Title:Reproducible measures of correlative and causal brain connectivity

Abstract:Patterns of interregional brain connectivity characterize the function of the human nervous system in both health and disease. Such connectivity can be inferred non-invasively by analyzing the blood-oxygen-level-dependent signal from functional magnetic resonance imaging. However, approaches to quantify this connectivity must solve an under-constrained problem with several potential solutions due to noise and spatial resolution limitations during imaging. Previous research has insufficiently evaluated the reproducibility of the traditional approaches and these approaches have been insufficiently regularized to increase solution stability. We propose three new measures of correlative and causal connectivity which incorporate explicit regularization through a structural connectivity prior from diffusion MRI. These new measures, which exploit a machine learning formulation for efficient computation, are evaluated against traditional measures for reproducibility and predictive power. In particular, correlation, partial correlation, spectral density, and a novel measure of functional connectivity based on machine learning feature importance (MLFC) are compared. Additionally, measures of effective connectivity are compared to two proposed connectivity measures: a machine learning approach and a Granger-causal approach incorporating a low-dimensional projection with a diffusion MRI-derived tractography prior. In addition to reproducibility, we quantify the capacity each measures has to infer three study participant traits: a physiologic trait, a cognitive trait, and a combined physiologic and cognitive trait. The proposed measures prove to be both reproducible and predictive, demonstrating the strong potential for the proposed connectivity measures to be broadly used in future neuroimaging studies.