Abstract
ABSTRACT
Functional MRI analyses commonly reduce high-dimensional brain activity using predefined spatial summaries, such as region-of-interest (ROI) averages, or temporally selective representations, such as co-activation patterns (CAPs) and high-amplitude events. These approaches provide useful summaries but do not explicitly characterize information distribution across the continuous magnitude of spatial and temporal network expression. We introduce an amplitude-resolved framework for systematically interrogating this continuum using independent component analysis (ICA). Resting-state fMRI data in the FBIRN dataset were decomposed using spatially-constrained ICA with NeuroMark. For the temporal representation, time points were stratified according to global ICA amplitude and frequency bands, with functional network connectivity (FNC) estimated for each amplitude range. For the spatial representation, we introduce an approach that stratifies voxels according to ICA loading magnitude and generates amplitude-specific spatial similarity matrices. Linear support vector machine (SVM) classification of schizophrenia (SZ) and healthy control (HC) participants was used to probe the information retained by each representation. Discriminative information varied non-monotonically across amplitude: intermediate ranges frequently matched or exceeded the highest amplitudes, while combinations of amplitude ranges further improved classification. These findings demonstrate that informative fMRI structure is distributed across the ICA expression continuum and motivate representations that complement fixed ROI summaries and high-amplitude approaches.