Abstract
Magnetoencephalography (MEG) data is traditionally analyzed using linear models like Independent Component Analysis (ICA), assuming that sensor signals are simple weighted sums of brain sources. This study evaluates the validity of this linearity assumption using Hierarchical Nonlinear Factor Analysis (HNFA) within a variational Bayesian framework. By employing an automated pruning mechanism that discards unnecessary model complexity, the researchers found that the model consistently favored linear mappings over nonlinear ones. The results provide strong empirical evidence that MEG signals are effectively linear mixtures. Furthermore, the study identifies a total signal space of 64 dimensions, within which approximately 20–25 stable components represent true biological and environmental sources.
Introduction
Magnetoencephalography (MEG) is a non-invasive neuroimaging technique that measures the magnetic fields produced by neuronal activity. For decades, the standard approach for processing MEG data has been Linear Blind Source Separation (BSS), which relies on the assumption that observed signals are linear combinations of underlying sources. While Maxwell’s equations provide a theoretical basis for this linearity, the inherent complexity of the brain suggests the possibility of nonlinear interactions. This research seeks to empirically validate the linearity assumption and determine the intrinsic dimensionality of the "true" signal space versus high-dimensional noise.
Methods
The researchers utilized Hierarchical Nonlinear Factor Analysis (HNFA), implemented via Variational Bayesian learning. HNFA is designed as a "superset" of linear models; it begins with a linear mapping and allows for the introduction of nonlinear hidden nodes.
A key feature of this approach is the use of a cost function based on Kullback-Leibler (KL) divergence, which facilitates "ensemble learning." This mechanism acts as an automated Occam’s Razor, pruning hidden nodes and weights that do not significantly contribute to explaining the data. This "self-simplifying" property allows the model to collapse into a purely linear form if nonlinearities are not supported by the data.
Experimental Setup
The study employed a benchmark dataset consisting of 122-channel whole-scalp MEG recordings from 61 locations. To facilitate source identification, the data included controlled artifacts:
- Biological signals: Eye blinks, horizontal eye movements, and the cardiac cycle.
- Environmental signals: Magnetic interference from a digital watch and muscle contractions (biting).
The data was bandpass-filtered (0.75 Hz to 45 Hz) and downsampled to 10,000 samples for analysis.
Key Results
The most significant finding was that the HNFA model, when given the capacity for nonlinearity, chose to remain linear. During the learning process, the algorithm pruned away nearly all nonlinear hidden nodes, suggesting that a linear model is the mathematically optimal representation for MEG data.
Quantitative analysis revealed two critical thresholds:
- Optimal Dimensionality: The cost function reached a clear minimum at 64 sources, representing the total dimensionality required to describe both signals and noise.
- Signal Consistency: By correlating results across multiple runs, the researchers identified a stable core of 20–25 generative signals. These consistent components were easily mapped to known physical phenomena, such as muscle artifacts, eye movements, and the digital watch interference. The remaining ~40 components were identified as Gaussian noise.
Discussion and Conclusion
The empirical evidence strongly supports the continued use of linear separation tools in neuroimaging. While a potential "Bayesian bias" might favor simpler models in low signal-to-noise environments, the pruning of nonlinearities in this study suggests they are not necessary for an accurate representation of MEG data. Future research may investigate whether subtle nonlinearities emerge in different experimental conditions, but for standard MEG analysis, the linear assumption remains robust.
Summary updated