Background: Large-scale synchronous neural activity produces electrical fields that can be measured by electrodes outside the head, and volume conduction ensures that neural sources can be measured by many electrodes. However, most data analyses in M/EEG research are univariate, meaning each electrode is considered as a separate measurement. Several multivariate linear spatial filtering techniques have been introduced to the cognitive electrophysiology literature, but these techniques are not commonly used; comparisons across filters would be beneficial to the field.
New method: The purpose of this paper is to evaluate and compare the performance of several linear spatial filtering techniques, with a focus on those that use generalized eigendecomposition to facilitate dimensionality reduction and signal-to-noise ratio maximization.
Results: Simulated and empirical data were used to assess the accuracy, signal-to-noise ratio, and interpretability of the spatial filter results. When the simulated signal is powerful, different spatial filters provide convergent results. However, more subtle signals require carefully selected analysis parameters to obtain optimal results.
Comparison with existing methods: Linear spatial filters can be powerful data analysis tools in cognitive electrophysiology, and should be applied more often; on the other hand, spatial filters can latch onto artifacts or produce uninterpretable results.
Conclusions: Hypothesis-driven analyses, careful data inspection, and appropriate parameter selection are necessary to obtain high-quality results when using spatial filters.
Keywords: Data analysis; Dimensionality reduction; EEG; Eigendecomposition; Neural oscillations; Spatial filtering.
Copyright © 2016 Elsevier B.V. All rights reserved.