Machine learning models predicted cannabis use disorder transitions with 74% accuracy using demographics, wearables, and social factors

For cannabis users, both elastic net and random forest models achieved AUC of about 0.74 (no significant difference). Demographic variables were the strongest predictors across both models. Social determinants of health, particularly income, contributed substantially. Wearable-derived metrics (activity and sleep data) provided incremental value in linear models but limited independent contribution in random forests.

Data from the All of Us Research Program, a nationwide cohort integrating electronic health records, surveys, wearable data, and social determinants. Individuals with baseline cannabis use were followed for incident SUD diagnoses. Elastic net logistic regression and random forest models were trained and compared using AUC on independent test sets.

Predicting who will progress from cannabis use to a use disorder could enable targeted prevention. This study shows that readily available demographic and social data already provide moderate predictive power, with wearable technology adding incremental value.

This represents a step toward precision prevention in substance use. While 74% accuracy is moderate, combining easily collected demographic data with emerging wearable technology could eventually enable proactive clinical intervention.

Moderate predictive accuracy limits clinical utility. All of Us cohort may not be fully representative. Electronic health record diagnoses may undercount SUD. Wearable data had limited contribution, possibly due to data quality or relevance.