Exploring Alzheimer's disease subtype replicability

Alexei Taylor

doi:10.17918/00010692

Alzheimer's disease (AD)-related tau pathology, which is highly correlated with AD-specific cognitive decline, has been suggested to follow a typical pattern initiating from the medial temporal lobe and spreading into the lateral temporal, parietal, and frontal and occipital lobes, at varying levels of severity. However, atypical patterns which disproportionately affect medial temporal regions relative to cortical regions, or vice versa, have previously been identified through neuropathologically confirmed postmortem studies. These typical and atypical AD subtypes have distinct demographic and clinical characteristics, broadly replicated using hypothesis-driven methods and data-driven clustering of in-vivo imaging data. However, the consistency of individual-level subtype assignment is poor. Further, how tau-defined versus atrophy-defined subtypes differ is unclear given their complex spatiotemporal relationship. While many AD subtyping studies focus on novel subtype discovery, the current study proposed a method to evaluate individual-level subtype assignment replicability, by defining individual-level subtype assignment consistency as being assigned the same subtype by a majority of hypothesis-driven methods. Assignment consistency and stability were examined using a nested cross-validation procedure to compare reference data-driven clustering methods to the proposed semi-supervised clustering informed by hypothesis-driven findings. These comparisons were explored under 3 different aims: (i) Examine constraint vs non-constraint clustering methods using single-modal atrophy or tau imaging features, (ii) Assess the benefit of unsupervised feature selection on these clusterings, and (iii) Assess the benefit of multimodal clustering with constraints from either modality. In reference to this project's definition of an AD subtype, it was determined that (i) constraint methods had higher performance and stability than non-constraint methods, (ii) unsupervised feature selection offered slight improvements in performance and stability, and (iii) multimodal clustering offered greater gains in interpretability based on subtype features and clinical characteristics. Overall, findings from this dissertation project contribute to the literature on AD subtypes by recommending an objective framework for validating subtype replicability.

Exploring Alzheimer's disease subtype replicability

Files and links (1)

Abstract

Metrics

Details

Exploring Alzheimer's disease subtype replicability

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media