Presentation (CTI)

OBJECTIVES: The Centers for Medicare and Medicaid Services (CMS) launched the Enhanced Medication Therapy Management (MTM) model in 2017 to improve enrollment and enhance program flexibility. Participating plans began adopting predictive modeling approaches, often utilizing machine learning to identify beneficiaries at greatest risk of higher future healthcare costs or medication-related issues. Transfer learning, which leverages patterns learned from majority populations to improve predictions for minority populations, can improve machine learning performance among racial/ethnic minority groups.
METHODS: This study evaluated machine learning performance in identifying high-cost Medicare beneficiaries under cost-based eligibility criteria for medication therapy management services across racial and ethnic groups. Among a population of 10% of Medicare beneficiaries in 2019, this study predicted whether a beneficiary’s spending ranked within the top 25%. A perceptron neural network was trained for total medication costs and total healthcare costs. Transfer learning was tested to determine whether it could enhance model performance for racial/ethnic minority groups.
RESULTS: This analysis included 1,848,654 Medicare beneficiaries. In pooled prediction models, the model demonstrated stronger discriminative ability among all minority groups compared to White beneficiaries. Transfer learning generally led to performance gains for the minority populations. Black, Hispanic, and Asian beneficiaries all experienced significant increases in the area under the receiver operating characteristic curve (range: 0.0024 to 0.0069), the area under the precision-recall curve (range: 0.0087 to 0.0164), and improvements in Brier scores (range: -0.0015 to -0.0031).
CONCLUSIONS: Racial disparities in predictive performance were not identified. Predictive models generally performed better for minority populations, particularly Black and Hispanic beneficiaries, compared with White beneficiaries. Transfer learning produced only modest gains for minority groups, underscoring its limited added value in this context. These results suggest that transfer learning is a valuable tool, but its application may be limited to situations where sample sizes are small.