Using machine learning to classify patients on opioid use.

Area of Science:

• Health Services Research
• Computational Medicine
• Pharmacovigilance

Background:

• High-frequency opioid use elevates risks of opioid use disorder, overdose, and mortality.
• Predicting individual opioid use frequency is crucial for optimizing opioid prescription outcomes.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) techniques in predicting high-frequency opioid use.
• To compare the performance of various ML models against penalized logistic regression.

Main Methods:

• Utilized the Medical Expenditure Panel Survey (MEPS) data from 2016-2018.
• Applied five ML models (SVM, random forest, neural network, gradient boosting, XGBoost) and penalized logistic regression.
• Assessed prediction performance using AUROC and AUPRC, identifying key patient characteristics for high-frequency opioid use.

Main Results:

• Random forest and gradient boosting models demonstrated superior performance in predicting high-frequency opioid use.
• These ML models outperformed penalized logistic regression and other ML techniques.
• Patient age, number of chronic conditions, public insurance, and self-perceived health status were significant predictors in the best-performing random forest model.

Conclusions:

• Machine learning techniques show significant promise for predicting opioid use frequency.
• These predictive capabilities can contribute to improved patient outcomes and safer opioid prescription practices.