PaperlayerView exact revision on arXiv

Source-linked AI summary

Untangling cross-frequency coupling in neuroscience

Juhan Aru, Jaan Aru, Viola Priesemann, Michael Wibral, Luiz Lana, Gordon Pipa, Wolf Singer, Raul Vicente

arXiv:1405.7965v2q-bio.NC

TL;DR

Standard CFC analyses can mistake measurement and algorithmic artifacts for biological coordination, complicating mechanistic interpretation. The paper organizes CFC approaches by biophysical interpretability and proposes controls to identify confounds and improve interpretation, while acknowledging unresolved limitations.

  • Problem

    CFC analyses face confounds that can generate apparent coupling without biological interactions and have been overlooked in many publications.

  • Method

    The paper organizes statistical and modeling approaches to CFC according to their biophysical interpretability and inference strategy.

  • Results

    The paper identifies measurement, algorithmic, nonlinear, and input-driven sources of spurious CFC and proposes controls for CFC analysis.

  • Takeaways & Limitations

    Improved mechanistic interpretation of CFC requires evaluating confounds and applying controls alongside biophysically interpretable modeling and statistical approaches.

  • Takeaways & Limitations

    The paper notes that apparent smoothness in amplitude dynamics can result from filtering rather than meaningful phase.

Abstract

from arXiv · show

Cross-frequency coupling (CFC) has been proposed to coordinate neural dynamics across spatial and temporal scales. Despite its potential relevance for understanding healthy and pathological brain function, the standard CFC analysis and physiological interpretation come with fundamental problems. For example, apparent CFC can appear because of spectral correlations due to common non-stationarities that may arise in the total absence of interactions between neural frequency components. To provide a road map towards an improved mechanistic understanding of CFC, we organize the available and potential novel statistical/modeling approaches according to their biophysical interpretability. While we do not provide solutions for all the problems described, we provide a list of practical recommendations to avoid common errors and to enhance the interpretability of CFC analysis.

Cross-frequency coupling: How much is that in real money?

Cross-frequency coupling may coordinate neural activity across spatial and temporal scales, but standard analyses cannot reliably distinguish genuine interactions from generic signal properties. The section therefore motivates biophysically interpretable approaches and practical recommendations for improving CFC analysis.

  • Motivation: CFC could coordinate neural activity across spatial and temporal scales by linking slower large-population dynamics with faster local activity.Phase-amplitude CFC has been associated with neural information processing, cognition, and neurological or mental disorders.
  • Methodological caveats: Standard CFC signatures do not necessarily reflect interactions between physiological processes at different frequencies.Abrupt signal changes and generic nonlinear responses can produce spurious CFC, including when faster components are short-lived relative to slower components.
  • Methodological caveats: A Van der Pol oscillator can show strong CFC even though its spectral correlations arise from nonlinear characteristics of a single oscillator.Interpreting this pattern as modulation or causal interaction between frequency components is misleading.
  • Classical CFC analysis: The classical workflow extracts band-limited phase and amplitude, quantifies their dependence, and evaluates coupling statistically against suitable surrogate data.Phase-amplitude coupling can be quantified by deviations of mean amplitude across phase from a uniform distribution, followed by a p-value assessment.
  • Methodological caveats: Current phase-to-amplitude CFC measures are not specific enough to conclude automatically that low-frequency phase modulates high-frequency power.The same interpretive limitation applies to amplitude-to-amplitude and phase-to-phase coupling measures.
  • Paper aims: The paper reviews methodological caveats, organizes CFC approaches by biophysical interpretability and statistical inference, and provides practical recommendations while acknowledging that not all problems are solved.The authors aim to alert the community and encourage novel solutions.

Caveats and confounds of the CFC analysis

Classical phase-amplitude CFC can classify conditions but lacks clear physiological interpretation unless oscillatory components and analysis parameters satisfy basic requirements. Non-stationarities and filtering choices can generate misleading or false-negative CFC results.

  • Interpretive limits: Classical CFC results may classify conditions, but they are markers without concrete physiological interpretation.Physiological interpretation requires identifying the mechanisms responsible for neural coupling.
  • Interpretive limits: A clear power-spectrum peak in the low-frequency component is a prerequisite for meaningfully interpreting any CFC pattern.The literature review found that this requirement has not always been met, leading to over-interpretation of phase and amplitude.
  • Filtering confounds: If the high-frequency amplitude bandwidth excludes side peaks produced by the low-frequency component, CFC cannot be detected even when present.Thus, commonly selected parameter values can bias CFC measures toward false-negative results.
  • Non-stationarity confounds: Unspecific non-stationarities, including common drive, can create spectral correlations that resemble interactions between frequency bands without neural coupling.For example, non-stationary input can simultaneously affect low-frequency phase and high-frequency activity, producing phase-amplitude CFC without interaction between rhythms.
  • Non-stationarity confounds: The current phase-amplitude CFC measure is constitutive for non-stationary responses of driven systems and is therefore not a specific marker of biophysical coupling.Surrogates that fully destroy both specific and unspecific non-stationarities can make significantly larger original-data CFC difficult to interpret.

Organization of modeling/statistical approaches to CFC

The paper organizes CFC approaches by biophysical interpretability and statistical inference to constrain how physiological findings are interpreted. Marker-level analyses may classify conditions, whereas mechanistic interpretation requires interventions, generative models, or formal model comparison.

  • Organization of modeling/statistical approaches to CFC: CFC methods are organized by biophysical interpretability and statistical inference to reduce over-interpretation of analysis results.The framework is presented in Figure 4 as a way to understand the physiological implications of different approaches.
  • Organization of modeling/statistical approaches to CFC: Correlation-based or classical CFC measures may suffice as markers for classifying conditions, but stronger mechanistic claims require additional evidence.The paper identifies generative models, external information from direct perturbations, or formal comparison of biophysical models as routes toward physiological interpretation.
  • Organization of modeling/statistical approaches to CFC: The measurement process can generate apparent CFC without biological CFC, so physiological coupling cannot be inferred from a non-zero CFC measure alone.Alternative explanations include biological processes unrelated to direct coupling and methodological pitfalls.
  • Organization of modeling/statistical approaches to CFC: When interventions are unavailable, formal comparison of computational models with and without biophysical CFC mechanisms can test which better explains observed data.The paper gives Dynamical Causal Models and Bayesian model comparison as examples.
  • Organization of modeling/statistical approaches to CFC: If neither intervention nor formal model comparison is feasible, interpretation should be limited to treating observed CFC patterns as markers.The hierarchy of approaches is reflected in the methods’ arrangement in Figure 4.

Practical recommendations

The recommendations frame phase-amplitude CFC analysis as a checklist for reducing technical pitfalls and over-interpretation. They emphasize verifying oscillations, choosing appropriate bandwidths, validating phase estimates, testing nonlinear and input-related confounds, characterizing temporal structure, and controlling surrogate and power-related effects.

  • 1 Presence of oscillations: Verify clear oscillatory peaks in the time-resolved power spectrum before defining instantaneous phase.The phase-defining frequency component should include one of the spectral peaks.
  • 2 Selection of bandwidths: Choose bandwidths that isolate the oscillatory component, capture higher-frequency sidebands, and preserve a meaningful lower-frequency phase.Adaptive rather than fixed bandwidths may be necessary when scanning the modulating frequency.
  • 3 Interpretation of instantaneous phase: Check that instantaneous phase grows monotonically, and justify any phase slips, reverses, or negative instantaneous frequencies.Phase interpretation is meaningful only when these irregularities are assessed rather than ignored.
  • 5 Testing for non-linearities: Test whether harmonics and nonlinear signal transduction contribute to CFC, using bicoherence and partialization of phase-phase and amplitude-amplitude coupling.Nonlinear responses to input or during signal transduction can generate phase-amplitude CFC.
  • 6 Testing for input-related non-stationarities; 7 Temporal structure: Use input timing and temporal-structure analyses to distinguish possible CFC origins and characterize whether coupling is sustained or transient.Relative locking among phase, amplitude, and input can inform correlation origins, while the modulation index alone provides only an average measure.
  • 8 Surrogates; 9 Specificity of effects: Minimize analytical confounds by using minimally disruptive surrogates and controlling CFC differences for power at the presumed interaction bands.For continuous recordings, random point block-swapping is preferred; trial subsets should be stratified when their power distributions differ across conditions.

Conclusions

CFC may help coordinate neural dynamics and illuminate learning, memory, brain function, and neurological or psychiatric disorders, but confounds can produce over-interpretations. The paper recommends stricter controls, canonical procedures, biophysically informed modeling, and reserving “coupling” for unequivocally demonstrated interactions.

  • Conclusions: CFC may coordinate neural dynamics and help investigate learning, memory, and neurological or psychiatric disorders.The authors describe CFC analysis as potentially useful for unraveling brain function and some pathologies.
  • Conclusions: The paper reviews confounds in phase-amplitude CFC analysis that have often been overlooked and may have contributed to over-interpretations.The authors characterize this as a serious issue because CFC could reveal fundamental features of neural computations.
  • Conclusions: The authors recommend stricter standards, canonical procedures, and routine checks for CFC analysis.They present the proposed list of controls as probably incomplete.
  • Conclusions: They organize statistical and modeling approaches to clarify their advantages, pitfalls, and areas requiring further methodological advances.The organization aims to improve identification of where current approaches are useful and limited.
  • Conclusions: The authors suggest using “cross frequency correlation” rather than “coupling” unless coupling is unequivocally demonstrated.This terminology recommendation is intended to distinguish observed correlations from demonstrated coupling.

Supplementary materials for

The supplementary materials detail the signal-processing procedures, experimental recordings, and literature-review criteria used to assess cross-frequency coupling analyses. They cover analytical-signal extraction, turtle and human recordings, and evaluation of methodological safeguards against common CFC problems.

  • Signal processing: Signals were band-filtered with a two-way least-squares FIR filter, then Hilbert-transformed to estimate instantaneous amplitudes and phases.Analytical-signal polar coordinates defined the instantaneous amplitudes and phases.
  • Signal processing: Modulation indices and mean-amplitude histograms used 20 equally sized phase bins, following Tort et al.’s original modulation-index formulation.Time-dependent power locked to slow-component phase troughs was also extracted for several supplementary figures.
  • Turtle recordings: Turtle retina and tectum recordings examined stimulus effects on CFC using low-pass-filtered, downsampled data divided into 60 trials of 50 seconds.Visual stimulation used red LED flashes with random durations and inter-pulse intervals.
  • Human recordings: Human electrocorticograms came from two epilepsy subjects with visual-cortex strip electrodes sampled at 1000 Hz during noisy-image perception tasks.Analyses covered one-second pre- and post-stimulus windows that excluded motor responses.
  • Literature review: The literature review evaluated 22 articles against criteria covering phase interpretability, bandwidth selection, non-stationarity, surrogate construction, and spectral changes across conditions.The review targeted recent publications from major neuroscience journals and assessed whether analyses addressed these methodological issues.

Results of the literature review

The review shows that nonzero CFC can arise without interactions between distinct frequency components. A single nonlinear oscillator and a smoothed spike train both produce apparent CFC through waveform shape or event-driven structure.

  • Nonlinear oscillator: A single nonlinear oscillator can exhibit CFC solely because of its nonlinear properties.In the Van der Pol example, high-frequency power depends on the phase of the low-frequency fundamental.
  • Nonlinear oscillator: The resulting spectral correlations need not reflect two causally interacting oscillatory subsystems.They can instead arise from the shape of the oscillatory orbit and state-dependent nonlinear damping.
  • Point-process example: A prime-number spike train convolved with a 5 ms alpha-function produces high CFC estimates.The spiking events simultaneously anchor slow-band phase and the kernel’s high-frequency components, automatically generating CFC.
  • Interpretive caution: Thus, nonzero CFC estimates can result from event timing and signal construction rather than cross-frequency interactions.The review also reports CFC effects for intervals between consecutive zeros of the Riemann Zeta function, further illustrating that unusual time series can yield such effects.

Effect of non-stationary input on CFC · Phase-amplitude coupling for event-related potentials (ERPAC)

Event-related changes in phase and amplitude can produce significant CFC without cross-frequency interaction, because high-frequency activity may lock more precisely to stimulus timing than to low-frequency phase. ERPAC targets event-related non-stationarity but still detects purely input-driven CFC when responses have even modest timing jitter.

  • Effect of non-stationary input on CFC: Stimulus-related low-frequency phase locking and increased high-frequency power can produce significant standard CFC without cross-frequency interaction.In turtle recordings, high-frequency bands included 80-200 Hz, and the combination was sufficient for significant CFC measures.
  • Effect of non-stationary input on CFC: In turtle recordings, jitter between high-frequency activity and stimulus onset was always smaller than jitter between high-frequency activity and any low-frequency phase.This indicates that high-frequency activity was more affected by the stimulus than by low-frequency phase.
  • Effect of non-stationary input on CFC: Human intracranial recordings likewise showed increased power, predominantly low-frequency phase reorganization, and a non-uniform post-stimulation amplitude-phase histogram.Pairing amplitude and phase across different trials rendered the histogram uniform, supporting a stimulus-related source for the apparent coupling.
  • Effect of non-stationary input on CFC: For human ECoG, high-frequency power-to-stimulus jitter was around 80 ms, and across a frequency range it was smaller than power-to-phase jitter.The greater locking to stimulus onset makes common drive from time-varying input a plausible explanation for frequency-component correlations.
  • Phase-amplitude coupling for event-related potentials (ERPAC): ERPAC was designed to address spurious CFC from event-related non-stationarities by sampling Hilbert amplitudes and phases at fixed stimulus-referenced times across repeated trials.The approach relies on repeatability to make locally non-stationary processes quasi-stationary.
  • Phase-amplitude coupling for event-related potentials (ERPAC): ERPAC overcomes input-related non-stationarity only for very precisely repeated responses, a condition generally limited to subcortical and sensory input stages.Slight jitter in input arrival times means the known non-stationarity problems leading to spurious CFC remain elsewhere despite a clear ERP.
  • Phase-amplitude coupling for event-related potentials (ERPAC): In the same example, low-frequency amplitude explained some high-frequency amplitude variance, suggesting that multivariate models could better delineate dependencies between frequency components.This extends beyond the bivariate generalized linear model used in the referenced ERPAC method.

Atmospheric noise shows CFC after squaring the signal · Small static non-linearity in ECoG data generates CFC · Mathematical example of a non-linearity

Static non-linearities can generate strong, structured CFC even when the underlying signal contains no evident phase-amplitude coupling or significant CFC. A harmonic-wave example shows mathematically how a high-frequency amplitude can depend on a low-frequency phase.

  • Atmospheric noise shows CFC after squaring the signal: Squaring atmospheric noise produced clear phase-amplitude modulation, although the original random noise showed no evident coupling structure.The noise comprised 10000 samples recorded at a nominal sampling rate of 1000 Hz.
  • Atmospheric noise shows CFC after squaring the signal: These results show that non-linearities can confound CFC measures by creating apparent coupling in random data.The relevant non-linearities may arise from neuronal processing, tissue properties, or deviations at any stage of signal transduction.
  • Small static non-linearity in ECoG data generates CFC: Adding a statistically negligible squared-signal contribution to an ECoG signal without significant CFC yielded strong CFC.The original and distorted signals had a Pearson correlation coefficient of approximately 0.99.
  • Small static non-linearity in ECoG data generates CFC: Adding 10% of the signal’s square produced a much stronger modulation index despite visually similar ECoG signals.This demonstrates that even small non-linearities can create spurious CFC patterns.
  • Small static non-linearity in ECoG data generates CFC: Small non-linearities create harmonics that spread across the spectrum and generate long-distance spectral correlations.They can also produce CFC patterns in which high-frequency amplitude is related to a specific phase of a low-frequency component.
  • Mathematical example of a non-linearity: For a harmonic wave x(t) with frequency f, a second-order non-linear approximation applies when a << 1, meaning the non-linear contribution is small.The example analyzes the phase of x and amplitude of the non-linear signal y using frequency-specific filtering.
  • Mathematical example of a non-linearity: The mathematical example yields perfect phase-amplitude coupling because y’s high-frequency amplitude is a function of x’s low-frequency phase.CFC analysis between x and y therefore produces perfect phase-amplitude coupling.

Supplementary Discussion · Conditions for a meaningful phase

A meaningful phase requires a signal with a natural concentration of power around a center frequency, because phase–amplitude interpretation is restricted outside narrow-band signals. Filtering broad-band or 1/f^−α activity can create smooth but uninterpretable dynamics and misleading CFC.

  • Conditions for a meaningful phase: Phase indexes position within a cycle and grows monotonically as the completed fraction of that cycle increases.For simple harmonic motion, amplitude is constant while phase tracks position along the cycle.
  • Conditions for a meaningful phase: For irregular signals, infinitely many amplitude–phase pairs can represent the same signal, motivating Gabor’s analytical-signal approach as a unique solution.The method decomposes Fourier components, applies 90-degree shifts through the Hilbert transform, and forms a complex analytical signal.
  • Conditions for a meaningful phase: Analytical phase and amplitude are clearly interpretable only for narrow-band signals that remain close to smooth periodic functions.Moderate noise and smooth frequency fluctuations can still satisfy the narrow-band condition.
  • Conditions for a meaningful phase: Broad-band signals do not preserve separable amplitude and phase, especially when the spectra of a(t) and cos(ϕ(t)) overlap.The analytical estimate then mixes the underlying amplitude and phase, and instantaneous frequency can even become negative.
  • Conditions for a meaningful phase: A two-dimensional phase–amplitude description can compromise interpretability when irregular signals require more dimensions to represent their degrees of freedom.Projecting all degrees of freedom into two variables is not a defect specific to one analytical method.
  • Conditions for a meaningful phase: Narrow-band filtering of 1/f^−α spectral regions can create smooth phase and amplitude dynamics without indicating an underlying meaningful process.The apparent smoothness is a filtering artifact, so spectral correlations in such regions require extreme caution.
  • Conditions for a meaningful phase: Only a natural concentration of power around a center frequency in a time–frequency decomposition enables meaningful phase interpretation and CFC analysis.This criterion distinguishes intrinsically meaningful phase structure from smoothness introduced by filtering.

The importance of the bandwidth

Bandwidth selection is central to whether extracted phase and amplitude meaningfully represent oscillatory components and whether CFC is observed. Poorly chosen or fixed bandwidths can reduce interpretability and produce false-positive, false-negative, or frequency-dependent results.

  • A meaningful phase interpretation requires spectral power concentrated around a frequency, so the selected center frequency and bandwidth should include a spectral peak.
  • Bandwidths that are too narrow produce smooth but unrepresentative phase, whereas overly broad filters can introduce phase slips or reversals through 1/f components.Both choices can reduce analysis sensitivity and interpretability.
  • A bandwidth smaller than 2f1 around f2 yields higher-component amplitude constantly equal to 1, whereas a bandwidth larger than 2f1 produces CFC.Thus bandwidth determines component isolation and can generate false-positive or false-negative CFC findings.
  • Because the critical bandwidth for detecting sidebands depends on f1, fixed bandwidth scans favor low modulating frequencies and bias against higher-frequency phase modulation.

Different model/statistical approaches to assess phase-amplitude CFC

The paper organizes phase-amplitude CFC methods by their modeling approach, contrasting projection-based analyses with explicit dynamical, statistical, non-parametric, and causal models. These approaches trade off interpretability, flexibility, data efficiency, and practical limitations.

  • I) Fourier/wavelet analyses: Fourier/wavelet methods project data onto oscillatory bases, while higher-order spectral quantities capture correlations between complex components as putative CFC.These approaches lack explicit temporal-evolution equations, and higher-order functionals are difficult to estimate, especially for multivariate extensions.
  • II) Classical time series techniques: Classical time-series techniques fit regression models such as autoregressive moving-average processes, offering data-efficient multivariate extensions but restricting the range of interactions they can represent.Cyclo-stationary models and generalized linear models can quantify particular phase-amplitude effects, including nonlinear functions of phase.
  • III) Non-linear systems analyses: Nonlinear systems analyses estimate coupling parameters in explicit dynamical models, enabling direct interpretation and multivariate extensions, but they usually lack a physiological foundation.Examples include Kuramoto phase-coupling models and Stuart-Landau formulations for phase-amplitude interactions.
  • IV) Non-parametric approaches: Non-parametric approaches estimate relationships without imposing a specific model structure, supporting exploratory analysis and nonlinear information-theoretic functionals but requiring substantial data.Transfer entropy quantifies increased predictability from another variable’s present and past states, while multivariate estimation faces the curse of dimensionality.
  • V) Causal statistical modeling: Causal statistical modeling compares evidence for generative hypotheses, often using Bayesian inference and measurement functions; dynamical causal models are being developed with biophysical neuronal descriptions.A practical method for inferring phase-amplitude cross-frequency interactions had not yet been established in the described framework.

The transitivity of correlation between phase and amplitude · Supplementary discussion on causality methods

Phase and amplitude are intrinsically coupled, making observed phase–amplitude relationships difficult to interpret as direct interactions. Although causality methods could add directionality, current approaches face limitations from nonlinear effects, data requirements, temporal resolution, and filtering or signal-to-noise differences.

  • The transitivity of correlation between phase and amplitude: Nonlinear oscillators intrinsically couple amplitude and phase, so generic perturbations can change both and selective modification of either is nearly impossible.Amplitude and phase may also differ in susceptibility to perturbations or inertia, complicating causal inference from their timing.
  • The transitivity of correlation between phase and amplitude: Correlation transitivity makes it difficult to determine whether phase–amplitude relationships are direct or mediated by phase–phase or amplitude–amplitude coupling.For example, low-frequency phase can influence high-frequency phase, while intrinsic high-frequency phase–amplitude coupling produces an observed cross-frequency relationship.
  • The transitivity of correlation between phase and amplitude: Partialing out indirect coupling pathways is advisable before assigning a functional role to a specific phase–amplitude coupling type.Phase–phase coupling is given as an example of an indirect pathway.
  • The transitivity of correlation between phase and amplitude: Analytical-signal estimation can itself couple phase and amplitude, because nominal changes in one quantity perturb the other.The issue applies to phase and amplitude defined using Hilbert transforms.
  • Supplementary discussion on causality methods: Observational causality methods could add directionality to CFC analysis and constrain possible explanations for spurious CFC, but current approaches face unresolved difficulties.These methods therefore are not yet ready for straightforward application to interactions between neurophysiological frequency components.
  • Supplementary discussion on causality methods: Linear Granger formalism is blind to cross-frequency effects, and nonlinear interactions among extracted phases and amplitudes may remain invisible to it.Transfer entropy captures all orders of nonlinear interactions but typically requires long stretches of data.
  • Supplementary discussion on causality methods: Harmonic-analysis uncertainty limits temporal localization, causing low-frequency onsets to be advanced relative to high-frequency onsets with non-causal filtering.Unequal signal-to-noise ratios can also impair interpretation even with causal filters because causality measures rely on temporal order.
  • Supplementary discussion on causality methods: Unequal signal-to-noise ratios across components hamper interpretation of most causality measures, including when causal filters are used.The difficulty arises because these measures depend on the temporal order of component dynamics.
Loading 1405.7965v2…