Background <p>Circadian rest–activity rhythm disruption and sleep–wake fragmentation are common in Alzheimer’s disease (AD) and may precede clinical diagnosis. Wrist-worn accelerometry enables objective, scalable circadian phenotyping in population cohorts. We examined whether AD polygenic risk and <i>APOE</i> ε4 dose are reflected in accelerometry-derived circadian phenotypes and whether these phenotypes predict subsequent incident dementia/AD.</p> Methods <p>We analyzed 94,241 UK Biobank participants with 7-day wrist accelerometry and genetic data. Circadian phenotypes included relative amplitude (RA), L5 and M10 start hour, interdaily stability (IS), and intradaily variability (IV). Genetic exposures were a standardized AD polygenic risk score (AD-PRS) corresponding to the dataset-provided UK Biobank AD PRS variable standardized within the analytic sample, and APOE ε4 dose. Cross-sectional associations were tested with linear regression. Incident dementia/AD were modeled with Cox proportional hazards models with time zero defined as the accelerometry end date. Prespecified sensitivity analyses included lagged exclusions, competing risk of death, selection into accelerometry (inverse probability weighting), additional adjustment for overall activity, strict versus broad AD definitions, and RA × genetic-risk interaction terms in incident models.</p> Results <p>AD-PRS showed no robust association with RA or phase timing (fully adjusted standardized beta = 0.0028; <i>p</i> = 0.358). APOE ε4 dose was associated with earlier M10 timing (standardized beta = -0.0074; <i>p</i> = 0.032). AD-PRS showed a nominal association with IS (standardized beta = 0.0073; <i>p</i> = 0.018) that did not survive multiple-testing correction. In incident analyses (<i>n</i> = 80,378; 949 dementia and 371 <span>AD</span> events), AD-PRS and APOE ε4 dose strongly predicted incident dementia and incident AD. Each 1-SD higher RA was associated with lower risk of incident dementia (HR = 0.80 per SD; <i>p</i> = 4.99 × 10^–19) and incident AD (HR = 0.90; <i>p</i> = 0.031), meaning that lower RA corresponded to higher risk. The RA–dementia association remained stable across sensitivity analyses, whereas the RA–AD association attenuated after additional adjustment for overall activity. We found no evidence that RA associations differed by AD-PRS or APOE ε4 dose in incident interaction analyses (all interaction <i>p</i> values ≥ 0.362).</p> Conclusions <p>In this population cohort, AD polygenic risk showed minimal cross-sectional association with circadian amplitude or phase, whereas <i>APOE</i> ε4 dose was linked to slightly earlier peak daytime activity timing. Lower circadian amplitude (RA) consistently predicted higher subsequent dementia risk across multiple bias-oriented sensitivity analyses. For incident AD, the RA association was weaker after additional control for overall activity, supporting cautious interpretation of RA as a robust behavioral risk marker for dementia rather than an AD-specific genetic interaction signal.</p>