Developing and Validating a Machine Learning Algorithm to Predict the Risk of Incident Opioid Use Disorder Among...

Area of Science:

• Clinical Informatics
• Public Health
• Machine Learning in Healthcare

Background:

• Opioid use disorder (OUD) is a major public health crisis in the U.S.
• Current screening for OUD risk is limited and lacks individualization.
• Machine learning (ML) offers potential for improved OUD risk prediction.

Purpose of the Study:

• Develop and validate an ML model to predict 3-month incident OUD risk in adults initiating opioid therapy.
• Stratify patients into clinically actionable risk groups using electronic health record (EHR) data.

Main Methods:

• Utilized OneFlorida+ EHR data (2017-2022) for model development and UPMC data for external validation.
• Included 182,083 adults without prior OUD, cancer, or overdose history receiving opioid prescriptions.
• Developed and compared elastic net, LASSO, GBM, and random forest models using 183 predictors.

Main Results:

• The Gradient Boosting Machine (GBM) model demonstrated strong performance (C-statistic=0.879) in validation, identifying key predictors like age and pain history.
• The top risk decile identified ~68% of OUD cases with a positive predictive value of 3.26%.
• External validation on UPMC data confirmed the GBM model's discriminative ability (C-statistic=0.756).

Conclusions:

• An ML algorithm developed using EHR data effectively predicts and stratifies incident OUD risk.
• The model shows promise for cross-health system application to inform early OUD interventions.
• This approach can enhance OUD prevention strategies in clinical settings.