Link, transport, integrate: a Bayesian data integration framework for scalable algorithmic dementia classification in population-representative studies

Project Summary Nationally representative cohorts are crucial for monitoring population trends in incidence, prevalence, and disparities in Alzheimer’s disease (AD) and Alzheimer’s disease-related dementias (ADRD), as well as for understanding determinants of AD/ADRD. Clinical dementia diagnosis is a time- and resource- intensive

process that is impossible to perform in large population-representative cohorts. Algorithmic dementia classification methods are often used as alternatives to this costly process. Current algorithms, however, cannot be developed in cohorts that do not contain a subset of clinically diagnosed dementia cases, such as

the nationally representative National Health and Aging Trends Study (NHATS). Further, available methods can only incorporate measures available for all participants they aim to classify. Thus, existing models cannot be adapted to include newly available and more comprehensive cognitive data such as data from the 2016

Harmonized Cognitive Assessment Protocol (HCAP) Study. The goal of this proposal is to fill the need for scalable algorithmic dementia ascertainment in population-representative cohort studies. We propose a flexible Bayesian framework for algorithmic dementia classification, accomplished through the following aims: (1)

transport the HCAP detailed cognitive assessment battery to (a) the full HRS population and (b) the NHATS population through data linkage and production of synthetic datasets and (2) develop a scalable model for inferring person-specific dementia probabilities through Bayesian data integration of multiple data sources. In

Aim 1, we will create synthetic versions of HCAP cognitive assessment outcomes for each participant in HRS and NHATS by modeling main effects of socio-demographic and health characteristics and their interaction effects on cognitive test performance. In Aim 2, we will use a Bayesian framework to incorporate data from

multiple sources to model the main effects of socio-demographic, health characteristics, and cognitive test performance (including synthetic data from Aim 1) and their interaction effects on dementia classifications. Prior distributions will be specified for the effects of these predictors on the probability of dementia. Person-

specific dementia probabilities based on Bayesian inference will be used to estimate dementia incidence, prevalence, and inferences about disparities in dementia patterns in the HRS and NHATS populations. I am submitting this proposal to support my dissertation research which will produce a foundational body of

work for my career as a researcher in AD/ADRD. During this fellowship, I will receive specialized training in advanced biostatistical methods and neuropsychological perspectives of AD/ADRD in both the clinical and research settings. I will contribute to the literature on AD/ADRD with advancements in statistical methods and

create accessible statistical computing tools to aid efforts in accurate trend monitoring and building a comprehensive understanding of risk factors and disparities in AD/ADRD. Advancing these aims is central to the goal of developing effective strategies to prevent AD/ADRD and reduce disparities in the disease.