Abstract
Functional connectivity (FC) is a cornerstone of resting-state fMRI analysis. While traditional methods focus on pairwise temporal synchronization (Low-Order FC or LOFC), recent "High-Order" FC (HOFC) methods capture more complex, multi-region interactions. This study systematically evaluates the test-retest reliability of two types of HOFC: topographical profile similarity-based (tHOFC/aHOFC) and dynamics-based (dHOFC). Using data from 25 healthy adults across seven repeated scans, the researchers found that HOFC metrics demonstrate fair-to-good reliability, establishing them as stable and promising tools for clinical biomarker discovery.
Introduction: Beyond Pairwise Correlations
Traditional resting-state fMRI analysis typically relies on Low-Order Functional Connectivity (LOFC). LOFC measures simple pairwise correlations between the BOLD signals of two brain regions. While LOFC has been the gold standard, it often fails to capture the hierarchical and dynamic ways that multiple brain networks interact.
To address this, researchers proposed High-Order Functional Connectivity (HOFC), which essentially calculates "correlations of correlations." While HOFC shows promise for detecting neurodegenerative diseases, its clinical utility depends on its test-retest reliability—the stability of the metric across multiple scans of the same individual. This paper provides the first systematic evaluation of that reliability.
Defining High-Order Metrics
The study evaluates three primary HOFC metrics:
- tHOFC (topographical profile similarity-based): Instead of comparing raw signals, it compares the "connectivity fingerprints" of two regions—how they each connect to the rest of the brain.
- aHOFC (associated HOFC): A hybrid metric that integrates low-level connectivity (LOFC) with high-level topographical patterns (tHOFC).
- dHOFC (dynamics-based HOFC): This captures "quadruple" relationships among four regions by measuring the temporal correlation of dynamic LOFC fluctuations rather than raw BOLD signals.
Methodology
The researchers utilized a high-quality dataset from the Consortium for Reliability and Reproducibility (CoRR), featuring 25 young healthy adults who underwent repeated resting-state fMRI sessions. Reliability was quantified using the Intra-Class Correlation (ICC), which distinguishes true individual differences from measurement noise.
For dHOFC, the analysis employed a sliding window approach to capture how connectivity between region pairs fluctuates over time. These fluctuations are then correlated between different pairs to define the high-order relationship.
Key Findings
1. Stability in Inter-Network Communication
A significant finding was that tHOFC and aHOFC were often more reliable than traditional LOFC for weak connections between high-level cognitive networks (e.g., the Default Mode Network) and primary sensory networks. This suggests HOFC is superior at capturing subtle but stable inter-network communication that LOFC might miss.
2. Reliability Benchmarks
- tHOFC and aHOFC: Demonstrated fair-to-good reliability (ICC > 0.4), performing comparably to LOFC but excelling in inter-network connectivity.
- dHOFC: Showed fair-to-moderate reliability, particularly for connections with high connectivity strength.
- Regional Gains: In high-level cognitive areas like the prefrontal cortex, aHOFC showed a significant reliability gain over LOFC, highlighting these regions as prime targets for future HOFC-based biomarker research.
3. The Impact of Dynamics
For dHOFC, the length of the sliding window is a critical parameter. Shorter windows (~40 seconds) appeared more reliable in this cohort, though researchers must balance this against the risk of overestimating connectivity due to noise in extremely short windows.
Conclusion
This study confirms that HOFC metrics are reliable and stable across time, validating their potential for clinical applications. By capturing "modulatory" connectivity—where the state of one system influences another—HOFC provides a more nuanced view of brain organization than simple pairwise connections. Future work should explore the reliability of graph-theoretic properties derived from these high-order matrices to further unveil the complexities of the human connectome.
Summary updated