Brain activity is highly dynamic even at rest, and this feature has started to be captured in functional magnetic resonance imaging studies through dynamic functional connectivity (dFC) approaches. Total activation (TA), when coupled to the generation of innovation-driven co-activation patterns (iCAPs), is a state-of-the-art dFC methodological pipeline that can accurately disentangle spatially overlapping resting-state networks, and extract their time courses of activity.
To go one step further in characterising brain dynamics, we propose to look at the interactions between those extracted networks with a novel sparse coupled hidden Markov model (SCHMM) framework. To do so, we parameterise their activity profiles while enabling a sparse set of cross-network couplings.
On artificial data, our SCHMM framework outperforms standard correlational approaches in accurately retrieving cross-network relationships, and predicts network dynamics more accurately than with parallel uncoupled HMMs. On real resting-state data, we observe a set of directional cross-network modulations largely conserved across two independent datasets.