Predicting substance use behaviors with machine learning using small sets of judgment and contextual variables

Abstract Substance use disorder (SUD) is characterized by behaviors of impaired control, physical dependence, social impairments, and risky use, regardless of the substance used—yet no prior work has predicted these behaviors directly. This study of 3476 adults used 15 judgment variables, derived from a picture rating task, with contextual variables to predict the SUD-defining behaviors, recency of four categories of substances being used, and SUD severity. This study achieved up to 83% accuracy and 0.74 AUC ROC for SUD behavior and moderate-high prediction for substance use with a balanced random forest approach, and 84% accuracy for predicting SUD severity. Judgment variable profiles revealed that participants with higher SUD severity are more risk-seeking, less resilient to losses, have more approach behavior, and have less variance in preference. This study argues that distinct constellations of 15 judgment variables yield a scalable system for addiction assessment, which can support research across a broad range of addictions.