Disparity implications of machine-learning-based MTM eligibility criteria

Although Medicare Medication Therapy Management (MTM) programs have demonstrated clinical and economic benefits, racial/ethnic minority groups face challenges meeting eligibility criteria for enrollment. In 2017, the Centers for Medicare and Medicaid Services launched the 5-year Enhanced MTM demonstration, granting Part D plans flexibility in identifying eligible beneficiaries. However, because participating plans adopted predictive modeling to determine eligibility, concerns persist that this approach may perpetuate existing racial/ethnic disparities.

To assess whether health cost-based MTM eligibility differs across race/ethnicity and whether machine learning models reproduce observed disparities in predicted eligibility.

This study analyzed 2019 Medicare administrative data linked to the Area Health Resource File for a 10% random sample of fee-for-service beneficiaries. Outcomes were binary indicators of top-quartile medication and health care costs, each measured from the Medicare and health care system perspectives. Multivariable logistic regression was employed to assess racial/ethnic disparities in top-quartile costs, using 6 algorithms - regularized logistic regression, random forest, gradient boosted trees, support vector machine, multilayer perceptron, and a consensus model. Predicted probabilities were computed to assess disparities in model outputs using multivariable fractional logistic regression.

Among 1,848,654 Medicare beneficiaries, Black and Hispanic individuals had significantly lower adjusted odds of top-quartile costs across all cost outcomes compared to their non-Hispanic White counterparts. For instance, the odds of being in the top quartile for total medication costs were 28% lower for Black beneficiaries (odds ratio [OR] = 0.72, 95% confidence interval [CI] = 0.70-0.75) and 21% lower for Hispanic beneficiaries (OR = 0.79, 95% CI = 0.74-0.84). Machine learning models reproduced these disparities in predicted probabilities, mirroring patterns in the empirical data.

Implementing cost-based MTM eligibility through predictive algorithms may perpetuate racial/ethnic disparities in MTM program access. Future research should explore strategies to mitigate such a potential when using such modeling to determine MTM eligibility.