Beyond binary classification: A tiered claims-based algorithm for Alzheimer's disease and related dementias in Medicare
Sara Knox, Mary Dooley, Maxwell Cutty, Kit N. SimpsonBackground
Administrative data is widely used to identify Alzheimer's disease and related dementias (ADRD), yet commonly applied algorithms (e.g., the Chronic Conditions Warehouse [CCW] algorithm) rely on binary classification that may obscure heterogeneity in likelihood and care needs.
Objective
To examine whether a tiered ADRD classification algorithm using Medicare home health data yields cohorts with more clearly distinguishable clinical and sociodemographic characteristics compared with the CCW-27 algorithm.
Methods
We conducted a retrospective cohort study of 2,252,040 Medicare Fee-for-Service beneficiaries with home health encounters. The novel algorithm applied episode-level and individual-level criteria to classify individuals into one of three mutually exclusive groups (ADRD-highly likely, ADRD-possible, ADRD-unlikely). Using descriptive statistics, we compared socio-demographic characteristics and OASIS-based cognitive and functional measures across groups defined by the novel algorithm and the CCW-27 algorithm.
Results
The CCW-27 algorithm classified 34.3% of individuals as ADRD-positive, whereas the novel algorithm classified 7.6% as ADRD-highly likely, 29.3% as ADRD-possible, and 63.1% as ADRD-unlikely. The ADRD-highly likely group demonstrated greater cognitive and functional impairment than the CCW-27 ADRD-positive group. Discordant classifications revealed important sociodemographic differences: individuals classified as ADRD-highly likely by the novel algorithm but ADRD-negative by the CCW-27 algorithm were more frequently Medicare/Medicaid dual-eligible (52.9%) and non-White (39.5%) than concordant and alternative discordant groups.
Conclusions
A tiered claims-based ADRD classification provides greater resolution than binary approaches and reveals clinically and sociodemographically distinct groups within the Medicare home health population. Such approaches may support more nuanced characterization of ADRD when clinical gold standards are unavailable.