Resting-State fMRI Dynamics and Null Models: Perspectives, Sampling Variability, and Simulations

Robyn L. Miller, Anees Abrol, Tülay Adalı, Yuri Levin-Schwarz, Vince D. Calhoun

TL;DR

This paper argues that current null models for analyzing resting-state fMRI (rs-fMRI) dynamics are often poorly formulated, failing to adequately distinguish functionally relevant brain activity from noise. The authors propose that metrics focusing on spectrally distinguishable temporal epochs, rather than kurtosis, are more promising for identifying true brain dynamics.

What Problem Does This Paper Address?

The field of rs-fMRI is increasingly interested in "dynamics" – how brain activity changes over short timescales, rather than just looking at average activity over the entire scan. This shift is motivated by the understanding that the brain is inherently dynamic. To rigorously study these dynamics, researchers need robust statistical methods, specifically hypothesis testing frameworks and "null models." A null model represents a scenario where the phenomenon of interest (e.g., functionally relevant brain dynamics) is absent. The challenge is that existing null models for rs-fMRI dynamics are often flawed. They may not accurately capture the statistical properties of real brain data, or they might be too easily rejected by noise or artifacts, leading to false positives. This makes it difficult to confidently identify genuine brain dynamics.

Methods

The authors critically examine existing null models for rs-fMRI dynamics, particularly those based on stationary Gaussian processes and measures like kurtosis. They argue that these models often fail because:

  • Stationarity Assumption: Assuming statistical stationarity (where statistical properties don't change over time) in a null model doesn't guarantee the absence of "temporal epochs" – distinct periods of signal variation that could be indicative of brain dynamics.
  • Kurtosis Limitations: Kurtosis, a measure of "tailedness" in a distribution, is overly sensitive to extreme outliers (like motion artifacts) and less effective at capturing the gradual, spectrally distinct temporal variations that are more likely to reflect functional brain dynamics.

To address these limitations, the paper introduces and discusses a novel metric, denoted by the Greek letter Phi ($\Phi$), which is designed to capture "spectrally distinguishable temporal epochs" (SDTEs). This metric uses wavelet transforms to analyze time-frequency information and is extended to multivariate signals (representing multiple brain regions). The authors propose that $\Phi$ offers a more targeted measurement of characteristics that could form functionally relevant brain dynamics.

The paper also presents a series of simulation regimes designed to mimic different aspects of rs-fMRI data, including:

  • SCC Gaussians: Spectrally and covariance constrained multivariate Gaussian processes.
  • Noisy SCC Gaussians: SCC Gaussians with added high-amplitude spikes.
  • Spectrally and Statistically Non-stationary (SS Non-stationary): Simulating task-responsiveness with distinct windows for task-positive, task-negative, and non-responder networks.
  • Covariance-Dynamic SS Non-stationary: Similar to SS Non-stationary but with explicitly different covariance constraints for different time windows.

These simulations are used to evaluate the performance of multivariate kurtosis and the proposed $\Phi$ metric.

Datasets

The study utilizes previously published resting-state fMRI (rs-fMRI) data from a large multisite clinical study (Allen et al., 2014). This "Real Data" consists of network timecourses from 314 subjects. The authors also create five simulation regimes, described in the Methods section, which serve as synthetic datasets for evaluating their proposed metric and critiquing existing null models.

Key Insights and Ideas

  • Brain Dynamics are Ubiquitous: The paper strongly suggests that functionally relevant brain dynamics are not rare but are a constant feature of the awake human brain. This challenges the premise of null models that assume their absence.
  • Stationarity is Not Enough: Statistical stationarity in a null model does not preclude the presence of spectrally distinguishable temporal epochs, which are considered signatures of brain dynamics.
  • Kurtosis is a Poor Indicator of Dynamics: Multivariate kurtosis is highly sensitive to noise and artifacts (like motion) and is not a reliable measure for identifying functionally relevant brain dynamics. It tends to flag extreme outliers rather than meaningful temporal shifts.
  • Temporal Epochs as a Better Metric: Metrics that capture "spectrally distinguishable temporal epochs" (SDTEs) are more promising for identifying functionally relevant brain dynamics because these epochs are less likely to be solely due to nuisance factors.
  • The Need for Targeted Null Models: Null models should be built on assumptions that are complementary to the phenomenon of interest and should be designed to rarely produce signals indicative of that phenomenon. Current models often fail this test.

Key Results

  • Kurtosis Fails to Distinguish: In simulations, multivariate kurtosis identified a very high percentage (93.4%) of "Noisy SCC Gaussians" (which contain artificial spikes) as non-stationary or dynamic, while correctly identifying very few of the more complex, dynamically structured non-stationary regimes. Real data was also not statistically distinguishable from stationary Gaussian processes by kurtosis.
  • Phi Shows Promise: The proposed $\Phi$ metric, designed to capture SDTEs, showed more plausible ordering of dynamism across different simulation regimes and real data, distinguishing between noise-driven artifacts and more structured temporal variations.
  • Simulation Overlap: The study highlights that even with sophisticated simulations, the distributional tails of occupancy rates for connectivity states can overlap significantly between stationary and non-stationary models, making it difficult to draw clear inferences.

Limitations and Future Work

  • Metric Tuning and Validation: The authors acknowledge that their proposed $\Phi$ metric, while promising, requires further tuning and validation across a wider range of spectral/amplitude variations and noise types.
  • Exhaustive Statistical Measure: Identifying a single, exhaustive statistical measure that is appropriate for all possible types of spectral and amplitudinal variations, and that reliably validates as statistically significant, remains a challenge.
  • Understanding Brain Dynamics: A fundamental limitation is our incomplete understanding of what specific signal features truly reflect functionally relevant brain dynamics and how they manifest across different temporal, spatial, and functional scales.
  • Narrower Hypotheses: The authors advocate for a shift towards formulating narrower, more specific hypotheses about brain dynamics, which would allow for the development of more targeted null models and metrics.
  • Broader Framework: A comprehensive framework is needed to not only detect potential dynamic behavior but also to pinpoint specific timepoints where signals indicate underlying functional shifts.

Summary updated

Abstract

Studies of resting state functional MRI (rs-fRMI) are increasingly focused on “dynamics”, or on those properties of brain activation that manifest and vary on timescales shorter than the scan’s full duration. This shift in focus has led to a flurry of interest in developing hypothesis testing frameworks and null models applicable to the dynamical setting. Thus far however, these efforts have been weakened by a number of crucial shortcomings that are outlined and discussed in this short paper. We focus here on aspects of recently proposed null models that, we argue, are poorly formulated relative to the hypotheses they are designed to test, i.e. their potential role in separating functionally relevant BOLD signal dynamics from noise or intermittent background and maintenance type processes is limited by factors that are fundamental rather than merely quantitative or parametric. In this short position paper, we emphasize that (1) serious care must be exercised in building null models for rs-fMRI dynamics from distributionally stationary univariate or multivariate timeseries, i.e. timeseries whose values are each independently drawn from one pre-specified probability distribution; and (2) measures such as kurtosis that quantify over-concentration of observed values in the far tails of some reference distribution may not be particularly suitable for capturing signal features most plausibly contributing to functionally relevant brain dynamics. Other metrics targeted, for example, at capturing the epochal temporal variation that contributes heavily to dynamic functional connectivity estimates and is and often taken as a signature of brain responsiveness to stimuli or experimental tasks, could play a more scientifically clarifying role. As we learn more about the phenomenon of functionally relevant brain dynamics and its imaging correlates, scientifically meaningful null hypotheses and well-tuned null models will naturally emerge. We also revisit the important concept of distributional stationarity, discuss how it manifests within realizations versus across multiple realizations, and provide guidance on the benefits and limitations of employing this type of stationarity in modeling the absence of functionally relevant temporal dynamics in resting state fMRI. We hope that the discussions herein are useful, and promote thoughtful consideration of these important issues.
BibTeX
@inproceedings{miller2018resting,
  author = {Robyn L. Miller and Anees Abrol and Tülay Adalı and Yuri Levin-Schwarz and Vince D. Calhoun},
  title = {{Resting-State fMRI Dynamics and Null Models: Perspectives, Sampling Variability, and Simulations}},
  booktitle = {Frontiers in Neuroscience},
  year = {2018},
  doi = {10.3389/fnins.2018.00551},
}
Figure