The unfolding symphony: unraveling brain dynamics via data-driven multi-scale functional connectivity in task-state fMRI

Task-state functional magnetic resonance imaging (fMRI) can characterize both regional responses and changes in statistical dependence among distributed brain systems. This narrative review examines dynamic and multiscale connectivity within a broader framework that distinguishes task-modulated functional connectivity, trial-wise covariation, effective connectivity, and data-driven state analysis. Sliding-window correlation, hidden Markov models, and co-activation patterns are compared by their data representations, temporal assumptions, and computational procedures; graph representations can be combined with each of these strategies. We discuss supervised recurrent, transformer, and graph models alongside unsupervised state discovery, self-supervised representation learning, and emerging brain foundation models. Explainability requires evaluation of model attributions, including graph masks and perturbation-based analyses, rather than anatomical plausibility alone. The spatial framework includes whole-brain networks, subnetworks, atlas-defined regions of interest, and voxel- or layer-level measurements. Proposed links across these levels remain hypotheses: faster sampling of the blood oxygenation level-dependent (BOLD) signal does not directly resolve millisecond neural interactions, and resting-state or animal findings do not establish the corresponding mechanism during human tasks. Shared task-evoked responses, hemodynamic variability, motion, and analytical choices can all alter apparent connectivity. Progress therefore depends on matching the estimator to the scientific question, comparing against activation and static-connectivity baselines, and validating results across participants, sessions, and acquisition settings.