State-Guided ICA of Functional Network Connectivity Reveals Temporal Signatures of Alzheimer's Disease.

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.