ChapterPDF Available

Electroencephalography (EEG): neurophysics, experimental methods, and signal processing



Content may be subject to copyright.
APA Citation (BibTeX last page):
Nunez, M. D., Nunez, P. L., & Srinivasan, R. (2016). Electroencephalography (EEG):
neurophysics, experimental methods, and signal processing. In Ombao, H., Linquist, M.,
Thompson, W. & Aston, J. (Eds.) Handbook of Neuroimaging Data Analysis (pp. 175-
197), Chapman & Hall/CRC. Advance online publication. doi:
Electroencephalography (EEG): neurophysics, experimental methods,
and signal processing
Michael D. Nunez1, Paul L. Nunez2, and Ramesh Srinivasan1,3
Department of Cognitive Sciences, University of California, Irvine1
Cognitive Dissonance, LLC2
Department of Biomedical Engineering, University of California, Irvine3
Address for Correspondence
Ramesh Srinivasan,
Acknowledgments: This work was supported by a grant from the NIH 2R01-MH68004
1. Introduction
Electroencephalography (EEG) is the measurement of the electric potentials on
the scalp surface generated (in part) by neural activity originating from the brain. The
sensitivity of EEG to changes in brain activity on such a millisecond time scale is the
major advantage of EEG over other brain imaging modalities such as functional magnetic
resonance imaging (fMRI) or near-infrared spectroscopy (NIRS) that operate on time
scales in the seconds to minutes range. Over the past 100 years, neuroscientists and
clinical neurologists have made use of EEG to obtain insight into cognitive or clinical
disease state by applying a variety of signal processing and statistical analyses to EEG
time series. More recently there has been growing interest in making use of statistical
modeling of EEG signals to directly control physical devices in Brain-Computer
Interfaces. In this chapter we provide an introduction to EEG generation and
measurement as well as the experimental designs that optimize information acquired by
EEG has statistical properties that vary over time and space. That is, we assume
the data recorded are observations of a spatio-temporal stochastic process. The starting
point of most EEG analysis is to consider the properties of the time series at each
electrode, such as spectral power, in relation to sensory stimulation, cognitive processing,
or clinical disease state. EEG power is typically split up into bands which correspond to
different spectral peaks that relate to behavior or cognitive state. These bands are
typically defined as the delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), beta (13-20 Hz),
and gamma (>20 Hz) bands and can have high-power over spatially distinct regions on
the scalp, as shown by an EEG recording of a subject at rest in Figure 1. Because this
EEG was recorded using a high-density, 128 electrode net, the topographic scalp maps
could be found by interpolating values between electrodes. Modern EEG systems
typically use a large number of electrodes (ranging from 64 to 256) to provide coverage
over most of the scalp, enabling the analyses of spatial properties of EEG, such as
correlation or coherence between electrode sites.
In this chapter we first review physical properties of EEG recordings, in order to
model the relationship between potentials on the scalp and current sources in the brain.
Models of volume conduction of current passing through the head imply that EEG signals
are strongly influenced by the synchronization of neural current sources; EEG is as much
a measure of neural synchrony as neural activity. We then introduce issues in EEG
recording and preprocessing, with particular emphasis on the problem of artifacts. The
last two sections consider the types of EEG analyses appropriate for different
experimental designs.
2. The neurophysics of EEG
Scalp potentials are believed generated by millisecond-scale modulations of synaptic
current sources at neuron surfaces (Lopes da Silva and Storm van Leeuwen 1978; Nunez
1981, 1995; 2000a,b; Lopes da Silva 1999), while single neuron firings of action
potentials are mainly absent in scalp activity due to other, inactive neurons contributing
to low-pass temporal filtering (Nunez and Srinivasan, 2006; Buzsáki, 2006). Action
potential time scales are typically on the order of less than 1 ms, while synaptic potentials
occur on the order of 10 ms or morea time scale more consistent with the oscillations
observed in EEG. Any patch of cortex 3-5 mm in diameter and traversing all cortical
layers contains ~ 106 neurons and perhaps 1010 synapses (Nunez, 1995). Excitatory and
inhibitory synapses inject current into the cell bodies and induce extracellular potentials
(via return currents) of opposite polarity. These extracellular currents yield the electrical
potentials recorded on the scalp with EEG.
When measurement of the potential is taken at “large distance away from the
source region, a complex current distribution in a small volume can be approximated by a
dipole or more accurately, a dipole moment per unit volume (Nunez and Srinivasan,
2006). The (current) dipole moment per unit volume is an intermediate scale vector
function based on the distribution of positive and negative micro-current sources in each
local tissue mass, typically applied to cortical columns. The dipole approximation to
cortical current sources provides a basis for realistic models of EEG signals. A “large”
distance in this case is at least 3 or 4 times the distance between the effective poles of the
dipole. In the context of EEG recording, the dipole approximation appears valid for
potentials in superficial cortical tissue with a maximum extent in any dimension of
roughly 0.5 cm or less. This is because superficial gyral surfaces are located at roughly
1.5-2 cm from scalp electrodes, separated by a thin layer of passive tissue, including
cerebrospinal fluid (CSF), skull, and scalp.
Thus, the current sources in the brain that generate EEG can be modeled in terms
of dipole moment per unit volume P(r', t) for any time t where the location vector r'
spans the volume of the brain. For convenience of this discussion, the brain volume may
be parceled into N small tissue masses of volume V (e.g., 3 mm x 3mm x 3mm), each
producing its vector dipole moment, such that, in an adult brain, N ~ 105 to 106. The
strength and orientation of each vector depends on the distribution and synchrony of
excitatory and inhibitory post-synaptic potentials within the tissue mass (Nunez and
Srinivasan, 2006). The potential on the scalp surface can then be expressed as a weighted
sum (or integral) of contributions from all these sources. In most models, each volume
element V(r') is located only within the superficial cortex since current sources in deeper
tissues, such as the thalamus or midbrain typically contribute very little to scalp potentials
(Nunez and Srinivasan, 2006). Thus, the volume integral may be reduced to a surface
integral over the folded cortical surface
( , ) ( , ) ( , ) ( )
t G t dV
r r r' P r' r'
The weighting term GH is the Green’s function for volume conduction of current passing
through the tissues of the head. It depends on both the location of source r' and the
location of the scalp electrode r. The Green’s function can be thought of as the impulse
response function between sources and surface locations and contains all geometric and
conductive information about the head as a volume conductor. GH will be larger for
superficial sources in the visible gyral crowns of the cortex than for deeper sources, such
as in the sulcal walls (folded surfaces) or sources on the mesial (underside) of the brain.
Any model of head volume conduction, i.e., any form of the function GH, is only
an approximation. Magnetic Resonance Imaging (MRI) can provide geometric
information by imaging the boundaries between tissue compartments with different
electrical conductivity (the inverse of resistivity). Numerical methods such as the
Boundary Element Method (BEM) or Finite Element Methods (FEM) may then be used
to estimate GH, by employing MRI to determine tissue boundaries as shown in the
examples in Figure 2 (b and c). However, the geometric model obtained from MRI is
still only approximate due to limits in spatial resolution (typically 2-5 mm). And even if
we were able to obtain perfect geometric information, the head model would still only be
approximate due to the substantial uncertainty in our knowledge of tissue conductivities
(Nunez and Srinivasan, 2006).
The poor conductivity of the skull is the feature that most strongly determines
volume conduction in the head. Estimates of the conductivity of the skull vary widely
depending on whether the estimate is in-vivo or in-vitro and differ between skull samples
from different regions of the head. Skull itself is composed of three layers of different
conductivity which vary in thickness across the head (Nunez and Srinivasan, 2006) and
are not easily measured with MRI (Srinivasan, 2006).
Despite these uncertainties in the geometry and conductivity of the head, the gross
features of volume conduction are captured by any model that includes a poorly
conducting skull layer in between conductive soft tissue. A simple estimate of GH can be
provided by models which consist of three concentric spherical shells (brain, skull, and
scalp) or four shells when including a CSF layer as shown in Figure 2a, such that these
models have been useful in a number of simulation studies (Nunez et al., 1994;
Srinivasan et al., 1998; Nunez and Srinivasan, 2006). These models also have the
advantage of easy checking of computational accuracy as analytic solutions for these
models have been obtained (Nunez and Srinivasan, 2006). The following gross features
of head volume conduction are captured by this model: (1) the poor conductivity of the
skull results in very little current entering the skull from the brain. (2) Current is expected
to mostly flow radially through the skull into the scalp as current follows the path of least
resistance. Exceptions are holes in the skull like the nasal passages. (3) All of the current
is contained in the scalp, as no current can enter the surrounding air. (4) Very little
current is expected to enter the body because of the high resistance of the neck, such that
the head can be considered a closed object to first approximation. In all models that
contain these essential features, the tangential spread of current within the scalp leads to
the “smearing” of the scalp potential, i.e., low-pass spatial filtering, resulting in the low
spatial resolution of EEG as compared to direct recordings on the brain surface. Thus this
model captures the fact that EEG is a direct measurement of the current flowing in the
There is considerable interest in the EEG literature in developing methods to
estimate the current source distribution in the brain P(r', t) from the EEG recording and a
volume conduction model of the head. However, this inverse problem is ill-posed; given
the potential distribution on the surface of the scalp, it is not possible to estimate the
source distribution without additional assumptions (Nunez and Srinivasan, 2006). That is,
for any given GH estimate and true scalp potential ΦS(r, t) there are a large number of
solutions for P(r', t). Although in some cases, for example an epileptic focus in the
cortex, it may be reasonable to assume a single isolated source P(r', t) in order to find a
solution. Another popular approach that is widely adopted is to use Tikhanov
regularization to obtain a minimum L2 norm estimate of P(r', t) (Hauk, 2004). While
this approach is mathematically tractable, there is no apparent theoretical reason why
neuroscientists should seek solutions with a minimum L2 norm. More recently, methods
based on Bayesian inference have been developed which have the advantage of making
assumptions explicit. These methods allow for the possibility of model validation and
allow for comparisons between models based on different assumptions (Baillet and
Garnero, 1997; Wipf and Nagarajan, 2009). They also allow for the use of prior
information, for example by making use of fMRI information to influence the source
solutions (Henson et al., 2010).
3. Synchronization and EEG
The magnitude of the scalp potential recorded with EEG can change for several
reasons related to source synchronization. The large changes in scalp amplitude that
occur when brain state changes are believed to be due mostly to distributed
synchronization changes. That is, large-scale synchronization increases (or decreases)
over cm scales in the tangential direction across the cortex will cause increases (or
decreases) in scalp potential if there are no other changes. With this knowledge, EEG
scientists and clinicians have adopted the label desynchronization to indicate large
amplitude reductions (Pfurtscheller and Lopes da Silva 1999). Although, at any one
location r' in the brain, small-scale changes in synaptic source synchronization will also
change the magnitude (or source strength) of P(r', t) (Nunez and Srinivasan, 2006), we
will concentrate in this section on large-scale synchronization that does not change the
magnitude of each individual dipole moment P(r', t). Instead, we will show that large-
scale changes across different locations r' cause the scalp potential in Eq. 1 to decrease;
this is because the integral approaches zero as more random positive and negative dipole
moments P(r', t) at different cortical locations r' cancel. However if multiple locations r'
have similar dipole moments per unit volume P(r', t) then synchronization occurs which
leads to larger observed scalp potentials.
We expect that the source of any EEG signal will never simply correspond to
source activity in only one of the volume elements. Because neurons are highly
interconnected, most EEG signals are generated by sources with spatial extent, i.e.,
patches of cortical tissue. Figure 3 shows examples of scalp potentials simulated in a
concentric spheres volume conduction model due to a single dipole (a; corresponding to a
patch of diameter < 3 mm) as well as dipole layers of diameter ranging from 3-5 cm (Fig.
3c, 3e, & 3g). For simplicity, we only make use of radial dipole sources in this example
as similar effects could be found with dipole layers of arbitrary orientation. Each dipole
layer is composed of dipole sources with time series that are constructed by adding a 6
Hz sinusoid of fixed amplitude A=15 to a Gaussian random processes with mean =0 and
standard deviation =150. The 6 Hz components are synchronized across the dipole
layers, whereas all other frequencies will have random phases. Each source signal is an
independent random time series representing the potential across the cortical surface
given by the dipole layer. The source time series of a single dipole source (i.e. the dipole
layer of very small size) is plotted in Figure 3a. The magnitude of the 6 Hz sinusoid is
only 1% of the total variance of each dipole source, and the sinusoid is not observable in
the dipole time series. Figure 3b shows the estimated potential measured at an electrode
on the scalp directly above the center of a dipole layer of diameter 3 cm, based on a four
concentric spheres model of the head (see Figure 3 caption for details of the head model).
The time series exhibits a smoother appearance compared to the source time series. And
as the diameter of the dipole layer is increased from 3 to 4 to 5 cm (i.e. increasing
synchrony), the calculated surface potential becomes more obviously sinusoidal (Fig. 3c
through 3h).
Clearly, spatial synchrony is a least as important as the strength of the source in
the generation of scalp potentials. Most (99%) of the source activity in these examples is
uncorrelated Gaussian noise across the dipoles in a patch of width 5 cm. Yet, the scalp
potential will appear smooth and periodic reflecting mostly relatively small magnitude
(1% of) source activity that is synchronous across all sources in the dipole layer. The
effect of volume conduction is to sum the source activity at the scalp electrode, so the
asynchronous source time series contribute minimally due to noise cancellation. The
synchronized 6 Hz signal is emphasized and the scalp potential is remarkably sensitive to
the size of the dipole layer.
We have previously quantified this effect as spatial filtering by volume
conduction (Srinivasan et al, 1998; Nunez and Srinivasan, 2006; Srinivasan et al, 2007).
One important implication is that spatial filtering by volume conduction can generally be
expected to filter the temporal structure of source activity in the scalp EEG. If sources
with different time series take place in dipole layers of different sizes, EEG favors signals
that are synchronized broadly over the cortical surface. The magnitude of any scalp EEG
signal is determined not only by the source strength but also by spatial properties of the
source such as its size and synchrony. Thus, we anticipate that EEG recorded within the
brain (known as electrocorticography or ECoG) will have quite different properties than
EEG recorded on the scalp. Neither signal is a more accurate representation of brain
activity; instead they emphasize different spatial scales of synchronization in the brain.
4. Recording EEG
Every EEG recording involves at least 3 electrodes, two measurement electrodes
and a ground electrode. Brain sources P(r, t) (current dipole moments per unit volume)
and biological artifacts generate the majority of scalp potential differences V2(t) V1(t).
Environmental electric and magnetic fields also contribute to the measured scalp
potential due mostly to capacitive coupling of body and electrode leads to power line
fields. However, the amplifier ground electrode placed on the scalp, nose, or neck
provides a reference voltage to the amplifier to prevent amplifier drift and facilitate better
common mode rejection by serving as a reference for the differential amplifier (Nunez
and Srinivasan, 2006).
Typically one electrode is singled out as the “reference electrode”; the remaining
electrodes are characterized as “recording” electrodes. But electrode pairs are always
required to measure scalp potentials because such recording depends on current passing
through a measuring circuit (Nunez and Srinivasan, 2006). There are no monopolar
recordings in EEG; all recordings are bipolar. Every EEG recording depends on the
location of both recording and “reference” electrodes. Therefore any particular choice of
reference placement offers possible advantages and disadvantages depending on actual
source locations.
But, in general, we do not know the location of the sources prior to recording
EEG, so no ideal reference location is likely to be found in advance. Reference strategies
have often been adopted in EEG laboratories without a clear understanding of the
attendant biases imposed on the recording. The linked-ears or linked-mastoids reference,
a historically popular reference choice with cognitive scientists, is one such idea with
minimal theoretical justification, but nevertheless persists in a number of laboratories. In
EEG, we generally measure potential differences between two locations on the head, and
these differences depend on both electrode locations, as well as on all brain generator
configurations and locations.
In most EEG practice, the potentials at all the other electrode sites (typically 32-
256) are recorded with respect to the reference electrode. The position of these electrodes
varies considerably across laboratories. Standard electrode placement strategies make
use of the 10-20, 10-10, and 10-5 electrode placement systems (Oostenveld and
Praamstra, 2001). These systems are widely but not universally used. For larger numbers
of channels (> 64), other electrode placement systems have been developed in order to
obtain more uniform sampling of scalp potential, which is advantageous for source
localization and high resolution EEG methods (Tucker, 1993). The reference point is
largely arbitrary; it is special only because we choose to record potential differences with
respect to one fixed location. But we do have the option of changing the effective
reference to another recording site further down the processing chain by simple
The average reference (also called common average reference or global average
reference) has become commonplace in EEG studies and has some theoretical
justification (Bertrand et al. 1985). When recording from N electrodes located at scalp
locations rn, n = 1, 2,…, N , the measured potentials V(rn) are related to the true scalp
(rn) (measured with respect to “infinity”) by
( ) ( ) ( )
n n R
V r r r   
where rn is the position of the nth electrode and rR is the reference electrode site. If we
sum over all N electrodes, the potential with respect to infinity at the reference site can be
written in terms of the scalp potentials as
( ) ( ) ( )
R n n
r r V r
 
The first term on the right side of Eq. 3 is the average of the scalp surface potential at all
recording sites. Theoretically, this term vanishes if the mean of the potentials
approximates a surface integral over a closed surface containing all current within the
volume. Only minimal current flows from the head through the neck even with reference
electrode placed on the body, so a reasonable approximation considers the head to be a
closed volume that confines all current. The surface integral of the potential over a
volume conductor containing dipole sources must be zero as a consequence of current
conservation (Bertrand et al., 1985). If we make this assumption, the reference potential
can be estimated by the second term on the right side of Eq. 3; that is, by averaging the
measured potentials at all electrodes and changing the sign of this average. This reference
potential (i.e. the average across electrodes) can thus be added to each measurement
V(rn), thereby estimating the reference-free potential
(rn) (potential with respect to
“infinity”) at each location rn.
However since we cannot measure the potentials on a closed surface surrounding
the brain, the first term on the right side of Eq. 3 will not generally vanish. The
distribution of potential on the underside of the head (within the neck region) cannot be
measured. Furthermore, the average potential for any group of electrode positions, given
by the second term on the right side of Eq 3, is only an approximation of the surface
integral. For example, the average potential is expected to be a very poor approximation
if applied with the standard 10-20 electrode system with 21 electrodes. As the number of
electrodes increases to 64 or more, the error in the approximation is expected to decrease.
Thus, like any other choice of reference, the average reference provides biased estimates
of reference-independent potentials. Nevertheless, when used in studies with large
numbers of electrodes (say 128 or more), we have found that the average reference
performs reasonably well as an estimate of reference independent potentials (Srinivasan
et al., 1998).
5. Preprocessing EEG
Measured EEG signals have been amplified and filtered by analog circuits to
remove both low and high frequency noise as well as power at frequencies greater than
the Nyquist limit, established by the sampling rate of the analog to digital converter
(ADC). The discrete sampling of continuous signals is a well-characterized problem in
time series acquisition and analysis (Bendat and Piersol, 2001). The central concept is the
Nyquist criterion: fdig > 2fmax where fdig is the digitization rate or sampling rate and fmax is
the highest frequency present in the time series. For instance, if the highest frequency in a
signal is 20 Hz (cycles/sec), a minimum sampling rate of 40 Hz (one sample every 25
ms) is required to record the signal discretely without aliasing. Aliasing is the
misrepresentation of a high-frequency signal as a low-frequency signal because the
sampling rate used during analog-to-digital conversion is lower than the Nyquist limit. If
a time series has been aliased by under-sampling, no digital signal processing method can
undo the aliasing because the necessary information for this procedure has been lost. In
conventional EEG practice, a sampling rate is selected and the aliasing error is avoided
by applying (in hardware) a low-pass filter to the analog signal that eliminates power at
frequencies greater than the maximum frequency determined by the Nyquist limit. The
low-pass filter is typically applied with a cut-off frequency 2.5 times smaller than the
sampling rate. This more restrictive limit, known as the Engineer’s Nyquist criterion,
accounts for the possibility of phase-locking between the sampling and high-frequency
components of the signal (Bendat and Piersol, 2001). The analog signal from each
channel is sampled at perhaps 200 to 1000 times per second, assigned numbers
proportional to instantaneous amplitude (digitized), and converted from ADC units to
volts. These samples can then be stored digitally in conventional EEG practice or further
processed online (e.g., using an FFT) in certain clinical or BCI applications.
The choice of filter settings requires some care. Clearly a low-pass filter must be
set to insure removal of power at the very high frequencies determined by the Nyquist
criterion. However, severe low pass filtering runs the risk of removing obvious muscle
artifact at high frequencies (which would indicate time segments of data that potentially
needs to be discarded), while passing muscle artifact at frequencies overlapping with
EEG that can be easily mistaken for EEG (Fisch, 1999). For example, imagine using an
analog filter to remove most power at frequencies greater than 20 Hz, thereby obtaining a
much cleaner looking signal. However, the remaining signal might well contain
significant muscle artifact in roughly the 15-30 Hz range (beta band), which is much
harder to identify without the information at higher frequencies. Such subtle artifact
could substantially reduce the signal to noise ratio in the beta band. Some EEG systems
have notch filters to remove power line interference (60 Hz in the Americas; 50 Hz in
Europe, Australia, and Asia). However, the presence of power line noise in the recorded
EEG signal is an easy way to detect electrodes that develop high contact impedances (or
come off entirely) during the recording. If EEG processing and analysis is based on FFT
or other spectral analysis methods, the presence of moderate 60 Hz noise will have no
practical effect on results at lower frequencies which contain most of the EEG
6. Artifact removal
A substantial portion of the electrical signals recorded from EEG systems
originate from outside the brain (Nunez and Srinivasan, 2006; Whitham et al., 2007). For
example, in some areas on the head, close to the ears, eyes, and neck, we expect electrical
signals originating in the cortex to have magnitude as much as 200 times lower than
electrical signals from muscle activity (Fitzgibbon et al., 2015). Furthermore, movement
of the head will generate artifacts over a large number of electrodes. Potentials generated
from sources other than cortical activity are dubbed artifact, and a major challenge in
EEG analysis is to detect these signals and remove them from EEG recordings.
Artifact can either be biological in nature, such as muscle activity, or due to
environmental factors such as electric fields caused by the common AC standard and
temporary potential shifts due to movement. Biological artifact is typically caused by
electrical potentials generated by muscle activity. The recording of muscle activity is
known as electromyography (EMG) and typically originates from the eyes, face, and
neck, but also from muscles all over the body (Whitham et al., 2007). Another source of
common artifact is the rhythmic beating of arteries in the temples or neck and potentials
from distant but large muscles in the heart (electrocardiography or EKG). Transient
muscle artifact can be due to head movements, eye blinks, lateral eye-movement, or jaw
clenching; all of which may display different spatio-temporal patterns of potentials on the
In order to better draw inference about brain activity, multiple procedures have
been developed to reduce the contribution of artifact in EEG recordings. Ocular artifact
such as eye blinks and lateral eye movements can be automatically removed using
regression methods (Gratton et al., 1983). In a typical ocular regression method,
electrodes are placed near the eyes to record electrooculographic (EOG) signals,
potentials generated by musculature associated with the eyes. The effect of these EOG
signals on the other EEG channels is then estimated with linear regression. The total
influence of the EOG signals on the EEG is then removed by subtracting the product of
the EOG signals and the regression coefficient estimates (Schlögl et al., 2007).
Independent Component Analysis (ICA; Bell & Sejnowski, 1995) has become an
important tool for identifying and removing artifact. Independent component analysis
(ICA) refers to a class of blind source-separation algorithms used to decompose linear
mixtures of data. For example, some ICA algorithms find linear mixtures of variables that
are maximally non-Gaussian by searching for mixtures with either minimum mutual
information or maximum kurtosis (Makeig et al. 1996; Jung et al., 1998). In practice
these methods often yield non-normal mixtures that have distributions with outliers. The
two most widely used algorithms in the EEG literature are the FastICA (Hyvärinen &
Oja, 1997) and InfoMax ICA (Bell & Sejnowski, 1995; Delorme & Makeig, 2002).
ICA assumes that there is a linear mixture of the EEG data V (a channels c by
time t matrix) such that the independent components M (a component k by time t matrix)
are given by
where W-1 is the matrix consisting of k by c weights. Some of
the resulting components have been shown to well represent some specific types of
artifact (Delorme et al. 2007). The components evaluated to reflect artifact can then be
removed from the data by inverting the equation using a reduced matrix WL to remove the
artifact components.
There are two caveats with this approach. First, the identification of the artifact
component is inherently a subjective judgement. Some artifact sources are easy to
identify such as eye blinks, eye movements, and temporary electrical discontinuities
(perhaps due to a reference electrode or ground electrode displacement during head
movement). But artifacts due to muscle are far more subtle. Second, the effect of
reducing the number of sources in M is to reduce the rank of the data matrix, which
potentially influences further analysis by reducing the amount of possible EEG mixtures.
To perform an ICA based artifact removal procedure, the continuous recording is
first split into 1-3 second epochs, usually based on the trial structure of the experiment in
cognitive experiments. Epochs that obviously contain artifact rather than EEG, usually
due to gross movements by the subject, can then be rejected by visual inspection, or by
examining trials with high variance compared to other trials. Not removing this one-off
data hinders the ability of the ICA algorithm to isolate typical artifacts such as eye blinks.
After this “precleaning” step, an algorithm is run with the EEG data as input to obtain an
ICA decomposition.
Typical graphical representations of Independent Components (ICs) are
topographic maps of the inverse weights, component spectra, and component time series
or average component time series across epochs. Figure 4 provides typical graphical
representations of the 12 components that describe the most variance in a subject's EEG
data using InfoMax ICA. The EEG was collected from a subject at rest who fixated on a
cross on a monitor for 42 seconds. Due to properties of the weight matrix W (columns
represented as the circular head plots in Figure 4 corresponding to each component), the
component spectra and the component time series, we identified 4 components that could
be indicative of artifact. Component 1 (IC1) most likely captures the electrical potentials
due to eye blinks. Indicative of eye blinks, the channel weights indicate that all the
component information is located near the eyes. The power spectrum has one peak in a
low frequency band because the time series has high amplitude waveforms located
sparsely in time (which occur once per blink). IC12 is probably muscle artifact, perhaps
due to facial tension. It contributes to the EEG recording mainly at peripheral electrodes,
and its spectrum has high power at high frequencies and low power at typical resting
EEG frequencies (such as alpha rhythm, around 10 Hz). Furthermore, the topography’s
spatial frequency is too high (i.e. too focal) as this spatial frequency is near impossible
for EEG to obtain due to the properties of head volume conduction which acts as a low-
pass spatial filter. Similarly, IC6 captures data that cannot be due to brain activity
because its weight is only at one electrode and the power spectrum exhibits a 1/f
frequency falloff (a property of electrical “pink” noise). As indicated by its time course,
this component is probably a mix of a temporary electrical discontinuity at about 6
seconds and a horizontal eye movement at about 12 seconds. IC7 is similar in its
properties to IC6 and captures only a temporary electrical discontinuity at 41 seconds.
The rest of the ICs most likely reflect cortical electrical activity or mixtures of cortical
electrical activity and muscle artifact. These ICs contain peaks in alpha (8-13 Hz) and/or
beta frequency bands (13-20 Hz) and have lower spatial frequency distributions typical of
We recommend keeping EEG and artifact mixtures in the data unless very
specific properties of the EEG are of interest a priori. There is empirical evidence to
suggest that ICA algorithms do not isolate many types of muscle artifact, especially task
related artifact, and thus rejecting ICs that do not clearly represent artifact becomes very
subjective (Shackman et al., 2009). Furthermore, the efficacy of ICA to reduce all EMG
artifact remains controversial at best (Olbrich et al., 2011; McMenamin et al., 2011).
However if one must analyze a dataset that has a large quantity of muscle artifact, there
may be a few indicators of EEG data that do not originate in the brain. For instance, an IC
representing EEG or an EEG-EMG mixture may have a constant distribution of sample
variances over all trials if the subject is in the same cognitive state and is doing the same
task. In contrast, irregular EMG components will typically only have large variances on
only a few trials.
In order to reduce subjectivity of the artifact independent component (IC) removal
process and reduce the time demand on performing artifact removal, some progress has
been made on automatic rejection of artifact components. ADJUST is an algorithm that
uses properties of the components such as spatial weight distributions on the scalp,
variance, and kurtosis of the components’ potentials to automatically label components as
eye blinks, vertical eye movements, horizontal eye movies, or generic potential
discontinuities so that they can be subtracted from the recording (Mognon et al., 2011)
No known modern artifact correction technique is perfect for muscle artifact
removal, and no EEG recording is completely immune to muscle artifact (Whitham et al.,
2007). This is particularly the case for the neck and face muscle variety; thus, good
recording and analysis practices are still the best approach for reducing artifact in EEG
recordings. Subjects should be told to remain still and minimize jaw clenching, and the
electrode cap or net should be positioned tightly (but comfortably) on the subject. Muscle
artifact exhibits broadband frequency spectra with substantial relative power above 15
Hz; therefore analyses of the delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz) and mu (11-
14 Hz) bands are typically more robust to muscle artifact contamination.
7. Stationary Data Analysis
The starting point of most EEG data analysis is spectral analysis to assess
statistical properties of amplitude and phase of multiple EEG frequency bands. Even
when the final goal of the analysis does not involve spectral analysis, examining the
spectrum of the EEG is a useful starting point for evaluating data quality and for
communication of more complex methods. The spectrum obtained by applying the
Fourier transform to a single EEG epoch or time window provides information about its
frequency content. Fourier transform algorithms yield estimates of Fourier coefficients
that reflect both the amplitude and phase of the oscillations within one frequency band.
Fast Fourier Transforms (FFT) are one class of algorithms that are particularly useful,
and a number of important issues in practical FFT analysis are detailed in several texts
(see Bendat & Piersol 2011 for examples). Other Fourier analysis or spectra-like
algorithms such as multi-taper analysis (Percival & Walden, 1993), autoregressive
models (Ding et al., 2000), wavelet analysis (Lachaux et al. 2002), and Hilbert transforms
(Bendat & Piersol 2011; van Quyen et al. 2001; Deng & Srinivasan, 2010) have potential
applications in EEG, particularly in the analysis of short epochs characterizing EEG
behavior after an experimental stimulus. Any of these algorithms can be used to carry out
spectral analysis of time series, but an FFT based analysis provides a quick and easy
assessment of the spectrum.
The amplitude spectrum of one epoch of EEG is an exact representation of the
frequency content of that particular time window, but only provides one observation
about the random process generating the signal. The full ensemble of K epochs {Vk(t)}
can be used to estimate statistical properties of the random process generating the EEG
under the assumption of weak stationarity (Bendat and Piersol, 2011). Weak stationarity
is obtained if the mean and variance of the signal do not change with time. This can be
verified by obtaining an estimate of mean and variance at each time point across epochs.
Typically, the weak stationarity assumption is reasonable in the analysis of spontaneous
EEG in resting-state experiments; it is not reasonable in any experiment where a sensory
stimulus is presented and/or a motor response is obtained from the subject.
Estimating the power spectrum from an ensemble of epochs yields an estimate of
the variance of the signal as a function of frequency. This is a particularly useful
approach because EEG contains oscillatory activity in distinct frequency bands that are
associated with different brain states. First, for each epoch Vk(t), Fourier coefficients
Fk(fn) are obtained by applying a Fourier Transform, perhaps using the FFT. Then the
power spectrum may be estimated from the ensemble of observations by summing over K
epochs, given in Eq. 4.
       
n k n k n k n
P f F f F f F f
1,2,..., 2 1nN
When applying the FFT, the frequency resolution f = 1/T of the resulting power
spectrum depends inversely on the length of each observation T as
f n T
where n
indexes the frequency band. The equation for the power spectrum is multiplied by a
factor of two because the Fourier transform provides amplitudes split between positive
and negative complementary phases and only amplitudes at positive phases are usually
calculated. If the mean value of the signal is zero, the power spectrum summed over all
frequencies is equal to the variance in the signal, a relationship known as Parseval’s
theorem (Bendat and Piersol 2011). The square root of the power spectrum, the amplitude
spectrum, places more emphasis on non-dominant spectral peaks. Any algorithm used to
obtain Fourier coefficients can be used to approximate Eq. 4; although the definition of
frequency bands depends on the algorithm. Eq. 4 provides a definition of the EEG power
spectrum in units that depend on the frequency resolution f. In order for the results to
be compared across all choices of epoch length, the power spectrum is sometimes
normalized by the frequency resolution f to express power in units of V2 per Hz.
Before the rise of widespread access to computational power and use of the Fast
Fourier Transform, the power spectrum of a time series was typically calculated in a two
stage procedure. First the autocorrelation function was estimated and then the Fourier
transform of the autocorrelation function was calculated. The result is equivalent to the
power spectrum of the signal. The autocorrelation function is the covariance of the signal
with itself as a function of lag:
Like the power spectrum, the true autocorrelation function is an unknown statistical
property of the time series and can only be estimated. In Eq. (5) the lag variable is
defined over positive and negative values. The autocorrelation function contains exactly
the same spectral information as the time series of epoch length T if the domain of is [-
T/2, T/2]. The Fourier transform of the autocorrelation function is then equal to the power
spectrum of the signal (Bendat and Piersol, 2001). However in modern spectral analysis
the Fourier transform is usually directly calculated before calculating the power
Estimation of the power spectrum involves tradeoffs in frequency resolution,
statistical power, and weak stationarity. For example, consider the choices involved in
analyzing a 60 second EEG record. Figure 5 demonstrates power spectra of two EEG
channels, one occipital and one frontal, recorded with the subject’s eyes closed and at
rest. The power spectra were obtained using an epoch length T = 60 seconds (f =0 .017
Hz) and no epoch averaging (K = 1 epochs). With this choice, the FFT of the entire
record is obtained (exact spectra of the two EEG signals), but no information about the
statistical properties of the underlying random process is gained. Note that the power
spectrum of the occipital channel (Fig. 5a) contains two peaks, one below 10 Hz and a
larger peak above 10 Hz. The frontal channel (Fig. 5b) shows a larger peak below 10 Hz.
By examining the other channels it was found that the two peaks have distinct spatial
distributions over the scalp, suggesting they have different source distributions. Each
peak is surrounded by power in sidebands (adjacent frequency bins) of the two peak
frequencies. The signals are stochastic processes occupying relatively narrow bands in
the frequency spectrum.
To analyze the 60 sec signal properly, we must decide how to divide the record
into epochs to implement Eq. 4. The choice is a compromise between the advantage of
good frequency resolution yielded by long epochs (large T and small K) and the statistical
power of our estimate gained by using a larger number of epochs (small T and large K). If
a frequency resolution of f = 0.5 Hz is chosen, the record is segmented into K = 30
epochs of length T = 2 sec. If frequency resolution is reduced to f =1 Hz, we divide the
record into K = 60 epochs of length T = 1 sec. Figures 5c and 5d show the power spectra
of the frontal and occipital channels with f = 1 Hz (grey circles) and
f = 0.5 Hz (black
circles). The power spectra at the occipital and frontal channels are both dominated by
alpha rhythm oscillations. At the occipital electrode (Fig. 5c) two separate peak
frequencies are at 9.5 and 10.5 Hz are evident with f = 0.5 Hz, but this separation is not
revealed with f = 1 Hz, where only a single peak frequency at 10 Hz is evident.
Lowering frequency resolution has a similar effect at the frontal channel (Fig. 5d), but
since there is very little power at 10.5 Hz, the only clear peak appears at 9 Hz. Thus, by
choosing a lower frequency resolution we observe different peak frequencies at the two
sites, while choosing higher frequency resolution results in pairs of frequency peaks at
both sites but with different magnitudes.
By examining the power spectra for the occipital (Fig. 6a) and frontal sites (Fig.
6b) for individual epochs with f = 0.5 Hz, evidence is found for two different
oscillations within the alpha band. At the occipital channel, individual epochs display two
distinct peaks at 9.5 Hz and 10.5 Hz. The first 15 epochs show a strong response at 10.5
Hz but the later epochs show a stronger response at 9.5 Hz. The dominant frequency in
each epoch is summarized in the peak power histograms in Fig. 5c showing that
individual epochs displayed peak frequencies at both 9.5 Hz and 10.5 Hz. By contrast,
very few epochs have a peak frequency of 10.5 Hz at the frontal site (Fig. 5d); most
epochs have peak frequencies either at 9.5 Hz or in the delta band (< 2 Hz). Note that
during most epochs with strong delta activity in Fig. 5d the alpha peaks are attenuated.
In the previous example, electrodes at different locations show different
magnitudes of two distinct oscillations with center frequencies at 9.5 Hz and 10.5 Hz.
The natural next step is to measure correlation between electrode sites to assess spatial
statistics of the EEG. This is motivated by the idea that correlation of EEG signals should
reflect functional connectivity of the brain. Neurons in distant (and nearby) cortex are
connected by axons which form the white-matter beneath the gray-matter consisting more
of cell bodies (Nunez, 1995). Coherence between two electrodes is a correlation
coefficient (squared) that measures the consistency of relative phase between signals in a
specific frequency band (and is weakly dependent on large amplitude changes). In EEG
signals, coherence mainly reflects the consistency of phase differences across epochs,
potentially reflecting axon transmission delays (Nunez and Srinivasan, 2006). Coherence
ranges from 0 (indicating random phase differences) to 1 (indicating identical phase
differences). Like the Pearson correlation coefficient, coherence is a unknown statistical
property whose true value depends on the entire theoretical population of all similar
epochs and can only be estimated using the given sample of epochs.
One way to estimate coherence between an electrode pair is to first calculate the
cross spectrum (Bendat and Piersol, 2011). The cross spectrum, provided in Equation 6,
is a measure of the joint spectral properties of two channels (i.e. a measure of the Fourier
coefficient covariance). Fourier coefficients at a certain frequency fn at each electrode are
multiplied and averaged over K epochs to calculate the cross spectrum.
   
uv K
uv n uv uk n vk n
C f A e F f F f
1,2,..., 2 1nN
Where refers to the complex conjugate of the Fourier coefficient at channel v. Note
that the cross spectrum between a channel and itself (u = v) is equal to the power
spectrum (Eq. 4) and is real valued.
In general, the cross spectrum is complex valued and can be thought of in terms of
its magnitude (or cross-power) Auv and phase uv information. The phase of the cross
spectrum is the average phase difference between the two channels. This average phase
difference is called the relative phase. The magnitude information Auv is analogous to
the ordinary covariance between two time series, i.e. the covariance across observations
between two channels in one frequency band. Coherence at fn is defined by the squared
magnitude of the cross spectrum standardized by the product of the power spectra (which
are variances) of each signal
   
   
2uv n
uv n u n v n
fP f P f
1,2,..., 2 1nN
Coherence can be thought of as the fraction of variance at frequency fn in channel u that
can be explained by a constant amplitude and phase change (i.e. a linear transformation
of Fourier coefficients) of the data at frequency fn obtained at channel v.
Coherence between channels is most sensitive to phase differences. For instance,
if the phase difference is constant over epochs between channels u and v, coherence is
maximized and is equal to 1. If the relative phase between channels u and v varies across
epochs, coherence will be less than one. Furthermore, if the phase difference is random
across all epochs, the coherence estimate will approach zero as the number of epochs (K)
increases. A coherence of 0.4 in frequency band fn indicates that at that 40% of the
variance at one channel can be explained by a linear transformation of the other channel.
However this does not imply that a linear relationship actually exists, only that the
relationship (or part of the relationship) can be approximated by a linear transformation.
If the activity at the two channels is truly related nonlinearly (e.g. have a quadratic
relationship), the coherence would provide information as to how well the nonlinear
relationship can be approximated by a linear relationship.
If the goal is to estimate phase synchronization independent of amplitude
fluctuations, Fourier coefficients in Equation 7 could be normalized by amplitude to
obtain phase-only coherence. Phase synchronization can also be evaluated with measures
of the relative phase distribution across or within epochs (Tass et al., 1998). On the other
hand, one reason to include amplitude as in Eq. 6 and 7 is that coherence measures are
weighted in favor of epochs with large amplitudes. Trials with large amplitudes are
preferred if large amplitudes are indicative of high signal to noise ratios. If only epoch
phase information were used, equal emphasis would be placed on low and high amplitude
epochs in estimates of phase synchronization. This would potentially reduce the overall
signal-to-noise ratio of the analysis.
Typically the goal of EEG coherence studies is to estimate the functional
connectivity of the brain. Coherent activity between pairs of electrode sites may be either
synchronous (zero phase difference) or asynchronous (constant phase difference). In
other words asynchronous coherent signals are essentially “synchronous with a time lag.”
Figure 7 shows the coherences between an electrode x and a ring of electrodes at
progressively greater distances from x, labeled 1-9. The subject is at rest with eyes
closed, a state in which coherence is usually high in the alpha band (Nunez, 1995). The
estimated coherence between electrode x and electrode n is labeled x:n. The electrode
positions are shown and the distance between adjacent electrodes is about 2.7 cm. At the
closest electrode pair x:1, coherence is very high (above 0.75) at all frequencies (i.e.
coherence is generally independent of temporal frequency). This effect was predicted by
the observed effect of volume conduction of current spreading through the head; the
effect of volume conduction is to mix the brain signals at each electrode, which correlates
them at all frequencies. This theoretical prediction of coherence due to volume
conduction suggests that the main effect of volume conduction is to artificially inflate
coherences at short to moderate distances, and that this effect is independent of
As the electrode separation is increased, the pair x:2 shows lower coherences, but
at most frequencies coherence is still above 0.4 suggesting a strong component of
coherence that is independent of frequency. A peak is visible in the alpha band at 9.5 Hz,
but it is difficult to evaluate this peak, since there is also a broad elevation of coherence.
As the sensor separation is further increased (x:3) the floor of the coherences reduces
further to about 0.2, and a peak becomes more evident in the 18 Hz range. This electrode
is about 10 cm away from electrode x consistent with model predictions of volume
conduction effects on EEG coherence for electrodes separated by less than 10 cm
(Srinivasan et al., 1998). For pairs of electrodes involving sites over the temporal lobes,
x:4 and x:5, the floor of the coherences approaches zero at most frequencies except for
the alpha band where a second peak at 10.5 Hz is visible. At a very long distance (x:8)
the coherences are again elevated across all frequencies, suggesting a very small volume
conduction effect at long distances as also observed in modeling studies (Srinivasan et
al., 1998; Srinivasan et al., 2007).
Spatial statistics, like coherence, will always be influenced by volume conduction
effects. In the case of coherence, a simple rule of requiring a separation distance of 10 cm
or more can be used to minimize volume conduction effects. Another approach is to use
the surface Laplacian to spatially high-pass filter the EEG signals, minimizing the effect
of volume conduction (Nunez et al., 1994; Srinivasan et al., 1998). And as new methods
are deployed to evaluate EEG spatial statistics (e.g. to make inferences about
connectivity), more modeling studies are needed to evaluate the impact of volume
conduction in order to interpret these new measures.
8. Nonstationary Data Analysis
A stationary time series is a random process whose statistical distribution is
invariant over time. A weakly stationary time series is one in which the mean, variance,
and autocorrelation of the random process are invariant to shifts in the time at which the
sample records are obtained. Many EEG analyses assume weak stationarity in order to
evaluate EEG in different brain states such as resting with eyes closed or eyes open,
during mental calculations, or during different stages of sleep.
However, in many experiments, especially those related to cognitive functions, a
stimulus is presented and/or a motor response is elicited and often both. This structure of
events in an experimental trial implies that the EEG time series is nonstationary, as we
can reasonably expect that the statistics of the signals will depend on time relative to the
events. Clearly the EEG before a stimulus is presented has a different statistical structure
than EEG following the stimulus. In fact we are interested in uncovering these
Many experimenters analyze the time-varying mean of EEG data epochs
following an experimental stimulus. The mean responses are called evoked potentials
(EP) or event related potentials (ERP). These types of experimental designs are
ubiquitous in the cognitive neuroscience literature. Because we may postulate that only
the signals relevant to the stimulus to remain after we average many trials of EEG
response data, the EP or ERP is calculated as the ensemble mean of observations across
epochs. If the subject views the same stimulus (or stimulus category) in all epochs (e.g.
separate experimental trials) then we can expect the average EEG signal to approximate
the response of ongoing EEG activity to the stimulus, such that the mean of the potential
at each electrode varies as a function of time and is given by
( ) ( )
t V t
Note that in order to calculate an evoked response, EEG is averaged across the dimension
of epochs, not across the dimension of time.
Figure 8 shows an example of a visual evoked potential, or the time-varying mean
at 124 electrodes of a high-density EEG cap, following presentation of a visual stimulus.
Much of the ERP literature focuses on the peaks and troughs of this waveform (Fig. 8a)
and its relationship to perception and cognition. These are labeled P1, N1, P2, etc,
reflecting the direction and order of the peaks (Luck et al., 2000). The inset topographic
maps of the potential at the peaks indicate some spatial distribution, although naturally
most of the response is over the back of the head, where visual cortex is located. Visual
inspections suggests that this signal is an oscillation with a period of about 120 ms, which
is confirmed by the wavelet spectrogram (Fig. 8b) at one channel.
Spectral analysis methods as defined in section 7 can be used to evaluate time-
varying power and coherence of the EEG signals. The assumption in these analyses is
that over narrow windows of time, e.g., 200 ms, the EEG can be considered stationary.
This window is then moved over the epoch. A more convenient approach to calculate
time-varying power and coherence is to use the complex Morlet wavelet transform to
estimate time-varying Fourier coefficents (Lachaux et al., 2002). Two examples of time-
varying power calculated using the Morlet wavelet transform are presented in Figures 8b
and 9. Fig. 8b shows the phase-locked change in alpha power after a visual stimulus is
presented. In order to calculate phase-locked power, power of the time-average of epochs
given in Eq. 8 can be calculated such that K=1 in Eq. 4, which removes all signal that is
not phase-locked. Fig. 9 shows a non-phase-locked “desynchronization” (i.e. decrease) in
alpha power in electrode C3 (an electrode close to the left motor-cortex) before the
subject responded with a button press. Power in the alpha band at electrodes over motor
cortex are sometimes called the “mu rhythm” to differentiate it from spectral changes in
the alpha band not associated with a motor response.
Steady-state evoked potentials (SSEPs) are another type of evoked potential,
associated with experimental designs known as “frequency-tagging” (Ding et al., 2006;
Deng and Srinivasan, 2010). SSEPs refer to a frequency-tagging paradigm where EEG
responses are observed in narrow frequency bands corresponding to the frequencies and
harmonics of a stimulus (Regan, 1977). Steady-state responses can be evoked with
flickering visual stimuli (steady-state visual evoked potentials; SSVEPs), modulated
auditory stimuli (steady-state auditory evoked potentials; SSAEPs), or rhythmic
somatosensory stimuli (steady-state somatosensory evoked potentials, SSSEPs). SSEP
paradigms are preferred by many neuroscientists due to their large signal to noise ratios,
minimizing the problem of artifacts which are typically broadband. SSVEPs, SSAEPs,
and SSSEPs have also been shown to track attention (Tiitinen et al., 1993, Morgan et al.,
1996, and Giabbiconi et al., 2004). Our group has also shown that SSVEP data can be
used to classify video game players versus non-gamers (Krishnan et al., 2013),
differentiate individual differences in attention (Ding et al., 2006; Bridwell et al., 2013),
and predict individual differences in perceptual decision making (Nunez et al., 2015).
9. Summary
In this chapter we introduce the physical basis of EEG recording and its
implications for the type of brain processes observable with EEG. Practical EEG
recording poses additional challenges, especially in relation to physiological artifacts
such as muscle artifacts. We recommend that EEG analysis begin with spectral analysis
methods, which provide a foundation from which more advanced statistical models can
be developed and foster better communication with other EEG scientists. Spectral
analysis is ubiquitous in the EEG literature in both stationary and non-stationary
experimental designs. Finally, in research involving spatial statistics, the effect of
volume conduction must be explicitly considered to make meaningful inferences about
functional connectivity in the brain.
Bell, A. J., & Sejnowski, T. J. An information-maximization approach to blind separation
and blind deconvolution. Neural computation, 7(6), 1129-1159. 1995.
Belouchrani A, Abed-Meriam K, Cardoso JF, and Moulines E, A blind source separation
technique using second-order statistics. IEEE Transactions on Signal Processing.
45:434-444, 1997.
Bendat JS and Piersol AG, Random Data. Analysis and Measurement Procedures, Fourth
Edition. New York: Wiley, 2011.
Bertrand, O., Perrin, F., & Pernier, J. A theoretical justification of the average reference
in topographic evoked potential studies. Electroencephalography and Clinical
Neurophysiology/Evoked Potentials Section, 62(6), 462-464, 1985.
Baillet, S., & Garnero, L. A Bayesian approach to introducing anatomo-functional priors
in the EEG/MEG inverse problem. Biomedical Engineering, IEEE Transactions on,
44(5), 374-385, 1997.
Bridwell, D. A., Hecker, E. A., Serences, J. T., & Srinivasan, R. Individual differences in
attention strategies during detection, fine discrimination, and coarse discrimination.
Journal of neurophysiology, 110(3), 784-794, 2013.
Buzsáki, G, Rhythms of the Brain. New York: Oxford University Press, 2006.
Cardoso JF Blind signal separation: Statistical principles. Proc IEEE, 9:2009-2025, 1998.
Delorme, A., & Makeig, S. EEGLAB: an open source toolbox for analysis of single-trial
EEG dynamics including independent component analysis. Journal of neuroscience
methods, 134(1), 9-21, 2004.
Delorme, A, Sejnowski, T, Makeig, S. Enhanced detection of artifacts in EEG data using
higher-order statistics and independent component analysis. Neuroimage. 34(4):1443-49,
Deng, S., Srinivasan, R., Lappas, T. and D'Zmura, M. EEG classification of imagined
syllable rhythm using Hilbert spectrum methods. Journal of neural engineering, 7(4),
046006. 2010.
Deng, S. and Srinivasan, R. Semantic and acoustic analysis of speech by functional
networks with distinct time scales. Brain Research, 1346, 132 - 144. 2010.
Ding, J., Sperling, G. and Srinivasan, R. Attentional modulation of SSVEP power
depends on the network tagged by the flicker frequency. Cerebral cortex, 16(7), 1016-
1029. 2006.
Ding, M., Bressler, S.L., Yang, W. and Liang, H. Short-window spectral analysis of
cortical event-related potentials by adaptive multivariate autoregressive modeling: data
preprocessing, model validation, and variability assessment. Biological cybernetics,
83(1), 35-45. 2000.
Fisch, B.J. and Spehlmann, R., Fisch and Spehlmann's EEG primer: basic principles of
digital and analog EEG. Elsevier Health Sciences. 1999.
Fitzgibbon, S.P., DeLosAngeles, D., Lewis, T.W., Powers, D.M.W., Grummett, T.S.,
Whitham, E.M., Ward, L.M., Willoughby, J.O. and Pope, K.J. Automatic determination
of EMG-contaminated components and validation of independent component analysis
using EEG during pharmacologic paralysis. Clinical Neurophysiology. 2015.
Gratton, G., Coles, M. G., & Donchin, E. A new method for off-line removal of ocular
artifact. Electroencephalography and clinical neurophysiology, 55(4), 468-484, 1983.
Giabbiconi, C. M., Dancer, C., Zopf, R., Gruber, T., & Müller, M. M. Selective spatial
attention to left or right hand flutter sensation modulates the steady-state somatosensory
evoked potential. Cognitive brain research, 20(1), 58-66, 2004.
Hauk, O. Keep it simple: a case for using classical minimum norm estimation in the
analysis of EEG and MEG data. Neuroimage, 21(4), 1612-1621, 2004.
Henson, R. N., Flandin, G., Friston, K. J., & Mattout, J. A Parametric Empirical Bayesian
framework for fMRIconstrained MEG/EEG source reconstruction. Human brain
mapping, 31(10), 1512-1531, 2010.
Hyvärinen, A., & Oja, E. A fast fixed-point algorithm for independent component
analysis. Neural computation, 9(7), 1483-1492, 1997.
Jung, T. P., Humphries, C., Lee, T. W., Makeig, S., McKeown, M. J., Iragui, V., &
Sejnowski, T. J. Extended ICA removes artifacts from electroencephalographic
recordings. Advances in neural information processing systems, 894-900, 1998.
Krishnan, L., Kang, A., Sperling, G., & Srinivasan, R. Neural strategies for selective
attention distinguish fast-action video game players. Brain topography, 26(1), 83-97,
Lachaux J. P., Lutz, A., Rudrauf, D., Cosmelli, D., Le Van Quyen, M., Martinerie, J.,
&Varela, F. Estimating the time-course of coherence between single-trial brain signals:
an introduction to wavelet coherence. Electroencephalography and Clinical
Neurophysiology 32(3):157-74, 2002.
Le Van Quyen, M., Foucher, J., Lachaux, J., Rodriguez, E., Lutz, A., Martinerie, J.,
Varela, F. J. Comparison of Hilbert transform and wavelet methods for the analysis of
neuronal synchrony. J Neurosci Methods. 111(2):83-98, 2001.
Lopes da Silva FH and Storm van Leeuwen W, The cortical alpha rhythm in dog: the
depth and surface profile of phase. In: MAB Brazier MAB and H Petsche (Eds)
Architectonics of the Cerebral Cortex. New York: Raven Press, pp 319-333, 1978.
Lopes da Silva, F. H. Dynamics of EEGs as signals of neuronal populations: models and
theoretical considerations. In: E Niedermeyer and FH Lopes da Silva (Eds)
Electroencephalography. Basic Principals, Clinical Applications, and Related Fields.
Forth Edition. London: Williams and Wilkins, pp 76-92, 1999.
Luck, S. J., Woodman, G. F., & Vogel, E. K. Event-related potential studies of attention.
Trends in cognitive sciences, 4(11), 432-440, 2000.
Makeig, S., Bell, A. J., Jung, T. P., & Sejnowski, T. J. Independent component analysis
of electroencephalographic data. Advances in neural information processing systems,
145-151, 1996
Makeig S, Debener S, Onton J, & Delorme A, Mining event-related brain dynamics,
Trends in Cognitive Science, 8(5):204-210, 2004.
McMenamin, B.W., Shackman, A.J., Greischar, L.L. and Davidson, R.J.
Electromyogenic artifacts and electroencephalographic inferences revisited. Neuroimage,
54(1), 4-9. 2011.
Morgan, S. T., Hansen, J. C., & Hillyard, S. A. Selective attention to stimulus location
modulates the steady-state visual evoked potential. Proceedings of the National Academy
of Sciences, 93(10), 4770-4774, 1996.
Mognon, A., Jovicich, J., Bruzzone, L., & Buiatti, M. ADJUST: An automatic EEG
artifact detector based on the joint use of spatial and temporal features.
Psychophysiology, 48(2), 229-240, 2011.
Murias, M., Swanson, J. M., & Srinivasan, R. Functional connectivity of frontal cortex
in healthy and ADHD children reflected in EEG coherence. Cerebral Cortex. 17:1788-
1799, 2007.
Nunez, M. D., Srinivasan, R., & Vandekerckhove, J. Individual differences in attention
influence perceptual decision making. Frontiers in psychology, 8, 2015.
Nunez, P.L. A study of origins of the time dependencies of scalp EEG: I-theoretical basis.
Biomedical Engineering, IEEE Transactions on, (3), 271-280, 1981.
Nunez, P. L., Silberstein, R. B., Cadusch, P. J., Wijesinghe, R. S., Westdorp, A. F., &
Srinivasan, R. A theoretical and experimental study of high resolution EEG based on
surface Laplacians and cortical imaging. Electroencephalography and clinical
neurophysiology, 90(1), 40-57, 1994.
Nunez, P. L. Neocortical Dynamics and Human EEG Rhythms. New York: Oxford
University Press, 1995.
Nunez, P. L. Toward a quantitative description of large scale neocortical dynamic
function and EEG. Behavioral and Brain Sciences 23: 371-398 (target article), 2000a.
Nunez, P. L. Neocortical dynamic theory should be as simple as possible, but not simpler.
Behavioral and Brain Sciences 23: 415-437 (response to commentary by 18
neuroscientists), 2000b.
Nunez, P. L. & Srinivasan, R. Electric Fields of the Brain: The Neurophysics of EEG,
2nd Edition. New York: Oxford University Press, 2006.
Olbrich, S., Jödicke, J., Sander, C., Himmerich, H. and Hegerl, U.. ICA-based muscle
artefact correction of EEG data: What is muscle and what is brain?: Comment on
McMenamin et al. Neuroimage, 54(1), 1-3. 2011.
Oostenveld, R., & Praamstra, P. The five percent electrode system for high-resolution
EEG and ERP measurements. Clinical neurophysiology, 112(4), 713-719, 2001.
Percival, D. B. & Walden, A. T., Spectral Analysis for Physical Applications: Multitaper
and Conventional Univariate Techniques, Cambridge U Press, Cambridge, UK, 1993.
Petsche, H. & Etlinger, S. C. EEG and Thinking. Power and Coherence Analysis of
Cognitive Processes. Vienna: Austrian Academy of Sciences, 1998.
Pfurtscheller, G. & Lopes da Silva, F. H. Event-related EEG/MEG synchronization and
desynchronization: Basic principles. Electroencephalography and Clinical
Neurophysiology 110: 1842-1857, 1999.
Phillips, C., Rugg, M. D., & Friston, K. J. Anatomically informed basis functions for
EEG source localization: combining functional and anatomical constraints. NeuroImage,
16(3), 678-695, 2002.
Polich, J., & Kok, A. Cognitive and biological determinants of P300: an integrative
review. Biological psychology, 41(2), 103-146, 1995.
Pope, K. J., Fitzgibbon, S. P., Lewis, T. W., Whitham, E. M., & Willoughby, J. O.
Relation of gamma oscillations in scalp recordings to muscular activity. Brain
topography, 22(1), 13-17, 2009.
Regan, D. Steady-state evoked potentials. JOSA, 67(11), 1475-1489, 1977
Schlögl, A., Keinrath, C., Zimmermann, D., Scherer, R., Leeb, R., & Pfurtscheller, G.
(2007). A fully automated correction method of EOG artifacts in EEG recordings.
Clinical neurophysiology, 118(1), 98-104.
Shackman, A. J., McMenamin, B. W., Slagter, H. A., Maxwell, J. S., Greischar, L. L., &
Davidson, R. J. Electromyogenic artifacts and electroencephalographic inferences. Brain
topography, 22(1), 7-12, 2009.
Srinivasan, R., and Deng, S. Multivariate spectral analysis of the electroencephalogram:
power, coherence, and second-order blind identification. 2012
Srinivasan, R., Nunez, P.L., Tucker, D.M., Silberstein, R.B., and Cadusch, P.J. Spatial
sampling and filtering of EEG with spline Laplacians to estimate cortical potentials.
Brain Topography, 8:355 366, 1996.
Srinivasan, R., Nunez, P.L., and Silberstein, R.B. Spatial filtering and neocortical
dynamics: estimates of EEG coherence. IEEE Transactions on Biomedical Engineering,
45:814 826, 1998.
Srinivasan, R. Anatomical constraints on source models for high-resolution EEG and
MEG derived from MRI. TCRT 5:389-399, 2006.
Srinivasan, R., Winter, W. R., Ding, J., Nunez, P.L., EEG and MEG coherence: Measures
of functional connectivity at distinct spatial scales of neocortical dynamics. Journal of
Neuroscience Methods 166:41-52., 2007.
Tang, A. C., Sutherland, M. T., McKinney, C. J., Validation of SOBI components from
high-density EEG, NeuroImage, 25:539-553, 2004.
Tass, P., Rosenblum, M. G., Weule, J., Kurths, J., Pikovsky, A., Volkmann, J.,
Schnitzler, A. & Freund, H. J. Detection of n: m phase locking from noisy data:
application to magnetoencephalography. Physical review letters, 81(15), 3291, 1998.
Tiitinen, H., Sinkkonen, J., Reinikainen, K., Alho, K., Lavikainen, J., and Näätänen, R.
Selective attention enhances the auditory 40-Hz transient response in humans. Nature,
364(6432), 59-60, 1993.
Tucker, D. M. Spatial sampling of head electrical fields: the geodesic sensor net.
Electroencephalography and clinical neurophysiology, 87(3), 154-163, 1993.
Wipf, D., & Nagarajan, S. A unified Bayesian framework for MEG/EEG source imaging.
Neuroimage, 44(3), 947-966, 2009.
Whitham, E. M., Pope, K. J., Fitzgibbon, S.P., Lewis, T., Clark, C. R., Loveless, S.,
Broberg, M., Wallace, A., DeLosAngeles, D., Lillie, P., Hardy, A., Fronsko, R.,
Pulbrook, A. and Willoughby, J. O., Scalp electrical recording during paralysis:
quantitative evidence that EEG frequencies above 20Hz are contaminated by EMG.
Clinical Neurophysiology, 118(8): 1877-1888, 2007.
Whitham, E. M., Lewis, T., Pope, K. J., Fitzgibbon, S. P., Clark, C. R., Loveless, S.,
DeLosAngeles, D., Wallace, A. K., Broberg, M., and Willoughby, J. O. Thinking
activates EMG in scalp electrical recordings. Clinical neurophysiology: official journal of
the International Federation of Clinical Neurophysiology, 119(5), 1166, 2008.
Wu, J., Quinlan, E. B., Dodakian, L., McKenzie, A., Kathuria, N., Zhou, R. J.,
Augsburger, R., See, J., Le, V. H., Srinivasan, R., and Cramer, S. C. Connectivity
measures are robust biomarkers of cortical function and plasticity after stroke. Brain,
138(8), 2359-2369, 2015.
Yuval-Greenberg, S., Tomer, O., Keren, A. S., Nelken, I., and Deouell, L. Y. Transient
induced gamma-band response in EEG as a manifestation of miniature saccades. Neuron.
58(3): 429-441, 2008.
Ziehe, A., Laskov, P., Muller, K.R., Nolte, G. A linear least-squares algorithm for
joint diagonalization. Proceedings ICA2003, 469474. 2003.
Figure 1. Typical power spectra of 124 EEG channels of 66 seconds (epoch length T = 2
sec with K = 33 epochs) of data from a subject (male, 25 yrs) who fixated on a computer
monitor with his eyes open. While EEG spectral band definitions vary from lab to lab and
across different fields, bands are typically defined as follows: delta 1-4 Hz, theta 4-8 Hz,
alpha 8-13 Hz, beta 13-20 Hz, and gamma > 20 Hz. Some groups also identify the mu
rhythm which exists as a peak either in the alpha or beta bands and typically has high
power over the motor cortex. In the eyes-open resting data, with some artifact power
removed using Independent Component Analysis (ICA), we see peaks in the delta, theta,
and alpha bands and some power in the beta band. However the dominant peaks in the
spontaneous EEG are in the theta and alpha bands, which have different spatial
distributions over the electrodes and are associated with different cognitive functions.
Topographic scalp maps were generated by summing power across frequencies in the
theta (left) and alpha (right) frequency bands and interpolating between electrodes, such
that brighter values correspond to higher power. Alpha, which is empirically associated
with the resting state, has maximum power over parietal channels as indicated by the
right topographic scalp map.
Figure 2. Volume conduction models for EEG. (a) A dipole is shown in the inner sphere
of a 4-concentric spheres head model consisting of the inner sphere (brain) and three
spherical shells representing CSF (cerebral spinal fluid), skull and scalp. The parameters
of model are the radii (r1, r2, r3, r4) of each shell and the conductivity ratios
(). Typical values are: radii (8, 8.1, 8.6, 9.2 cm) and conductivity ratios
(0.2, 40, 1). This model is used in the simulations in this chapter. (b) A realistic shaped
boundary element model (BEM) of the head. The brain and scalp boundaries were found
by segmenting the images with a threshold, and opening and closing operation
respectively, while the outer skull boundary was obtained by dilation of the brain
boundary (ASA, Netherlands). Although geometrically more accurate than the spherical
model, the (geometrically) realistic BEM may be no more accurate than a concentric
spheres model because tissue resistivities are poorly known. (c) A realistic finite element
model (FEM) obtained from MRI. This model has potentially better accuracy than the
BEM model because the skull is subdivided into three layers corresponding to hard
(compact) and spongy (cancellous) bone layers.
Figure 3. (a) Time series of a dipole meso-source P(r, t) composed of a 6 Hz, 15 V
sine wave added to Gaussian random noise with a standard deviation of
= 150 V. The
Gaussian random noise was low pass filtered at 100 Hz. The sine wave has variance
(power) equal 1% of the noise. (b) Power spectrum of the time series shown in Part (a).
The power spectrum has substantial power at many frequencies other than 6 Hz. (c)
Time series recorded by an electrode on the outer sphere (scalp) of a four concentric
spheres model above the center of a dipole layer of diameter 3 cm. The dipole layer is
composed of 32 dipole sources P(r, t) with time series constructed similar to Part (a) with
independent Gaussian noise (uncorrelated) at each dipole source. Scalp potential was
calculated for a dipole layer at a radius rz = 7.8 cm in a four concentric spheres model.
The model parameters were radii (r1, r2, r3, r4) = (8, 8.1, 8.6, 9.2) and conductivity ratios
(1/2, 1/3, 1/4) = (0.2, 40, 1). Notice that the time series is smoother than in the
case of the individual dipole source. (d) Power spectrum of the time series shown in Part
(c). Note the peak at 6 Hz. (e) Time series similar to Part (c), but due to a dipole layer
of diameter of 4 cm composed of 68 dipole sources. (f) Power spectrum of the time series
shown in Part (e). (g) Similar time series to Part (c), but with a dipole layer of diameter 5
cm composed of 112 dipole sources. The presence of the 6 Hz sinusoid is obvious from
the time series. (h) Power spectrum of the time series shown in Part (g). A large spectral
peak at 6 Hz is evident.
Figure 4. (From top-left, clockwise) Power spectra, time courses, and spatial loading
topographies of the first twelve independent components (ICs) from an Independent
Component Analysis (ICA) of an EEG recording while a subject (male, 25 yrs) was
fixating on a computer monitor. The ICs are ordered by their contribution to the total
variance in the raw data. ICs that are likely to reflect artifact contribution can be removed
from the raw EEG data. IC1 is indicative of an eye blink. IC6 and IC7 are indicative of
temporary electrical discontinuities. IC12 is indicative of muscle artifact.
Figure 5. Example power spectra from a single subject (female, 22 yrs). The subject is
at rest with eyes closed. (a) Power spectrum of a midline occipital channel with epoch
length T = 60 sec and K = 1 epochs. The power spectrum appears to have two distinct
peaks one below 10 Hz and one above 10 Hz. (b) Power spectrum at a midline frontal
channel with epoch length T = 60 sec and K = 1 epochs. Here only the peak below 10 Hz
is visible. (c) Power spectra of a midline occipital channel calculated with two different
choices of epoch length T and number of epochs K. The grey circles indicate the power
spectrum with T = 1 sec and K = 60 epochs. The black circles indicate the power
spectrum with T = 2 sec and K = 30 epochs. (d) Power spectra of a midline frontal
channel calculated as in Part (c).
Figure 6. (a) Plots of 30 (individual epoch) power spectra for the occipital channel
shown in Figs. 5a and 5c. (b) Plots of the same 30 individual epoch spectra for the
frontal channel shown in Figs. 5b and 5d. (c) Peak power histograms show the
distribution of peak frequencies for the 30 epochs shown in Part (a). (d) Peak power
histograms for the 30 epochs shown in Part (b).
Figure 7. Scalp potential coherence spectra from a subject (female, 22 yrs) at rest with
eyes closed in order to maximize alpha coherence. Coherence was estimated with T = 2
sec (f = 0.5 Hz) in a 60 sec record. The head plot shows the location of 9 electrodes,
labeled x and 1 through 8. Coherence spectra between electrode x and each of the other
electrodes 1-8 are shown, with increasing separations along the scalp. Note that very
close electrodes have higher coherence independent of frequency as predicted by the
theoretical volume conductor model. Alpha band coherence is high for large electrode
separations, apparently reflecting the large cortical source coherence.
Figure 8. (a) A typical Visual Evoked Response (VEP; also known as an Event Related
Potential; ERP) to a large, high contrast sinusoidal grating stimulus recorded at 124
electrodes of a high-density 128 Electrical Geodesics, Inc. (EGI) cap. The VEP was
calculated by averaging low-pass Butterworth filtered data (with a 20 Hz passband)
across all trials in one subject (male, 23 yrs) and by subtracting each trial by the time
average of 200 milliseconds before that trial’s stimulus onset (to baseline the VEP).
Topographies of traditional local peaks (Luck et al., 2000) are labeled with P1, N1, P2,
and N2 indicating the first and second positive and negative peaks over posterior
electrodes. The N1 and P2 components evoke typical bilateral responses over parietal
electrodes. The P1 and N2 reflect other network behavior related to processing of the
visual stimulus. (b) A time-varying, phase-locked power spectrum of the same average
data at an electrode over the left parietal cortex calculated with a Morlet wavelet
transform. As shown by the wavelet, the VEP can also be thought of as a phase-locked
alpha response to the visual stimulus.
Figure 9. A non-phase-locked time-varying power spectrum from the same subject and
visual task as in Fig. 8 such that power was calculated using all K=157 epochs time-
locked to the motor response (button press given by the right or left hand). Data from a
left-central electrode C3 in the 10-20 electrode placement system is presented. Mu power
“desynchronizes” (i.e. decreases) approximately 200 ms before the button press, likely
reflecting cognitive control over the motor response. Similar magnitude of mu power
before the desynchronization is observed after the motor response.
author = {Nunez, Michael D. and Nunez, Paul L. and Srinivasan, Ramesh},
title = {Electroencephalography ({EEG}): neurophysics, experimental methods, and
signal processing},
booktitle = {Handbook of Neuroimaging Data Analysis},
year = {2016},
editor = {Ombao, Hernando and Linquist, Martin and Thompson, Wesley and Aston,
volumes = {Advance online publication},
publisher = {Chapman \& Hall/CRC},
isbn = {9781482220971},
pages = {175--197},
doi = {10.13140/rg.2.2.12706.63687},
url =

Supplementary resource (1)

... However, for most electrophysiology and brain imaging, a perfect measure of the brain is unrealistic. This is especially true for EEG where derived measures of EEG are expected to be influenced by muscle and electrical artifacts, even when the measure of EEG is somewhat robust to artifact (Nunez et al., 2016). While other forms of integrated neurocognitive models could be beneficial for studying human cognition (see Appendix), the key benefit of our frameworks is that these models contain at least one source of measurement noise for the EEG measures, σ 2 , as well as variance in the underlying cognition that can change across trials s 2 . ...
... While other forms of integrated neurocognitive models could be beneficial for studying human cognition (see Appendix), the key benefit of our frameworks is that these models contain at least one source of measurement noise for the EEG measures, σ 2 , as well as variance in the underlying cognition that can change across trials s 2 . By measurement noise, we mean the combination of electrical noise, artifacts due to EMG, and unrelated cognitive processes embedded in all measures of EEG (Whitham et al., 2007;Nunez et al., 2016). Measurement noise is especially relevant for EEG on single trials (Bridwell et al., 2018) because we are often not using measures that are robust to EEG artifacts, like trialaveraged ERPs (although even trial-averaged ERPs are not completely robust to artifact either) (Nunez et al., 2016). ...
... By measurement noise, we mean the combination of electrical noise, artifacts due to EMG, and unrelated cognitive processes embedded in all measures of EEG (Whitham et al., 2007;Nunez et al., 2016). Measurement noise is especially relevant for EEG on single trials (Bridwell et al., 2018) because we are often not using measures that are robust to EEG artifacts, like trialaveraged ERPs (although even trial-averaged ERPs are not completely robust to artifact either) (Nunez et al., 2016). Note also that preprocessing steps for EEG measures often have many degrees of freedom (Clayson et al., 2021), and therefore models should provide some quantification of the amount of relevant cognition described by singletrial EEG measures compared to the amount of extraneous information. ...
Full-text available
Despite advances in techniques for exploring reciprocity in brain-behavior relations, few studies focus on building neurocognitive models that describe both human EEG and behavioral modalities at the single-trial level. Here, we introduce a new integrative joint modeling framework for the simultaneous description of single-trial EEG measures and cognitive modeling parameters of decision-making. As specific examples, we formalized how single-trial N200 latencies and centro-parietal positivities (CPPs) are predicted by changing single-trial parameters of various drift-diffusion models (DDMs). We trained deep neural networks to learn Bayesian posterior distributions of unobserved neurocognitive parameters based on model simulations. These models do not have closed-form likelihoods and are not easy to fit using Markov chain Monte Carlo (MCMC) methods because nuisance parameters on single trials are shared in both behavior and neural activity. We then used parameter recovery assessment and model misspecification to ascertain how robustly the models’ parameters can be estimated. Moreover, we fit the models to three different real datasets to test their applicability. Finally, we provide some evidence that single-trial integrative joint models are superior to traditional integrative models. The current single-trial paradigm and the simulation-based (likelihood-free) approach for parameter recovery can inspire scientists and modelers to conveniently develop new neurocognitive models for other neural measures and to evaluate them appropriately.
... However, for most electrophysiology and brain imaging, a perfect measure of the brain is unrealistic. This is especially true for EEG where derived measures of EEG are expected to be influenced by muscle and electrical artifacts, even when the measure of EEG is somewhat robust to artifact (Nunez et al., 2016). While other forms of integrated neurocognitive models could be beneficial for studying human cognition (see Appendix), the key benefit of our frameworks is that these models contain at least one source of measurement noise for the EEG measures, σ 2 , as well as variance in the underlying cognition that can change across trials s 2 . ...
... While other forms of integrated neurocognitive models could be beneficial for studying human cognition (see Appendix), the key benefit of our frameworks is that these models contain at least one source of measurement noise for the EEG measures, σ 2 , as well as variance in the underlying cognition that can change across trials s 2 . By measurement noise we mean the combination of electrical noise, artifacts due to EMG, and unrelated cognitive processes embedded in all measures of EEG (Whitham et al., 2007;Nunez et al., 2016). Measurement noise is especially relevant for EEG on single trials (Bridwell et al., 2018) because we are often not using measures that are robust to EEG artifacts, like trialaveraged ERPs (although even trial-averaged ERPs are not completely robust to artifact either) (Nunez et al., 2016). ...
... By measurement noise we mean the combination of electrical noise, artifacts due to EMG, and unrelated cognitive processes embedded in all measures of EEG (Whitham et al., 2007;Nunez et al., 2016). Measurement noise is especially relevant for EEG on single trials (Bridwell et al., 2018) because we are often not using measures that are robust to EEG artifacts, like trialaveraged ERPs (although even trial-averaged ERPs are not completely robust to artifact either) (Nunez et al., 2016). Note also that preprocessing steps for EEG measures often have many degrees of freedom (Clayson et al., 2021), and therefore models should provide some quantification of the amount of relevant cognition described by single-trial EEG measures compared to the amount of extraneous information. ...
Full-text available
Despite advances in techniques for exploring reciprocity in brain-behavior relations, few studies focus on building neurocognitive models that describe both human EEG and behavioral modalities at the single-trial level. Here, we introduce a new integrative joint modeling framework for the simultaneous description of single-trial EEG measures and cognitive modeling parameters of decision making. The new framework can be utilized for the evaluation of research questions as well as the prediction of both data types concurrently. In the introduced joint models, we formalized how single-trial N200 latencies and Centro-parietal positivities (CPP) are predicted by changing single-trial parameters of various drift-diffusion models (DDMs). These models do not have clear closed-form likelihoods and are not easy to fit using Markov chain Monte Carlo (MCMC) methods because nuisance parameters on single trials are shared in both behavior and neural activity. We trained deep neural networks to learn the Bayesian posterior distributions of unobserved neurocognitive parameters based on model simulations. We then used parameter recovery assessment and model misspecification to ascertain how robustly the models’ parameters can be estimated. Moreover, we fit the models to three different real datasets to test their applicability. Our results show that the single-trial integrative models can recover their latent parameters. Finally, we provide some pieces of evidence that single-trial integrative joint models are superior to traditional integrative models. The current single-trial paradigm and the likelihood-free approach for parameter recovery can inspire scientists and modelers to conveniently develop new neuro-cognitive models for other neural measures and to evaluate them appropriately.
... Although epoching is not necessary for resting-state datasets, the signals can be segmented into non-overlapping segments of a few seconds to facilitate the analysis and reduce the effect of the nonstationary property of EEG (Cohen, 2014;Sakkalis, 2011). 5. Removing epochs because of artefacts: Epochs that obviously contain artefacts rather than EEG, such as ocular artefact, can be automatically rejected through the EEGLAB toolbox, for example, by examining epochs with high variance compared to other epochs. Such artefacts can also be identified by visual inspection (Nunez et al., 2016). Not removing these artefacts hinders the ability of the independent components analysis algorithm, described next, to isolate typical artefacts such as eye blinks (Nunez et al., 2016). ...
... Such artefacts can also be identified by visual inspection (Nunez et al., 2016). Not removing these artefacts hinders the ability of the independent components analysis algorithm, described next, to isolate typical artefacts such as eye blinks (Nunez et al., 2016). ...
... ICA is limited by the fact that the identification of the artefact component is inherently a subjective judgement (Nunez et al., 2016). In addition, the components are likely to contain both signal and noise, and thus, one should be cautious about removing a component that seems to contain a signal. ...
Neurodevelopmental disorders (NDDs) are a group of neurological disorders emerging during early development and impacting higher-level brain function. Early identification of NDDs raises the possibility of improving outcomes. However, NDDs are still clinically detected through a subjective evaluation approach that lacks biological evidence and requires an extended period to identify the disorder. This thesis aims to identify biomarkers from the electroencephalogram (EEG) for abnormal brain dynamics and utilise these biomarkers for NDDs prediction. Nonlinear time series analysis methods, mainly complexity (entropies) and synchronisation (node degree of the phase-lag index) features are investigated for this purpose due to their ability to reflect brain dynamics. A machine learning framework that combines advanced nonlinear EEG processing methods for NDDs identification has been developed and evaluated. The proposed framework was first used with a dataset of children with autism spectrum disorder (ASD) and controls to validate its classification efficacy. This exploration resulted in high performance, suggesting complexity features as biomarkers for ASD prediction. The framework was then used to explore at-birth EEG characteristics of infants with Hypoxic-Ischemic Encephalopathy (HIE) who later developed cerebral palsy (CP). High performance has been reached, and the proposed features were reported as potential biomarkers for early CP prediction. A regression model has also been developed to explore the correlation between the EEG characteristics of the HIE infants and their two-year cognitive scores. The results suggest the proposed features as biomarkers for early cognitive function prediction. Finally, the framework was used with a dataset of individuals with a major depressive disorder to validate its ability to predict their depression severity. Compared to the state-of-the-art research, an acceptable regression performance has been reached, and the proposed features have been suggested as biomarkers for depression severity prediction. This work lays the foundation for evidence-based decision-making applications for early prediction of CP and cognitive outcomes of infants with HIE, paving the way for establishing tailored intervention programs at an appropriate point during development to improve the outcomes.
... In EEG analysis, a common objective is to perform inference on the signals that are decomposed/filtered into different frequency bands (e.g., delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), beta (12)(13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30), and gamma (30-100 Hz)) (Ombao et al., 2005;Nunez et al., 2016;Guerrero et al., 2021) which have well-known associated cognitive functions (Harmony, 2013;Bjørge and Emaus, 2017). To increase the applicability of TE in neuroscience, its development in the frequency domain becomes relevant. ...
Full-text available
Brain connectivity reflects how different regions of the brain interact during performance of a cognitive task. In studying brain signals such as electroencephalograms (EEG), this may be explored via Granger causality (GC) which tests if knowledge of the past values of a channel improves predictions on future values of another channel. However, the common approach to investigating GC is the vector autoregressive (VAR) model which is limited only to linear lead-lag relations. An alternative information-theoretic causal measure, transfer entropy (TE), becomes more appropriate since it does not impose any distributional assumption on the variables and covers any form of relationship (beyond linear) between them. To improve utility of TE in brain signal analysis, we propose a novel methodology to capture cross-channel information transfer in the frequency domain. Specifically, we introduce a new measure, the spectral transfer entropy (STE), to quantify the magnitude and direction of information flow from a certain frequency-band oscillation of a channel to an oscillation of another channel. In contrast with previous works on TE in the frequency domain, we differentiate our work by considering the magnitude of filtered series (frequency band-specific), instead of using the spectral representation (frequency-specific) of a series. The main advantage of our proposed approach is that it allows adjustments for multiple comparisons to control family-wise error rate (FWER). One novel contribution is a simple yet efficient estimation method based on vine copula theory that enables estimates to capture zero (boundary point) without the need for bias adjustments. We showcase the advantage of our proposed measure through some numerical experiments and provide interesting and novel findings on the analysis of EEG recordings linked to a visual task.
Full-text available
We compared Spanish (L1)-English (L2) bilinguals and Spanish monolinguals in a semantic judgment relationship task in L1 that produced within-language conflict due to the coactivation of the two meanings of a Spanish homophone (e.g., "hola" and "ola" meaning "hello" and "a wave" in English). In this task, participants indicated if pairs of words were related or not ("agua-hola," "water-hello"). Conflict arose because a word ("agua," "water") not related to the orthographic form of a homophone ("hola," "hello") was related to the alternative orthographic form ("ola," "wave"). Compared to a control condition with unrelated word pairs ("peluche-hola," "teddy-hello"), the behavioral results revealed greater behavioral interference in monolinguals compared to bilinguals. In addition, electrophysiological results revealed N400 differences between monolinguals and bilinguals. These results are discussed around the impact of bilingualism on conflict resolution.
Full-text available
Trance is an altered state of consciousness characterized by alterations in cognition. In general, trance states induce mental silence (i.e., cognitive thought reduction), and mental silence can induce trance states. Conversely, mind-wandering is the mind's propensity to stray its attention away from the task at hand and toward content irrelevant to the current moment, and its main component is inner speech. Building on the previous literature on mental silence and trance states and incorporating inverse source reconstruction advances, the study's objectives were to evaluate differences between trance and mind-wandering states using: (1) electroencephalography (EEG) power spectra at the electrode level, (2) power spectra at the area level (source reconstructed signal), and (3) EEG functional connectivity between these areas (i.e., how they interact). The relationship between subjective trance depths ratings and whole-brain connectivity during trance was also evaluated. Spectral analyses revealed increased delta and theta power in the frontal region and increased gamma in the centro-parietal region during mind-wandering, whereas trance showed increased beta and gamma power in the frontal region. Power spectra at the area level and pairwise comparisons of the connectivity between these areas demonstrated no significant difference between the two states. However, subjective trance depth ratings were inversely correlated with whole-brain connectivity in all frequency bands (i.e., deeper trance is associated with less large-scale connectivity). Trance allows one to enter mentally silent states and explore their neurophenomenological processes. Limitations and future directions are discussed.
We explored the underlying latent process of spatial prioritisation in perceptual decision processes, based on the drift-diffusion model, and subsequent nested model comparison. Our hierarchical cognitive modelling analysis revealed that spatial attention changed the non-decision time parameter across experimental conditions, quantified using the deviance information criterion score (DIC) and R-squared. We also constructed joint models with embedded neural covariates to discover which of contralateral and ipsilateral EEG measures could most manipulate spatial attention in perceptual decision making. Using multiple regression analysis, it can be concluded that poststimulus N2nc can predict mean response times and non-decision time parameters related to spatial prioritisation. However, the contralateral minus neutral alpha power at parieto-occipital electrodes can only predict the mean RTs and not the non-decision time relating to spatial prioritisation. Results suggest that individual differences in spatial attention are encoded more by contralateral (and not ipsilateral) N2 oscillations and non-decision times.
Full-text available
Applications of quantum mechanics and physics-based calculus allow for advanced mathematical modeling of source localization estimations. Because EEG waveforms can be modeled as continuous or discrete quantum matter, algorithmic models that estimate current source density must also consider the statistical properties of the dipole fields that are etiologically relevant to the reflected waveforms. Further applications of quantum physics to the electroencephalogram (EEG) suggest that neurodynamic behavior that originates in thalamo-cortical regions compared to cortico-cortical regions yield divergent 3-dimensional dispersions of wave forms and quantum energy. Evaluations of the dispersion of quantum energy and dipole magnetic fields according to classical physics and electromagnetism indicate that the area of tissue by which the oscillatory mechanisms are thought to originate inherently influences the algorithmic modeling and estimations current source density. Principles and algorithms related to classical physics are included within this paper to evaluate limitations of algorithmic models of source localization and the inverse problem.
Full-text available
Objective: High frequency oscillations (HFOs) recorded by intracranial electrodes have generated excitement for their potential to help localize epileptic tissue for surgical resection. However, the number of HFOs per minute (i.e. the HFO "rate") is not stable over the duration of intracranial recordings; for example, the rate of HFOs increases during periods of slow-wave sleep. Moreover, HFOs that are predictive of epileptic tissue may occur in oscillatory patterns due to phase coupling with lower frequencies. Therefore, we sought to further characterize between-seizure (i.e. "interictal") HFO dynamics both within and outside the seizure onset zone (SOZ). Approach: Using long-term intracranial EEG (mean duration 10.3 hours) from 16 patients, we automatically detected HFOs using a new algorithm. We then fit a hierarchical Negative Binomial model to the HFO counts. To account for differences in HFO dynamics and rates between sleep and wakefulness, we also fit a mixture model to the same data that included the ability to switch between two discrete brain states that were automatically determined during the fitting process. The ability to predict the SOZ by model parameters describing HFO dynamics (i.e. clumping coeffcients and coeffcients of variation) was assessed using receiver operating characteristic curves. Main results: Parameters that described HFO dynamics were predictive of SOZ. In fact, these parameters were found to be more consistently predictive than HFO rate. Using concurrent scalp EEG in two patients, we show that the model-found brain states corresponded to (1) non-REM sleep and (2) awake and rapid eye movement sleep. However, the brain state most likely corresponding to slow-wave sleep in the second model improved SOZ prediction compared to the first model for only some patients. Significance: This work suggests that delineation of seizure onset zone with interictal data can be improved by the inclusion of time-varying HFO dynamics.
Full-text available
Sequential sampling decision-making models have been successful in accounting for reaction time (RT) and accuracy data in two-alternative forced choice tasks. These models have been used to describe the behavior of populations of participants, and explanatory structures have been proposed to account for between individual variability in model parameters. In this study we show that individual differences in behavior from a novel perceptual decision making task can be attributed to (1) differences in evidence accumulation rates, (2) differences in variability of evidence accumulation within trials, and (3) differences in non-decision times across individuals. Using electroencephalography (EEG), we demonstrate that these differences in cognitive variables, in turn, can be explained by attentional differences as measured by phase-locking of steady-state visual evoked potential (SSVEP) responses to the signal and noise components of the visual stimulus. Parameters of a cognitive model (a diffusion model) were obtained from accuracy and RT distributions and related to phase-locking indices (PLIs) of SSVEPs with a single step in a hierarchical Bayesian framework. Participants who were able to suppress the SSVEP response to visual noise in high frequency bands were able to accumulate correct evidence faster and had shorter non-decision times (preprocessing or motor response times), leading to more accurate responses and faster response times. We show that the combination of cognitive modeling and neural data in a hierarchical Bayesian framework relates physiological processes to the cognitive processes of participants, and that a model with a new (out-of-sample) participant's neural data can predict that participant's behavior more accurately than models without physiological data.
Objective: Validate independent component analysis (ICA) for removal of EMG contamination from EEG, and demonstrate a heuristic, based on the gradient of EEG spectra (slope of graph of log EEG power vs log frequency, 7-70Hz) from paralysed awake humans, to automatically identify and remove components that are predominantly EMG. Methods: We studied the gradient of EMG-free EEG spectra to quantitatively inform the choice of threshold. Then, pre-existing EEG from 3 disparate experimental groups was examined before and after applying the heuristic to validate that the heuristic preserved neurogenic activity (Berger effect, auditory odd ball, visual and auditory steady state responses). Results: (1) ICA-based EMG removal diminished EMG contamination up to approximately 50Hz, (2) residual EMG contamination using automatic selection was similar to manual selection, and (3) task-induced cortical activity remained, was enhanced, or was revealed using the ICA-based methodology. Conclusion: This study further validates ICA as a powerful technique for separating and removing myogenic signals from EEG. Automatic processing based on spectral gradients to exclude EMG-containing components is a conceptually simple and valid technique. Significance: This study strengthens ICA as a technique to remove EMG contamination from EEG whilst preserving neurogenic activity to 50Hz.
Valid biomarkers of motor system function after stroke could improve clinical decision-making. Electroencephalography-based measures are safe, inexpensive, and accessible in complex medical settings and so are attractive candidates. This study examined specific electroencephalography cortical connectivity measures as biomarkers by assessing their relationship with motor deficits across 28 days of intensive therapy. Resting-state connectivity measures were acquired four times using dense array (256 leads) electroencephalography in 12 hemiparetic patients (7.3 ± 4.0 months post-stroke, age 26-75 years, six male/six female) across 28 days of intensive therapy targeting arm motor deficits. Structural magnetic resonance imaging measured corticospinal tract injury and infarct volume. At baseline, connectivity with leads overlying ipsilesional primary motor cortex (M1) was a robust and specific marker of motor status, accounting for 78% of variance in impairment; ipsilesional M1 connectivity with leads overlying ipsilesional frontal-premotor (PM) regions accounted for most of this (R(2) = 0.51) and remained significant after controlling for injury. Baseline impairment also correlated with corticospinal tract injury (R(2) = 0.52), though not infarct volume. A model that combined a functional measure of connectivity with a structural measure of injury (corticospinal tract injury) performed better than either measure alone (R(2) = 0.93). Across the 28 days of therapy, change in connectivity with ipsilesional M1 was a good biomarker of motor gains (R(2) = 0.61). Ipsilesional M1-PM connectivity increased in parallel with motor gains, with greater gains associated with larger increases in ipsilesional M1-PM connectivity (R(2) = 0.34); greater gains were also associated with larger decreases in M1-parietal connectivity (R(2) = 0.36). In sum, electroencephalography measures of motor cortical connectivity-particularly between ipsilesional M1 and ipsilesional premotor-are strongly related to motor deficits and their improvement with therapy after stroke and so may be useful biomarkers of cortical function and plasticity. Such measures might provide a biological approach to distinguishing patient subgroups after stroke. © The Author (2015). Published by Oxford University Press on behalf of the Guarantors of Brain. All rights reserved. For Permissions, please email: