IMPORTANCE Disruptions in the sleep-wake cycle have been reported in the preclinical period of dementia; whether they contribute to dementia prediction remains unclear. OBJECTIVE To examine associations of accelerometer-derived sleep-wake cycle metrics with incident dementia and their contribution to dementia risk prediction in models containing age and known risk factors. DESIGN, SETTING, AND PARTICIPANTS This study included 2 prospective UK population-based cohort studies: (1) UK Biobank (derivation study) and (2) Whitehall II (validation study). A UK Biobank accelerometer substudy was undertaken from 2013 to 2015, yielding accelerometer data on 103 278 participants. A Whitehall II accelerometer substudy was undertaken from 2012 to 2013 that provided data on 4267 participants. Analyses were performed between August 2024 and November 2025. Included participants were 60 years and older, without dementia, and with valid accelerometer and covariate data. EXPOSURES Thirty-six accelerometer-derived sleep-wake cycle metrics were extracted. A machine learning approach identified and combined metrics predicting dementia risk. MAIN OUTCOME AND MEASURE Incident all-cause dementia, ascertained from electronic health records. RESULTS Analyses were based on 53 448 UK Biobank participants (mean [SD] age, 67.5 [4.2] years; 28 448 female [54.2%]; mean [SD] follow-up, 7.8 [1.1] years) and 3965 Whitehall II participants (mean [SD] age, 69.4 [5.7] years; 1025 female [25.9%]; mean [SD] follow-up, 10.6 [2.4] years). In UK Biobank, 9 sleep-wake cycle metrics were combined in 2 components. Higher values in component 1 represented shorter durations and less frequent bouts of moderate to vigorous physical activity, more time in low-intensity activity, lower diversity of activity intensities, and higher probabilities to transition from activity to rest during daytime. Higher component 2 corresponded to more extreme sleep durations, longer wake bouts during sleep, lower probabilities to tran