Content area
Background: Opioid Use Disorder (OUD) remains a major public health crisis in the U.S., and while buprenorphine is an effective treatment, early discontinuation undermines its impact. Existing statistical and machine learning (ML) models often overlook key social and healthcare-related predictors and fail to capture subgroup disparities, limiting their utility for equitable intervention design. This study addresses these gaps using a nationwide, longitudinal dataset and a two-stage modeling approach focused on disparity identification.
Methods: We analyzed buprenorphine treatment episodes for OUD from the IQVIA Pharmetrics® Plus database (2006–2023), defining discontinuation as episodes lasting under 180 days. A Cox proportional hazards model was used to assess a wide range of predictors, including demographics, comorbidities, health service utilization, and contextual socioeconomic factors. We then implemented a two-stage Virtual Twins model to identify disparity-driven population subgroups by estimating counterfactual changes in discontinuation risk based on key attributes (e.g., gender, insurance type). Predictor sets incorporated Social Vulnerability Index (SVI), buprenorphine provider density, mental health service density, and early adherence (3-month proportion of days covered, or PDC). ML models evaluated in the first stage of virtual twins included logistic regression, decision trees, gradient boosting machines, neural networks, and random forest, with the best-performing model used for the second stage.
Results: We identified 214,735 episodes from 109,784 patients (median duration: 56 days). Key protective factors included early adherence (PDC ≥ 0.80), longer initial days’ supply, high provider and mental health service density, and outpatient/psychiatric care. Risk factors included high SVI, co-occurring substance use and psychiatric disorders, and younger or older age groups. Model performance improved substantially with the inclusion of contextual SDOH and early adherence. Random forest yielded the highest predictive accuracy. Second-stage Virtual Twins analysis revealed subpopulations at heightened risk of discontinuation, including females in high-SVI regions, privately insured individuals with alcohol use in the West, and younger patients with non-opioid drug use.
Conclusion: Our integrated statistical and ML framework offers a better understanding of buprenorphine treatment retention and disparities, informing policies and interventions to improve equitable access and outcomes in OUD care.