Predicting opioid dependence from electronic health records with machine learning.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Public Health

Background:

• The United States faces an opioid epidemic with over 100 daily overdose deaths.
• Physicians prescribe over 200 million opioid prescriptions annually due to their effectiveness in pain management and patient demand.
• Understanding the biomedical profile of opioid-dependent patients is crucial for intervention.

Purpose of the Study:

• To predict opioid substance dependence using electronic health records (EHR).
• To analyze the clinical profile of opioid-dependent patients.
• To identify associations between clinical factors and substance dependence.

Main Methods:

• Trained a machine learning model to classify patients based on EHR data.
• Utilized data from diagnosed substance dependence patients and matched controls.
• Included lab tests, vital signs, procedures, and prescriptions in the analysis.

Main Results:

• The top machine learning classifier achieved an AUROC of approximately 92%.
• Identified associations between basic clinical factors and substance dependence.
• Analyzed pre-diagnosis data to elucidate the clinical profile of dependent patients.

Conclusions:

• The predictive model can identify patients at risk of developing opioid dependence.
• The model may help identify patients at risk of overdose.
• Useful for recognizing opioid-seeking behavior in emergency room visits.