Comprehensive Summary
This study examines how common chronic non-cancer pain (CNCP) is among Medicaid primary care patients, the frequency of opioid prescription for CNCP, and whether machine learning models (ML models) can reliably predict opioid prescribing. Researchers analyzed 2017-2019 Medicaid paid claims data from Indiana, Georgia, Colorado, Ohio, New Mexico, and Nevada to estimate CNCP and opioid prescribing patterns. Using ICD-10 codes, CNCP was identified and categorized into low- and high-impact groups, CP1 (single visit within 90 days), and CP2 (≥ 2 pain visits ≥ 90 days apart), and ran eight ML model classifiers (75/25 training and testing) with accuracy and AUC metrics set in place. Researchers also evaluated prescription characteristics such as morphine milligram equivalents (MME) and duration, and applied machine learning models to predict opioid prescribing; among 7.3 million Medicaid beneficiaries with at least one primary care visit, 24.1% had CNCP with 6.5% being classified as CP2, and the mean age of 33.0 years was lower than expected, with 64% being female. CNCP prevalence differed by state (Georgia 14.8% to Nevada 36.5%, p < 0.0001). 21.1% of CNCP patients received opioid prescriptions, and 59.5% of those prescriptions were filled within 180 days, with 15.3% receiving ≥50 MME/ day, while 40.2% of patients had long-term supply ≥90 days with higher doses and durations in CP2. Machine learning models performed well, with particularly XGBoost reaching 81% accuracy (AUCs > 0.80), and key predictors including number of pain claims, age, and mental health comorbidities. These findings emphasize the high and uneven burden of chronic pain and substantial reliance on opioids in Medicaid primary care populations, specifically of younger and reproductive-aged women.
Outcomes and Implications
This study highlights the large CNCP burden in younger, reproductive-aged, and socioeconomically vulnerable patient populations, and documents substantial reliance on opioids in primary care, including frequent long-term use in younger female patients. The geographic (state) disparities in prevalence and prescription patterns also suggests that socioeconomic status and policy differences strongly shape pain management and patient safety. These findings can guide for more refined guidelines for safer prescription, as multiple key features discovered by the ML models can help to define risks and support decisions toward multimodal pain care and conservative opioid prescribing. ML models offer a practical approach to identifying patients at risk for prolonged opioid use, enabling more targeted interventions. While further validation in electronic health record systems and future trials is necessary, these methods can be integrated into clinical decision making to reduce opioid-related harms in primary care.