A Machine Learning Model to Improve Risk Adjustment Accuracy in Medicare

This prognostic study developed "Franklin," a machine learning algorithm designed as a plug-and-play replacement for the Hierarchical Condition Category score, the tool currently used by the Centers for Medicare and Medicaid Services to risk-adjust payments for more than 65 million Americans enrolled in Medicare Advantage, Accountable Care Organizations, and Affordable Care Act marketplace plans. Trained on the same 20% sample of 2018 traditional Medicare claims data used to build the current HCC model, Franklin was evaluated for predictive accuracy using R-squared log cost, Spearman rho, and sensitivity and specificity measures, demonstrating meaningfully improved performance over the existing model while remaining designed for practical implementation within existing CMS infrastructure.

The stakes of this work extend well beyond technical modeling: inaccurate risk adjustment creates financial incentives for insurers to selectively enroll healthier patients and for providers to engage in diagnostic upcoding, behaviors that drive billions of dollars in overpayments and ultimately undermine the equitable allocation of healthcare resources. By demonstrating that a machine learning approach can improve on the current standard using the same underlying data, this study makes a compelling case for modernizing the risk adjustment infrastructure that quietly governs how healthcare dollars flow across the American system.