The Construction of a Multidomain Risk Model of Alzheimer’s Disease and Related Dementias

Abstract

Background: Alzheimer’s disease (AD) and related dementia (ADRD) risk is affected by multiple dependent risk factors; however, there is no consensus about their relative impact in the development of these disorders. Objective: To rank the effects of potentially dependent risk factors and identify an optimal parsimonious set of measures for predicting AD/ADRD risk from a larger pool of potentially correlated predictors. Methods: We used diagnosis record, survey, and genetic data from the Health and Retirement Study to assess the relative predictive strength of AD/ADRD risk factors spanning several domains: comorbidities, demographics/socioeconomics, health-related behavior, genetics, and environmental exposure. A modified stepwise-AIC-best-subset blanket algorithm was then used to select an optimal set of predictors. Results: The final predictive model was reduced to 10 features for AD and 19 for ADRD; concordance statistics were about 0.85 for one-year and 0.70 for ten-year follow-up. Depression, arterial hypertension, traumatic brain injury, cerebrovascular diseases, and the APOE4 proxy SNP rs769449 had the strongest individual associations with AD/ADRD risk. AD/ADRD risk-related co-morbidities provide predictive power on par with key genetic vulnerabilities. Conclusion: Results confirm the consensus that circulatory diseases are the main comorbidities associated with AD/ADRD risk and show that clinical diagnosis records outperform comparable self-reported measures in predicting AD/ADRD risk. Model construction algorithms combined with modern data allows researchers to conserve power (especially in the study of disparities where disadvantaged groups are often grossly underrepresented) while accounting for a high proportion of AD/ADRD-risk-related population heterogeneity stemming from multiple domains.