Resting-state functional connectivity

Method at a glance

• Modality: multimodal (task-free whole-brain fMRI)
• Measures: The correlation structure of spontaneous low-frequency (~0.01–0.1 Hz) BOLD fluctuations while a participant lies at rest without a task — used to identify intrinsic connectivity networks (default-mode, salience, executive/frontoparietal) and to quantify how strongly regions co-fluctuate, independent of any stimulus or instruction.
• Procedure: Participant rests in the scanner (eyes open or closed) for several minutes while functional volumes are collected. Time-series are preprocessed (motion correction, spatial normalization, band-pass filtering to the low-frequency window, nuisance regression of motion/WM/CSF), then analysed by one or more of: seed-based correlation (pick a region, map everywhere it correlates with), independent components analysis (ICA — data-driven decomposition into spatial networks), local-signal metrics (fALFF for low-frequency power, REHO for local coherence), or graph/multivariate methods (PLS).
• Also called: resting-state fMRI, RSFC, intrinsic connectivity networks, resting-state analysis

Strengths

• Task-free — usable in populations who cannot comply with task paradigms (dementia, young children, severe aphasia), which is precisely why it is the method of choice in clinical neurodegeneration work.
• Captures intrinsic network architecture that recurs across individuals and is stable enough to serve as a biomarker; the same canonical networks (DMN, salience, executive) appear across cohorts.
• Multiple analytic routes (ICA / seed / fALFF / REHO / PLS) can be run on the same data as a convergence check, so a finding need not depend on one pipeline’s assumptions.

Limitations

• Correlational and cross-sectional by nature — connectivity co-fluctuation does not establish direction or causation, and a resting session cannot say what the participant was doing mentally while ‘at rest’.
• Highly sensitive to head motion and physiological noise (cardiac, respiratory), which can masquerade as connectivity if not regressed out; standards for nuisance removal have shifted over time.
• Network-decomposition methods (ICA) can miss focal local abnormalities that local-signal metrics (fALFF, REHO) detect — the two see different things, so ‘no network difference’ is not ‘no difference’.
• ‘Resting state’ is not a defined cognitive state; what drifts through it (mind-wandering, body-wandering, rumination) varies and is only partly controlled.

Connections

• Key studies: Farb et al. (2012), Banellis et al. (2026), Weng et al. (2021), Schwarzlose et al. (2023), Treves et al. (2025), Cauda et al. (2011), Chang et al. (2013), Deen et al. (2011), Vercelli et al. (2015), Gong et al. (2015), Zeharia et al. (2019)
• Related methods: Multidimensional Experience Sampling (MDES), Dynamic functional connectivity (DFC), Breath-counting task (BCT)

The task-free imaging method behind the wiki’s intrinsic-network constructs — the default-mode, salience, and executive/frontoparietal networks are all defined by the correlation structure of spontaneous BOLD fluctuations that this method measures. Given its own page because it is used across several studies here and had been referenced without one.

Why it recurs in this wiki

Three uses, three purposes:

• Clinical, because it needs no task. Farb et al. (2012) use it in frontotemporal dementia — a population where task compliance is unreliable and some patients are too aphasic for standard testing — to show frontolimbic disconnection plus prefrontal hyperconnectivity (“prefrontal isolation”). The task-free property is the whole reason the study is possible.
• To parcellate a structure by its intrinsic connectivity. Cauda et al. (2011) use seed-based rsFC plus three clustering methods (k-means, hierarchical, fuzzy c-means) to partition the insula into a ventral-anterior salience network and a dorsal-posterior sensorimotor network — recovering the insula’s anterior/posterior functional dichotomy from spontaneous BOLD alone, no task and no cytoarchitecture. The wiki’s example of rsFC as a parcellation tool (connectivity-based subdivision) rather than a network-mapping or biomarker one, and of using method convergence (three clustering algorithms agreeing) as the check. Deen et al. (2011) apply the same connectivity-clustering approach voxelwise (k-means, k=3) to both insulae and add a second kind of validation the others lack — carrying the resting-defined clusters into an independent task dataset (disgust vs neutral images) to show they are functionally distinct. Chang et al. (2013) push the tool further still: k-means parcellation of the right insula with a validity index selecting cluster number (k=3), then cross-validated against an entirely different data class — meta-analytic coactivation across ~4400 studies — which recovers the same three networks. Three grades of validation for the same tripartition: cross-algorithm convergence (Cauda), independent-task validation (Deen), and cross-data-class replication (Chang). Vercelli et al. (2015) then push the same tool to its high-dimensional limit — 12 fuzzy-c-mean nodes per insula, where the fuzzy algorithm detects parcels and their borders in one pass — and show the finer nodes carry the “echoes” of the default-mode and dorsal-attentional networks nested inside the two dominant insular patterns.
• To characterize spontaneous thought. Banellis et al. (2026) map the intrinsic connectivity supporting body-oriented mind-wandering, finding a thalamus/striatum–somatomotor mode distinct from the DMN.
• As a treatment target. Weng et al. (2021) frame meditation as engaging interoception/salience networks and disengaging the DMN — network changes stated in the vocabulary of resting-state connectivity.
• As the rest baseline against which a task is contrasted — statically and dynamically. Treves et al. (2025) use resting-state as the comparison condition for a breath-counting task, and their study is the wiki’s clearest demonstration of the static/dynamic split: static FC gave the expected coarse picture (breath focus raises salience↔executive coupling, anticorrelates the DMN), while the more informative — and cautionary — findings lived in the time-varying brain states.
• As a large-N, preregistered biomarker screen. Schwarzlose et al. (2023) use it in 7,760+ ABCD children to link sensory over-responsivity to network-level FC — and model the method’s chief liability by design: exploratory FC screening in one dataset (Y0), preregistered replication in an independent dataset (Y2) collected 2 years later, retaining only pairs significant in both. It is the wiki’s example of the field’s answer to correlational-and-cross-sectional FC: hold-out replication instead of a single fishing expedition, at the cost (the authors note) of being unable to model FC change over time.

The analytic toolkit, briefly

The Farb (2012) paper is a compact tour of the method’s variants, run together for convergence: group ICA (data-driven network maps), fALFF (fractional amplitude of low-frequency fluctuation — local low-frequency power), REHO (regional homogeneity — local coherence with neighbouring voxels), seed-based univariate connectivity, and PLS as a multivariate replication. A methodological lesson recorded there: the local metrics (fALFF/REHO) flagged a right anterior-insula abnormality that the whole-brain ICA missed — so the choice of analysis, not just the data, determines what is found.