Machine learning evaluation of racial and ethnic inclusion under alternative cost-based approaches to Medicare Medication Therapy Management eligibility

BACKGROUND: The eligibility framework for the Medicare Medication Therapy Management (MTM) program has been associated with a lower likelihood of meeting enrollment criteria among racial and ethnic minority beneficiaries compared with their non-Hispanic White (White) counterparts. To optimize the MTM program design, the Centers for Medicare and Medicaid Services launched the Enhanced MTM demonstration, granting Part D plans greater flexibility to test alternative eligibility strategies. Many plans have adopted predictive modeling and cost-based criteria; however, their equity implications remain unclear. OBJECTIVE: To assess and compare how alternative cost-based eligibility schemes, implemented using machine learning models, influence predicted racial and ethnic inclusion in potential MTM eligibility. METHODS: A retrospective, cross-sectional analysis was conducted using a 10% national sample of Medicare fee-for-service beneficiaries in 2019. Two binary outcomes were defined using the top quartile of total health care costs and combined hospitalization and emergency department (ED) costs as eligibility thresholds. Five machine learning algorithms, including regularized logistic regression, random forest, gradient-boosted trees, support vector machines, and multilayer perceptron, were developed for each cost-based scheme. Their predicted probabilities were combined using a soft-voting ensemble to generate consensus estimates of potential eligibility. A multinomial logistic regression model was then used to assess associations between race and ethnicity and predicted inclusion across the 2 cost-based schemes, adjusting for predisposing, enabling, and need factors based on the Gelberg-Andersen model. RESULTS: The analytic sample included 1,848,654 Medicare beneficiaries, with 75% designated as the training population and the remaining 25% as the test population. Racial and ethnic differences in predicted eligibility varied by cost-based scheme. Compared with White beneficiaries, Black (relative risk ratio [RRR] = 4.31; 95% CI = 3.92-4.75) and Hispanic (RRR = 2.28; 95% CI = 2.04-2.56) beneficiaries had a higher likelihood of inclusion under the hospitalization and ED-based scheme only, relative to inclusion under the total health care cost-based scheme only, whereas Asian (RRR = 0.24; 95% CI = 0.19-0.29) and Other (RRR = 0.43; 95% CI = 0.36-0.52) racial groups had lower likelihoods. CONCLUSIONS: Alternative cost-based definitions of MTM eligibility identify different beneficiary populations across racial and ethnic groups. Incorporating hospitalization and ED costs may improve representation of Black and Hispanic beneficiaries in potential MTM eligibility. These findings highlight the need for transparency and continued evaluation of algorithmic eligibility frameworks to ensure equitable access to medication management services.