Causal Discovery for fMRI data: Challenges, Solutions, and a Case Study

Eric Rawls, Bryan Andrews, Kelvin Lim, Erich KummerfeldarXiv 2023 · DOI · OpenReview

Introduction

Functional Magnetic Resonance Imaging (fMRI) has long been a cornerstone of neuroimaging, primarily used to identify brain regions that exhibit synchronized activity. This "functional connectivity" relies on correlation, which identifies associations but fails to establish directionality. To understand the brain's information processing architecture and develop targeted clinical interventions, researchers must move toward "effective connectivity"—the directed, causal influence of one region over another. However, fMRI data presents unique hurdles for causal discovery algorithms, including high dimensionality, feedback loops, and temporal undersampling.

The Nine Challenges of Causal Discovery in fMRI

The authors identify a framework of nine fundamental challenges (C1-C9) that complicate causal inference in neural data. These include:

  • C2 (Cycles): The brain is inherently cyclical, yet many algorithms assume acyclic structures.
  • C3 (Undersampling): fMRI samples signals much slower than the underlying neural firing rates.
  • C8 (Scale-free structure): Brain networks contain highly connected "hubs" that can confound standard discovery methods.
Comparison of causal discovery algorithms across nine fundamental challenges (C1-C9).

As shown in the comparison table, traditional methods like GES often fail to address critical biological realities like feedback loops or non-Gaussianity, whereas newer frameworks like GANGO are specifically designed to navigate these constraints.

Methodology: The GANGO Algorithm

The centerpiece of this research is GANGO (Greedy Adjacencies with Non-Gaussian Orientations). GANGO addresses the high dimensionality of fMRI data through a two-step process:

  1. Skeleton Discovery: Utilizing a fast greedy search (fGES) to identify connections between regions without assuming direction.
  2. Orientation: Leveraging the "non-Gaussianity" (the skewness and kurtosis of the data distribution) to determine causal direction.

A key insight is that standard preprocessing often removes non-Gaussian "noise," which the authors argue is actually the vital signal needed to distinguish cause from effect.

Case Study: The Human Connectome Project

The GANGO framework was applied to resting-state fMRI data from 442 healthy individuals in the Human Connectome Project (HCP). The brain was parcellated into 360 cortical regions, providing a high-resolution map of neural interactions.

Key Results

The application of GANGO yielded a sparse, directed graph (approximately 2.25% edge density) that remained fully connected, offering a clearer organizational picture than thresholded correlation maps.

  • Physiological Plausibility: Hubs were identified in regions associated with executive control and attention.
  • Network Topology: The causal maps exhibited "small-world" and "scale-free" properties, indicating high efficiency for information transfer.
  • Directional Dissociation: The analysis revealed significant differences between "indegree" (receiving) and "outdegree" (sending) connections, a distinction impossible to achieve with traditional functional connectivity.
Analysis of network topology showing scale-free degree distributions and small-world efficiency metrics.

Limitations and Future Directions

While GANGO represents a significant advancement, challenges remain. Latent variables—unmeasured influences from the peripheral nervous system or subcortical regions—can still introduce bias. Furthermore, while cortical signals are strongly non-Gaussian, subcortical regions often appear more Gaussian due to lower signal-to-noise ratios, complicating orientation. Future work may involve applying these causal methods to EEG data to leverage its superior temporal resolution and capture millisecond-scale brain dynamics.

Summary updated

Abstract

Designing studies that apply causal discovery requires navigating many researcher degrees of freedom. This complexity is exacerbated when the study involves fMRI data. In this paper we (i) describe nine challenges that occur when applying causal discovery to fMRI data, (ii) discuss the space of decisions that need to be made, (iii) review how a recent case study made those decisions, (iv) and identify existing gaps that could potentially be solved by the development of new methods. Overall, causal discovery is a promising approach for analyzing fMRI data, and multiple successful applications have indicated that it is superior to traditional fMRI functional connectivity methods, but current causal discovery methods for fMRI leave room for improvement.
BibTeX
@misc{rawls2023causal,
  author = {Eric Rawls and Bryan Andrews and Kelvin Lim and Erich Kummerfeld},
  title = {{Causal Discovery for fMRI data: Challenges, Solutions, and a Case Study}},
  year = {2023},
  eprint = {2312.12678},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.12678},
}
Figure