Abstract
Identifying robust neuroimaging biomarkers for Alzheimer's disease (AD) and mild cognitive impairment (MCI) is essential for early diagnosis and intervention. In this study, we introduce a novel, fully automated, guided dynamic functional connectivity (dFNC) framework for extracting multiple dynamic measures to distinguish MCI/AD from cognitively normal (CN) individuals. Resting-state fMRI data were used to extract subject-specific brain networks via spatially constrained independent component analysis (scICA), using a multi-objective optimization framework to ensure alignment with known functional networks while preserving individual variability. Using these components, dFNC was computed through a sliding-window approach. ICA was then applied to the concatenated dFNC matrices from the UK Biobank (UKBB) dataset to identify five canonical brain states, each representing a replicable, independent pattern of connectivity. These states served as biologically informed priors in a state-constrained ICA (St-cICA), which was applied to each subject in the combined OASIS-3 and ADNI datasets to guide individual-level decomposition and ensure interpretable connectivity states guided by state priors derived from the normative UKBB sample. St-cICA extracted subject-specific dFNC features and associated weighted timecourses. To characterize dFNC patterns, we computed metrics from the most strongly expressed (primary) state and introduced estimation of the second-most expressed (secondary) state at each time point, including dwell time, occupancy rate, and transition probabilities. Group comparisons using two-sample t-tests revealed widespread and significant alterations in AD/MCI compared to CN individuals. AD/MCI participants exhibited higher dwell times and increased self-transitions, indicating reduced neural flexibility and a tendency to remain in specific connectivity states. In contrast, CN individuals showed more diverse and recurrent transitions, reflecting greater adaptability. Secondary transitions revealed widespread selective switching in CN, whereas AD/MCI showed reduced cross-state engagement. A classification model trained on 6,960 dynamic features achieved strong performance in distinguishing AD/MCI from CN (mean AUC ≈ 0.85). These findings highlight the potential of guided dFNC as a biomarker framework for early-stage AD detection using resting-state fMRI.