A Machine Learning Application to Classify Patients at Differing Levels of Risk of Opioid Use Disorder:...

Area of Science:

• Medical Informatics
• Clinical Decision Support
• Public Health

Background:

• Opioid use disorder (OUD) remains a significant public health challenge despite current management guidelines.
• Machine learning (ML) presents a potential solution for identifying and alerting clinicians to OUD.
• Early identification supports improved clinical decision-making for OUD treatment.

Purpose of the Study:

• To evaluate the clinical validity of an ML application for OUD risk detection.
• To compare ML-generated OUD risk alerts against structured clinician reviews of medical records.

Main Methods:

• An ML application analyzed outpatient data for 649,504 patients across two medical centers (2010-2013).
• A random sample of 180 patients was selected across three OUD risk categories.
• Clinicians performed systematic reviews to establish a consensus OUD risk level, compared against ML predictions.

Main Results:

• The ML application demonstrated good agreement with clinician assessments (weighted kappa = 0.62).
• For combined high-risk and OUD categories, ML achieved 56.6% sensitivity and 94.2% specificity.
• Key discrepancies between ML alerts and clinician reviews were identified and analyzed.

Conclusions:

• The ML application provides clinically valid alerts for varying OUD risk levels.
• ML tools show promise in identifying patients at risk for OUD.
• ML is expected to complement existing rule-based systems for opioid safety alerts.