アルツハイマー病におけるタウ病理のトポグラフィーのグラフィカルモデリングによるサブタイピング

Hyperphosphorylated tau tangles are essential hallmarks of Alzheimer's disease (AD) and their propagation across brain regions was often considered to follow the classic Braak stages. Recent post-mortem and in vivo tau positron emission tomography (PET) studies, however, revealed the frequent presence of tau pathology heterogeneity. Clustering or event-based methods were proposed previously for the subtyping to tau pathology in AD, but they often lack robustness to varying distributions of disease severity across cohorts. To robustly discover and model tau pathology subtypes in AD, we propose in this work a novel graphical modeling framework that can disentangle the phenotypical differences of tau PET imaging due to disease heterogeneity from the spatiotemporal variations of disease stages. First, we propose a novel Reeb graph representation at the individual level to characterize the topographic patterns of salient tau pathology on cortical surfaces. Next, we use only cross-sectional tau PET data to develop a graphical model at the population level to encode the inter-subject spatiotemporal relationships, which enables us to robustly derive subtypes based on the topographic patterns of tau pathology and hence achieve increased generalization power to new samples with distinct tau pathology severity from the training data. Using synthetic and large-scale tau PET imaging data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Anti-Amyloid Treatment in Asymptomatic Alzheimer's (A4) studies, we compare with the state-of-the-art SuStaIn method and demonstrate the improved generalization performance of the proposed method. In addition, we validate both methods on a cohort of autosomal dominant Alzheimer's disease (ADAD) patients with known tau pathology patterns to show that our method has more robust performance in testing data with large deviations from training data. Furthermore, for preclinical patients of the A4 cohort, we demonstrate more significant differences in clinical cognitive measures can be observed across subtypes discovered by our method.