for better interpretation of how influential factors may affect DUD risk at both the sample and case levels. Cross-sectional data were analyzed from a sample of 600 youth aged 14-24 with past 6-month illicit drug use (59 % male, 58 % Black; 57 % meeting DUD criteria). Global assessments revealed that risk factors aligned with deviance proneness and stress/negative affect were associated with higher likelihoods of a DUD, whereas academic achievement and later drug use initiation were associated with lower likelihoods. Local assessment methods highlighted how these broad predictive patterns can be applied to individual risk and protective factor profiles to inform precision-based preventive strategies. Overall, these analytic approaches and findings may help inform the development of indicated preventive interventions for DUDs in adolescent and young adult populations.