A Machine Learning Model to Improve Risk Adjustment Accuracy in Medicare

To develop a machine learning (ML) algorithm that improves accuracy compared to the Hierarchical Condition Category (HCC) score used by the Centers for Medicare and Medicaid Services to risk-adjust payments for > 65 million Americans.

Prognostic study using Medicare claims data to train "Franklin", an ML algorithm predicting one-year costs, trained using identical data to HCC. Predictive accuracy was evaluated using R² log cost, Spearman rho, and sensitivity and specificity.

Random sample of 2018-2019 Part A and B claims from aged, community-based enrollees in Traditional Medicare who were not dually eligible and did not have end-stage renal disease.

The sample consisted of 4,176,666 Medicare beneficiaries (mean [SD] age 74.9 [7.2] years, 55.9% women; 85.9% Non-Hispanic white, 5.6% African-American, 3.4% Hispanic). Franklin was more accurate than HCC (R² log cost 0.44 vs. 0.15; Spearman rho 0.61 vs. 0.41, p < 0.001 for both). Accuracy improved for the 47% of beneficiaries with 0 HCCs and the 27% of beneficiaries with one HCC (Spearman rho 0.59 vs. 0.08 and 0.46 vs. 0.16, respectively; p < 0.001 for both). Franklin outperformed HCC in detecting the 20% lowest-cost beneficiaries (sensitivity 0.60 vs. 0.34, specificity 0.90 vs. 0.83; p < 0.001 for both). Franklin improved accuracy over HCC for racial/ethnic minorities and rural-dwelling beneficiaries (R² log cost Black 0.48 vs. 0.14, Hispanic 0.55 vs. 0.09, rural 0.36 v. 0.11; p < 0.001 for all), although Franklin disproportionately classified Black (15.8% vs. 10.1%) and Hispanic (22.9% vs. 12.2%) beneficiaries in the lowest predicted cost decile.

Franklin is an ML risk adjustment model that significantly improves risk-adjustment accuracy for Medicare beneficiaries compared to HCC. Franklin could generate improvement in payment accuracy, reduction in selection incentives, and financial savings to Medicare. Clarifying the equity impacts of more accurate risk adjustment is necessary.