Identification of functional dynamic brain states based on graph attention networks

Investigation of the functional dynamics of the human brain can help to unveil inherent cognitive systems. In this study, we adopted a graph attention network-based anomaly detection technique to identify abrupt changes in functional time series. We used the resting-state functional magnetic resonance imaging data of 1010 participants from the Human Connectome Project. By applying multivariate time series anomaly detection using the graph attention network approach, we identified three distinct brain states, termed S1, S2, and S3. We further generated low-dimensional representations of functional connectivity (i.e., gradients) for each brain state and compared these gradients among brain states. S1 and S3 exhibited segregated network patterns, whereas S2 displayed more integrated patterns. A topological analysis based on the graph measures revealed that the integrated state (S2) exhibited strong inter-regional connectivity. Further, the two segregated states exhibited distinct patterns, with S1 being more involved in the somatomotor network and S3 being related to higher-order association areas. When we assessed the transitions between brain states, transitions between the low-level sensory (S1) and higher-order default mode states (S3), as well as between the sensory-focused segregated state (S1) and integrated state (S2), were associated with sensory/motor and memory-related tasks. In contrast, the transitions between the integrated (S2) and segregated states with higher centrality in the default mode region (S3) were found to be related to language and reward tasks. These findings indicate that the proposed approach captures changes in individual participant-level brain dynamics, thereby enabling the assessment of inherently dynamic brain systems.