Source-linked AI summary
Scale-Free and Multifractal Time Dynamics of fMRI Signals during Rest and Task
P. Ciuciu, G. Varoquaux, P. Abry, S. Sadaghiani, A. Kleinschmidt
TL;DR
Prior fMRI scaling analyses often focused on functional networks and a single second-order exponent, leaving artifact specificity and richer non-Gaussian dynamics unresolved. This study combines MSDL with WLMF across rest and task, finding that long-memory decreases are widespread whereas task-related multifractality is specific to functional networks. The results identify multifractal modulation as the key property distinguishing functional networks from artifacts within this dataset.
Problem
Prior work often used univariate analyses and single scaling exponents, without systematically comparing functional-network and artifact dynamics or capturing departures from Gaussianity and self-similarity.
Method
The study combines MSDL spatial decomposition with Wavelet Leader-based Multifractal analysis of fMRI signals from functional and artifactual components during rest and task.
Results
Long memory decreases under task throughout the brain, while task-related multifractal modulation is significant only in functional networks.
Takeaways & Limitations
Task-related multifractal modulation, rather than long-memory reduction, distinguishes functional networks from artifacts in this analysis.
Takeaways & Limitations
The relationship between scale-free brain topology and scale-free temporal dynamics remains outside the study’s scope.
Abstract
from arXiv · showhide
Scaling temporal dynamics in functional MRI (fMRI) signals have been evidenced for a decade as intrinsic characteristics of ongoing brain activity (Zarahn et al., 1997). Recently, scaling properties were shown to fluctuate across brain networks and to be modulated between rest and task (He, 2011): notably, Hurst exponent, quantifying long memory, decreases under task in activating and deactivating brain regions. In most cases, such results were obtained: First, from univariate (voxelwise or regionwise) analysis, hence focusing on specific cognitive systems such as Resting-State Networks (RSNs) and raising the issue of the specificity of this scale-free dynamics modulation in RSNs. Second, using analysis tools designed to measure a single scaling exponent related to the second order statistics of the data, thus relying on models that either implicitly or explicitly assume Gaussianity and (asymptotic) self-similarity, while fMRI signals may significantly depart from those either of those two assumptions (Ciuciu et al., 2008; Wink et al., 2008). To address these issues, the present contribution elaborates on the analysis of the scaling properties of fMRI temporal dynamics by proposing two significant variations. First, scaling properties are technically investigated using the recently introduced Wavelet Leader-based Multifractal formalism (WLMF; Wendt et al., 2007). This measures a collection of scaling exponents, thus enables a richer and more versatile description of scale invariance (beyond correlation and Gaussianity), referred to as multifractality. Also, it benefits from improved estimation performance compared to tools previously used in the literature. Second, scaling properties are investigated in both RSN and non-RSN structures (e.g., artifacts), at a broader spatial scale than the voxel one, using a multivariate approach, namely the Multi-Subject Dictionary Learning (MSDL) algorithm (Varoquaux et al., 2011) that produces a set of spatial components that appear more sparse than their Independent Component Analysis (ICA) counterpart. These tools are combined and applied to a fMRI dataset comprising 12 subjects with resting-state and activation runs (Sadaghiani et al., 2009). Results stemming from those analysis confirm the already reported task-related decrease of long memory in functional networks, but also show that it occurs in artifacts, thus making this feature not specific to functional networks. Further, results indicate that most fMRI signals appear multifractal at rest except in non-cortical regions. Task-related modulation of multifractality appears only significant in functional networks and thus can be considered as the key property disentangling functional networks from artifacts. These finding are discussed in the light of the recent literature reporting scaling dynamics of EEG microstate sequences at rest and addressing non-stationarity issues in temporally independent fMRI modes.
1. Introduction
The study addresses limitations of prior fMRI scale-invariance analyses by combining multivariate decomposition with Wavelet Leader-based Multifractal analysis across functional and artifactual components and rest-task states.
- Motivation: Prior analyses often used voxelwise or regionwise univariate methods, leaving the specificity of scale-free dynamics to functional networks unresolved.Systematic comparison with non-RSN components such as artifacts had not been undertaken.
- Approach: MSDL extracts spatial components and corresponding subject-specific time series spanning functional networks and artifacts.Its sparsity-promoting regularization produces less noisy spatial maps than earlier approaches.
- Approach: WLMF estimates a collection of scaling exponents, extending analysis beyond correlation and Gaussianity to multifractality.The formalism offers improved mathematical grounding and estimation performance relative to previously used tools.
- Approach: The combined tools assess spatial and rest-task modulation of scale-free and multifractal properties in functional and artifactual components.Analyses use resting-state and activation datasets with group-level statistical comparisons.
- Findings: Long memory decreases under task throughout the brain rather than specifically in functional networks, whereas task-related multifractality changes only significantly in functional networks.Most resting-state signals are multifractal except in non-cortical regions, making multifractality the key discriminator from artifacts.
2. Data acquisition and analysis
The study used an ethically approved 3T fMRI experiment with 12 right-handed normal-hearing participants, comprising resting-state and auditory detection runs.
- Participants and acquisition: Twelve right-handed normal-hearing subjects aged 19–30 participated after informed consent and institutional ethics approval.Two participants were female.
- Participants and acquisition: Resting-state data comprised 820 volumes acquired with eyes closed before experimental runs of 820 volumes each.The resting-state dataset had been published previously.
- Task: Experimental runs used an auditory detection task with sparse supra-threshold stimuli and a motor response.Participants responded by right-hand key press to target sounds despite scanner noise.
- Task: The target stimulus was a 500 ms noise burst with frequency modulation at 2 Hz and unpredictable 20–40 s inter-stimulus intervals.Each specific interval was used only once.
- Preprocessing: Preprocessing used SPM5 for realignment, coregistration, normalization to MNI space, and 5 mm isotropic Gaussian smoothing.Subsequent analyses used in-house software, scikit-learn for MSDL, and WLBMF for multifractal analysis.
3. Multivariate decomposition of resting state networks
MSDL separates mixed voxel-level processes into group and subject-specific spatial maps with associated time series, enabling joint analysis of functional, artifactual, and undefined components across rest and task.
- MSDL framework: The fMRI voxel signal mixes neural BOLD activity with cardiac, respiratory, movement, and scanner contributions.MSDL estimates spatial maps and time series jointly to separate these processes.
- MSDL framework: MSDL extracts a group-level atlas and subject-specific maps while accounting for individual specificities.The procedure is a multi-subject adaptation of dictionary learning.
- Model: The model represents each subject map as the group map plus subject-specific variability, Vs = V + Fs.Maximum a posteriori estimation combines squared-error likelihood with sparsity and smoothness priors.
- Model: The objective uses an ℓ1 sparsity penalty and a Laplacian gradient penalty, with λ selected by cross-validation and µ by intra- versus inter-subject variance.These priors regulate map sparsity and smoothness.
- Resting-state maps: The decomposition yielded K = 42 maps from 12 subjects, 820 time points, 3 mm voxels, and approximately 50,000 brain voxels.Maps were classified as 25 functional, 13 artifactual, and 4 undefined.
- Rest-task comparison: The same spatial decomposition was projected onto task data, producing map-level rest and task time series for subsequent univariate scale-free analysis.Task time series were obtained by least-squares projection onto the inferred maps.
4.1. Intuition
Scale-free brain dynamics distribute activity across frequencies according to a power law, so scaling exponents summarize temporal organization without privileging a specific frequency.
- Scale-free intuition: Resting-state fMRI intrinsic activity is characterized by scale-free behavior rather than concentration in predefined frequency bands.All frequencies jointly contribute to the dynamics.
- Scale-free intuition: A scale-free signal has a power spectral density that decreases as a power law across a broad frequency range.This implies that no frequency in that range plays a specific role.
- Scale-free intuition: The scaling exponent β is used as a key descriptor of scale-free activity.Scale-free, scale-invariant, and scaling are treated as equivalent terminology.
4.2. Scale-free models
Scale-free fMRI dynamics can be modeled beyond Gaussian 1/f assumptions using self-similar processes with stationary increments and multifractal processes. Strictly concave scaling exponents indicate multifractality and encode dependence beyond second-order statistics.
- Model limitations: Second-order scale-free models do not capture marginal distributions or higher-order dependence, including departures from joint Gaussianity.Their limitation motivates models that describe more than covariance or power-spectrum behavior.
- Self-similar processes: H-sssi processes provide a non-Gaussian model for fMRI signals as increments of a self-similar process with stationary increments.This model encompasses the simpler 1/f-spectrum formulation while avoiding a Gaussianity requirement.
- Scaling exponents: The self-similarity exponent H describes amplitude scaling under time dilation, whereas multifractal ζ(q) characterizes scale dependence across statistical orders.The latter therefore reflects both temporal dependence and distributional structure.
- Self-similar processes: Under joint Gaussianity, the H-sssi model reduces to fractional Brownian motion and its increment process, fractional Gaussian noise.The paper notes that choosing between these formulations can be practically challenging for brain activity.
- Multifractal processes: Strictly concave ζ(q) rules out H-sssi models and motivates multifractal processes with richer, non-Gaussian dependence structures.In the cascade model, ζ(q) is tunable and departures from linearity quantify departures from strict joint Gaussianity.
4.3. Scale-free analysis
The analysis compares Welch and wavelet spectral estimation before extending scale-free analysis across statistical orders with wavelet leaders. This produces multifractal descriptions that better distinguish linear from strictly concave scaling behavior.
- From spectrum to wavelet analysis: Welch estimation splits a signal into blocks and averages independently computed squared Fourier transforms to estimate spectral scaling.For scale-free data, this targets the power-law spectrum parameter β.
- From spectrum to wavelet analysis: Wavelet spectra estimate scale-free exponents from wavelet coefficients and are more robust and accurate than Welch-based estimates under several forms of non-stationarity.They can help separate environmental non-stationarity from genuine long memory in fMRI signals.
- From 2nd to other statistical orders: Wavelet leaders: Wavelet leaders are used instead of simply extending wavelet-coefficient structure functions across q because the latter poorly determine whether ζ(q) is linear or strictly concave.This is the key methodological transition from second-order scaling to multifractal estimation.
- From 2nd to other statistical orders: Wavelet leaders: Wavelet-Leader multifractal analysis estimates ζ(q,γ) across statistical orders, with concavity distinguishing H-sssi-like linear scaling from multifractal behavior.The formalism relates ζ(q,γ) to ζ(q) through ζ(q,γ) = ζ(q) + γq and uses polynomial coefficients to summarize curvature.
- From 2nd to other statistical orders: Wavelet leaders: Multifractal analysis describes temporal variation in local regularity through Hölder exponents and their global spectrum D(h).Figure 2 presents corresponding multifractal spectra alongside Welch and wavelet spectra for rest and task signals.
5.1. Single subject analysis
Single-subject analyses show task-related reductions in self-similarity and multifractality-related spectral changes, while multifractality is generally present in fMRI signals across brain states.
- Analysis setup: Scale-free properties are assessed over four octaves corresponding to 0.008–0.063 Hz using Daubechies wavelets with Nψ = 3.This range is consistent with the hemodynamic boundary commonly associated with fMRI scaling.
- Self-similarity: β decreases during task in the example functional map, yielding a lower Hurst exponent H = (β − 1)/2 and reduced self-similarity.For f18, the task-related estimate is bH_R ≃ 0.5.
- Self-similarity: The task-related reduction in self-similarity appears as a leftward shift of the multifractal-spectrum maximum bc1.The paper relates bc1 to the self-similarity parameter and the Hurst exponent.
- Multifractal spectrum: Strictly negative c2 values indicate multifractality in fMRI signals, with c2 quantifying the width of the multifractal spectrum.The spectrum width can decrease from rest to task, although opposite fluctuations are also observed among functional maps.
- Multifractal spectrum: The analysis uses c1 and c2 as the sufficient descriptors of self-similarity and multifractality in subsequent analyses.The superscript γ is omitted for conciseness, with γ set to 2.
5.2. Group-level analysis
Group-level analyses show that task reduces self-similarity across functional, artifactual, and undefined maps, whereas task-related multifractality changes are concentrated in functional networks.
- Group-level scale-free properties: Group-level self-similarity decreases under task across functional, artifactual, and undefined maps, with mean changes of −0.125, −0.11, and −0.13, respectively.The decrease is therefore not specific to functional components.
- Group-level scale-free properties: Multifractality remains evident across map types, but task-related changes are larger in functional maps than in artifactual or undefined maps.Changes in irrelevant maps remain below 0.03 for A/U maps, compared with below 0.08 for F maps.
- One-sample statistical tests: At rest, most functional components are self-similar, but task reduces the number of significant self-similar and multifractal components across map classes.For functional maps, 22/25 components are self-similar at rest, versus four during task; multifractal components also decline under task.
- One-sample statistical tests: At the macroscopic level, all averaged map groups are multifractal at rest, whereas only functional maps remain multifractal during task.All averaged groups retain significant self-similarity in both states, though task reduces its statistical significance.
- Two-sample statistical tests: ANOVA results indicate that self-similarity cannot distinguish functional from artifactual or undefined maps, while multifractality provides the distinguishing interaction for functional maps.The functional-map interaction is significant for multifractal parameters, whereas self-similarity changes occur more broadly.
- Two-sample statistical tests: Functional networks show significant task-related multifractality interactions, whereas artifacts do not show significant multifractality differences across observation levels.Non-cortical regions account for a major increase in multifractality, while artifacts show no major change.
6. Discussion
The discussion interprets multifractal fMRI dynamics as more discriminative of functional networks than self-similarity, while emphasizing scale-range, sample-size, and generative-mechanism limitations.
- Results interpretation: Multifractality distinguishes functional components from artifacts because rest–task modulation was significant only in functional maps and networks.This contrasts with self-similarity, whose task-related decrease also occurred in artifacts and undefined maps.
- Results interpretation: Task-related multifractality changes were directionally heterogeneous, decreasing in cortical f18 but increasing in non-cortical f4 and f7.The study found no systematic increase-or-decrease pattern across regions.
- Results interpretation: The small sample of 12 subjects and difficulty estimating parameters on short time series limited power to establish general multifractality trends.The authors call for larger groups and more scans to obtain reliable estimates.
- Monofractal scale-free EEG microstate sequences vs multifractal dynamics for RSN: Comparisons with monofractal EEG microstate dynamics are constrained because fMRI and EEG studies analyzed different frequency ranges.The fMRI range was [.008, .063] Hz, whereas the EEG microstate result used [.063, 3.9] Hz.
- Monofractal scale-free EEG microstate sequences vs multifractal dynamics for RSN: A linear time-invariant filter is unlikely to create fMRI multifractality from monofractal EEG, but nonlinear filtering can produce multifractality.Theoretical understanding of filtering and multifractality remains limited.
- Monofractal scale-free EEG microstate sequences vs multifractal dynamics for RSN: Alternative EEG microstate definitions may alter the comparison because correlated sequences can lose singularities present in temporally independent microstates.Joint EEG/fMRI analyses identified thirteen temporally independent microstates rather than four grouped patterns.
- Results interpretation: Wavelet analysis supports stationary step-process modeling and is robust to several forms of non-stationarity, including smooth trends.The estimated Hurst coefficient remained below 1 in these fMRI time series.
7. Conclusion
The study identified multifractal scale-free dynamics in both functional networks and artifacts, then used rest–task modulation to distinguish the two. Only functional components showed significant multifractal modulation between resting and task-related activity.
- 7. Conclusion: Multifractal scale-free dynamics appeared in functional networks and artifacts across four octaves spanning 15–125 seconds.The analysis then separated functional from artifactual components using their modulation patterns.
- 7. Conclusion: Only functional components showed significant modulation of multifractal attributes between rest and task-related activity.This provided a robust and significant basis for disentangling functional components from artifacts.