Machine learning models struggled to predict who would respond to cannabis use disorder treatment | Research

The takeaway

Using data from a multi-site clinical trial, machine learning and traditional models achieved only modest accuracy in predicting cannabis use disorder treatment response, suggesting better predictors are needed.

Addiction researchers, clinical trial designers, precision medicine advocates, and clinicians treating cannabis use disorder.

All prediction models achieved only modest accuracy for CUD treatment response

What the researchers found

Both multivariable logistic regression and machine learning models (random forest, gradient boosting) had limited ability to classify CUD treatment responders versus non-responders. Prediction accuracy was modest, indicating that commonly measured variables do not strongly predict treatment response.

Why it matters

CUD treatments have limited efficacy overall. If clinicians could identify who will respond to which treatment approach, they could personalize care and improve outcomes. This study shows current predictive tools fall short.

The numbers in context

Multi-site clinical trial data used. Multiple machine learning approaches tested (random forest, gradient boosting, logistic regression). All achieved modest classification accuracy for treatment response.

How the study worked

Secondary analysis of a National Drug Abuse Treatment Clinical Trials Network multi-site outpatient trial. Adult CUD patients were assessed with multivariable logistic regression and machine learning models (random forest, gradient boosting) to predict treatment response.

What this study cannot tell us

Secondary analysis limited to variables collected in the original trial. Treatment was a specific multi-component protocol that may not generalize. Machine learning models can overfit to training data. Modest sample sizes may limit model performance.

How to read the evidence

Secondary analysis of a well-conducted multi-site trial with appropriate statistical methods. Limited by available variables and inherent prediction difficulty.

When this study was published

Published 2023.

The bigger picture

Precision medicine for substance use disorders lags behind other fields. The inability to predict treatment response for CUD suggests either better biomarkers are needed or that treatment response depends on factors not typically measured in clinical trials.

Questions still open

Would biological markers (genetics, neuroimaging) improve prediction accuracy? Are there treatment-matching variables that differ from treatment-response variables?

Common questions

Can we predict who will benefit from cannabis use disorder treatment?

Not well, according to this study. Even with advanced machine learning techniques, prediction accuracy was modest. This suggests that the variables typically measured in clinical settings (demographics, use history, co-occurring conditions) do not capture the full picture of what makes someone respond to treatment.

Why is predicting treatment response important?

If clinicians could identify likely responders before treatment begins, they could match patients to the most effective interventions, allocate intensive resources to those who need them most, and spare others from ineffective treatments.

Read the original research

Who responds to a multi-component treatment for cannabis use disorder? Using multivariable and machine learning models to classify treatment responders and non-responders.

Addiction (Abingdon, England), 118(10), 1965-1974

View on PubMed ↗Publisher’s paper ↗

Citation

Tomko, Rachel L; Wolf, Bethany J; McClure, Erin A; Carpenter, Matthew J; Magruder, Kathryn M; Squeglia, Lindsay M; Gray, Kevin M. (2023). Who responds to a multi-component treatment for cannabis use disorder? Using multivariable and machine learning models to classify treatment responders and non-responders.. Addiction (Abingdon, England), 118(10), 1965-1974. https://doi.org/10.1111/add.16226