ArticlePDF Available

EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics

Authors:

Figures

ERP image construction. (A) ERP-image plots are constructed by color-coding (grey bars) potential variations occurring in single-trial epochs (black traces). (B) Vertically stacking thin color-coded horizontal bars, each representing a single trial in an event-related dataset, produces an ERP image. Here, trials were sorted vertically according to the subject reaction-time (right curving black trace) before applying a 10epoch vertical moving average. The trace below the ERP image shows the ERP average of the imaged data epochs. The dot on the scalp map (top) indicates the scalp position of the channel whose data are imaged. (C) The erpimage() function automates several methods of sorting trials. Here, EEG phase in a given time/frequency window was used as the sorting variable. For each trial, a 10-Hz wavelet was applied to measure oscillatory activity in a 3-cycle window centered at time 0. Trials were then sorted (top to bottom) in order of their alpha band frequency phase values (-π to π) relative to stimulus onset and were displayed as an ERP image, again smoothed by a 10-trial moving average. The data were not otherwise filtered. The partial inter-trial phase coherence of the data following the stimulus onset is then visible as a change in the slope of the imaged activity wave fronts to near-vertical after 200 ms. Inter-trial phase coherence (bottom trace) shows that the distribution of alpha activity phase across trials is non-random (i.e., is partially phase-reset) between 200 and 450 ms (dotted line in lower trace shows p=0.01), resulting in same alpha activity appearing in the ERP average trace (top panel). The middle trace shows that mean changes in alpha power (in 'dB') did not change significantly (dotted lines) during the epochs. The baseline power level at the analysis frequency (25.9 dB, relative units) is indicated for possible comparison with other conditions.
… 
Content may be subject to copyright.
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 1
EEGLAB: an open source toolbox for analysis of single-trial EEG
dynamics including independent component analysis
Arnaud Delorme, Scott Makeig
Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San
Diego, La Jolla CA 92093-0961; {arno,scott}@sccn.ucsd.edu
Abstract: We have developed a toolbox and graphic user interface, EEGLAB, running under the cross-platform
MATLAB environment (The Mathworks, Inc.) for processing collections of single-trial and/or averaged EEG data
of any number of channels. Available functions include EEG data, channel and event information importing, data
visualization (scrolling, scalp map and dipole model plotting, plus multi-trial ERP-image plots), preprocessing
(including artifact rejection, filtering, epoch selection, and averaging), Independent Component Analysis (ICA) and
time/frequency decompositions including channel and component cross-coherence supported by bootstrap statistical
methods based on data resampling. EEGLAB functions are organized into three layers. Top-layer functions allow
users to interact with the data through the graphic interface without needing to use MATLAB syntax. Menu options
allow users to tune the behavior of EEGLAB to available memory. Middle-layer functions allow users to customize
data processing using command history and interactive ‘pop’ functions. Experienced MATLAB users can use
EEGLAB data structures and stand-alone signal processing functions to write custom and/or batch analysis scripts.
Extensive function help and tutorial information are included. A ‘plug-in’ facility allows easy incorporation of new
EEG modules into the main menu. EEGLAB is freely available (http://www.sccn.ucsd.edu/eeglab/) under the GNU
public license for noncommercial use and open source development, together with sample data, user tutorial and
extensive documentation.
INTRODUCTION
Though computing capabilities of nearly every
electrophysiology laboratory are now sufficient to
allow advanced signal processing of biophysical
signals including high-density electroencephalographic
(EEG) recordings, many researchers continue to rely
on amplitude and latency measures of peaks in EEG
trial averages, termed event related potentials (ERPs).
Historically, the response averaging method was
developed under technical constraints imposed by
hardware initially available for psychophysiological
experiments in 1950s and 1960s. Before digital
computers were available, researchers had to find a
way to summarize event-related activity across several
EEG trials representing brain responses to sensory
stimulations. For this purpose, they first used analog
registers to sum activity across EEG data trials. The
first computerized response averaging computer, the
computer of average transients (CAT, ca. 1962) helped
promote the use of response averaging, called at first
sensory ‘evoked potentials’ (EPs) and later the sensory
/ cognitive ‘event-related potentials’ (ERPs).
Using the fast and low-cost digital computers now
available, technical limitations that constrained
researchers to confine their EEG data analysis to
simple ERP measures and parametric statistics are no
longer relevant. The rationale used to justify response
averaging is that the single-trial EEG data time locked
to some class of experimental events consists of an
average ERP, whose time course and polarity is fixed
across the trials, plus other EEG processes whose time
courses are completely unaffected by the experiment
events. The cortical sources of ERP features may be
assumed to be spatially distinct from sources of
spontaneous EEG activities. However, as we have
demonstrated recently, focusing data analysis on
response averages alone ignores, first, event-related
dynamics that do not appear in, or are poorly
represented in response averages, and second, ignores
ongoing EEG processes that may be partially time and
phase-locked by experimental events, thereby
contributing portions of response averages (Delorme et
al., 2002; Makeig et al., 2002).
In the past decades, pioneer researchers have tried to
apply to EEG data analysis techniques developed in
electrical engineering and information theory,
including time/frequency analysis (Pfurtscheller and
Aranibar, 1979; Bressler and Freeman, 1980; Makeig,
1993; Neuenschwander and Varela, 1993; Tallon-
Baudry et al., 1996; Weiss and Rappelsberger, 1996)
and Independent Component Analysis (ICA) (Makeig
et al., 1996; Makeig et al., 1997; Makeig et al., 1999;
Jung et al., 2001). These techniques have revealed EEG
processes whose dynamic characteristics are also
correlated with behavioral changes, though they cannot
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 2
FIG. 1. Sample EEGLAB session. Screen capture of an EEGLAB user session running under Linux. Users call EEGLAB functions from the
main window menus (center) via ‘pop-up’ parameter selection windows (upper left). Warnings and data processing messages are shown in the
Matlab command line window (lower left), which can also be used to call EEGLAB or other data processing functions directly.
be seen in the averaged ERP. For example, short-term
changes in spectral properties of the ongoing EEG in
specific frequency bands may be correlated with
cognitive processes, e.g. expectancy of a target
stimulus (Makeig et al., 1999) and with visual
awareness (Rodriguez et al., 1999). The sufficiency of
studying average ERPs has also been questioned by
Makeig et al. (2002), who showed that some average
ERP peaks may result from partial synchronization of
oscillatory EEG processes to time locking events in
single data trials.
Currently, most EEG researchers still interpret their
data by measuring peaks in event-locked ERP
averages. Free availability of more general and easy-to-
use signal processing software for EEG data may
encourage the wider adoption of more inclusive
approaches. Our EEGLAB software toolbox for Matlab
(freely available from
http://www.sccn.ucsd.edu/eeglab/) allows processing
of collections of single EEG data epochs using ICA
and spectral analysis as well as data averaging
techniques. Using this toolbox, we have demonstrated
the advantages of combining ICA, time-frequency
analysis, and multi-trial visualization in several
publications (e.g., Makeig et al., 1999; Delorme et al.,
2002; Makeig et al., 2002; Delorme and Makeig,
2003). In EEGLAB, all these functions are available
under a common graphic interface under Matlab, a
widely used multi-platform computing environment.
EEGLAB extends the collection of publicly available
Matlab packages for brain imaging including SPM
(Friston, 1995) and FRMLAB (Duann et al., 2002) for
functional MRI studies and Brainstorm (Baillet et al.,
1999) for EEG/MEG source analysis.
METHODS AND RESULTS
1. Basic functions
The ICA/EEG toolbox of Makeig and colleagues
(1997) included a collection of Matlab functions for
signal processing and visualization of EEG data
including runica(), a function for automated infomax
ICA decomposition (Makeig et al., 1997), ERP-image
plotting (Jung et al., 1999; Makeig et al., 1999), a
method of visualizing time-locked potential variations
across sets of single trials, and time-frequency
decomposition (Makeig, 1993). By 2002, over 5,000
researchers from over 50 countries had downloaded the
ICA/EEG toolbox. However the provided tools could
only be used for EEG analysis by knowledgeable users
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 3
FIG. 2. Data scrolling. The EEGLAB scrolling data review
function, eegplot(), allows the user to review and reject data by
visual inspection. Here, five data epochs (separated by dashed lines)
are plotted at 31 electrode sites (channel names on the left). Other
channels in the dataset can be accessed using the vertical slider on
the left. The arrow buttons (lower left) scroll horizontally through
the data. The user may zoom in on a selected time range and/or
electrode group and may change the plotting parameters using menu
options (upper left). Values of the data point closest to the cursor are
continuously displayed at the bottom of the display. In this example,
two (central) data epochs have been automatically marked for
rejection for out-of-bounds values set by the user in the EEGLAB
data-rejection pop-up window (not shown). The rejection routine
here highlights (in white) the channels containing the outlier values.
The user can further mark (or unmark) data epochs for rejection by
clicking on them. Pressing ‘Update Marks’ (lower right) saves the
a
ccu
m
u
l
ated
r
eject
i
o
n m
a
rkin
gs.
who were prepared to write custom data analysis
scripts. EEGLAB, by contrast, includes a
comprehensive graphic user interface for interactively
calling and viewing results of enhanced and extended
ICA/EEG toolbox functions while further facilitating
the development of custom analysis scripts by prepared
users. Fig.1. shows a screen capture of an EEGLAB
user session running under Linux.
Data preprocessing. EEGLAB allows reading of data,
event information, and channel location files in several
different formats including binary, Matlab, ASCII,
Neuroscan, EGI, Snapmaster, European standard BDF,
and Biosemi EDF. Standard data analysis functions
available in EEGLAB include data filtering, data epoch
extraction, baseline removal, average reference
conversion, data resampling and extraction of data
epochs time locked to specified experimental events
from continuous or epoched data. EEGLAB also
includes methods allowing users to remove data
channels, epochs, and/or components dominated by
non-neural artifacts, by accepting or rejecting visually-
cued EEGLAB recommendations derived from signal
processing and information measures. EEG scalp maps
and channel locations can be converted between
several widely-used Cartesian, polar and spherical
coordinate systems and then visualized in two or three
dimensions. Continuous data and data epochs of any
number of channels can also be scrolled (both
vertically and horizontally).
Data structures and events. EEGLAB uses a single
structure (‘EEG’) to store data, acquisition parameters,
events, channel locations, and epoch information as an
EEGLAB dataset. This structure can also be accessed
directly from the Matlab command line. Text files
containing event and epoch information can be
imported via the EEGLAB menu. The user can also use
the menu to import event and epoch information in any
of several file formats (Presentation, Neuroscan, ASCII
text file), or can read event marker information from
the binary EEG data file (as in, e.g., EGI, Neuroscan,
and Snapmaster data formats). The menu then allows
users to review, edit or transform the event and epoch
information. Event information can be used to extract
data epochs from continuous EEG data, select epochs
from EEG data epochs, or to sort data trials to create
ERP-image plots (Jung et al., 1999; Makeig et al.,
1999). EEGLAB also provides functions to compute
and visualize epoch and event statistics.
To illustrate the utility of EEGLAB, below we employ
a small set of EEG data trials (also available from
http://www.sccn.ucsd.edu/eeglab/) drawn from an
experiment in which the subject covertly attended a
cued location on the computer screen, responding
quickly with a thumb button press each time a target
(filled square) was briefly presented at this location
(Makeig et al., 1999). In different trial blocks, the
attended location was any one of five positions
arranged horizontally on the computer screen above a
fixation cross. The sample dataset consists of 80 3-
second EEG epochs time-locked to targets presented in
the left visual field between 3 and 1.5 degrees of visual
angle. Data from thirty-one scalp electrodes (referred
to the right mastoid) were sampled at 500 Hz (later
reduced for compactness to 125 Hz). Fig. 2 shows five
sample data epochs and illustrates the capabilities of
eegplot(), the EEGLAB data scrolling function.
2. Multi-trial visualization
ERP-image plotting. The field of electrophysiological
data analysis has been dominated by analysis of one-
dimensional averaged event-related potential (ERP)
time series (single channel values across latencies).
The ERP-image is a more general two-dimensional
representation of the data (single channel values within
epochs across latencies) sorted in order of some
relevant measure (e.g., collection time, subject
response, amplitude or phase, etc.). Fig. 3(A) illustrates
the process of constructing ERP-image plots. An ERP
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 4
FIG. 3. ERP image construction. (A) ERP-image plots are constructed by color-coding (grey bars) potential variations occurring in single-trial
epochs (black traces). (B) Vertically stacking thin color-coded horizontal bars, each representing a single trial in an event-related dataset,
produces an ERP image. Here, trials were sorted vertically according to the subject reaction-time (right curving black trace) before applying a 10-
epoch vertical moving average. The trace below the ERP image shows the ERP average of the imaged data epochs. The dot on the scalp map
(top) indicates the scalp position of the channel whose data are imaged. (C) The erpimage() function automates several methods of sorting trials.
Here, EEG phase in a given time/frequency window was used as the sorting variable. For each trial, a 10-Hz wavelet was applied to measure
oscillatory activity in a 3-cycle window centered at time 0. Trials were then sorted (top to bottom) in order of their alpha band frequency phase
values (-π to π) relative to stimulus onset and were displayed as an ERP image, again smoothed by a 10-trial moving average. The data were not
otherwise filtered. The partial inter-trial phase coherence of the data following the stimulus onset is then visible as a change in the slope of the
imaged activity wave fronts to near-vertical after 200 ms. Inter-trial phase coherence (bottom trace) shows that the distribution of alpha activity
phase across trials is non-random (i.e., is partially phase-reset) between 200 and 450 ms (dotted line in lower trace shows p=0.01), resulting in
same alpha activity appearing in the ERP average trace (top panel). The middle trace shows that mean changes in alpha power (in ‘dB’) did not
change significantly (dotted lines) during the epochs. The baseline power level at the analysis frequency (25.9 dB, relative units) is indicated for
possible comparison with other conditions.
image is a colored rectangular image in which each
horizontal line represents a potential time series during
a single experimental trial. Instead of plotting activity
in single trials as left-to-right traces in which potential
is encoded by the ordinate of a data trace, trials are
represented as horizontal lines whose changing color
values indicate the potential at each time point in the
trial. Trials may be plotted in any sorting order of
interest, and a moving average of across adjacent single
trials may be used to highlight trial-to-trial consistency.
Fig. 3(B) illustrates the process of sorting the data trials
by the subject reaction time.
Some features of the visual ERP may be produced by
partial phase-resetting of ongoing EEG activities
following stimulus presentation (Makeig et al., 2002).
Fig. 3(C) illustrates a phase-sorted ERP-image plot, a
visualization tool used to assess whether partial phase
synchronization may account for ERP features. Sorting
by value or spectral amplitude in a given time window,
or by an auxiliary variable are also supported. The
erpimage() function can also plot the response average
ERP, changes in signal power and inter-trial coherence
(as defined below) at a selected frequency, the mean
signal spectrum, and a representative scalp topography.
Although a set of event-locked data trials has just one
ERP average, the number of possible ERP images of a
set of trials is very large since the trials can be sorted,
optionally smoothed, and imaged along any path
(linear or nonlinear), through the possibly high-
dimensional space of trial attributes and/or event
values. However, not all trial sorting orders give equal
insights into the brain dynamics expressed in the data.
It is therefore up to the user to decide which ERP
images to study.
ERP images can also be misinterpreted. For example,
using phase-sorting at one frequency (see Fig. 3(C))
can obscure the presence of oscillatory phenomena at
other frequencies. It is important not to lose sight of the
fact that nearly all activity recorded from scalp
electrodes is the volume conducted sum of activities
originating within a number of cortical domains.
EEGLAB uses independent component analysis (see
below) to separate out these activities under the
assumption that their activities are temporally
independent or at least more temporally independent
than any linear combinations of their signals.
3. Independent component analysis (ICA)
A primary tool of EEGLAB is to facilitate the process
of applying and evaluating the results of independent
component analysis (ICA) of EEG data. ICA
algorithms have proven capable of isolating both
artifactual and neurally generated EEG sources
(Makeig et al., 1999; Jung et al., 2000) whose EEG
contributions, across the training data, are maximally
independent of one another. ICA was first applied to
EEG by Makeig et al. (1996) and is now widely used in
the EEG research community, most often to detect and
remove stereotyped eye, muscle, and line noise
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 5
artifacts (Jung et al., 1999; Jung et al., 2000). The
temporal independence assumption of ICA is readily
understood as a basis for separating artifact sources,
since their activities will ordinarily not be reliably
phase-locked to one another, given enough training
data. In practice, however, ICA also has proved
capable of separating biologically plausible brain
sources whose activity patterns are distinctly linked to
behavioral phenomena. In fact, many of the
biologically plausible sources ICA identifies in EEG
data have scalp maps nearly fitting the projection of a
single equivalent current dipole (Jung et al., 2001;
Makeig et al., 2002), and are therefore quite compatible
with the projection to the scalp electrodes of
synchronous local field activity within a connected
patch of cortex.
EEGLAB contains an automated version, runica()
(Makeig, 1997), of the infomax ICA algorithm (Bell
and Sejnowski, 1995) with several enhancements
(Amari et al., 1996; Lee et al., 1999) both as a Matlab
function and as a stand-alone binary C program that
allows faster and less memory-intensive computation.
The toolbox also allows the user to select any of over
20 available ICA algorithms including JADE (Cardoso
and Souloumiac, 1993) and fixed-point ICA
(Hyvarinen and Oja, 2000).
Though it is not our goal here to describe ICA in detail,
we will try to give some insight about its nature. In
short, ICA finds a coordinate frame in which the data
projections have minimal temporal overlap. The core
mathematical concept of ICA is to minimize the mutual
information among the data projections or maximize
their joint entropy. ICA can be viewed as an alternative
linear decomposition to principal component analysis
(PCA). PCA applied in the temporal domain would
specifically make each successive component account
for as much as possible of the activity uncorrelated
with previously determined components – whereas
ICA seeks maximally independent sources.
This difference in goals leads to dramatic differences in
their results. PCA components are both temporally and
spatially orthogonal, a constraint unrealistic for actual
EEG sources, which arise in domains (spatial regions)
of partially synchronous activity in electrically oriented
cortical neurons (and possibly glia). Because the
density of cortical connections is weighted towards
local connections (<<1 cm), particularly in the network
of inhibitory cells that sustain cortical oscillations
(Pauluis et al., 1999), the partially synchronous
domains giving rise to EEG activity recorded on the
scalp should be mainly compact – though the extent
and density of these partially synchronous activities are
not known. Through simple volume conduction, the
projection of synchronous activity within nearly any
patch of cortex will be widespread on the scalp. Any
electrode will therefore sum contributions of EEG
sources in a large portion of cortex. EEG source
contributions to scalp electrode potentials depend on
source strengths and orientations as much as source
locations. The scalp projections of actual brain EEG
sources, therefore, are nearly always overlapping and
non-orthogonal, contrary to the assumption of PCA.
Indeed, because of the spatial orthogonality constraint,
projections of smaller principal components to the
scalp typically resemble checkerboard maps that could
not represent coherent activity within a connected
patch of cortex.
Therefore, to find biologically plausible sources, PCA
must be followed by an axis rotation procedure.
Previously proposed procedures, such as Promax and
Varimax, were drawn from the factor analysis
literature. ICA can be viewed as a more powerful
rotation method, though in practice ICA is usually
applied to the original data without PCA pre-
processing (for details, see Makeig et al., 1999). ICA
seeks to find component time courses that are mutually
independent, meaning that component cross-
correlations as well as all the higher-order moments of
the signals are zero. ICA is free to adapt to the actual
projection patterns of EEG generators if their activity
time courses are (near) independent of one another.
ICA is now being applied to many biomedical signal
processing problems including decomposing fMRI data
(Duann et al., 2002b) and speech and noise separation
(Park et al., 1999). Performing ICA decomposition is
most appropriate when sources are linearly mixed in
the recorded signals, without differential time delays.
These assumptions are precisely met for brain (and
non-brain) generator processes summed by volume
conduction in scalp EEG data. Because ICA does not
attempt to maximize the variance of each component,
ICA components may account for more equal portions
of the total signals than PCA components. For
example, in 32-channel decompositions ICA
component activities typically account for near 0% to
about 5% of the total signals. ICA may usefully be
applied to data with 128 or 256 channels, though
meaningful results are also possible using 32 or fewer
channels (Makeig et al., 2002).
Some earlier studies applied ICA to collections of ERP
data averages (Makeig et al., 1997; Makeig et al.,
1999). However, this approach requires care and
caution in interpretation of results. To separate two or
more processes, ICA requires that their independence
be expressed in the data. A small set of data averages
may not include enough conditions in the training set to
demonstrate the independence of the underlying
processes. If, for example, several processes are
partially phase reset in similar ways, the resulting
event-locked response averages may not express their
underlying functional and temporal independence. Data
averages, by their nature, contain sums of activities
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 6
FIG. 4. Visualizing independent components. (A) Topographical 2-D scalp maps of the nine independent components (ICs) accounting for the
most EEG variance of the 32 components returned by the ICA algorithm for the sample dataset. The component scalp map values returned by
ICA are proportional to µV (scaling is distributed between the component maps and activity time courses). From its far-frontal scalp map, IC3
appears to account for eye movement artifacts. (B) The ‘Component Properties’ display for IC3 verifies that it accounts for eye artifacts since its
activity spectrum is smoothly decreasing (bottom panel), and prominent eye movement artifacts appear in its activity ERP image (top right
panel). By removing this and other eye movement components (not shown) from the dataset, the user can remove most evidence of eye
movements from the data without removing other activity of interest (Jung et al., 2000).
occurring at similar latencies relative to some class of
events. When two or more sources invariably
contribute to a set of response averages at the same
latency, ICA, trained on these averages, may assign
their summed activities to a single component. Trained
on the unaveraged data however, ICA may use their
relative variability in single trials to separate them. A
second problem with applying ICA to data averages is
that the averaging process nearly cancels out the
activity of many of the EEG sources. Thus applying
ICA to the unaveraged EEG data also allows ICA to
separate ongoing activity of EEG sources even if they
are only partially phase-locked for brief time periods.
This is most useful when there are a sufficient number
of channels to fit the most active EEG and artifact
processes.
Theoretical assumptions underlying the use of ICA to
decompose EEG data include: (1) The data must
contain enough data points for the temporal
independence of the underlying sources to be expressed
(see Discussion). (2) No electrode activity should be a
linear mixture of other electrode activities (as may
occur for, e.g., average-reference data). If so, before
running ICA training, EEGLAB runica() function
automatically performs PCA pre-processing to reduce
the number of data dimensions to the rank of the input
data. (3) ICA assumes that each data source is spatially
stationary throughout the training data. This restriction
may be partially relaxed in more recent ICA methods
(Anemüller et al., 2003). (4) ICA assumes that the
distributions of activation values for each EEG source
are not precisely Gaussian. When a source distribution
is sub-Gaussian (e.g., as with line noise), the extended
option of infomax ICA must be used to separate it. The
current distribution of EEGLAB therefore focuses on
applying ICA directly to continuous EEG data or,
typically, to concatenated collections single EEG data
trials. Fig. 4 illustrates the use of infomax ICA applied
to the 80 EEG epochs of the EEGLAB sample dataset.
The lower the component index returned from runica(),
the more EEG data (neural and/or artifactual) it
accounts for.
To determine which components are behaviorally
relevant and should be selected for further
investigation, EEGLAB allows the user to plot
component contributions to the raw data spectrum
and/or to the trial-average ERP at all (or specified)
channels. Fig. 5(A) shows component contributions at
an alpha frequency to channel POz during the sample
epochs. The function returns the amount contributed by
each component as a percentage of total data power.
Another EEGLAB function for estimating component
contributions to the data, depicted in Fig. 5(B), shows
component contributions to the trial-average ERP in
the -500 to 1000 ms latency range. These and other
visualization functions help users to select which
components to process further using ERP-image
plotting (as described above) or using a variety of
spectral decomposition techniques (discussed below).
4. Time/frequency analysis
To assess event-related spectral amplitude, phase and
coherence perturbations in data recorded from single
electrodes and/or in ICA components, EEGLAB
employs custom spectral decomposition techniques.
Our primary measures are the baseline or epoch-mean
power spectrum and three event-related time/frequency
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 7
FIG. 5. Evaluating independent component contributions. (A) An EEGLAB spectopo() plot showing the components accounting for the
largest portions of 10-Hz activity at electrode POz (middle scalp map). The figure shows the power spectrum of the selected channel (top black
trace), the activity spectra of the projection to that channel of each of the 32 components (lower traces), and the scalp power maps of the four
largest-contributing components (4, 5, 7, 10). (B) An envtopo() plot showing the envelopes (i.e., the min and max values, over all channels, at
each time point) of the five independent components making the largest potential contributions to the ERP. The black thick traces show the
envelope of the (all channel) ERP data and the thin traces, the envelopes of the depicted component contributions to the ERP.
measures: (1) the event-related spectral perturbation
(ERSP), measuring mean event-related changes in the
power spectrum at a data channel or component
(Makeig, 1993), (2) inter-trial coherence (ITC
magnitude and phase, also called phase-locking factor)
at single channels or components, and (3) event-related
cross-coherence (ERCOH, magnitude and phase)
between two data channels or components.
ERSP. Plots of the baseline-normalized spectrogram or
the event-related spectral perturbation (ERSP) are
increasingly used in the EEG literature to visualize
mean event-related changes in spectral power over time
in a broad frequency range. They generalize the
narrow-band event-related desynchronization (ERD)
and synchronization (ERS) measures introduced by
Pfurtscheller and colleagues (Pfurtscheller and
Aranibar, 1979).
Calculating an ERSP requires computing the power
spectrum over a sliding latency window then averaging
across data trials. The color at each image pixel then
indicates power (in dB) at a given frequency and
latency relative to the time locking event. Typically,
for n trials, if ),( tfFk is the spectral estimate of trial
k at frequency f and time t
=
=
n
k
ktfF
n
tfERSP
1
2
),(
1
),( (1)
To compute ),( tfFk EEGLAB uses either the short-
time Fourier transform, a sinusoidal wavelet (short-
time DFT) transform, or a Slepian multitaper
decomposition (Thompson, 1982) that provides a
specified time and frequency resolution. In our
experience, there are no dramatic differences between
these decompositions (though the number of cycles in
each data window can be critical). Most often we use a
version of sinusoidal wavelets in which the number of
cycles is increased slowly with frequency (Fig. 6). This
feature allows us to obtain better frequency resolution
at higher frequencies than a conventional wavelet
approach that uses constant cycle length. This method
is also better matched to the linear scale we use to
visualize frequencies. To visualize power changes
across the frequency range, we subtract the mean
baseline log power spectrum from each spectral
estimate, producing the baseline-normalized ERSP.
Significance of deviations from baseline power is
assessed using a bootstrap method. A surrogate data
distribution is constructed by selecting spectral
estimates for each trial from randomly selected latency
windows in the specified epoch baseline (e.g., prior to
stimulus onset), and then averaging these. Applying
this process several hundred times (default: N=200)
produces a surrogate ‘baseline’ amplitude distribution
whose specified percentiles are then taken as
significance thresholds. If sufficient pre-stimulus data
are not available, the surrogate data may be drawn
from any other part or from the whole epoch. Figs.
6(A) and 6(B) show significant ERPS phenomena for
two independent EEG components.
ITC. Inter-Trial Coherence (ITC) is a frequency-
domain measure of the partial or exact synchronization
of activity at a particular latency and frequency to a set
of experimental events to which EEG data trials are
time locked. The measure was introduced by Tallon-
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 8
FIG. 6. Time/frequency decompositions of independent component activities. Time/frequency decomposition was applied to the activities of
two independent EEG components using sinusoidal wavelet transforms, 3 cycles in length at the lowest frequency (6 Hz), increasing linearly
with frequency up to 9 cycles at the highest plotted frequency (35 Hz). Using this approach, it is possible to obtain reasonable time and frequency
stability at all frequencies. (A-B) Event-related spectral perturbation (ERSP) plots showing mean changes in spectral power during the epoch,
relative to a 1-s pre-stimulus baseline (plotted vertically on the left). Component IC4 shows a transient increase near 12 Hz centered at 500 ms,
while component IC9 shows a power decrease in this range following 500 ms. (C-D) Phase cross-coherence (ERPCOH) magnitude and phase
delay between the two components shown in panels A-B, zero-masked in regions in which cross-coherence magnitude was not significant
(p>0.01). The components appear to become partially synchronized above 10 Hz (coherence 0.53) during the period 400 to 1000 ms with a
phase offset near -120 degrees. Under the minimum phase assumption, this implies that high-alpha activity of IC9 tends to lead that of IC4
during this period by about 30 ms.
Baudry et al. (1996) and termed a ‘phase locking
factor.’ The term ‘inter-trial coherence’ refers to its
interpretation as the event-related phase coherence
(ITPC) or event-related linear coherence (ITLC)
between recorded EEG activity and an event-phase
indicator function (e.g. a Dirac or cosine function
centered on the time locking event). Using the same
notation as above
Inter-trial phase coherence (2)is defined by
=
=
n
kk
k
tfF
tfF
n
tfITPC
1),(
),(1
),(
and Inter-trial linear coherence (3) by
=
=
=n
k
k
n
i
k
tfFn
tfF
tfITLC
1
2
1
),(
),(
),(
where represents the complex norm. The most
common (and default) version is inter-trial phase
coherence (called ‘phase-locking factor’ by Tallon-
Baudry et al., (1996)). The ITC measure takes values
between 0 and 1. A value of 0 (not expected in practice
based on a finite number of epochs) represents absence
of synchronization between EEG data and the time-
locking events; a value near 1 indicates their perfect
synchronization (i.e. near perfect EEG phase
reproducibility across trials at a given latency). In the
complex 2-D Cartesian coordinate frame, spectral
estimates at given frequencies and times are returned as
complex vectors in the 2-D phase space. The norm and
phase angle of each vector are represented by the
magnitude and phase of the spectral estimate. To
compute inter-trial phase coherence (ITPC), we first
normalize the lengths of each of the trial activity
vectors to 1 and then compute their complex average.
Thus, only the information about the phase of the
spectral estimate of each trial is taken into account.
For linear inter-trial coherence (ITLC), the initial
normalization step is omitted: the vector sum is
computed and then normalized by RMS power in the
single-trial estimates. EEGLAB function erpimage()
computes ITPC at a single frequency for display
beneath an ERP image (Fig. 3(C)); function timef()
computes color-coded ITPC or ITLC images across
frequencies (not shown). As for the ERSP, ITC
significance levels are assessed using surrogate data by
randomly shuffling the single-trial spectral estimates
from different latency windows during the baseline
period.
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 9
ERCOH. EEGLAB function crossf() computes event-
related coherence (ERCOH) between two channel or
component activities in sets of trials to determine the
degree of synchronization between the two activity
measures. As for ITC, both phase coherence
(ERPCOH) and linear coherence (ERLCOH) measures
are supported. Other phase coherence measures have
not (yet) been included in EEGLAB (e.g., Lachaux et
al., 1999). In EEGLAB, for two signals, a and b, and
using the same notation as above
Phase cross-coherence (4) is defined by
=
=
n
kb
k
a
k
b
k
a
k
ba
tfFtfF
tfFtfF
n
tfERPCOH
1
*
,
),(),(
),(),(1
),(
and linear cross-coherence (5) by
==
=
=n
k
b
k
n
k
a
k
b
k
n
k
a
k
ba
tfFtfF
tfFtfF
tfERLCOH
1
2
1
2
*
1
,
),(),(
),(),(
),(
Here, *
),( tfF b
k is the complex conjugate of
),( tfF b
k. The magnitude of cross-coherence varies
between 0 and 1, a value of 0 again indicating a
complete absence of synchronization at the given
frequency f in the time window centered on t, and 1
indicating perfect synchronization. As for ITPC, the
normalizing factor in the ERPCOH denominator
ensures that only the relative phase of the two spectral
estimates at each trial is taken into account. Linear
ERCOH (ERLCOH), by contrast, estimates the extent
of complex linear relationship between the two signals
(proportional amplitudes at a fixed delay).
When ERCOH magnitude (i.e., norm of the complex-
valued ERCOH vector) is significantly above its
expected baseline value, the phase of the ERCOH
vector may indicate, under the minimum phase
assumption, which of the two component activities
tends to lead the other at the analysis frequency. The
minimum phase assumption means that the actual
phase lag is les than ±180°. Fig. 6(C) illustrates
significant ERPCOH synchronization between two
components. Even though independent components
were identified by ICA as being (maximally)
independent over the whole time range, they may
exhibit partial but statistically significant
synchronization, within specific event-related
time/frequency windows (Delorme et al., 2002). Here
again, crossf() can assess significance of the observed
ERCOH using the method of surrogate data by
computing the expected ERCOH distribution using
randomly selected data windows from the ‘baseline’
portion of each epoch. Different surrogate data
selection methods are used to estimate ERCOH for the
two processes, either including or excluding any
common spectral amplitude changes and/or partial
phase-locking related to the time-locking experimental
events. These four methods are referred to in EEGLAB
as linear or phase coherence, with or without removal
of common ITC. The preferable method may depend
on several factors that we do not detail here.
5. Menu calls and script writing
The EEGLAB graphic user interface (GUI) is designed
to allow non-experienced Matlab users to apply
advanced signal processing techniques to their data.
However, more experienced users can also use the GUI
to save time in writing custom and/or batch analysis
scripts in Matlab by incorporating menu shortcuts and
EEGLAB history functions. Table I provides examples
of EEGLAB scripts of different levels of complexity.
EEGLAB functions may be roughly divided into three
layers designed to increase ease-of-use for different
types of users:
I. GUI-based use. Naive Matlab users may choose to
interact only with the main EEGLAB window menu,
first to import data into EEGLAB (in any of several
supported formats), and then to call any of a large
number of available data processing and visualization
functions by selecting main-window menu items
organized under five headings: ‘File’ menu functions
read/save data file and data information files. ‘Edit’
menu functions allow editing a dataset, changing its
properties, reviewing and modifying its event and
channel information structures. ‘Tools’ menu functions
extract epochs from continuous data (or sub-epochs
from data epochs), perform frequency filtering,
baseline removal, and ICA, and can assist the user in
performing semi-automated artifact data rejection
based on a variety of statistical methods applied to
activity in the raw electrode channels or their
independent components. ‘Plot’ menu functions allow
users to visualize the data in a variety of formats, via
(horizontally and vertically) scrolling displays or as
trial (ERP), power spectrum, event-related
time/frequency averages, etc. A large number of
visualization functions are dedicated to the display and
review of properties of scalp data channels and
underlying independent data components. The user can
make use of standard Matlab capabilities to edit, print,
and/or save the resulting plots in a variety of formats.
Finally, the user can use ‘Help’ menu functions to call
up documentation on EEGLAB functions and data
structures.
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 10
1. >> pop_erpimage(EEG);
2. >> figure; pop_erpimage(EEG, 1, [1], [], 'Channel 1
erpimage', 10, 1);
3. >> erpimage(EEG.data(1,:), ones(1, EEG.trials ) *
EEG.xmax * 1000, linspace( EEG.xmin * 1000,
EEG.xmax * 1000, EEG.pnts), 'Channel 1 ERP
image', 10, 1, 'topo', {1 EEG.chanlocs }, 'erp',
'cbar');
TABLE I. Sample EEGLAB processing scripts. All scripts assume
that the Matlab data structure ‘EEG’ contains the sample EEGLAB
dataset (described in the EEGLAB tutorial and available for
download). Script 1 calls the EEGLAB (‘pop’) interface function
that in turn calls the erpimage() processing function to compute and
draw an ERP image plot (Jung et al., 1999; Makeig et al., 1999) of a
selected single-channel time record for each trial. Additional plotting
parameters can then be entered manually by the user in the resulting
pop-up window. Script 2 performs the same action, but now the
ERP-image ‘pop’ function is called with specific arguments. The
ERP-image plot then appears directly, with no intervening ‘pop’
window. Each time the user selects an operation from the EEGLAB
menu, the resulting Matlab function call (including all input
parameters) is appended to the EEGLAB session command history.
Subsequently, the user can simply copy and paste commands from
the command history to repeat the same actions. Thus, in Script 3,
the ‘pop’ ERP-image function is bypassed and the eponymous
EEGLAB data processing function, erpimage(), is called directly by
the user script referencing parameters stored in the EEG data
structure. The erpimage() function requires no knowledge of the
EEG data structure used by EEGLAB, and can be applied to any
user-defined data array. If the user selects the supplied default
parameters in the pop_erpimage() pop-up data entry window, the
three scripts will all have the same effect. See the Matlab help
messages for the meaning of the pop_erpimage() and erpimage()
function arguments (also available as HTML pages linked to the
main EEGLAB website).
II. EEGLAB command history. Intermediate level users
may first use the menu to perform a series of data
loading, processing and visualization functions, and
then may take advantage of the EEGLAB command
history functions to easily produce batch scripts for
processing similar data sets. Every EEGLAB menu
item calls a Matlab function that may also be called
from the Matlab command line. These interactive
functions, called ‘pop’ functions, work in two modes.
Called without (or in some cases with few) arguments,
an interactive data-entry window pops up to allow
input of additional parameters. Called with additional
arguments, ‘pop’ functions simply call the eponymous
data processing function, without creating a pop-up
window. For example, function pop_erpimage() calls
erpimage(). When a ‘pop’ function is called by the user
by selecting a menu item in the main EEGLAB
window, the function is called without additional
parameters, bringing up its GUI pop-up window to
allow the user to enter computation parameters. When
the processing function is called by EEGLAB, its
function call is added as a command string to the
EEGLAB session history variable. By copying history
commands to the Matlab command line or embedding
them in Matlab text scripts, users can easily apply
actions taken during a GUI-based EEGLAB session to
a different data set. A comprehensive help message for
each of the ‘pop’ functions allows users to adapt the
commands to new EEG data.
III. Custom EEGLAB scripting. More experienced
Matlab users can take advantage of EEGLAB functions
and dataset structures to perform computations directly
on datasets using their own scripts that call EEGLAB
and any other Matlab functions while referencing
EEGLAB data structures. Since all the EEGLAB data
processing functions are fully documented, they can be
used directly. Experienced users should benefit from
using all three modes of EEGLAB processing: GUI-
based, history-based, and autonomously scripted data
analyses. Such users can take advantage of the data
structure (‘EEG’) in which EEGLAB datasets are
stored. The GUI interface uses a single Matlab
variable, a structure named ‘EEG’ that contains all
dataset information and is always available at the
Matlab command line. This variable can easily be used
and/or modified to perform custom signal processing or
visualizations. Finally, while EEGLAB ‘pop’ functions
(described above) assume that the data are stored in an
EEG data structure, most EEGLAB signal processing
functions accept standard Matlab array arguments.
Thus, it is possible to bypass the EEGLAB interface
and data structures entirely, and directly apply the
signal processing functions to data matrices.
6. Distribution, documentation and support
The EEGLAB toolbox is distributed under the GNU
General Public License (for details see
http://www.gnu.org/licenses/gpl.txt). The source code,
together with web tutorials and function description
help pages, is freely available for download from
http://sccn.ucsd.edu/eeglab/. As the toolbox currently
includes approximately 300 Matlab functions
comprising 50,000 lines of Matlab code, it is not
possible to describe all of its functionality in a journal-
length paper. An extensive user tutorial explains in
detail how to import and process data using EEGLAB,
including the derivation and evaluation of its
independent components. We also provide ‘Frequently
Asked Questions (FAQ)’ and ‘Known Bugs’ web
pages, a support email (eeglab@sccn.ucsd.edu), a
dedicated mailing list for software updates
(eelagbnews@sccn.ucsd.edu), and a discussion mailing
list (eeglablist@sccn.ucsd.edu) which currently reaches
over a thousand EEG researchers.
Open-source EEGLAB functions are not precompiled;
users can read and modify the source code of every
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 11
function. Each EEGLAB function is also documented
carefully using a standardized help-message format and
each function argument is described in detail with links
to related functions. We have attempted to follow
recognized best practice in software design for
developing EEGLAB. The source code of EEGLAB is
extensively documented and is internally under the
Linux revision control system (RCS), which allows us
to easily collaborate with remote researchers on the
development of new functions. Matlab allows
incremental design of functions, so adding new
features to a function can be easily accomplished while
preserving backward compatibility. The EEGLAB
history feature also makes it easy to generate test
scripts that we now launch nightly to maintain
EEGLAB stability.
DISCUSSION
We have developed EEGLAB, a complete interactive
environment for processing EEG (or MEG) data under
Matlab, to provide both standard and advanced EEG
processing functions developed in our own and other
laboratories. EEGLAB is strongly oriented towards
single-trial visualization techniques, ICA and event-
related time/frequency analysis. Because the software
was developed by and for ERP/EEG researchers, we
have taken care to make the data processing as
transparent as possible and to allow users to tune their
parameters as easily as possible. We will now briefly
review a few limitations of EEGLAB and, because the
methods incorporated into EEGLAB are not yet widely
practiced, some limitations of ICA applied to high-
density EEG data.
Limitations of time/frequency decomposition
Filtering methods implemented in EEGLAB take
advantage of linear filtering implemented in the Matlab
Signal Processing toolbox. One of drawback of using
linear filters is that the signal roll-off at the cut-off
frequency is weaker than what it would be using
nonlinear filters. However, with linear filtering, data
phase information is preserved across frequencies.
Time-frequency decomposition in EEGLAB is limited
to FFTs, multi-taper analysis, and a single type of
sinusoidal wavelet, as is standard for EEG analysis.
Other methods, for example the Hilbert method, are not
currently implemented. However quantitative
comparisons show that results on EEG data using
Hilbert transforms do not differ dramatically from
applying sinusoidal wavelets (Le Van Quyen et al.,
2001). Also, bi-coherence between frequencies cannot
yet be assessed within EEGLAB (e.g., von Stein and
Sarnthein, 2000; Lachaux et al., 2003). We intend in
the future to include functions to assess
synchronization (1) of phase at one frequency with
amplitude at another frequency, (2) of phase
synchronization between frequencies, and (3) of
amplitude correlation between frequencies. We
welcome further open source contributions
implementing other time-frequency approaches, and
have added an EEGLAB plug-in facility to promote
and ease development of such contributions.
Significance and statistical comparisons across
subjects or conditions
To assess significance of within-subject measures,
EEGLAB uses non-parametrical methods that do not
assume a known activity distribution. A null hypothesis
distribution, used to determine significance thresholds,
is estimated by accumulating surrogate data, shuffling
the data across latencies alone, latencies and trials, or
trials alone. To compensate for multiple comparisons,
significance thresholds may need to be decreased (e.g.
Bonferroni, 1950; Holm, 1979). Since it is not
reasonable to compute an unlimited amount of
surrogate data to estimate very low probability
thresholds heuristically, we have implemented a
method to fit the observed surrogate data distribution
using a fourth order distribution fit (Ramberg et al.,
1979). This feature will be available in a near-term
release of EEGLAB.
To test significance across conditions or subjects, we
either use parametrical tests or accumulated
significance results from each subject. Our ERP
function pop_comperp() currently uses a t-test to
compare two conditions for several subjects. When
processing spectral decompositions of one channel (or
component class) from different subjects (already been
masked for significance), our tftopo() function applies
a threshold derived by simple statistics on the number
of subjects for which the spectral decomposition is
significant at a give time-frequency point. If not
enough subjects show a significant change at the
specified point, this point is considered non-significant
in the group average. This is a statistically conservative
approach. For further statistical assessment, raw data,
ERP, or independent component weights and activity
can be exported as ASCII to statistical packages such
as Statview (SAS Institute Inc.), SPSS (SPSS Inc.), or
the Matlab Statistics Toolbox (The Mathworks, Inc.).
ICA Stability
Because the infomax ICA algorithm begins with a
random unmixing matrix and then randomly shuffles
the order of the data time points before each training
step, the results of successive ICA decompositions may
be slightly different even when ICA is performed on
the same data. In particular, the scalp maps and activity
time courses of the independent components (and their
order), may differ slightly across runs. Therefore, we
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 12
advise that features of the decomposition that do not
remain stable across decompositions of the same data
should not be interpreted except as irresolvable ICA
‘uncertainty.’ Differences between decompositions
trained on somewhat different data subsets may have
additional causes. We are currently investigating the
stability of ICA methods applied to typical datasets
(Delorme et al., in preparation).
Difference between ICA algorithms
Which is the best ICA algorithm to use for EEG
decomposition? From a theoretical point of view, all
ICA algorithms maximize independence in an
approximate sense (Lee et al., 2000), while the degree
to which EEG data actually fit ICA assumptions is
unknown. Applied to simulated, relatively low
dimensional data sets for which the ICA assumptions
are exactly fulfilled, leading ICA algorithms (including
infomax, JADE, and FastICA) return near-equivalent
components. However, the physiological significance
of any differences in the results of the same or different
ICA algorithms (or of different parameter choices for
the various algorithms) has not been systematically
tested and reported - neither by us nor, as far as we
know, by others. Therefore, different ICA
decompositions may give slightly different results, as
has been shown for neural ensemble data (Laubach et
al., 1999) and fMRI data (Duann et al., 2001; Esposito
et al., 2002). Each ICA algorithm has its own
particularities. The infomax algorithm in its native
form can only separate sources with super-Gaussian
(i.e. peaky, thick-tailed) activity distributions. If there
are strong electrical artifacts in data, it is preferable to
use the ‘extended’ ICA option of runica() (Lee et al.,
1999), to allow the algorithm to detect sources with
sub-gaussian activity distribution, such as line current
artifacts and/or slow activity.
Whereas infomax implicitly uses a combination of
higher-order moments of the data to find independent
components, the JADE algorithm (Cardoso and
Souloumiac, 1993) diagonalizes all the fourth-order
moments explicitly. Although for low numbers of data
channels the JADE algorithm is fast and stable, the
memory required to manipulate all the fourth-order
moments becomes quite impractical with high numbers
of channels. Whereas both infomax and JADE
algorithms find and return all the independent
components at once, the default setting of the fixed-
point ICA algorithm of Hyvärinen (2000) computes
and returns components one by one. The order of the
components it returns, however cannot be known in
advance, and performing a complete decomposition is
not faster than with infomax. Also, in our experience
(see also Esposito et al., 2002) the fixed-point ICA
algorithm may be more less robust than infomax ICA
when applied high-dimensional real data. To
decompose EEG data, therefore, we most often use
infomax or extended infomax ICA. The infomax
algorithm reliably finds independent components that
are physiologically plausible, functionally distinct, and
often have spatial and functional similarities across
data sets, sessions, and subjects (Delorme et al., 2002;
Makeig et al., 2002).
Insufficient data for running ICA
A chief case in which ICA algorithms may not return
reliable results is when too few data are provided to
them. ICA being a statistical method, if the
independence of the functionally distinct EEG
processes is not adequately exhibited in the data, ICA
cannot separate them.
The size of the weight matrix being the square of the
number of channels, a number of time points at least a
few times the square of the number of channels is
usually needed to obtain reliable decompositions.
These data points may be drawn from continuous data
or from several data epochs Of course, additional data
points can only improve the decomposition - when and
if relative stationarity of the spatial structure of the
EEG sources set can be assumed.. In our experience,
using short baseline-zeroed data epochs that include
task-related behavior may give qualitatively more
consistent results than using longer data epochs. Using
short epochs constrains ICA to focus on the task-
relevant portion of the data.
Nonlinearities
Another case in which ICA will fail to extract all the
involved sources occurs when the data are not a linear
sum of the underlying source projections – this chiefly
occurs when the amplifiers become ‘railed’ at high
signal levels, leading to signal ‘clipping’, or when high
signal levels exceed the input range of the A/D
converter, leading to signal ‘wrap-around.’ In either
case, the severe nonlinearity involved will cause linear
ICA algorithms to give spurious results, so such data
epochs must be carefully rejected from the data before
running ICA.
Noise
Finally, when the data contains many more strong
spatial sources than the number of recording channels,
the additional sources must be mixed into the output
components. In particular, this may occur during
‘paroxysmal’ noise which may for instance be
introduced into EEG data during strong head
movements. Else, a loose electrode may introduce a
large noise signal not linearly related to any of the
other electrode signals. In this case, ICA may dedicate
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 13
a single component to the electrode noise, thus
unnecessarily reducing the number of components
available to separate other neural and artifact sources.
Therefore, we find it best to train ICA on carefully
pruned ‘clean’ data epochs, which can, however, retain
spatially stereotyped artifact activity such as eye blinks
and eye movements, repeated muscle activity, etc.
Processing speed
Matlab offers a powerful environment for processing
biophysical data because of (1) the simplicity of its
command line language, (2) the many Matlab functions
made available by The Mathworks, Inc and by
independent researchers, and (3) its high-level
visualization capabilities. However, there are two
possible problems in using Matlab for processing EEG
data. First, though ever-increasing speed of current
workstations continues to make processing time of less
limiting importance for data analysis, interpreted
computer languages are inherently slower than
compiled languages. Matlab has facilities for compiling
and running binary versions of scripted functions, but
their speed may still be suboptimal. For this reason, we
converted to C the most time consuming EEGLAB
function, runica(). Both the Matlab ICA function -
runica() - and the binary C-language ICA function -
binca() - can be called from the EEGLAB GUI. Under
Matlab v4, we observed a speed-up of about 10 for
binica() compared to runica(). However, under Matlab
v6 the speed-up factor seems to be much smaller
(<100%). Most other EEGLAB functions are less
compute intensive.
Memory requirements
Another relative disadvantage of using Matlab to
process high-density EEG data is that Matlab currently
converts all floating-point numbers to 64-bit double-
precision, thus requiring large amounts of main
memory to process large data sets. Though hopefully
some future Matlab versions may allow the option of
processing data in 32-bit floating-point format, we
have taken care to address this issue in EEGLAB by
including various options to minimize memory usage,
such as constraining EEGLAB to work on a single
dataset, or computing the ‘activation’ time courses of
independent components only as needed. However, this
issue remains a serious problem for large datasets:
parts of the toolbox may have to be updated to allow
very large (e.g., long 256-channel) datasets to be
analyzed within the current Linux 2GB/process limit.
One possibility is to use the Matlab MEX language, an
interface between C and Matlab that allows a wider
variety of data types including single precision.
Another possibility is to have EEGLAB load into main
memory only a part of the dataset at a time. However,
as 64-bit processors become more available, the current
data space limits of operating systems and Matlab
should increase, in which case the remaining problem
would only be the burden of purchasing the necessary
RAM.
Current development of EEGLAB focuses on
processing of large datasets (>1 Gb), semi-
automatically grouping independent component across
subjects, and component source localization. EEGLAB
will also be linked to our FMRLAB toolbox
(http://www.sccn.ucsd.edu/fmrlab) to process
simultaneously recording EEG and fMRI data (Duann
et al., 2002a). We also have begun working with co-
developers to increase the range of EEGLAB functions
using the ‘plug-in’ facility, whereby contributors may
easily contribute optional EEGLAB code that is readily
incorporated into the EEGLAB menu. The plug-in
facility is designed so that plug-in functions can be
used and distributed both within EEGLAB and
independently. By this mechanism we hope to
encourage the open source development of
comprehensive EEG (and MEG) signal processing
tools under EEGLAB.
ACKNOWLEDGEMENTS
The authors acknowledge the contributions to
EEGLAB and its hundreds of functions by many
contributors. Principal among these were Colin
Humphries, who wrote the topographic plotting
functions and the first version of the data scrolling
display function, Sigurd Enghoff, who wrote the first
versions of the time/frequency functions and translated
the MATLAB-coded runica() infomax ICA function to
binary, and Tzyy-Ping Jung, who contributed the first
version of the erpimage() function. The runica()
function itself was written by one of us (SM) building
on core ICA code contributed by Tony Bell and Te-
won Lee. Several other members of the Computational
Neurobiology Laboratory at The Salk Institute
contributed other functions. We gratefully
acknowledge the support, collaboration and
encouragement of Terrence Sejnowski at Salk Institute
throughout this research process. We thank Stefan
Debener and anonymous reviewers for suggestions on
this report, which was supported by the National
Institutes of Health USA and by The Swartz
Foundation.
Journal of Neuroscience Methods 134 (2004), 9-21
Delorme and Makeig 14
References
Anemüller J, Sejnowski TJ, Makeig S. Complex Independent
Component Analysis of Frequency-Domain
Electroencephalographic Data. Neural Networks, 2003; in press.
Amari S-I, Cichocki A, Yang HH. A new learning algorithm for
blind source separation. Adv. Neural Inf. Process. Syst., 1996; 8:
757-63.
Baillet S, Mosher JC, Leahy RM, Shattuck DW. BrainStorm: a
Matlab Toolbox for the Processing of MEG and EEG Signals.
Proceedings of the 5th Int. Conf. on Human Brain Map.,
Neuroimage, 1999; 9: 246.
Bell AJ, Sejnowski TJ. An information-maximization approach to
blind separation and blind deconvolution. Neural Comput., 1995;
7: 1129-59.
Bressler SL, Freeman WJ. Frequency analysis of olfactory system
EEG in cat, rabbit, and rat. Electroencephalogr. Clin.
Neurophysiol., 1980; 50: 19-24.
Bonferroni CE. Sulle medie multiple di potenze. Bollettino
dell’Unione Matematica Italiana, 5 third series, 1950; 267–70.
Cardoso J-F, Souloumiac A. Blind beamforming for non gaussian
signals. IEE-Proc.-F, 1993; 140: 362-70.
Delorme A, Makeig S, Fabre-Thorpe M, Sejnowski T. From Single-
trials EEG to Brain Area Dynamics. Neurocomputing, 2002; 44-
46: 1057-64.
Delorme A, Makeig S. EEG changes accompanying learning
regulation of the 12-Hz EEG activity. IEEE Transactions on
Rehabilitation Engineering, 2003; 11: 133-6.
Duann JR, Jung TP, Makeig S, Sejnowski TJ. fMRLAB: An ICA
Toolbox for fMRI Data Analysis. In: Human Brain Mapping,
2002; Sendai, Japan.
Duann JR, Jung TP, Kuo WJ, Yeh S, Makeig S, Hsieh JC, Sejnowski
TJ. Measuring the variability of event-related BOLD signals. In:
Third International Workshop on Independent Component
Analysis and Signal Separation, 2001; San Diego, USA.
Duann JR, Jung TP, Kuo WJ, Yeh TC, Makeig S, Hsieh JC,
Sejnowski TJ. Single-trial variability in event-related BOLD
signals. Neuroimage, 2002b; 15: 823-35.
Esposito F, Formisano E, Seifritz E, Goebel R, Morrone R, Tedeschi
G, Di Salle F. Spatial independent component analysis of
functional MRI time-series: to what extent do results depend on
the algorithm used? Hum. Brain Mapp., 2002; 16: 146-57.
Friston KJ. Statistical parametric mapping: Ontology and current
issues. J. Cereb. Blood Flow Metab., 1995; 15: 361-70.
Holm S. A simple sequentially rejective multiple test procedure.
Scandinavian Journal of Statistics, 1979; 6: 65-70.
Hyvarinen A, Oja E. Independent component analysis: algorithms
and applications. Neural Netw., 2000; 13: 411-30.
Jung TP, Makeig S, Westerfield M, Townsend J, Courchesne E,
Sejnowski TJ. Analyzing and Visualizing Single-trial Event-
related Potentials. Adv. Neural Inf. Process. Syst., 1999; 11: 118-
24.
Jung TP, Makeig S, Westerfield M, Townsend J, Courchesne E,
Sejnowski TJ. Analysis and visualization of single-trial event-
related potentials. Hum. Brain Mapp., 2001; 14: 166-85.
Jung TP, Makeig S, Humphries C, Lee TW, McKeown MJ, Iragui V,
Sejnowski TJ. Removing electroencephalographic artifacts by
blind source separation. Psychophysiology, 2000; 37: 163-78.
Lachaux JP, Rodriguez E, Martinerie J, Varela FJ. Measuring phase
synchrony in brain signals. Hum. Brain Mapp., 1999; 8: 194-208.
Lachaux JP, Chavez M, Lutz A. A simple measure of correlation
across time, frequency and space between continuous brain
signals. J Neurosci Methods, 2003; 123: 175-88.
Laubach M, Shuler M, Nicolelis MA. Independent component
analyses for quantifying neuronal ensemble interactions. J.
Neurosci. Methods, 1999; 94: 141-54.
Le Van Quyen M, Foucher J, Lachaux J, Rodriguez E, Lutz A,
Martinerie J, Varela FJ. Comparison of Hilbert transform and
wavelet methods for the analysis of neuronal synchrony. J
Neurosci Methods, 2001; 111: 83-98.
Lee TW, Girolami M, Sejnowski TJ. Independent component
analysis using an extended infomax algorithm for mixed
subgaussian and supergaussian sources. Neural Comput., 1999;
11: 417-41.
Lee TW, Girolami M, Bell AJ, Sejnowski TJ. A Unifying
Information-theoretic Framework for Independent Component
Analysis. Comput. Math. Appl., 2000; 31: 1-21.
Makeig S. Auditory event-related dynamics of the EEG spectrum and
effects of exposure to tones. Electroencephalogr. Clin.
Neurophysiol., 1993; 86: 283-93.
Makeig S. Response: Event-related brain dynamics - unifying brain
electrophysiology. Trends Neurosci., 2002; 25: 390.
Makeig S, Bell AJ, Jung TP, Sejnowski TJ. Independent component
analysis of electroencephalographic data. In Touretzky D, Mozer
M, Hasselmo M, editors. Adv. Neural Inf. Process. Syst., 1996; 8:
145-51.
Makeig S, Jung TP, Bell AJ, Ghahremani D, Sejnowski TJ. Blind
separation of auditory event-related brain responses into
independent components. Proc. Natl. Acad. Sci. U.S.A., 1997; 94:
10979-84.
Makeig S, Westerfield M, Jung TP, Covington J, Townsend J,
Sejnowski TJ, Courchesne E. Functionally independent
components of the late positive event-related potential during
visual spatial attention. J. Neurosci., 1999; 19: 2665-80.
Makeig S, Westerfield M, Jung TP, Enghoff S, Townsend J,
Courchesne E, Sejnowski TJ. Dynamic brain sources of visual
evoked responses. Science, 2002; 295: 690-4.
Makeig S et al. Matlab Toolbox for analysis of electrophysiological
data. http://www.cnl.salk.edu/~scott/ica.html, 1997.
Neuenschwander S, Varela FJ. Visually triggered neuronal
oscillations in the pigeon: an autocorrelation study of tectal
activity. Eur. J. Neurosci., 1993; 5: 870-81.
Park H-M, Jung H-Y, Lee T-W, Lee S-Y. On Subband-Based Blind
Signal Separation for Noisy Speech Recognition. Electronic
Letters, 1999; 35: 2011-2.
Pauluis Q, Baker SN, Olivier E. Emergent oscillations in a realistic
network: the role of inhibition and the effect of the spatiotemporal
distribution of the input. J. Comput. Neurosci., 1999; 6: 27-48.
Pfurtscheller G, Aranibar A. Evaluation of event-related
desynchronization (ERD) preceding and following voluntary self-
paced movement.PG - 138-46. Electroencephalogr. Clin.
Neurophysiol., 1979; 46.
Ramberg JS, Dudewicz EJ, PTadikalama PR, Mykytka EF. A
probability distribution and its uses in fitting data. Technometrics,
1979; 21.
Rodriguez E, George N, Lachaux JP, Martinerie J, Renault B, Varela
FJ. Perception's shadow: long-distance synchronization of human
brain activity. Nature, 1999; 397: 430-3.
Tallon-Baudry C, Bertrand O, Delpuech C, Pernier J. Stimulus
specificity of phase-locked and non-phase-locked 40 Hz visual
responses in human. J. Neurosci., 1996; 16: 4240-9.
Thompson DJ. Spectrum estimation and harmonic analysis. IEEE
proc., 1982; 70: 1055-96.
von Stein A, Sarnthein J. Different frequencies for different scales of
cortical integration: from local gamma to long range alpha/theta
synchronization. Int J Psychophysiol, 2000; 38: 301-13.
Weiss S, Rappelsberger P. EEG coherence within the 13-18 Hz band
as a correlate of a distinct lexical organisation of concrete and
abstract nouns in humans. Neurosci. Lett., 1996; 209: 17-20.
... All electrode impedances were kept below 10kΩ. EEG signals were preprocessed using the open-source toolbox EEGLAB [23], running in the MATLAB environment (MathWorks, USA). Continuous EEG data were filtered with a 0.1-60 Hz band-pass filter, and a 49-51 Hz notch filter was applied to remove 50 Hz powerline interference. ...
... Specifically, the events of interest were the warning tones for Type-1 trials and the countdown challenges for Type-2 trials. All epochs were detrended to remove polynomial trends, and those contaminated with eyeblinks, movements, or other artifacts were corrected using an independent component analysis algorithm [23]. ...
... Spectrograms at each electrode were averaged across all trials at the single-subject level and then averaged across subjects within each group. The spectrum was divided into the following bands: δ (1-3 Hz), θ (4-7 Hz), α (8-13 Hz), β (14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30), and γ (31-60 Hz) [24]. To investigate the prestimulus attentional processes and movement preparation [25][26][27][28], the PSD for each frequency band was calculated as the average within respective band and compared between the two groups in regions of interest, that is, δ-and θ-bands at the frontal cortex (Fz electrode); α-, β-and γ-bands at the sensorimotor cortex (Cz electrode) and parietal cortex (Pz electrode), and α-and γ-bands at the occipital cortex (Oz electrode). ...
Article
Full-text available
Background Patients with migraine often experience not only headache pain but also cognitive dysfunction, particularly in attention, which is frequently overlooked in both diagnosis and treatment. The influence of these attentional deficits on the pain-related clinical characteristics of migraine remains poorly understood, and clarifying this relationship could improve care strategies. Methods This study included 52 patients with migraine and 34 healthy controls. We employed the Attentional Network Test for Interactions and Vigilance–Executive and Arousal Components paradigm, combined with electroencephalography, to assess attentional deficits in patients with migraine, with an emphasis on phasic alerting, orienting, executive control, executive vigilance, and arousal vigilance. An extreme gradient boosting binary classifier was trained on features showing group differences to distinguish patients with migraine from healthy controls. Moreover, an extreme gradient boosting regression model was developed to predict clinical characteristics of patients with migraine using their attentional deficit features. Results For general performance, patients with migraine presented a larger inverse efficiency score, a higher prestimulus beta-band power spectral density and a lower gamma-band event-related synchronization at Cz electrode, and stronger high alpha-band activity at the primary visual cortex, compared to healthy controls. Although no behavior differences in three basic attentional networks were found, patients showed magnified N1 amplitude and prolonged latency of P2 for phasic alerting-trials as well as an increased orienting evoked-P1 amplitude. For vigilance function, improvements in the hit rate of executive vigilance-trials were exhibited in controls but not in patients. Besides, patients with migraine exhibited longer reaction time as well as larger variability in arousal vigilance-trials than controls. The binary classifier developed by such attentional deficit features achieved an F1 score of 0.762 and an accuracy of 0.779 in distinguishing patients with migraine from healthy controls. Crucially, the predicted value available from the regression model involving attentional deficit features significantly correlated with the real value for the frequency of headache. Conclusions Patients with migraine demonstrated significant attentional deficits, which can be used to differentiate migraine patients from healthy populations and to predict clinical characteristics. These findings highlight the need to address cognitive dysfunction, particularly attentional deficits, in the clinical management of migraine.
... Following preprocessing, sleep stages were scored by an expert polysomnographic technologist following established guidelines 73 . The classification was performed using the EEGLABcompatible 74 Counting Sheep PSG toolbox (https://github.com/ stuartfogel/CountingSheepPSG) 75 . ...
... Slow waves were automatically detected from Fz, Cz, Pz, C3, and C4 during movement artifact-free NREM sleep (N2 and SWS) via a period amplitude analysis detection algorithm (https://github. com/stuartfogel/Period-Amplitude-Analysis) based on established methods 76 , adapted for EEGLAB 74 and written for MATLAB R2019b (The MathWorks Inc.). First, the EEG signal was band-pass filtered (32nd-order Chebyshev Type 2 low-pass filter, 80 dB stopband attenuation, 2.15 Hz frequency cut-off; 64th-order Chebyshev type 2 high-pass filter, 80 dB stopband attenuation, 0.46 Hz frequency cut-off). ...
Article
Full-text available
Sleep is essential for the optimal consolidation of newly acquired memories. This study examines the neurophysiological processes underlying memory consolidation during sleep, via reactivation. Here, we investigated the impact of slow wave - spindle (SW-SP) coupling on regionally-task-specific brain reactivations following motor sequence learning. Utilizing simultaneous EEG-fMRI during sleep, our findings revealed that memory reactivation occured time-locked to coupled SW-SP complexes, and specifically in areas critical for motor sequence learning. Notably, these reactivations were confined to the hemisphere actively involved in learning the task. This regional specificity highlights a precise and targeted neural mechanism, underscoring the crucial role of SW-SP coupling. In addition, we observed double-dissociation whereby primary sensory areas were recruited time-locked to uncoupled spindles; suggesting a role for uncoupled spindles in sleep maintenance. These findings advance our understanding the functional significance of SW-SP coupling for enhancing memory in a regionally-specific manner, that is functionally dissociable from uncoupled spindles.
... To provide a reference point for the start of the movement, the EMG electrodes were placed on the biceps and triceps of the upper arms to record EMG activity. EEG data was processed offline using EEGLAB [43] and generally followed past preprocessing pipeline [43]. First, all the signals were digitally filtered to 0.5-40 Hz and line noise was removed using CleanLine (cleanline function in EEGLAB, Computational Neuroscience Laboratory, Salk Institute, California). ...
... To provide a reference point for the start of the movement, the EMG electrodes were placed on the biceps and triceps of the upper arms to record EMG activity. EEG data was processed offline using EEGLAB [43] and generally followed past preprocessing pipeline [43]. First, all the signals were digitally filtered to 0.5-40 Hz and line noise was removed using CleanLine (cleanline function in EEGLAB, Computational Neuroscience Laboratory, Salk Institute, California). ...
Article
Full-text available
Stroke remains a leading cause of adult disability, underscoring why research continues to focus on advancing new treatment methods and neurophysiological indexes. While these studies may be effective, many lack a clear theoretical framework. The current study first determined the optimal combination effects of mirror therapy (MT) with transcranial direct current stimulation (tDCS) on the premotor or primary motor cortex on its short-term and sustained clinical outcomes. We then introduced electroencephalogram (EEG) indexes derived from the gating-by-inhibition model to explore the underlying therapeutic mechanisms. The EEG indexes used in this study focused on the functional involvement for motor generation: alpha power at temporal regions (inhibiting non-motor activity) and central-frontal regions (releasing motor regions from inhibition). Results showed that post-training benefits, measured by Fugl-Meyer Assessment (FMA), were similar across 3 tDCS interventions (premotor, primary motor, sham). EEG seemed more sensitive to the training, with notable responses in the premotor tDCS group. Three months after training, only the premotor tDCS group maintained the gains in FMA, with these improvements correlated with the EEG indexes. Again, this pattern was specific to premotor tDCS. Since the gating-by-inhibition model suggests that EEG index reflects an individual’s psychomotor efficiency, we also found that the baseline EEG index could predict FMA retention. Our findings demonstrate the superiority of combined premotor tDCS with MT and identify functionally oscillatory alpha-band activity in the temporal and central-frontal regions as potentially underlying the therapeutic mechanism. An individual’s spatial pattern of EEG may be effective in predicting upper extremity retention effect.
... The EEG data were sampled at 500 Hz without the application of any high or low pass filters. EEG data was analyzed offline using MATLAB with the EEGLAB [9] and FIELDTRIP [10] toolboxes. The data underwent baseline correction from 400 to 200 ms prior to stimulus onset, with the sampling rate reduced from 500 Hz to 250 Hz. ...
... Epochs with extreme voltages over 150 microvolts were automatically excluded, with further manual inspection for any remaining artifacts. Finally, time-frequency analysis was conducted using Morlet wavelets [9]. ...
Preprint
Sleep is crucial for episodic memory consolidation, yet the impact of habitual sleep quality on memory performance remains underexplored. This study investigates the relationship between sleep quality and episodic memory retrieval using EEG-based representational similarity analysis (RSA). Thirty-six participants wore wrist accelerometers for one week to capture habitual sleep patterns, including total sleep time and restlessness. Memory performance was assessed through a paired associate learning task, with EEG data recorded during encoding and retrieval phases. RSA was applied to EEG oscillatory power across time-frequency windows to examine the neural similarity between encoding and retrieval. The results showed both positive and negative correlations between sleep metric and memory performance, with sleep restlessness being linked to both increases and decreases in neural similarity across specific clusters. These findings emphasize the important role of sleep quality in shaping the neural processes underlying episodic memory retrieval, indicating a strong connection between sleep patterns and memory function.
... EEG and breath signals are preprocessed and analyzed in MATLAB using EEGLAB [18] and custom written functions. ...
Article
The interplay between the brain and lungs involves intricate physiological mechanisms operating bidirectionally. We study the neural dynamics during slow-symmetric breathing (SSB) at different rates, with and without incorporating breath-holds. Electroencephalogram (EEG) is recorded from 63 healthy adults while breathing at 10, 6, and 4 cycles per minute (cpm) guided by a visual metronome. Synchronization between breathing and very low frequency EEG is studied using coherence. A significant increase in coherence is observed during SSB at all breathing rates and resonance at 6 cpm. We also study EEG band power changes, coupling of the breathing phase with EEG amplitude, and EEG modulation during SSB. The power in alpha and beta bands is significantly higher during breath-hold than during inhalation and exhalation. Compared to the baseline, the modulation index increases significantly during SSB, illustrating the enhanced modulation of brain activity. Interestingly, coherence, phase-amplitude coupling, and modulation index depend mainly on the breathing rate, and the differences between symmetric breathing with and without breath-hold are not statistically significant. The observed commonalities in neural dynamics during different breathing patterns suggest the possibility of achieving very low breathing rates by symmetric breathing with hold, while retaining the reported beneficial effects of slow breathing.
... EEG preprocessing was performed using the EEGLAB toolbox [42]. The EEG signals were filtered through a bandpass filter between 1 Hz and 32 Hz, consistent with most previous AAD tasks [16], [27], [41]. ...
Preprint
Full-text available
Decoding the directional focus of an attended speaker from listeners' electroencephalogram (EEG) signals is essential for developing brain-computer interfaces to improve the quality of life for individuals with hearing impairment. Previous works have concentrated on binary directional focus decoding, i.e., determining whether the attended speaker is on the left or right side of the listener. However, a more precise decoding of the exact direction of the attended speaker is necessary for effective speech processing. Additionally, audio spatial information has not been effectively leveraged, resulting in suboptimal decoding results. In this paper, we observe that, on our recently presented dataset with 15-class directional focus, models relying exclusively on EEG inputs exhibits significantly lower accuracy when decoding the directional focus in both leave-one-subject-out and leave-one-trial-out scenarios. By integrating audio spatial spectra with EEG features, the decoding accuracy can be effectively improved. We employ the CNN, LSM-CNN, and EEG-Deformer models to decode the directional focus from listeners' EEG signals with the auxiliary audio spatial spectra. The proposed Sp-Aux-Deformer model achieves notable 15-class decoding accuracies of 57.48% and 61.83% in leave-one-subject-out and leave-one-trial-out scenarios, respectively.
... All the datasets provided in this paper are raw data. To process the data, the authors recommend using MATLAB R2021a (MathWorks, MA) 14 and/or the MATLAB toolbox EEGLAB 15 for processing and analysis of the multimodal data. De-noising algorithms have been proposed for identifying and removing physiological and non-physiological artifacts, including motion artifacts from the EEG data [16][17][18] . ...
Article
Full-text available
This report contains a description of physiological and motion data, recorded simultaneously and in synchrony using the hyperscanning method from two professional dancers using wireless mobile brain-body imaging (MoBI) technology during rehearsals and public performances of “LiveWire” - a new composition comprised of five choreographed music and dance sections inspired by neuroscience principles. Brain and ocular activity were measured using 28-channel scalp electroencephalography (EEG), and 4-channel electrooculography (EOG), respectively; and head motion was recorded using an inertial measurement unit (IMU) placed on the forehead of each dancer. Video recordings were obtained for each session to allow for tagging of physiological and motion signals and for behavioral analysis. Data recordings were collected from 10 sessions over a 4-month period, in which the dancers rehearsed or performed (in front of an audience) choreographed expressive movements. A detailed explanation of the experimental set-up, the steps carried out for data collection, and an explanation on the usage are provided in this report.
Article
Full-text available
Background/Objectives: Musical pleasure is considered to be induced by prediction errors (surprise), as suggested in neuroimaging studies. However, the role of temporal changes in musical features in reward processing remains unclear. Utilizing the Information Dynamics of Music (IDyOM) model, a statistical model that calculates musical surprise based on prediction errors in melody and harmony, we investigated whether brain activities associated with musical pleasure, particularly in the θ, β, and γ bands, are induced by prediction errors, similar to those observed during monetary rewards. Methods: We used the IDyOM model to calculate the information content (IC) of surprise for melody and harmony in 70 musical pieces across six genres; eight pieces with varying IC values were selected. Electroencephalographic data were recorded during listening to the pieces, continuously evaluating the participants’ subjective pleasure on a 1–4 scale. Time–frequency analysis of electroencephalographic data was conducted, followed by general linear model analysis to fit the power-value time course in each frequency band to the time courses of subjective pleasure and IC for melody and harmony. Results: Significant positive fits were observed in the β and γ bands in the frontal region with both subjective pleasure and IC for melody and harmony. No significant fit was observed in the θ band. Both subjective pleasure and IC are associated with increased β and γ band power in the frontal regions. Conclusions: β and γ oscillatory activities in the frontal regions are strongly associated with musical rewards induced by prediction errors, similar to brain activity observed during monetary rewards.
Article
Full-text available
Brainstorm is a collaborative open-source application dedicated to magnetoencephalography (MEG) and electroencephalography (EEG) data visualization and processing, with an emphasis on cortical source estimation techniques and their integration with anatomical magnetic resonance imaging (MRI) data. The primary objective of the software is to connect MEG/EEG neuroscience investigators with both the best-established and cutting-edge methods through a simple and intuitive graphical user interface (GUI).
Article
Full-text available
Most current analysis methods for functional magnetic resonance imaging (fMRI) data assume a priori knowledge of the time course of the hemodynamic response (HR) to experimental stimuli or events in brain areas of interest. In addition, they typically assume homogeneity of both the HR and the non-HR "noise" signals both across brain regions and across similar experimental events. When HRs vary unpredictably from area to area, or from trial to trial, different approaches are needed. Here we used infomax Independent Component Analysis (ICA) to detect and visualize variations in single-trial HRs in event-related fMRI data. ICA decomposition of the resulting BOLD data produced independent components with variable stimulus-locked HRs active in primary visual (V1) and medial temporal (MT/V5) cortices respectively. Contrary to expectation, in four of six subjects the HR of the V1 component contained two positive peaks in response to short-stimulus bursts, while nearly identical component maps were associated with single-peaked HRs in long-stimulus sessions from the same subject. Thus, ICA combined with single-trial visualization can reveal dramatic and unforeseen task-related HR variation not apparent to researchers analyzing the data with fixed HR templates.
Article
Full-text available
A four-parameter probability distribution, which includes a wide variety of curve shapes, is presented. Because of the flexibility, generality, and simplicity of the distribution, it is useful in the representation of data when the underlying model is unknown. A table based on the first four moments, which simplifies parameter estimation, is given. Further important applications of the distribution include the modeling and subsequent generation of random variates for simulation studies and Monte Carlo sampling studies of the robustness of statistical procedures.
Article
Full-text available
Independent component analysis (ICA) has proven useful for modeling brain and electroencephalographic (EEG) data. Here, we present a new, generalized method to better capture the dynamics of brain signals than previous ICA algorithms. We regard EEG sources as eliciting spatio-temporal activity patterns, corresponding to, e.g. trajectories of activation propagating across cortex. This leads to a model of convolutive signal superposition, in contrast with the commonly used instantaneous mixing model. In the frequency-domain, convolutive mixing is equivalent to multiplicative mixing of complex signal sources within distinct spectral bands. We decompose the recorded spectral- domain signals into independent components by a complex infomax ICA algorithm. First results from a visual attention EEG experiment exhibit: (1) sources of spatio-temporal dynamics in the data, (2) links to subject behavior, (3) sources with a limited spectral extent, and (4) a higher degree of independence compared to sources derived by standard ICA. q 2003 Elsevier Ltd. All rights reserved.
Article
An extension of the infomax algorithm of Bell and Sejnowski (1995) is presented that is able blindly to separate mixed signals with sub- and supergaussian source distributions. This was achieved by using a simple type of learning rule first derived by Girolami (1997) by choosing negentropy as a projection pursuit index. Parameterized probability distributions that have sub- and supergaussian regimes were used to derive a general learning rule that preserves the simple architecture proposed by Bell and Sejnowski (1995), is optimized using the natural gradient by Amari (1998), and uses the stability analysis of Cardoso and Laheld (1996) to switch between sub- and supergaussian regimes. We demonstrate that the extended infomax algorithm is able to separate 20 sources with a variety of source distributions easily. Applied to high-dimensional data from electroencephalographic recordings, it is effective at separating artifacts such as eye blinks and line noise from weaker electrical signals that arise from sources in the brain.
Article
Cortical activity and perception are not driven by the external stimulus alone; rather sensory information has to be integrated with various other internal constraints such as expectations, recent memories, planned actions, etc. The question is how large scale integration over many remote and size-varying processes might be performed by the brain. We have conducted a series of EEG recordings during processes thought to involve neuronal assemblies of varying complexity. While local synchronization during visual processing evolved in the gamma frequency range, synchronization between neighboring temporal and parietal cortex during multimodal semantic processing evolved in a lower, the beta1 (12-18 Hz) frequency range, and long range fronto-parietal interactions during working memory retention and mental imagery evolved in the theta and alpha (4-8 Hz, 8-12 Hz) frequency range. Thus, a relationship seems to exist between the extent of functional integration and the synchronization-frequency. In particular, long-range interactions in the alpha and theta ranges seem specifically involved in processing of internal mental context, i.e. for top-down processing. We propose that large scale integration is performed by synchronization among neurons and neuronal assemblies evolving in different frequency ranges.
Article
We show that different theories recently proposed for independent component analysis (ICA) lead to the same iterative learning algorithm for blind separation of mixed independent sources. We review those theories and suggest that information theory can be used to unify several lines of research. Pearlmutter and Parra [1] and Cardoso [2] showed that the infomax approach of Bell and Sejnowski [3] and the maximum likelihood estimation approach are equivalent. We show that negentropy maximization also has equivalent properties, and therefore, all three approaches yield the same learning rule for a fixed nonlinearity. Girolami and Fyfe [4] have shown that the nonlinear principal component analysis (PCA) algorithm of Karhunen and Joutsensalo [5] and Oja [6] can also be viewed from information-theoretic principles since it minimizes the sum of squares of the fourth-order marginal cumulants, and therefore, approximately minimizes the mutual information [7]. Lambert [8] has proposed different Bussgang cost functions for multichannel blind deconvolution. We show how the Bussgang property relates to the infomax principle. Finally, we discuss convergence and stability as well as future research issues in blind source separation.
Article
This article presents, for the first time, a practical method for the direct quantification of frequency-specific synchronization (i.e., transient phase-locking) between two neuroelectric signals. The motivation for its development is to be able to examine the role of neural synchronies as a putative mechanism for long-range neural integration during cognitive tasks. The method, called phase-locking statistics (PLS), measures the significance of the phase covariance between two signals with a reasonable time-resolution (<100 ms). Unlike the more traditional method of spectral coherence, PLS separates the phase and amplitude components and can be directly interpreted in the framework of neural integration. To validate synchrony values against background fluctuations, PLS uses surrogate data and thus makes no a priori assumptions on the nature of the experimental data. We also apply PLS to investigate intracortical recordings from an epileptic patient performing a visual discrimination task. We find large-scale synchronies in the gamma band (45 Hz), e.g., between hippocampus and frontal gyrus, and local synchronies, within a limbic region, a few cm apart. We argue that whereas long-scale effects do reflect cognitive processing, short-scale synchronies are likely to be due to volume conduction. We discuss ways to separate such conduction effects from true signal synchrony.
Article
In this study, a linear decomposition technique, independent component analysis (ICA), is applied to single-trial multichannel EEG data from event-related potential (ERP) experiments. Spatial filters derived by ICA blindly separate the input data into a sum of temporally independent and spatially fixed components arising from distinct or overlapping brain or extra-brain sources. Both the data and their decomposition are displayed using a new visualization tool, the “ERP image,” that can clearly characterize single-trial variations in the amplitudes and latencies of evoked responses, particularly when sorted by a relevant behavioral or physiological variable. These tools were used to analyze data from a visual selective attention experiment on 28 control subjects plus 22 neurological patients whose EEG records were heavily contaminated with blink and other eye-movement artifacts. Results show that ICA can separate artifactual, stimulus-locked, response-locked, and non-event-related background EEG activities into separate components, a taxonomy not obtained from conventional signal averaging approaches. This method allows: (1) removal of pervasive artifacts of all types from single-trial EEG records, (2) identification and segregation of stimulus- and response-locked EEG components, (3) examination of differences in single-trial responses, and (4) separation of temporally distinct but spatially overlapping EEG oscillatory activities with distinct relationships to task events. The proposed methods also allow the interaction between ERPs and the ongoing EEG to be investigated directly. We studied the between-subject component stability of ICA decomposition of single-trial EEG epochs by clustering components with similar scalp maps and activation power spectra. Components accounting for blinks, eye movements, temporal muscle activity, event-related potentials, and event-modulated alpha activities were largely replicated across subjects. Applying ICA and ERP image visualization to the analysis of sets of single trials from event-related EEG (or MEG) experiments can increase the information available from ERP (or ERF) data. Hum. Brain Mapping 14:166–185, 2001. © 2001 Wiley-Liss, Inc.