SpARCD: A Spectral Graph Framework for Revealing Differential Functional Connectivity in fMRI Data

TL;DR

SpARCD introduces a distance-correlation–based spectral framework to detect differential functional connectivity between conditions in high-dimensional fMRI data. By building condition-specific graphs, applying spectral filtering to remove shared structure, and evaluating region-level differences via a leading eigenvector, it achieves robust, network-level inference with a nonparametric permutation test. Simulations show superior power over edge-wise and seed-based methods across linear, nonlinear, and hybrid dependencies, while empirical analyses in PTSD-related EFMT and resting-state data reveal distributed, task- and group-specific connectivity alterations concentrated in visual networks. The method offers a scalable, interpretable approach for uncovering network reorganization and has broad applicability to other domains involving high-dimensional structured dependence.

Abstract

Identifying brain regions that exhibit altered functional connectivity across cognitive or emotional states is a key problem in neuroscience. Existing methods, such as edge-wise testing, seed-based psychophysiological interaction (PPI) analysis, or correlation network comparison, typically suffer from low statistical power, arbitrary thresholding, and limited ability to capture distributed or nonlinear dependence patterns. We propose SpARCD (Spectral Analysis of Revealing Connectivity Differences), a novel statistical framework for detecting differences in brain connectivity between two experimental conditions. SpARCD leverages distance correlation, a dependence measure sensitive to both linear and nonlinear associations, to construct a weighted graph for each condition. It then constructs a differential operator via spectral filtering and uncovers connectivity changes by computing its leading eigenvectors. Inference is achieved via a permutation-based testing scheme that yields interpretable, region-level significance maps. Extensive simulation studies demonstrate that SpARCD achieves superior power relative to conventional edge-wise or univariate approaches, particularly in the presence of complex dependency structures. Application to fMRI data from 113 early PTSD patients performing an emotional face-matching task reveals distinct networks associated with emotional reactivity and regulatory processes. Overall, SpARCD provides a statistically rigorous and computationally efficient framework for comparing high-dimensional connectivity structures, with broad applicability to neuroimaging and other network-based scientific domains.

Figures (10)

• Figure 1: Flowchart describing the steps in the SpARCD algorithm.
• Figure 2: An example for the blow-dependency structure in states $X$ and $Y$. One block in $X$ is divided into two blocks in $Y$.
• Figure 3: Linear simulation setting. Top panel: Observed test statistic $s(r)$ (red) compared with the mean (orange) and standard deviation (purple) of the permutation-based null distribution. Middle panel: Distance-correlation matrices for datasets $X$ and $Y$, highlighting the regions with altered connectivity in the first cluster. Bottom panel: Performance of SpARCD and competing methods in the linear simulation setting with block-diagonal covariance structure. Precision (left), recall (middle), and PR–AUC (right) are shown as functions of the signal-strength parameter $\gamma$. Higher $\gamma$ values correspond to stronger and more structured linear dependencies among clusters.
• Figure 4: Nonlinear simulation setting. Top panel: Observed test statistic $s(r)$ (red) compared with the mean (orange) and standard deviation (purple) of the permutation-based null distribution. Middle panel: Distance-correlation matrices for datasets $X$ and $Y$, demonstrating that the detected regions correspond to genuine nonlinear differences in connectivity. Bottom panel: Performance of SpARCD and competing methods in the nonlinear simulation setting under varying noise levels ($\sigma$). Precision (left), recall (middle), and PR–AUC (right) as functions of the noise level. The results show that SpARCD maintains high accuracy and robustness even as noise increases, whereas competing methods rapidly lose power after multiple-testing correction.
• Figure 5: Performance of SpARCD and competing methods in the hybrid simulation setting with varying degrees of linearity ($\alpha$). Precision (left), recall (middle), and PR–AUC (right) as functions of the mixing parameter $\alpha$, where smaller values indicate stronger nonlinear effects. SpARCD achieves the best overall performance for low-to-moderate $\alpha$, while maintaining competitive accuracy as dependencies become predominantly linear.