A multi-frequency ICA-based approach for estimating voxelwise frequency difference patterns in fMRI data.

In resting-state functional magnetic resonance imaging (fMRI) studies, blood oxygenation level dependent (BOLD) signals exhibit temporal correlation across voxels. Prior research has established resting-state functional connectivity (RSFC) across multiple frequency ranges, indicating that functional integration between brain regions occurs at various frequencies. While a few studies have studied RSFC data filtered into different frequency ranges, prior work has focused on extracting the main effects of frequency and comparing these posthoc, rather than focusing on estimating multivariate spatial patterns explicitly capturing frequency differences. Here, we propose a novel multi-stage independent component analysis (ICA)-based approach for estimating frequency difference patterns (FDPs) in fMRI data. Our novel approach involves separating fMRI images into four frequency sub-bands, concatenating them, and then applying group ICA to extract informative components. After removing non-gray matter components (edge effects, white matter, ventricles), we compute voxelwise differences between these sub-bands and perform a second ICA stage. This allows us to identify distinct covarying spatial patterns associated with FDPs. Understanding the frequency-dependent characteristics is crucial for uncovering the underlying spatial and temporal signatures of brain activity across different frequency bands. This method allows for a more comprehensive spatial analysis of frequency specific filtered fMRI data as it captures the frequency differences within maximally spatially independent spatial maps via a multivariate model. We applied our method to fMRI data from 90 subjects with schizophrenia (SZ) and 90 healthy controls. Our approach revealed structured spatial and temporal patterns which showed frequency-specific partial overlap with known resting-state networks (RSNs) but also exhibited unique spatial patterns. Our frequency-specific analysis unveils connectivity that might be overlooked by single frequency band methods, providing a new window into the brain's functional architecture. These findings suggest that RSFC is a spatially distributed multi-frequency band phenomenon and highlight the potential for further investigation of BOLD signals in relation to cognitive processes.