Independent EEG sources are dipolar.
Delorme A, Palmer J, Onton J, Oostenveld R, Makeig S. · 2012
Read Independent EEG sources are dipolar. by Delorme A, Palmer J, Onton J, Oostenveld R, Makeig S. (2012) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc3280242.txt.
Introduction
Arnaud Delorme and colleagues tested whether the components extracted from scalp EEG recordings correspond to single brain sources, in this 2012 paper in PLoS ONE. Independent component analysis is widely used to unmix the signals smeared together by conduction through the skull, but whether the separated components are physically meaningful was debated. The team decomposed thirteen seventy-one-channel datasets with twenty-two different algorithms and scored the results.
Likelihood and mutual-information methods, led by AMICA, performed best; plain principal component analysis performed worst. The key measure was dipolarity: how well each component's scalp pattern matches the projection of one compact source in the cortex. Components that separated most cleanly also scored most dipolar, consistent with many of them being genuine signals from small cortical patches. The authors released their data and software to spur more comparisons.
Catalog record
| CREATOR | Delorme A, Palmer J, Onton J, Oostenveld R, Makeig S. |
|---|---|
| YEAR | 2012 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 82 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Independent EEG Sources Are Dipolar By Arnaud Delorme, Jason Palmer, Julie Onton, Robert Oostenveld, Scott Makeig, Lawrence M Ward Published in: PLoS ONE DOI: 10.1371/journal.pone.0030135 ABSTRACT Abstract Independent component analysis (ICA) and blind source separation (BSS) methods are increasingly used to separate individual brain and non-brain source signals mixed by volume conduction in electroencephalographic (EEG) and other electrophysiological recordings. We compared results of decomposing thirteen 71-channel human scalp EEG datasets by 22 ICA and BSS algorithms, assessing the pairwise mutual information (PMI) in scalp channel pairs, the remaining PMI in component pairs, the overall mutual information reduction (MIR) effected by each decomposition, and decomposition ‘dipolarity’ defined as the number of component scalp maps matching the projection of a single equivalent dipole with less than a given residual variance. The least well-performing algorithm was principal component analysis (PCA); best performing were AMICA and other likelihood/mutual information based ICA methods. Though these and other commonly-used decomposition methods returned many similar components, across 18 ICA/BSS algorithms mean dipolarity varied linearly with both MIR and with PMI remaining between the resulting component time courses, a result compatible with an interpretation of many maximally independent EEG components as being volume-conducted projections of partially-synchronous local cortical field activity within single compact cortical domains. To encourage further method comparisons, the data and software used to prepare the results have been made available (http://sccn.ucsd.edu/wiki/BSSComparison). INTRODUCTION Brain-generated EEG data are generally considered to index synchronous aspects of local field potentials surrounding radially-arrayed cortical pyramidal cells [1], [2]. There are strong biological reasons to believe that under favorable circumstances ICA should separate signals arising from local field activities in physically distinct, compact cortical source areas: First, short-range (<100 µm) lateral connections between cortical neurons are vastly more dense than longer-range connections [3], [4], while inhibitory and glial cell networks have no long-range processes [3]. Also, thalamocortical connections that also play a strong role in cortical field dynamics [5], [6] are predominantly radial. For these reasons, synchronization of cortical field activities within sparsely connected distributed domains should be much weaker than whole or partial synchronization of field activity within compact domains supported by short-range anatomic connections. Thus, cortical field potentials contributing to scalp EEG should arise largely from near-synchronous field activities within cortical ‘patches’ whose net far-field signals are near-instantaneously volume conducted to and linearly summed at EEG scalp electrodes. Emergence of near-synchronous field activity within small cortical domains has been observed and modeled in vivo (‘phase cones’) [7], in vitro (‘neuronal avalanches’) [8], [9] and in silico [10]. The net far-field projection of such a cortical patch will nearly equal that of a single ‘equivalent’ current dipole located near and typically beneath the center of the generating patch [11], [12]. In practice, we have observed that linear decomposition of high-density scalp EEG data by Independent Component Analysis (ICA) may return up to dozens of maximally independent component processes whose scalp maps are generally compatible with their generation in such a cortical patch [13], [14], [15], [16], [17]. This suggests an approach to comparing the relative biological plausibility and utility for EEG source separation of the many ICA and other blind source separation (BSS) algorithms that have been introduced in the last two decades. BSS/ICA sources BSS and, in particular, ICA methods are now widely used for separating artifacts from scalp-recorded electroencephalographic (EEG) and related data [18], [19], [20], [21], [22] and, increasingly, to separate and study brain source activities [13], [14], [23], [24], [25], [26]. ICA identifies signals in recorded multi-channel data mixtures whose time courses are maximally independent of one other and in this sense contribute maximally distinct information to the recorded data. Instead of directly addressing the general EEG inverse problem of determining the time courses and spatial distributions of cortical (and other) source areas of recorded scalp signals using an electrical forward head model (estimating the projection weights of all possible sources to the scalp sensors), ICA directly models What distinct signals are contained in the volume-conducted scalp data, and returns the relative projection strength of each maximally independent source to the scalp sensors, thereby also greatly simplifying the problem of determining Where in the brain each EEG source signal is generated [12], [24]. Non-brain sources Scalp-recorded EEG data also include non-brain or ‘artifact’ signals that are linearly mixed with brain EEG source activities at the scalp electrodes. ICA has been found to efficiently separate out several classes of spatially stereotyped non-brain signals: scalp and neck muscle electromyographic (EMG) activities, electro-oculographic (EOG) activities associated with eye blinks [20], saccades, and ocular motor tremor [15] as well as electrocardiographic (ECG) signal and single-channel noise produced by occasional loose connections between electrodes and scalp. Spatially non-stereotyped artifacts associated with irregular scalp maps (for example, artifacts produced by extreme participant movements) cannot be parsed by ICA into one (or a few) component(s), so these are best removed from the data before decomposition. Decomposition differences Though ICA algorithms all have the same root goal [27] and generally produce similar results when used to unmix idealized source mixtures, since EEG brain and non-brain source signals are likely not perfectly independent and different algorithmic approaches to maximizing independence differ, different ICA algorithms may return somewhat different results when applied to the same EEG data. Unlike most ICA algorithms that attempt to minimize instantaneous dependence, some BSS algorithms attempt to reduce redundancy between lagged versions of the data. To date, the three ICA/BSS algorithms applied most often to EEG data are likely extended Infomax ICA [27], [28], so-called FastICA [29], and Second-Order Blind Identification (SOBI) [30]. Computer code for these and a variety of other proposed ICA and BSS algorithms are readily available, making of interest a comparison of their effectiveness for EEG data decomposition. Comparing decompositions To date, however, suitable measures have not been demonstrated for comparing the components returned by different ICA/BSS algorithms applied to actual (as opposed to simulated) EEG data for which ‘ground truth’ source signals and scalp projections are not available. In particular, components produced by ICA decompositions that minimize mutual information between simultaneously recorded signal values have not been much compared to components produced by BSS algorithms that simultaneously minimize component signal redundancy at multiple time delays [16], [21]. Here, we use three measures – the amount of mutual information reduction (MIR) between the recovered component time courses relative to the recorded data channels (in kbits/sec), the mean remaining pairwise mutual information (PMI) between pairs of component time courses (in kbits/sec) [31], and the ‘dipolarity’ of the decomposition defined as the number of returned components whose scalp maps can be fit to the scalp projection of a single equivalent dipole with less than a specified error threshold (specified as percent residual variance). We applied these measures to 71-channel EEG data we had collected from 14 subjects (roughly 300,000 time points for each) as they performed a modified Sternberg visual working memory task [25] (see Methods). The motivation for the first two measures (MIR and PMI) is clear; they test how well the results of the decomposition approach the instantaneous independence objective. MIR, introduced here, is a direct (and as we show, easily computed) measure of the statistical distinctness of the activities of the resulting components, the absence of