Increasing robustness of pairwise methods for effective connectivity in magnetic resonance imaging by using fractional moment series of BOLD signal distributions

Estimating causal interactions in the brain from functional magnetic resonance imaging (fMRI) data remains a challenging task. Multiple studies have demonstrated that all current approaches to determine direction of connectivity perform poorly when applied to synthetic fMRI datasets. Recent advances in this field include methods for pairwise inference, which involve creating a sparse connectome in the first step, and then using a classifier in order to determine the directionality of connection between every pair of nodes in the second step. In this work, we introduce an advance to the second step of this procedure, by building a classifier based on fractional moments of the BOLD distribution combined into cumulants. The classifier is trained on datasets generated under the dynamic causal modeling (DCM) generative model. The directionality is inferred based on statistical dependencies between the two-node time series, for example, by assigning a causal link from time series of low variance to time series of high variance. Our approach outperforms or performs as well as other methods for effective connectivity when applied to the benchmark datasets. Crucially, it is also more resilient to confounding effects such as differential noise level across different areas of the connectome. Estimating causal interactions from functional magnetic resonance imaging (fMRI) data is a formidable task. Recent advances in this field include methods for pairwise inference. In the first step of this procedure, connections are revealed by means of functional connectivity. In the second step, every detected connection is analyzed separately to reveal the direction of the causal links. We introduce an advance to the second step of this procedure by building a classifier based on the novel concept of fractional moments of the BOLD distribution combined into cumulants. The classifier is trained on datasets generated under the dynamic causal modeling (DCM) generative model. Using fractional cumulants gives a measure resilient to confounding effects such as differential noise levels across different areas of the connectome.