A Generalized Framework of Antisymmetric Polyspectral Indices for Identifying High-Order Neural Interactions

Technical Summary: A Generalized Framework of Antisymmetric Polyspectral Indices for Identifying High-Order Neural Interactions

Problem Statement
Cross-frequency coupling (CFC) is a fundamental mechanism for integrating information across temporal scales in the brain. However, accurately identifying genuine high-order neural interactions is hindered by two primary challenges: complex multi-frequency nonlinearities and spurious, zero-lag artifacts caused by volume conduction and field spread in EEG/MEG recordings. Conventional connectivity metrics (e.g., PLV, ImCoh, wPLI) typically assume same-frequency interactions and fail to characterize dependencies where a frequency of interest $f_N$ arises from the combination of $N-1$ components ($f_N = \sum_{i=1}^{N-1} f_i$). Furthermore, existing cross-frequency measures often lack intrinsic robustness to linear instantaneous mixing, do not naturally generalize to interactions among more than two regions, and separate phase and amplitude in ways that may reduce signal-to-noise ratio. These limitations are particularly critical for emerging multi-locus transcranial magnetic stimulation (mTMS) protocols, which induce emergent drives at composite frequencies in overlapping cortical regions; monitoring only same-frequency connectivity in such scenarios risks missing the actual dynamical features mediating stimulation effects.

Methodology
The authors introduce a general family of antisymmetric cross-polyspectral indices designed to quantify harmonic dependencies (specifically $1:n$ interactions) while remaining robust to instantaneous linear mixing.

1. Theoretical Formulation:

   • The framework begins with the definition of a trispectrum for two time series $x(t)$ and $y(t)$, focusing on cubic dependencies ($1:3$ coupling where $f_1=f_2=f_3=f$ and $f_4=3f$).
   • Under a linear mixing model ($x = \sum a_i s_i$, $y = \sum b_i s_i$), the standard trispectrum contains contributions from both genuine sources and linear mixing artifacts.
   • To isolate genuine coupling, the authors define an antisymmetric cross-trispectrum by subtracting the permuted version of the estimator: $T[x|xx|y] = T_{xxxy} - T_{yxxx}$.
   • Theoretical derivation demonstrates that under the null hypothesis of independent sources linearly mixed at the sensors, the symmetric mixing contributions cancel out exactly, leaving only terms associated with genuine non-linear coupling from the same source.
   • This construction is normalized using fourth-order amplitude norms to yield the Antisymmetric Cross-Tricoherence (ACT), bounded within $[0, 1]$.
   • The framework is generalized to an arbitrary order $m$ (Antisymmetric Cross-Polyspectrum), defining the Antisymmetric Cross-Polycoherency (ACP).
   • The authors show that for $m=2$, this formulation reduces to the imaginary part of the coherency (ImCoh), and for $m=3$, it corresponds to the antisymmetric cross-bicoherence.

2. Validation Experiments:

   • Synthetic Data: A generative model was used to simulate time series containing a mix of genuine cubic coupling and artifactual instantaneous linear mixing, controlled by a parameter $\gamma$. The performance of ACT was compared against non-antisymmetrized trispectral measures (CT1 and CT2).
   • Real EEG Data: The indices were applied to a single-subject resting-state EEG dataset (61 channels). Statistical significance was assessed using a parametric test based on surrogate data (randomly permuting temporal segments) to generate p-values.

Key Results

• Synthetic Data:
  • In regimes dominated by genuine cubic coupling ($\gamma \approx 0$), both ACT and standard trispectral measures showed elevated values, confirming sensitivity to true non-linear dependencies.
  • In regimes dominated by instantaneous linear mixing ($\gamma \approx 1$), standard measures (CT1, CT2) were systematically inflated due to zero-lag correlations. In contrast, ACT values remained close to baseline, demonstrating robust immunity to spurious mixing artifacts.
  • In intermediate regimes, standard measures showed complex, confounded behaviors (e.g., U-shaped dependence on $\gamma$), whereas ACT maintained specificity to genuine coupling.
• Real EEG Data:
  • Standard trispectral maps (CT1, CT2) displayed pronounced, spatially confined high-value patches centered on seed sensors, a topography consistent with volume conduction rather than long-range interaction.
  • ACT maps did not exhibit these short-range "blooms." Instead, they revealed modest, spatially distributed increases toward frontal channels (for occipital seeds) and clear long-range inter-hemispheric patterns (for parietal seeds).
  • Statistical testing confirmed that while some local peaks in standard measures survived thresholding (likely artifactual), the long-range patterns identified by ACT were statistically robust after surrogate thresholding.
  • A direct comparison with ImCoh (linear connectivity) showed extremely low correlation coefficients ($\rho \approx 0.04-0.06$), indicating that ACT captures largely independent features of neural connectivity not explained by linear phase synchrony.

Significance and Claims
The paper claims to provide the first robust framework for identifying genuine high-order cross-frequency interactions among multiple time series that is intrinsically resistant to linear mixing artifacts. The authors posit that these indices are particularly timely for the analysis of multi-locus TMS (mTMS), where stimulation protocols induce composite spectral drives ($f_{overlap} = \sum f_i$) that cannot be monitored via same-frequency connectivity alone.

The study serves as a proof of concept, demonstrating that:

1. Antisymmetrization effectively cancels volume conduction artifacts in higher-order spectral analysis.
2. Genuine higher-order dependencies can be detected at the sensor level in single-subject EEG data.
3. These metrics offer a principled basis for guiding personalized, multi-site, multi-frequency stimulation strategies.

The authors acknowledge that while the method is robust, future work is required to validate these indices across larger subject cohorts to determine if observed coupling patterns represent general principles of cortical organization or individual characteristics. They also note the necessity of rigorous statistical characterization to optimize data length and signal-to-noise requirements for potential real-time applications in adaptive neuromodulation.