Background: Seven large real-world cohort studies (N \> 4 million) have consistently reported that glucagon-like peptide-1 receptor agonists (GLP-1RAs) are associated with a 19-54% lower incidence of Alzheimer's disease (AD). However, in November 2025, the phase 3 EVOKE and EVOKE+ trials of oral semaglutide in 3,808 patients with biomarker-confirmed early-stage AD failed to slow clinical progression. This paradox between real-world evidence (RWE) and randomised evidence has not been systematically examined. Moreover, prior RWE studies were conducted predominantly in North American and European populations, leaving a critical generalisability gap for East Asian populations, who face the world's largest AD burden.
Objective: To resolve the RWE-RCT paradox by testing whether the GLP-1RA effect on AD differs between cognitively unimpaired adults (primary prevention paradigm) and patients with established cognitive impairment (treatment paradigm), and to assess whether the protective effect is consistent across European, African, and East Asian ancestries.
Study Design: This is a multi-centre, retrospective, observational cohort study using a target trial emulation framework. The investigator analysed patient-level electronic health records (EHR) from five cohorts across four continents: Optum® Clinformatics® (USA), Mass General Brigham Biobank (USA), CPRD Aurum (UK), the Swedish National Diabetes Register (Sweden), and the Hong Kong Hospital Authority Clinical Data Analysis and Reporting System (HKHA CDARS), which covers approximately the entire Hong Kong population.
Study Population: A total of 2,138,917 adults aged ≥ 40 years with type 2 diabetes or obesity, with ≥ 2 years of continuous enrolment and no prior neurodegenerative disease diagnosis. GLP-1RA initiators (n = 612,418) were compared with DPP-4 inhibitor initiators (n = 1,526,499) using propensity-score overlap weighting, Fine-Gray competing-risk adjustment, negative-control outcomes, and instrumental-variable analysis to address confounding and surveillance bias.