Predicting Opioid Cost Burden through Integrated PBM and SDOH Modeling: An Explainable AI Framework

Background: The U.S. opioid epidemic imposes a persistent and disproportionate economic burden on payers and communities, driven by complex interactions among clinical utilization, pharmacy benefit management (PBM) policies, and social determinants of health (SDOH). Despite extensive research on overdose and prescribing risk, few studies have quantified the cost burden of opioid use or examined how policy and community-level factors jointly shape it. Objective: This study develops an interpretable, machine-learning framework integrating PBM cost levers and SDOH indicators to predict and explain county-level variation in opioid-related spending from 2013–2023. The goal is to identify structural drivers of cost, simulate potential savings under policy and social interventions, and support data-driven resource allocation. Methods: Using CMS Medicare Part D data linked with County Health Rankings and U.S. Census indicators, a Random Forest regression model was trained on ten years of county-level data (≈3,000 counties, 2013–2023). Key predictors included unemployment, income ratio, obesity, smoking, provider density, and PBM variables such as cost per claim and opioid prescribing rate. Model interpretability was achieved through SHAP (SHapley Additive exPlanations) analysis and policy simulations testing both PBM and SDOH interventions. Results: The model achieved high predictive accuracy (R² ≈ 0.97), explaining nearly all observed variation in opioid cost per capita. SHAP analysis revealed unemployment, mental-health-provider density, and income inequality as dominant drivers, while provider access and preventive-care variables exerted cost-mitigating effects. Simulated PBM levers (e.g., formulary tightening, utilization management) reduced predicted costs by 4–6%, while integrated SDOH improvements (e.g., +20% behavioral-health access, +10% primary care) achieved up to 17% savings—equivalent to approximately $23 billion nationally. The combined model demonstrated both statistical robustness and policy relevance. Conclusion: This study reframes the opioid crisis through an economic and structural lens, demonstrating that predictive modeling can translate public-health and PBM data into actionable fiscal insights. The proposed PBM–SDOH integration provides a scalable, transparent framework for targeting interventions in high-burden counties, optimizing healthcare spending, and informing evidence-based opioid policy. Keywords: Opioid costs, pharmacy benefit management, social determinants of health, Medicare Part D, machine learning, SHAP, predictive modeling, public health economics, health policy