Method at a glance
- Modality: multimodal (time-resolved whole-brain or ROI fMRI)
- Measures: How the correlation structure of BOLD signals changes over time within a single scan, rather than being averaged into one static connectivity matrix. Sliding time windows produce a sequence of connectivity matrices that are clustered (typically k-means) into a small set of recurring ‘brain states’; dwell time, number of episodes, proportion of time, and switching rate between states then quantify the temporal dynamics of network organization.
- Procedure: Preprocess and denoise the time series (motion, WM/CSF regression, band-pass filtering); slide a tapered window (e.g. 30–60 s) in steps across the scan, computing a regularized correlation matrix per window; concatenate windows across participants and conditions and apply k-means clustering (with the elbow criterion choosing k) to extract recurring connectivity states; assign each window to a state and compute dynamic measures (dwell time, episodes, proportion, switch rate). Sensitivity checks vary window width, hold out data, and correlate states across subsets.
- Also called: DFC, time-varying connectivity, dynamic FC, brain states, sliding-window connectivity
Strengths
- Sensitive to within-scan variation that static connectivity averages away — better suited than static FC to capturing on-task vs off-task fluctuation and, some evidence suggests, to trait/state self-report and test–retest reliability (Marusak et al. 2018).
- Yields interpretable recurring ‘brain states’ (hypoconnected, DMN-anticorrelated, hyperconnected, etc.) that replicate across datasets and window lengths, giving a common vocabulary for comparing conditions.
- Lets an analyst ask temporally specific questions — how long the brain dwells in a state, how often it switches — that static connectivity cannot pose.
Limitations
- Sliding-window DFC is analytically sensitive: window width, clustering choices, and k all shape the states, and short windows risk mistaking sampling variability for real dynamics.
- Brain states are correlational and their behavioural meaning is not guaranteed — Treves et al. (2025) found states that separated task from rest yet did not track within-task attention or individual differences.
- Scrubbing motion frames breaks the continuity DFC needs, so motion must be handled by regression instead — leaving residual motion a live concern; head motion also differs systematically between task and rest.
Connections
- Key studies: Treves et al. (2025)
- Related methods: Resting-state functional connectivity, Breath-counting task (BCT)
Where static functional connectivity averages a whole scan into one correlation matrix, DFC asks how that correlation structure changes moment to moment. A tapered window slides across the time series, producing a sequence of connectivity matrices; these are clustered into a handful of recurring brain states, and the brain’s traffic between them — how long it dwells in each, how often it switches — becomes the object of study (Allen et al. 2014; Mooneyham et al. 2017).
Why the wiki has it
DFC is the central method of Treves et al. (2025), the wiki’s first first-hand study of breath-focused attention, and it is distinct enough from static resting-state FC to warrant its own page. The distinction is the whole point of that study: static connectivity gives the expected coarse picture (focused attention engages control/salience networks, anticorrelates the DMN), but the dynamic picture is where the interesting — and cautionary — findings live.
The recurring states
Across breath-meditation and rest, the clustering repeatedly recovers a small, replicable set of states: a hypoconnected state (low connectivity overall, thought to be an amalgam of infrequent states), a DMN-anticorrelated state, a hyperconnected state, and network-specific reductions (e.g. an SN-reduced state). Mooneyham et al. (2017) identified several of these around a breath exercise; Treves et al. recovered the same three plus one, and found the DMN-anticorrelated and hyperconnected states more prevalent, longer-dwelt, and more frequently switched-to during breath counting than rest.
The dynamic measures
Once windows are assigned to states, the standard readouts are dwell time (mean length of a continuous period in a state), number of episodes, proportion of time, and switch rate (transitions per unit time). These are what let DFC distinguish conditions that look similar statically — Treves et al.’s clearest task-vs-rest effects were in dwell time and switching, not in which states existed.
The caution the wiki attaches
DFC’s sensitivity cuts both ways. Its states replicate and its dynamic measures separate task from rest — but a state that marks a condition need not mark the behaviour the condition is meant to elicit. In Treves et al. the DMN-anticorrelated state best separated breath counting from rest yet was unrelated to within-task on/off-task attention, and no state predicted individual differences in attentiveness. The method reliably finds structure; whether that structure is the psychological variable of interest is a separate, harder question. See Treves et al. (2025).