Interpreting temporal fluctuations in resting-state functional connectivity MRI

Raphaël Liégeois, Timothy O. Laumann, Abraham Z. Snyder, Juan Helen Zhou, B. T. Thomas YeoNeuroImage 2017 · DOI · OpenReview

Abstract

Resting-state functional connectivity is a powerful tool for studying human functional brain networks. Temporal fluctuations in functional connectivity, i.e., dynamic functional connectivity (dFC), are thought to reflect dynamic changes in brain organization and non-stationary switching of discrete brain states. However, recent studies have suggested that dFC might be attributed to sampling variability of static FC. Despite this controversy, a detailed exposition of stationarity and statistical testing of dFC is lacking in the literature. This article seeks an in-depth exploration of these statistical issues at a level appealing to both neuroscientists and statisticians. We first review the statistical notion of stationarity, emphasizing its reliance on ensemble statistics. In contrast, all FC measures depend on sample statistics. An important consequence is that the space of stationary signals is much broader than expected, e.g., encompassing hidden markov models (HMM) widely used to extract discrete brain states. In other words, stationarity does not imply the absence of brain states. We then expound the assumptions underlying the statistical testing of dFC. It turns out that the two popular frameworks - phase randomization (PR) and autoregressive randomization (ARR) - generate stationary, linear, Gaussian null data. Therefore, statistical rejection can be due to non-stationarity, nonlinearity and/or non-Gaussianity. For example, the null hypothesis can be rejected for the stationary HMM due to nonlinearity and non-Gaussianity. Finally, we show that a common form of ARR (bivariate ARR) is susceptible to false positives compared with PR and an adapted version of ARR (multivariate ARR). Application of PR and multivariate ARR to Human Connectome Project data suggests that the stationary, linear, Gaussian null hypothesis cannot be rejected for most participants. However, failure to reject the null hypothesis does not imply that static FC can fully explain dFC. We find that first order AR models explain temporal FC fluctuations significantly better than static FC models. Since first order AR models encode both static FC and one-lag FC, this suggests the presence of dynamical information beyond static FC. Furthermore, even in subjects where the null hypothesis was rejected, AR models explain temporal FC fluctuations significantly better than a popular HMM, suggesting the lack of discrete states (as measured by resting-state fMRI). Overall, our results suggest that AR models are not only useful as a means for generating null data, but may be a powerful tool for exploring the dynamical properties of resting-state fMRI. Finally, we discuss how apparent contradictions in the growing dFC literature might be reconciled.

Introduction: Is Your Brain Really "Switching States"?

In the world of neuroimaging, few topics are as hot—or as controversial—as dynamic Functional Connectivity (dFC). We’ve all seen the colorful movies of brain networks "flickering" and "switching" between discrete states while a subject simply rests in a scanner. But are these fluctuations representing a non-stationary "switching" of brain organization, or are they just the statistical echoes of a simpler, linear process?

A recent paper by Liégeois and colleagues provides a rigorous statistical deep-dive into this debate. Their findings suggest we might need to trade our "discrete state-switching" metaphors for the language of linear dynamical systems.

Defining the Terms

To understand the debate, one must first be clear on the terminology. The distinction between "static" and "dynamic" measures, and "stationary" versus "non-stationary" processes, is often blurred in the literature.

Table 1: Definitions of key statistical terms used in dFC research.

What Problem Does This Paper Address?

For years, the standard approach to functional connectivity (FC) was "static"—averaging correlations across an entire 10-minute fMRI scan. Recently, the field shifted toward "dynamic" FC, using sliding windows to see how connectivity changes over seconds. This led to the popular theory that the brain transitions through a repertoire of discrete "states."

However, a skeptical camp emerged, arguing that these "fluctuations" might just be sampling variability—the "noise" you get when you calculate correlations on short segments of data. The problem is that the field lacks a unified statistical framework to test these ideas. This paper steps in to clarify what "stationarity" actually means in a neuroscientific context and tests whether the "brain state" model actually fits the data better than simpler alternatives.

Key Insights: The Stationarity Misconception

One of the most important conceptual "aha" moments is that a process can be statistically stationary (its ensemble statistics don't change over time) and still exhibit "states." For example, a Hidden Markov Model can be stationary. Therefore, rejecting a null model doesn't automatically prove the brain is "non-stationary"; it might just mean the brain is non-linear or non-Gaussian.

Figure 1: Illustration of stationary (A) vs. non-stationary (B, C) processes. Stationarity depends on ensemble statistics across realizations.

As shown in Figure 1, Process $U_t$ is stationary because its mean is constant over time and across realizations. Process $V_t$ is non-stationary because its mean changes at time $t_0$. Process $W_t$ is non-stationary because the mean differs between realizations, even if it's constant within each one.

Methods and Datasets

The authors compared several frameworks for modeling fMRI time series:

  1. Phase Randomization (PR) & Multivariate Autoregressive Randomization (ARR): These are "null" models that generate surrogate data that is linear, stationary, and Gaussian but preserves the original data's correlation structure.
  2. Bivariate ARR: A common but limited version of ARR that looks at pairs of regions in isolation.
  3. Hidden Markov Models (HMM): The gold standard for the "discrete state-switching" hypothesis.
  4. AR(1) Models: Simple linear models where the current brain state is a linear function of the immediate previous state plus some noise.

The team used high-quality resting-state fMRI data from the Human Connectome Project (HCP) S900 release, focusing on the 100 participants with the lowest motion to minimize confounding effects.

Key Results

  • Failure to Reject: Using rigorous multivariate ARR and PR tests, the researchers could not reject the stationary linear Gaussian null hypothesis for 99% of the low-motion participants. For most people, their brain fluctuations look exactly like what you'd expect from a simple linear system.
  • The Bivariate Trap: The authors show that bivariate ARR is highly susceptible to false positives. Because it ignores the wider network context, it often "rejects" the null hypothesis even when the data is perfectly linear and stationary.
  • AR Models are the "Sweet Spot": First-order AR models (AR(1)) capture both the static correlation and the "one-lag" temporal dependency. The authors found that these simple models explain dFC fluctuations significantly better than static models and, surprisingly, better than HMMs ($p < 1e-30$).

Conclusion

By applying "Occam’s Razor" to brain dynamics, this paper suggests that before we reach for complex "state-switching" theories, we should first appreciate how much of the brain's complexity can be captured by simple linear elegance. While the brain is undoubtedly dynamic, the fluctuations measured by resting-state fMRI are often best described as linear and continuous rather than discrete and switching.

Summary updated

Abstract

Highlights•Space of stationary models bigger than expected; includes hidden Markov model (HMM).•Phase & autoregressive randomizations test for stationarity, linearity, Gaussianity.•Resting-state fMRI is mostly stationary, linear, and Gaussian.•1st order autoregressive (AR) model encodes static & one-lag FC.•1st order AR model explains sliding window correlations very well, better than HMM.
BibTeX
@article{liegeois2017interpreting,
  author = {Raphaël Liégeois and Timothy O. Laumann and Abraham Z. Snyder and Juan Helen Zhou and B. T. Thomas Yeo},
  title = {{Interpreting temporal fluctuations in resting-state functional connectivity MRI}},
  journal = {NeuroImage},
  year = {2017},
  doi = {10.1016/j.neuroimage.2017.09.012},
}

Loading related content...

Figure