Efficient coherence inference on complex time-frequency coefficients using a general linear model.

Area of Science:

• Neuroscience
• Signal Processing
• Statistical Modeling

Background:

• Statistical significance testing is crucial for identifying true neural signal coupling.
• Current surrogate-based methods are computationally intensive and yield unstable p-values, hindering scalability for large datasets like electroencephalography (EEG) and intracranial EEG (iEEG).

Purpose of the Study:

• To introduce a novel, computationally efficient parametric framework for testing neural coherence significance.
• To provide a robust alternative to computationally expensive surrogate resampling techniques.

Main Methods:

• A General Linear Model (GLM) framework was applied to complex-valued time-frequency coefficients.
• A likelihood ratio test was employed to derive continuous coherence significance estimates, bypassing the need for surrogate resampling.

Main Results:

• The GLM demonstrated sensitivity comparable to or exceeding surrogate testing, detecting lower coherence levels (C≈0.16 vs C≈0.31) with improved signal-to-noise ratio.
• The GLM achieved an approximately 190x speed increase compared to surrogate-based methods.
• Continuous and stable p-values were generated, eliminating the permutation floor observed in surrogate methods.

Conclusions:

• The GLM-based inference provides a statistically sound and computationally scalable approach for neural coherence testing.
• This method enables efficient analysis of large-scale EEG/iEEG data across multiple channels, frequencies, and participants.