Identifying the 'gray zone': Developing scalable methods to detect opioid misuse in veterans on long-term opioid...

Patients prescribed opioids who are at high risk for misuse but don't meet diagnostic criteria for opioid use disorder (OUD) fall into a clinical 'gray zone,' posing challenges for identification and intervention. Manual chart reviews are effective but resource intensive while electronic health records (EHR) are imprecise and lack specific codes to identify this high-risk group, limiting understanding of treatments and outcomes.

We conducted manual chart reviews of 741 US Veterans with long-term opioid prescriptions and at least two opioid-related ICD codes to identify 'gray zone' patients. Using this dataset we trained elastic net regression (ENR) models which then, using only structured EHR data (e.g., ICD codes, pharmacy claims), predicted gray zone status in a larger, unreviewed sample of Veterans creating a sample larger than using manual reviews and more reliable than using only structured data fields.

Of 741 patients reviewed, 541 (73 %) met gray zone criteria. The primary ENR model supplemented gray zone identification with a minimum positive predictive (PPV) value of 84 %, 11 % over identification using structured data alone (73 %). The primary (PPV) model augmented sample identified 4047 additional individuals who fell within the "gray zone." Key model variables included age, proportion of no-show appointments, opioid-related ICD code type, non-opioid substance use disorders, and average opioid dose.

Machine learning models can improve the identification of gray zone patients beyond what is possible through chart review using only structured EHR data. This approach may help facilitate population-level identification, enabling targeted research and clinical interventions.