Subjective cognitive decline (SCD) represents a critical window in the Alzheimer’s disease (AD) continuum, as it may reflect the earliest symptomatic manifestation of pathological processes preceding mild cognitive impairment and Alzheimer’s dementia. However, SCD is highly heterogeneous: while some individuals are on a trajectory toward progressive neurodegeneration, others exhibit decline for non-AD-related reasons. This heterogeneity makes individualized prediction of progression at the SCD stage critical for earlier risk stratification and timely intervention. Structural magnetic resonance imaging (MRI) and positron emission tomography (PET) have been shown to capture biomarkers associated with progression in SCD and early AD, with MRI characterizing anatomical degeneration and PET probing molecular and functional pathology. These complementary modalities provide a promising foundation for early progression modeling. Existing computational models are often limited by cross-modality heterogeneity and modality missingness, hindering their ability to capture the subtle, non-linear neurodegenerative signatures inherent in SCD, especially under conditions of extreme data scarcity. Motivated by these challenges, this dissertation develops artificial intelligence (AI)-driven frameworks for multimodal neuroimaging analysis and individualized prediction of cognitive progression in SCD. It encompasses three strategic pillars: (1) data-efficient prediction through knowledge transfer; (2) generative modeling for PET completion; and (3) multimodal longitudinal forecasting through integrated representation learning. First, a domain-aware transfer learning framework is developed to improve progression prediction in small-scale SCD cohorts by leveraging informative priors from large auxiliary neuroimaging datasets while mitigating cross-dataset distribution shift. This design improves data efficiency and reduces the reliance on extensive labeled SCD data. Second, three deep generative frameworks are proposed to mitigate incompleteness in PET imaging. These models synthesize clinically meaningful PET images from structural MRI, providing high-fidelity surrogate signals for downstream classification. Third, a hybrid multimodal multitask framework is designed to model progression trajectories in SCD by integrating transferred knowledge, modality-specific representation learning, multimodal fusion, and clinical outcome prediction within a unified pipeline. This dissertation establishes a robust methodological foundation for modeling early-stage cognitive decline. By addressing challenges in data scarcity and heterogeneity, it provides a translational framework for individualized progression modeling, advancing clinical decision support and longitudinal monitoring of neurodegenerative disorders.
Minhui Yu (2026) studied this question.