AbstractFlexible functional interactions among brain regions mediate critical cognitive functions. Such interactions can be measured from functional magnetic resonance imaging (fMRI) data with either instantaneous (zero-lag) or lag-based (time-lagged) functional connectivity; only the latter approach permits inferring directed functional interactions. Yet, the fMRI hemodynamic response is slow, and sampled at a timescale (seconds) several orders of magnitude slower than the underlying neural dynamics (milliseconds). It is, therefore, widely held that lag-based fMRI functional connectivity, measured with approaches like as Granger-Geweke causality (GC), provides spurious and unreliable estimates of underlying neural interactions. Experimental verification of this claim has proven challenging because neural ground truth connectivity is often unavailable concurrently with fMRI recordings. We address this challenge by combining machine learning with GC functional connectivity estimation. We estimated instantaneous and lag-based GC functional connectivity networks using fMRI data from 1000 participants, drawn from the Human Connectome Project database. A linear classifier, trained on either instantaneous or lag-based GC, reliably discriminated among seven different task and resting brain states, with over 80% cross-validation accuracy. With network simulations, we demonstrate that instantaneous and lag-based GC exploited interactions at fast and slow timescales, respectively, to achieve robust classification. With human fMRI data, instantaneous and lag-based GC identified distinct, cognitive core networks. Finally, variations in GC connectivity explained inter-individual variations in a variety of cognitive scores. Our findings show that instantaneous and lag-based methods reveal complementary aspects of functional connectivity in the brain, and suggest that slow, directed functional interactions, estimated with fMRI, provide robust markers of behaviorally relevant cognitive states.Author SummaryFunctional MRI (fMRI) is a leading, non-invasive technique for mapping networks in the human brain. Yet, fMRI signals are noisy and sluggish, and fMRI scans are acquired at a timescale of seconds, considerably slower than the timescale of neural spiking (milliseconds). Can fMRI, then, be used to infer dynamic processes in the brain such as the direction of information flow among brain networks? We sought to answer this question by applying machine learning to fMRI scans acquired from 1000 participants in the Human Connectome Project (HCP) database. We show that directed brain networks, estimated with a technique known as Granger-Geweke Causality (GC), accurately predicts individual subjects’ task-specific cognitive states inside the scanner, and also explains variations in a variety of behavioral scores across individuals. We propose that directed functional connectivity, as estimated with fMRI-GC, is relevant for understanding human cognitive function.
Design of Deep Learning Model for Task-Evoked fMRI Data Classification
