Predictive Risk Stratification in Medicare Advantage Using Interpretable Machine Learning: Reducing Cost While Improving Outcome Equity

Abstract

Background: Risk adjustment is fundamental to payment design in Medicare Advantage, where capitated reimbursements depend on diagnosisbased
risk scoring to predict beneficiary spending. Although current models improve financial alignment between plans and enrollee health
status, persistent concerns remain regarding cost growth, coding intensity, and uneven reimbursement patterns across demographic groups.
Problem: Continued expansion in Medicare Advantage expenditures has intensified scrutiny over whether existing risk stratification methods
adequately balance fiscal sustainability with equitable health outcomes. Incentives embedded in diagnosis coding and selection dynamics may
distort payments, while emerging evidence shows that predictive tools can unintentionally reproduce racial and socioeconomic disparities.
Objective: This study develops and evaluates an interpretable machine learning framework for predictive risk stratification in Medicare Advantage
designed to enhance cost prediction accuracy while explicitly promoting outcome equity.
Methods: We compare the traditional CMS-Hierarchical Condition Categories model with interpretable machine learning approaches, including
additive transparent models and feature-attribution techniques. Models are assessed using cost prediction metrics, calibration performance,
and fairness criteria aligned with equality of opportunity principles.
Results: Interpretable models demonstrate superior cost prediction performance relative to the baseline approach and meaningfully reduce
disparity gaps across racial and socioeconomic groups without sacrificing transparency.
Conclusion: Interpretable artificial intelligence provides a viable pathway for modernizing Medicare Advantage risk adjustment by simultaneously
strengthening payment accuracy, limiting distortion incentives, and advancing equity in population-based reimbursement systems.