rethinkTHC Search
Menu
Study breakdown

A Machine Learning Model Can Predict Which Young Cannabis Users Will Develop Use Disorder

LongitudinalStrong evidence
The takeaway

Using just five personality and behavioral factors — biological sex, delinquency, conscientiousness, neuroticism, and openness — a Bayesian machine learning model predicted cannabis use disorder within 5 years of first use with moderate accuracy.

Addiction medicine clinicians, school counselors, primary care providers screening young cannabis users, and prevention researchers.

What the researchers found

The model achieved AUC values of 0.68 (training), 0.64, and 0.75 (two validation datasets) for predicting CUD within 5 years of first cannabis use. The five risk factors were biological sex, delinquency, and personality traits of conscientiousness, neuroticism, and openness. Calibration was excellent (E/O ratios of 0.95–1.0).

Why it matters

Currently, clinicians have no validated tool to identify which young cannabis users are most likely to develop use disorder. A simple 5-factor model could be integrated into routine clinical screening to enable early, targeted intervention before problems develop.

The numbers in context

5 risk factors. Training AUC: 0.68. Validation AUCs: 0.64 and 0.75. E/O ratios: 0.95, 0.98, and 1.0 (excellent calibration). Predicts CUD risk within 5 years of first cannabis use. Personalized absolute risk output.

How the study worked

Bayesian machine learning model trained on the National Longitudinal Study of Adolescent to Adult Health. Five-fold cross-validation assessed performance (AUC and E/O ratio). Independent validation on two external datasets. Model provides personalized absolute risk scores for individual patients.

What this study cannot tell us

Moderate AUC (0.64–0.75) means the model misses some who develop CUD and flags some who don't. Based on an older cohort — cannabis products and patterns have changed. Limited to 5 factors for simplicity, potentially missing important predictors. Requires first cannabis use as entry point.

How to read the evidence

Well-validated machine learning model with independent external validation on two datasets, demonstrating good calibration and moderate discrimination.

When this study was published

Published 2025.

The bigger picture

Moving from population-level risk factors to personalized risk prediction represents a shift toward precision prevention in substance use. A brief personality and behavior assessment could identify the specific young cannabis users who would benefit most from intervention.

Questions still open

  • Would adding cannabis use patterns (frequency, product type) improve prediction? How would clinicians use personalized CUD risk scores in practice? Would this model perform differently in the current era of legal cannabis and high-potency products?

Common questions

Can you really predict who will become addicted to cannabis?
The model provides a probability estimate, not a certainty. With an AUC of 0.64–0.75, it correctly ranks risk better than chance but isn't perfect. It's designed as a screening aid, not a definitive diagnosis.
What personality traits increase CUD risk?
Higher neuroticism, lower conscientiousness, and higher openness were associated with increased CUD risk. These traits relate to emotional instability, impulsivity, and novelty-seeking — all established risk factors for substance use disorders.

Read the original research

Absolute Risk Prediction for Cannabis Use Disorder in Adolescence and Early Adulthood Using Bayesian Machine Learning.

Drug and alcohol review, 44(6), 1680-1690

Citation

Wang, Tingfang; Boden, Joseph M; Biswas, Swati; Choudhary, Pankaj K. (2025). Absolute Risk Prediction for Cannabis Use Disorder in Adolescence and Early Adulthood Using Bayesian Machine Learning.. Drug and alcohol review, 44(6), 1680-1690. https://doi.org/10.1111/dar.14098

Explore the wider topic