Fairness-aware, explainable clinical decision support for opioid use disorder risk stratification: development and internal validation of a dual-layer AI system

Opioid use disorder (OUD) remains a leading cause of preventable death in the United States, yet the tools used to assess OUD risk rely on episodic self-report, produce binary output, and exhibit documented performance disparities across demographic groups that can widen existing inequities in care. Machine-learning models for OUD risk are seldom evaluated for demographic fairness or designed for the transparency clinicians need to trust and act on them. We present a fairness-aware, explainable clinical decision support system for four-tier OUD risk stratification, developed and internally validated on a large electronic health record-derived cohort accessed via Mayo Clinic Platform_Discover. The system pairs an XGBoost classifier with a transparent Clinical Rules Engine that attributes risk across six clinical domains, providing clinician-interpretable explanations alongside each prediction. To address demographic disparity directly, we applied an iterative bias-mitigation strategy combining age-balanced resampling, removal of race as a model input, and cost-sensitive reweighting, and measured its effect using group-fairness metrics (demographic parity, equal opportunity, equalized odds, and calibration within groups). On a held-out internal test set, mitigation reduced the White-Black gap in high-risk detection from 30.3 to 7.4 percentage points (a 76% relative reduction) and the age-based accuracy gap from 6.6 to 2.7 percentage points (59% reduction), raising high-risk detection for Black patients from 58.3% to 75.0%, at a cost of fewer than two percentage points of overall accuracy; gender differences remained below three points. The system was independently qualified through Mayo Clinic Platform_Solutions Studio. This work offers an implementable, transparent blueprint for operationalizing fairness and explainability in clinical AI for high-risk prescribing, with external and prospective validation as the clear next steps.