On the nature of time-varying functional connectivity in resting fMRI

Daniel J. Lurie, Daniel Kessler, Danielle S. Bassett, Richard F. Betzel, Michael Breakspear, Shella Keilholz, Aaron Kucyi, Raphaël Liégeois, Martin A. Lindquist, Anthony R. McIntosh, Russell A. Poldrack, James M. Shine, William Hedley Thompson, Natalia Bielczyk, Linda Douw, Dominik Kraft, Robyn L. Miller, Muthuraman Muthuraman, Lorenzo Pasquini, Adeel Razi, Diego Vidaurre, Hua Xie, Vince D. CalhounPsyArXiv 2018 · DOI

TL;DR

This review tackles the complex and often debated topic of time-varying functional connectivity (TVC) in resting-state fMRI (rfMRI). It synthesizes current knowledge, clarifies terminology, and outlines key challenges and future directions for using TVC to understand brain dynamics, cognition, and neurological disorders.

What Problem Does This Paper Address?

The brain is a dynamic system, and understanding how its functional connections change over time (TVC) is crucial for grasping its architecture, cognitive processes, and the disruptions seen in illness. While TVC analysis of rfMRI data holds immense promise, there's significant debate about its statistical validity, physiological underpinnings, and relevance to cognition, especially during rest. This paper aims to clarify these issues, review the existing literature, and propose a path forward for rigorous TVC research.

Methods

The paper discusses a spectrum of analytical approaches for studying TVC, broadly categorized into:

  • Data-driven models: These methods directly analyze observed BOLD signals to estimate time-resolved connectivity. This includes:
  • Sliding-window approaches: The most common method, calculating correlations within moving time windows. Variations exist in window type, flexibility, and length.
  • Other methods: Including Hidden Markov Models (HMMs) for identifying brain states and time-frequency approaches.
  • Models of underlying neuronal dynamics: These approaches attempt to model the neural processes that generate BOLD TVC, often using biophysically derived differential equations. This involves fitting and inverting models to empirical data.

The paper also emphasizes the importance of rigorous statistical testing against null models and careful data cleaning to mitigate confounds like head motion.

Datasets

Not explicitly detailed as training or evaluation datasets. The paper discusses findings from numerous studies that have utilized resting-state fMRI (rfMRI) data, as well as multimodal neuroimaging data (e.g., EEG-fMRI, MEG).

Key Insights and Ideas

  • TVC is not a monolithic concept: Different analytical methods capture different aspects of brain dynamics, and the choice of method should be guided by the research question.
  • "Dynamic" vs. "Time-Varying": The authors advocate for "time-varying connectivity" (TVC) as a more broadly applicable term, acknowledging that "dynamic" can have varied interpretations.
  • The "resting state" is not quiescent: Even without explicit tasks, the brain exhibits complex, fluctuating patterns of activity and connectivity.
  • TVC might be more sensitive to pathology than static FC: Preliminary evidence suggests TVC could be a better marker for neurological and psychiatric conditions.
  • The brain's "stream of consciousness" may be reflected in TVC: Spontaneous fluctuations in mental states, even during rest, could be captured by TVC.
  • Head motion is a major confound: Robust data cleaning strategies are critical for reliable TVC estimation.
  • Arousal and sleep states significantly impact TVC: These factors need careful consideration and accounting for in TVC studies.
  • Static FC might mask underlying dynamics: Stable static FC patterns can arise from different temporal arrangements of underlying TVC, suggesting that static measures may not fully capture brain dynamics.

Key Results

  • A growing body of literature demonstrates that TVC can be reliably estimated from rfMRI data using various methods.
  • TVC has been associated with cognitive and behavioral traits, and altered resting TVC has been observed in various psychiatric and neurological conditions.
  • Studies using multimodal recordings (e.g., fMRI with EEG or intracranial recordings) suggest that BOLD TVC reflects underlying electrophysiological activity across different frequency bands.
  • TVC during anesthesia shows that some aspects of dynamic connectivity are present even in unconscious states, suggesting non-cognitive contributions.
  • TVC patterns are generally reproducible across individuals and sessions, indicating reliability.

Limitations and Future Work

  • Statistical validity: Debates persist regarding appropriate null models and statistical tests for TVC.
  • Physiological origins: While evidence points to neural activity, the exact contributions of neuromodulators, arousal, and physiological confounds (like head motion and cardiac/respiratory effects) are still being disentangled.
  • Cognitive relevance: The extent to which resting TVC reflects "spontaneous" cognition versus non-cognitive processes remains a key question. Distinguishing these is challenging.
  • Methodological standardization: Inconsistent terminology and operationalizations can muddy the waters.
  • Future directions:
    • Develop standardized terminology and frameworks.
  • Conduct more rigorous testing against null models.
  • Utilize multimodal approaches to better understand the neural basis of TVC.
  • Employ continuous task paradigms and causal manipulation (e.g., brain stimulation) to validate TVC findings.
  • Develop and refine data cleaning strategies specifically for TVC.
  • Investigate the role of arousal and sleep more thoroughly.
  • Explore the temporal dynamics of brain states ("meta-states") more deeply.
  • Integrate findings from data-driven and biophysical modeling approaches.

Summary updated

Abstract

The brain is a complex dynamical system composed of many interacting sub-regions. Knowledge of how these interactions reconfigure over time is critical to a full understanding of the brain’s functional architecture, the neural basis of flexible cognition and behavior, and how neural systems are disrupted in psychiatric and neurological illness. The idea that we might be able to study neural and cognitive dynamics through analysis of neuroimaging data has catalyzed substantial interest in methods which seek to estimate moment-to-moment fluctuations in functional connectivity (often referred to as “dynamic” or time-varying connectivity; TVC). At the same time, debates have emerged regarding the application of TVC analyses to resting fMRI data, and about the statistical validity, physiological origins, and cognitive relevance of resting TVC. These and other unresolved issues complicate the interpretation of resting TVC findings and limit the insights which can be gained from this otherwise promising research area. This article reviews the current resting TVC literature in light of these issues. We introduce core concepts, define key terms, summarize current controversies and open questions, and present a forward-looking perspective on how resting TVC analyses can be rigorously applied to investigate a wide range of questions in cognitive and systems neuroscience.
BibTeX
@misc{lurie2018nature,
  author = {Daniel J. Lurie and Daniel Kessler and Danielle S. Bassett and Richard F. Betzel and Michael Breakspear and Shella Keilholz and Aaron Kucyi and Raphaël Liégeois and Martin A. Lindquist and Anthony R. McIntosh and Russell A. Poldrack and James M. Shine and William Hedley Thompson and Natalia Bielczyk and Linda Douw and Dominik Kraft and Robyn L. Miller and Muthuraman Muthuraman and Lorenzo Pasquini and Adeel Razi and Diego Vidaurre and Hua Xie and Vince D. Calhoun},
  title = {{On the nature of time-varying functional connectivity in resting fMRI}},
  year = {2018},
  doi = {10.31234/osf.io/xtzre},
  howpublished = {PsyArXiv},
}
Figure