Methodological advances in claims-based dementia algorithms: integrating medication and clinical data for medicare populations

BACKGROUND: Accurate identification of Alzheimer's Disease and Related Dementias (ADRD) in Medicare data is essential for research, policy, and care planning. Existing approaches, including the Centers for Medicare & Medicaid Services (CMS) Chronic Conditions Warehouse (CCW) algorithm, rely primarily on International Classification of Diseases, Tenth Revision (ICD-10) codes and are subject to misclassification and potential bias across populations. METHODS: We evaluated whether incorporating neuropsychiatric symptom (NPS) ICD-10 codes and medication data improves prediction of algorithm-defined ADRD classificaitons. We analyzed 269,214 Medicare beneficiaries receiving home health services in 2019 using claims, assessment, and prescription drug files. Two classification algorithms were examined: (1) the CCW binary classification algorithm and (2) a clinically informed ADRD ternary classification algorithm (ADRD-highly likely, ADRD-possible, ADRD-unlikely). Predictive models were developed for each algorithm using stacked elastic net regression, incorporating demographics, healthcare utilization, psychiatric and cognitive indicators, and medication use. Models were trained on beneficiaries with definitive ADRD status and evaluated in an independent test set using area under the curve (AUC), sensitivity, specificity, and predictive values. Thresholds were selected to prioritize specificity (≥ 80%) while maximizing sensitivity. RESULTS: In the test dataset, the CCW predictive model achieved an AUC of 0.863 for predicting CCW binary-defined ADRD status. The ADRD predictive model achieved a higher AUC of 0.906 when predicting ADRD ternary-defined classifications. Both models demonstrated strong discrimination for their respective outcomes, but performance characteristics differed. Among beneficiaries classified as ADRD-possible, the ADRD predictive model more frequently classified individuals as ADRD-negative, indicating a more conservative approach to ambiguous cases. CONCLUSIONS: Incorporating NPS codes and medication data improves prediction of algorithm-defined ADRD classifications, particularly among individuals with uncertain status. Models aligned with the ADRD ternary algorithm demonstrated more specific and conservative classification patterns, suggesting that integrating behavioral, cognitive, and medication data may enhance the validity of claims-based phenotyping. Further validations against clinical diagnosis are needed.