Directed functional connectivity using dynamic graphical models

Simon Schwab, Ruth Harbord, Valerio Zerbi, Lloyd T. Elliott, Soroosh Afyouni, Jim Q. Smith, Mark W. Woolrich, Stephen M. Smith, Thomas E. NicholsNeuroImage 2018 · DOI

TL;DR

This paper introduces Dynamic Graphical Models (DGMs), a novel method for estimating directed functional connectivity in brain networks that can capture time-varying relationships and handle complex network structures, including feedback loops. DGMs show promising accuracy in simulations, even when faced with confounding hemodynamic delays, and reveal consistent patterns in human resting-state fMRI data.

What Problem Does This Paper Address?

Understanding how different brain regions communicate is crucial for deciphering brain function and dysfunction in neurological and psychiatric disorders. Traditional methods for measuring functional connectivity (FC) often rely on static correlations, which fail to capture the dynamic nature of brain activity. Furthermore, estimating the direction of information flow (effective connectivity) is a significant challenge, particularly in fMRI due to the slow hemodynamic response and its variability across brain regions and individuals. This variability can confound methods that rely on temporal precedence (like Granger causality), leading to incorrect directionality estimates. The authors aim to address these limitations by proposing a method that can estimate directed, dynamic functional connectivity, is robust to hemodynamic delays, and can handle complex network structures.

Methods

The core of this paper is the introduction of Dynamic Graphical Models (DGMs). DGMs are a collection of Dynamic Linear Models (DLMs), one for each node in the network. Each DLM models the time series of a "child" node as a regression on the time series of its "parent" nodes. The key innovations are:

  • Time-Varying Coefficients: The connectivity strengths (regression coefficients) between nodes are allowed to change over time, modeled as a random walk. This captures the dynamic nature of brain networks.
  • Directed Relationships: The regression framework inherently models directed influences from parent to child nodes.
  • No Acyclic Constraint: Unlike some previous methods, DGMs do not assume that brain networks are acyclic, allowing for the modeling of feedback loops and reciprocal connections.
  • Robustness to Hemodynamic Delays: DGMs estimate instantaneous relationships between nodes, making them less susceptible to confounds introduced by systematic variations in the hemodynamic response function (HRF) across brain regions.
  • Model Evidence for Selection: The method uses model evidence (ME) to select the most likely set of parents for each node, effectively performing network discovery.
  • Pruning Method: A post-hoc pruning process is introduced to simplify networks by converting bidirectional connections to unidirectional ones if the reduction in model evidence is minimal, promoting parsimony.

Datasets

The authors used two main types of datasets for their evaluation:

  • Simulated fMRI Data: Various synthetic datasets were generated to test the DGM's performance under controlled conditions. These included:
  • A 5-node network with both stationary and dynamic connection strengths.
  • Simulations with systematic offsets in the hemodynamic response function (HRF) between nodes, ranging from <0.4s to 1.9s, to specifically test robustness to lag confounds.
  • An 8-node network with reciprocal connections.
  • Longer (60 min) simulations to assess the impact of session duration.
  • Human Resting-State fMRI Data: Resting-state fMRI data from 500 participants from the Human Connectome Project (HCP) were used. Resting-state networks (RSNs) were treated as nodes in the analysis.
  • Mouse Resting-State fMRI Data: Data from two previously published mouse fMRI studies were analyzed to evaluate the plausibility of DGM-derived networks in a different species.

Key Insights and Ideas

  • Dynamic Connectivity is Key: The study strongly suggests that brain connectivity is inherently dynamic, and static measures miss crucial information. DGMs provide a way to capture these dynamic shifts.
  • Hemodynamic Lags are a Major Hurdle (and DGM's Strength): The variability in HRF across brain regions is a significant challenge for methods relying on temporal precedence. DGMs' focus on instantaneous relationships offers a robust solution, maintaining reasonable accuracy even with substantial lag confounds.
  • Networks are Complex (and Often Cyclic): The ability of DGMs to model cyclic connections is important, as real brain networks are not strictly acyclic due to reciprocal connections.
  • The Default Mode Network's Influence: In human resting-state data, the DMN consistently shows influence on other networks, particularly the cerebellum, limbic, and auditory/temporal regions.
  • Visual Network Reciprocity: A consistent reciprocal relationship was observed between the visual medial and visual lateral networks.

Key Results

  • Simulation Performance:
    • On dynamic network simulations, DGM achieved higher sensitivity (70-79%) and specificity (70-79%) compared to Patel's (42-82%) and Lingam's (41-80%) methods.
  • DGM demonstrated a sensitivity of 72%-77% for detecting true directions even with lag confounds of up to 0.8s. While stronger lags reduced sensitivity, they did not increase false positives.
  • For stationary data, Lingam performed best in sensitivity (70%), but DGM still showed competitive performance.
  • The method showed good reproducibility in human resting-state fMRI data, with consistent edges observed across participants and runs.
  • Human fMRI Findings:
    • The cerebellum consistently received input from other RSNs.
  • Reciprocal connections were found between visual medial and visual lateral networks.
  • Mouse fMRI Findings:
    • Highly reproducible directed connectivity was found, with evidence of the orbital cortex projecting to the cingulate cortex, and CA1/DG also feeding into the cingulate cortex, aligning with anatomical tracing studies.

Limitations and Future Work

  • Computational Cost: While DGMs are more efficient than some alternatives, estimating networks for very large numbers of nodes (e.g., >25) remains computationally intensive. The authors mention ongoing work on greedy search algorithms to speed this up.
  • Heuristic Approach: The authors acknowledge that DGM is a heuristic approach and requires thorough validation in different contexts.
  • Comparison with DCM: The paper focuses on comparing DGM to observational methods (Patel, Lingam). Future work will involve comparing DGM to Dynamic Causal Modeling (DCM), particularly spectral DCM (spDCM), to explore effective connectivity.
  • Construct Validity: Establishing construct validity between DGM and spectral DCM for resting-state fMRI is a stated future goal.
  • Generalizability: While the paper demonstrates applicability to fMRI, the authors suggest DGMs could be applied to other neuroimaging modalities with higher temporal resolution, such as EEG and MEG.
  • Mouse Data Interpretation: The authors note that comparing functional connectivity to anatomical connectivity has limitations, as they are conceptually different.

Summary updated

BibTeX
@article{schwab2018directed,
  author = {Simon Schwab and Ruth Harbord and Valerio Zerbi and Lloyd T. Elliott and Soroosh Afyouni and Jim Q. Smith and Mark W. Woolrich and Stephen M. Smith and Thomas E. Nichols},
  title = {{Directed functional connectivity using dynamic graphical models}},
  journal = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.neuroimage.2018.03.074},
}
Figure