Applying machine learning in predicting medication treatment outcomes for opioid use disorder

Area of Science:

• Clinical Informatics
• Machine Learning in Healthcare
• Addiction Medicine

Background:

• Medication for opioid use disorder (MOUD) is effective, but many patients relapse.
• Identifying factors for successful MOUD treatment is crucial for improving implementation strategies.
• Interpretable machine learning (ML) can aid in predicting treatment response for MOUD patients.

Purpose of the Study:

• To apply and compare various interpretable ML algorithms for predicting treatment failure in MOUD.
• To identify key predictive features for MOUD treatment outcomes using ML.

Main Methods:

• Applied 10 ML algorithms (KNN, logistic regression, MARS, SVM, CART, Random Forest, BART, Boosted Trees, Neural Networks) to predict treatment failure.
• Utilized data from 2478 individuals across three large pragmatic clinical trials of MOUD.
• Evaluated model performance using Receiver Operating Characteristic Area Under the Curve (ROC AUC).

Main Results:

• All models achieved ROC AUC estimates between 0.62-0.67; Random Forest performed optimally with an ROC AUC of 0.65.
• Commonly identified predictors included age, intravenous drug use frequency, study medication, and study site.
• Some algorithms highlighted smoking behaviors, while others identified complex non-linear trends from patient timelines.

Conclusions:

• While overall ML model performance was similar, different algorithms revealed distinct sets of predictive features.
• Several identified features were previously unrecognized in predicting MOUD treatment outcomes.
• A companion website offers resources for clinical investigators to understand and replicate ML applications in MOUD research.