Causal Probabilistic Graphical Models for Decoding Effective Connectivity in Functional Near InfraRed Spectroscopy

Uncovering effective relations from non-invasive functional neuroimaging data remains challenging because the physical truth does not match the modelling assumptions often made by causal models. Here, we explore the use of causal Probabilistic Graphical Models for decoding the effective connectivity from functional optical neuroimaging. Our hypothesis is that directions of arcs of the connectivity network left undecided by existing learning algorithms can be resolved by exploiting prior structural knowledge from the human connectome. A variant of the fast causal inference algorithm, seeded FCI, is proposed to handle prior information. For evaluation, we used an existing dataset from prefrontal cortical activity of a cohort of 62 surgeons of varying expertise whilst knot-tying was monitored using fNIRS. Seeded FCI is used to built the prefrontal effective networks across expertise groups to reveal expertise-dependent differences. As hypothesized, the incorporation of prior information from the connectome reduces the set of undecided links. Good nomological validity is achieved when data is retrospectively compared to the findings in the original publication of the dataset. We contribute to the analysis of effective connectivity in fNIRS with the incorportation of structural information, and contribute to the field of causal PGMs with a new structure learning algorithm capable of exploiting existing knowledge to reduce the number of links remaining undecided when only information from observations is used. This work has implications thus for both, the AI and the neuroscience communities.