Abstract
BACKGROUND: Alzheimer's disease (AD) affects over 6 million people and is the seventh-leading cause of death in the United States. This study compares wrist-worn accelerometry-derived PA measures against traditional risk factors for incident AD in the UK Biobank. METHODS: Of 42 157 UK Biobank participants 65 years and older who had accelerometry data and no prior AD diagnosis, 157 developed AD by April 1, 2021 (264 988 person-years or on average 6.2 years of follow-up). Twelve traditional predictors and 8 accelerometer-based physical activity (PA) measures were used in single- and multivariate Cox models. Their predictive performances for future AD diagnosis were compared across models using the repeated cross-validated concordance (rcvC). To account for potential reverse causality, sensitivity analyses were conducted by removing dropouts and cases within the first 6 months, 1 year, and 2 years. RESULTS: The best-performing individual predictors of incident AD were age (p < .0001, rcvC = 0.658) and moderate-to-vigorous PA (MVPA, p = .0001, rcvC = 0.622). Forward selection produced a model that included age, diabetes, and MVPA (rcvC = 0.681). Adding MVPA to the model containing age and diabetes improved its rcvC from 0.665 to 0.681 (p = .0030), more than all other potential risk factors considered. CONCLUSIONS: Objective PA summaries are the best single predictors among modifiable risk factors with a predictive performance close to that of age. Adding PA summaries to traditional risk factors for AD substantially increases the predictive performance of these models. Increasing MVPA by 14.5 minutes per day reduces the hazard substantially, equivalent to 2 years younger.