Bayesian modelling demonstrates clinically relevant heterogeneity in Tau PET patterns in Alzheimer's disease.

BACKGROUND: Traditionally, subgroups have been used to explore the effects of tau heterogeneity on cognition. However, categorization into rigid, exclusive subtypes, each with their own tau pattern, may overlook the fact that individual tau patterns are complex, and most individuals express features of multiple patterns. Mapping tau patterns in all their complexity has important clinical implications, as it may enable more accurate prognostication and support the development of personalized therapeutic strategies tailored to an individual’s unique tau profile. METHODS: We applied a data-driven Bayesian model using Latent Dirichlet Allocation (LDA) to identify four covarying tau PET patterns in amyloid-positive individuals with symptomatic Alzheimer’s disease (AD) from the Amsterdam Dementia Cohort (ADC, N = 93, mean age = 65.3). The four latent tau spatial patterns identified via LDA were designated as follows: Limbic (Factor 1), characterized by predominant involvement of the limbic regions; Left TPC (Factor 2), centered on the left temporo-parietal cortex (TPC); Posterior (Factor 3), reflecting a posterior neocortical distribution; and MTL-sparing (Factor 4), showing relative sparing of limbic regions with diffuse neocortical tau deposition. Associations between individual loadings on each of these factors and cognitive domain scores were assessed using linear regression analyses. We then applied the ADC-derived model to an independent validation sample from the Alzheimer’s Disease Neuroimaging Initiative (ADNI, N = 162, mean age = 72.7) to extract individual factor loadings. Associations between factor loadings and contemporaneous cognitive performance were again tested using linear regression, and linear mixed-effects models were used to additionally explore associations with cognitive decline. All analyses were adjusted for age, sex, education, and clinical diagnosis. RESULTS: In both the discovery and validation cohort, we identified distinct associations between tau factor loadings and cognitive performance. Higher loading on Factor 1: limbic tau was generally associated with relatively better baseline cognition and slower cognitive decline. Factor 2: left TPC tau was linked to worse baseline scores and faster decline in memory, MMSE and language scores but only in the ADNI cohort. Factor 3: posterior tau was associated with worse baseline MMSE in the ADC cohort and to worse visuospatial performance both cross-sectionally and longitudinally in the ADNI cohort. Factor 4: MTL-sparing tau was related to better longitudinal memory in the ADNI cohort. CONCLUSION: This data-driven approach identified overlapping tau patterns that relate to distinct cognitive domains. The findings highlight the value of continuous factor modeling in understanding heterogeneity in Alzheimer’s disease. Such a refinement of quantifying tau heterogeneity may improve patient stratification, refine prognostic accuracy, and ultimately guide the development of individualized treatment strategies. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s00259-026-07868-5.