Disambiguating brain functional connectivity

Eugene P. Duff, Tamar R. Makin, Michiel Cottaar, Stephen M. Smith, Mark W. WoolrichNeuroImage 2018 · DOI

TL;DR

This paper introduces a new method, Additive Signal Change (ASC), to better understand changes in brain functional connectivity (FC). By analyzing how both correlation and variance change between brain states, ASC can distinguish between true changes in neural coupling and changes due to signal amplitude or noise variations, offering a more nuanced interpretation of FC data.

What Problem Does This Paper Address?

Functional connectivity (FC) analysis, commonly using correlation, is a powerful tool for understanding how different brain regions interact. However, a significant limitation of correlation is its ambiguity: changes in correlation can arise from various underlying signal properties, not just alterations in neural coupling. For instance, increased noise levels or changes in the amplitude of neural signals can independently alter correlation values, making it difficult to infer the true nature of brain network interactions. This ambiguity hinders our ability to reliably interpret FC findings, especially in clinical contexts or when trying to understand complex network dynamics. While more advanced methods like effective connectivity exist, they often require strong assumptions. This paper aims to provide a more flexible and less assumption-laden approach to disambiguate the causes of observed FC changes.

Methods

The core of this paper is the Additive Signal Change (ASC) approach. It builds upon the observation that changes in functional connectivity (correlation) are often accompanied by changes in signal variance. ASC analyzes these simultaneous changes to infer the nature of the underlying signal modifications.

The method defines Additive Signal Changes (ASC) as a class of signal alterations where new signal is added to existing activity. This broad category includes:

  • Addition of uncorrelated signal: Signal added to one node that is independent of the signal in another node. This typically reduces correlation and increases variance.
  • Addition of common signal: The same latent signal component is added to both nodes. This can increase correlation and variance.
  • General additive changes: A broader category encompassing various additive signal modifications.

The key insight is that ASCs always produce changes in variance. This allows for hypothesis testing: if observed changes in variance are insufficient to explain the observed change in correlation, then the change is unlikely to be a pure ASC. The paper uses a Monte Carlo (MC)-based null-hypothesis approach to assess the likelihood that observed changes in correlation and variance can be explained by ASCs or their subclasses. This involves sampling from distributions of underlying true covariances to account for measurement uncertainty.

Datasets

The paper uses two types of data:

  • Simulated data: Used to demonstrate the effects of different types of additive signal changes on correlation and variance, and to validate the inference procedure.
  • Empirical fMRI data: Collected from 16 healthy volunteers scanned under five different steady-state conditions (rest, visual only, motor only, simultaneous visual and motor, and a visually-cued motor task). This dataset was used to apply the ASC analysis to real-world functional connectivity changes.

Key Insights and Ideas

  • Correlation is ambiguous: The paper strongly emphasizes that correlation alone is an insufficient metric for understanding changes in neural coupling.
  • Variance is key: Changes in signal variance are a crucial, often overlooked, indicator that can help disambiguate the causes of FC changes.
  • ASC provides a framework: The ASC approach offers a principled way to categorize and test hypotheses about the underlying causes of FC changes, moving beyond simple correlation.
  • Distinguishing noise from coupling: ASC can help differentiate between changes in FC due to increased noise or signal amplitude versus genuine alterations in neural interactions.
  • Rich diversity in FC changes: Applying ASC to fMRI data revealed that many observed FC changes can be explained by additive signal changes, suggesting that amplitude variations play a significant role in observed connectivity modulations.

Key Results

  • Simulations demonstrated that uncorrelated signals tend to decrease correlation, while common signals tend to increase it, with both affecting variance.
  • The ASC analysis applied to fMRI data showed that a substantial proportion of significant FC changes between resting and active states could be explained by additive signal changes (either uncorrelated, common, or general additive signals).
  • For visual stimulation, increased variance in visual regions was associated with increased correlation between visual nodes, explainable by common signals. Decorrelations with other regions were also explainable by additive changes.
  • For motor tasks, modest reductions in variance in motor regions were sufficient for additive changes (common signal) to explain correlation changes, including decreases between motor regions and increases with the cerebellum.
  • Some connections, particularly those involving the cerebellum during motor tasks, showed changes that could not be explained by additive signal changes, suggesting other mechanisms like synchronization without amplitude changes.

Limitations and Future Work

  • Gaussianity Assumption: The ASC inference procedure relies on assumptions of Gaussianity for the data and Wishart distributions for covariance matrices. This might limit its applicability to non-Gaussian data.
  • Scan Length: For fMRI, shorter scan durations can make it difficult to reject the null hypothesis for ASC, especially if variance changes are small. Longer scan durations (e.g., 13 minutes) are suggested to improve FC stability.
  • Global Signal Variations and Confounds: Like standard FC analysis, ASC can be sensitive to global signal variations and confounds, although it can help detect some of these effects (e.g., uncorrelated noise).
  • Future Work:
    • Developing random-effects or hierarchical modeling approaches for group-level ASC analysis.
  • Exploring more sensitive statistical correction methods for detecting changes in networks.
  • Extending ASC to partial correlation matrices and to characterize interactions among three or more regions simultaneously.
  • Applying ASC to dynamic functional connectivity analysis, particularly in interpreting transitions between states in dynamic FC models.

Summary updated

BibTeX
@article{duff2018disambiguating,
  author = {Eugene P. Duff and Tamar R. Makin and Michiel Cottaar and Stephen M. Smith and Mark W. Woolrich},
  title = {{Disambiguating brain functional connectivity}},
  journal = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.neuroimage.2018.01.053},
}
Figure