Predicting the fMRI signal fluctuation with echo-state neural networks trained on vascular network dynamics

Abstract

Resting-state functional MRI (rs-fMRI) studies have revealed specific low-frequency hemodynamic signal fluctuations (<0.1 Hz) in the brain, which could be related to oscillations in neural activity through several mechanisms. Although the vascular origin of the fMRI signal is well established, the neural correlates of global rs-fMRI signal fluctuations are difficult to separate from other confounding sources. Recently, we reported that single-vessel fMRI slow oscillations are directly coupled to brain state changes. Here, we used an echo-state network (ESN) to predict the future temporal evolution of the rs-fMRI slow oscillatory feature from both rodent and human brains. rs-fMRI signals from individual blood vessels that were strongly correlated with neural calcium oscillations were used to train an ESN to predict brain state-specific rs-fMRI signal fluctuations. The ESN-based prediction model was also applied to recordings from the Human Connectome Project (HCP), which classified variance-independent brain states based on global fluctuations of rs-fMRI features. The ESN revealed brain states with global synchrony and decoupled internal correlations within the default-mode network.