VALIDITY OF ALGORITHMIC DIAGNOSES OF DEMENTIA: A CASE STUDY OF APPLICATION TO RESEARCH ON RACE/ETHNIC DISPARITIES IN DEMENTIA

Abstract

Understanding of race/ethnic disparities in Alzheimer's disease and related dementia (ADRD) is limited by the paucity of diverse cohorts with formal dementia ascertainment. Algorithmic diagnosis of ADRD provides an attractive alternative. Here we evaluate the performance of algorithmic versus gold-standard ADRD classification for the purpose of quantifying and understanding race-ethnic disparities. We used data from the nationally representative Health and Retirement Study (HRS) and its sub-study, the Aging, Demographics, and Memory Study (ADAMS) study, which conducted gold-standard, study-based dementia ascertainment. We identified five existing algorithms for algorithmic classification of dementia using HRS data. We also developed a new algorithm for dementia classification that minimized differences in sensitivity and specificity by race/ethnicity. We classified ADAMS Wave A participants as normal or demented using all six algorithms. In a sample restricted to participants with sufficient data to apply all algorithms, we compared findings using gold-standard and algorithmic diagnoses when estimating (a) the overall prevalence of dementia, (b) the odds ratio (OR) comparing dementia prevalence in non-Hispanic blacks to whites, and (c) the OR quantifying the association between prevalent diabetes and prevalent dementia overall and within race-groups. Models were adjusted for age, age2, race, sex, and HRS proxy status, and analyses were weighted to recover nationally representative estimates. While overall prevalence of dementia using gold-standard diagnoses was 12% (95%CI: 10%, 15%), estimates obtained using algorithmic diagnoses ranged from 6% to 17%. With the exception of our new algorithm designed to minimize sensitivity and specificity differences by race-ethnicity, all of the existing algorithms substantially overestimated the OR comparing blacks to whites (OR: 1.9 (95%CI: 0.8, 4.1) using the gold-standard diagnoses, OR: 1.8 (95%CI: 0.8, 4.3) using our new algorithm, ORs ranging from 2.8 to 5.8 using the other 5 algorithms). In overall or race-stratified adjusted models, use of some, but not all algorithms produced associations between diabetes and dementia that were similar to those obtained using gold-standard diagnoses.