2D-11 - Automated Prediction of Medicare Member Medication Adherence Using Predictive Modeling and Real-World Data

Background: Four Medicare Part D Star measures evaluate medication adherences for diabetes, hypertension, cholesterol, and statin use in persons with diabetes (SUPD). Monitoring and predicting adherence are essential for: (1) assessing health plan performance (2) enabling proactive interventions to improve adherence and health outcomes.

Objectives: This study aims to identify key predictors and develop four automated predictive models to assess individual medication adherence across four drug classes among eligible Medicare members. The framework enables automated weekly updates using the latest claim data with minimal manual oversight.

Methods: The retrospective study used the Elevance Health databases to identify Medicare members eligible for the four Part D Star medication adherence measures in 2024. Cross-validation was performed using data of members eligible for 2023 measures.

Predictors used to build predictive models include demographics, medical and pharmacy claims, laboratory data, health plan details, intervention programs, social determinants of health (SDoH) data and medication adherence history from 2021 to 2023.

Univariate and multivariate logistic regression were used to identify key predictors and to build model for each cohort. The variable selection process was automated using SAS Output Delivery System (ODS) tables.

Results: The model-building was based on four cohorts, with 360,047 members in the diabetes group, 803,480 members in the hypertension group, 847,711members in the cholesterol group, and 281,885 in the SUPD group.

The predictive models demonstrated strong performance, with model AUC (95% CI) values of 0.80 [0.77,0.84] for diabetes, 0.78 [0.75,0.81] for hypertension, 0.76 [0.73,0.79] for cholesterol, and 0.94[0.90,0.98] for SUPD respectively.

Among the 20 to 40 selected predictors, common strong predictors include (1) member compliance history (Y/N): OR 24.42; 95% CI [18.16,32.83] for SUPD, (2) Comprehensive Medication Review (CMR) received (Y/N): OR 4.02; 95% CI [3.35,4.83] for diabetes, (3) medication home delivery (Y/N): OR 1.29; 95% CI [1.24,1.36] for diabetes. Additionally, members who use CarelonRx pharmacy, on multiple medications, or have a 90-day of supply are more likely to be adherent.

Conclusions: Four predictive models were developed for medication adherence among Medicare members, achieving strong performance with AUC values from 0.76 to 0.94. Key predictors were applied in clinical strategies that increased >1% in compliance rates across three drug classes in one year. For example, diabetes adherence improved from 85.5% in 2023 to 86.8% in 2024. These results highlight the use of predictive modeling in identifying at-risk members and informing interventions to improve adherence and health outcomes.