BACKGROUND: Previous cohort studies have investigated the relationship between self-reported physical activity (PA) and dementia. Evidence from objective device-measured PA data is lacking. This study aimed to explore the association of device-measured PA with the risk of dementia incidence and common subtypes (Alzheimer's disease [AD] and vascular dementia) using the UK Biobank study. METHODS: 84,854 participants (55.8% women), invited to participate in the device-measured PA between 2013 and 2015, were included in this prospective cohort study. Wrist accelerometers were used to measure light, moderate, vigorous, moderate-to-vigorous PA (MVPA) and total PA intensity and duration (MET/min/week). Incident dementia (fatal and non-fatal) was extracted from hospital episodes records for incidence and death register for mortality. Incidence follow-up was carried out until the end of March 2021in England and Scotland and the end of March 2018 in Wales. Mortality data were available until February 2021. Nonlinear associations were first investigated using penalised cubic splines fitted in the Cox proportional hazard models. In addition, using MVPA, five categories were created. Associations of these categories with the outcomes were investigated using Cox proportional hazard models. Analyses were adjusted for sociodemographic, lifestyle and health-related factors. RESULTS: After a median follow-up of 6.3 years, 678 individuals were diagnosed with dementia. Evidence of nonlinearity was observed for all PA modes and all-cause dementia. For categories of MVPA, there was a significant trend towards a low risk of overall dementia when higher levels of MVPA were achieved (HR 0.66 [95% CI 0.62 to 0.70]. The lowest risk was identified in individuals who performed more than 1200 MET/min/week, those who had 84% (95% CI 0.12 to 0.21) lower risk of incident dementia compared to those who performed < 300 MET/min/week. CONCLUSIONS: Participants with higher PA levels had a lower risk of incident dementia than those less active, independently of sociodemographic, lifestyle factors and comorbidity. Considering that the majority of previous studies have reported this association using self-reported data, our findings highlight the strong inverse association between PA objectively measured and incident dementia.