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

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.

Figure 1:

Using CDA on parcellated fMRI data is complicated by the likely presence of both cycles and latent variables. This figure illustrates a toy example of parcels P1, P2, and P3, with unmeasured confounding of P1 and P2 from the peripheral nervous system (PNS). P1 has a cycle of length 1, as activity in P1 directly impacts further activity in P1 (for example, the firing of interneurons in a brain area will suppress the firing of other neurons). There is a 2-cycle between P1 and P2, as each directly causes the other. There is also a 3-cycle between P1, P2, and P3. Many methods capable of learning causal cycles are limited to learning only cycles with length greater than 2, however cycles of all lengths are likely present in brain parcellations.

Figure 2:

Application of CDA methods to BOLD data allows estimating patterns of directed connectivity, separating incoming and outgoing connections. In these brain images, circles are plotted at the 3D centroids of cortical parcels and the size of the circle indicates the number of connections (the degree) of the parcel. Individual causal connectomes reveal a striking dissociation between brain regions with especially high indegree and outdegree. Different colors refer to different brain functional communities (resting-state networks).

Figure 3:

The case study that developed the GANGO framework found that the causal connectomes produced were both scale-free and small-world graphs. The scale-free nature of GANGO connectomes is demonstrated by degree distributions that were more skewed than random graphs. The small-world nature of GANGO connectomes is demonstrated by global efficiency that is nearly equal to random graphs but higher than lattice graphs, combined with local efficiency that is greater than random graphs. Panels illustrating scale-free connectivity are adapted with permission from Rawls et al. [2022].