CHAPTER FIFTEEN

Brief Introduction to
Electroencephalography
Alex Proekt1
Perelman School of Medicine, Department of Anesthesiology and Critical Care, University of Pennsylvania,
Philadelphia, PA, United States
1
Corresponding author: e-mail address: [email protected]

Contents
1. What Does EEG Measure? 257
2. Technical Considerations for Recording EEG Signals 260
2.1 Reference Electrode 260
2.2 Filtering and Sampling Frequency 261
2.3 Artifact Rejection 263
3. Brief Introduction to Spectral Analysis of the EEG 265
4. Multivariate Spectral Analysis: Coherence and Phase Lag Index 268
5. Effects of Anesthetics on the EEG 270
6. Concluding Remarks 273
References 273

Abstract
Electroencephalography (EEG) has a long history in neuroscience starting with its orig-
inal description in humans by Hans Berger in 1929 (Berger, 1932). Investigations of EEG
under anesthesia started soon after in the mid-1930s (Gibbs, 1937). No single method-
ology paper can credibly cover all of the issues relating to this rich field. The purpose of
this chapter is to introduce some caveats that complicate and inform analysis of the EEG.
Special emphasis will be given to common issues such as choice of reference electrode,
filtering, artifact rejection, and spectral analysis. We will specifically emphasize high-
density EEG recordings that have become the norm due to technological improvement
in electrode and data acquisition design methods. In the last section we will discuss
some applications of EEG analysis techniques to the study of the effects of anesthetics
on the nervous system.

1. WHAT DOES EEG MEASURE?

The neurophysiological basis of Electroencephalography (EEG) sig-
nals will not be covered here in any significant detail. Please consult
Niedermeyer and da Silva (2005) and Nunez and Srinivasan (2006) for a

Methods in Enzymology, Volume 603 # 2018 Elsevier Inc. 257

ISSN 0076-6879 All rights reserved.
https://doi.org/10.1016/bs.mie.2018.02.009
258 Alex Proekt

detailed discussion of the biophysical origins of EEG signals. In what follows

we will briefly introduce the key issues that inform and complicate the inter-
pretation of the EEG.
Electrical signals are fundamental to the functioning of the nervous sys-
tem. Neurons maintain a negative transmembrane potential (typically in the
neighborhood of
70 mV) relative to the extracellular space and can fire
action potentials (brief voltage impulses) depending on the synaptic inputs
(postsynaptic potentials). This transmembrane voltage (Vm) can be measured
as the potential difference between the inside of the cell and ground.
Recordings of Vm provide the most direct measure of neuronal activity.
Yet, these measurements are not easily obtained in the intact vertebrate brain
and are almost never available in human subjects.
Any ion movement across an excitable membrane be that of a soma, den-
drite, or an axon must be accompanied by movement of ions in the extra-
cellular space. Thus, transmembrane currents contribute to the electric field
that can be recorded extracellularly. The strength of this electric field (Ve)
can be recorded as a potential difference between two electrodes one of
which is defined as reference. EEG is a specific case of Ve that involves plac-
ing electrodes onto the skull. Many different electrode configurations exist
and are to various degrees used in clinical and research settings. Here we will
focus on the discussion of high-density EEG. Currently high density refers
to >100 channels of EEG recorded simultaneously from different points on
the skull.
There is a fundamental difference between Vm and Ve: while Vm
reflects the voltage state of a single neuron, Ve necessarily reflects a com-
plex mixture of contributions from myriad neuronal and glial sources
(Buzsáki, Anastassiou, & Koch, 2012; Dietzel, Heinemann, & Lux,
1989). Thus, for all extracellular signals including the EEG (see below)
it is crucial to distinguish between the sensors (i.e., electrodes where the
potential is recorded) and the sources (neuronal substrates and processes
that give rise to the observed extracellular potential).
If one knew the sources of neuronal activity, i.e., currents flowing
through each patch of membrane in each neuron, one could in principle
predict the voltage recorded by an extracellular electrode (Holt & Koch,
1999; Rall & Shepherd, 1968). This approach has recently been incorpo-
rated into a simulation environment (Lind
en et al., 2013). Going from
the microscopic variables such as current fluxes across patches of membrane
to macroscopic variables such as extracellular field is known as the forward
problem. The opposite—i.e., going from the macroscopic variables such as
Brief Introduction to Electroencephalography 259

the local field or EEG to microscopic variables—is known as the “inverse

problem.” Unlike the forward problem, the inverse problem does not have
a unique solution—many different microscopic current fluxes can give rise
to the same macroscopic potential. Thus, the EEG cannot inform us directly
about activity of individual neurons (Nunez & Srinivasan, 2006). That
being said, there is significant research effort aimed at constraining the
likely distribution of sources that could give rise to the observed EEG sig-
nals (Grech et al., 2008).
While Ve ultimately reflects many components of neuronal activity such
as action and postsynaptic potentials, under most conditions Ve in general
and EEG in particular is dominated by the contribution of postsynaptic
potentials. There are several reasons for this. The primary reason is that in
order to be discernible in the extracellular recording, transmembrane cur-
rents in many sources must overlap in time. Action potentials last approxi-
mately 1 ms and are thus extremely unlikely to have significant temporal
overlap between different neurons in a small volume of brain tissue. In con-
trast postsynaptic potentials can last 10s of milliseconds and are much more
synchronous across neurons (Buzsáki et al., 2012; Niedermeyer & da Silva,
2005). Furthermore, because the amplitude of the electric fields decays
rapidly as a function of distance from the source, extracellular voltage will
be dominated by sources closest to it. Because EEG is placed onto the surface
of the head, the closest electrical sources are located in the superficial cortical
layers which contain apical dendrites of pyramidal neurons. Apical dendrites
of pyramidal neurons have a preferred orientation normal to the cortical
surface. Consequently many different dendrites will have current flow
in approximately the same direction. Thus, contributions from multiple
dendrites add to the extracellular field rather than cancel each other out
(Kirschstein & K€ ohling, 2009). Synaptic inputs into apical dendrites come
from the so-called matrix projections originating in the thalamus (Jones,
1998, 2001) with some contribution from corticocortical synapses. Thus,
EEG signal is dominated by the fluctuations of the synaptic input into
apical dendrites of cortical pyramidal cells which reflect reverberating
activity of thalamocortical loops.
Last important issue to mention in the context of origins of EEG signals is
the issue of spatial scale. A small electrode placed into the brain parenchyma
will faithfully record extracellular fields dominated by the neuronal sources
in its immediate vicinity. For instance, using silicone multielectrode arrays
spacing less than 1 mm between electrodes can be readily attained in exper-
imental settings (Buzsáki et al., 2015). This allows recording of activity of
260 Alex Proekt

single neurons. In contrast, even with high-density EEG recordings each

channel of the EEG records electrical activity arising from 10 cm2 patch
of cortex or more. In order to be recorded in the EEG despite being far from
the source of the signal and separated by cerebrospinal fluid, skull, muscle,
and scalp, the local fields across areas of cortex many orders of magnitude
larger than a single neuron have to be correlated (Nunez & Srinivasan,
2010). EEG is therefore primarily suited for investigating fluctuations in this
macroscopic-scale brain activity. Thus, because spatial correlations between
local fields depend on the global brain states (Destexhe, Contreras, &
Steriade, 1999), it is not surprising that changes in global brain states such
as loss and recovery of consciousness induced with general anesthetics are
readily observed in the EEG (Gibbs, 1937).

2. TECHNICAL CONSIDERATIONS FOR RECORDING EEG

SIGNALS
2.1 Reference Electrode
EEG amplifier records potential difference between each electrode and
reference. Clearly, the choice of reference will influence the recorded sig-
nals. Ideally, reference should be electrically stable and have a zero potential.
Thus, it may seem appropriate to place the reference electrode far away from
the subject. In practice, however, the long wires that would be required for
this would serve as antennae that would pick up ambient electrical noise.
Because EEG signals are small (microvolts) recording noise can readily
overwhelm the signal. Thus, traditionally, reference is placed on the sub-
ject. Many different reference choices have been used in the literature such
as the tip of the nose (Andrew & Pfurtscheller, 1996), vertex (Lehmann,
Strik, Henggeler, Koenig, & Koukkou, 1998), mastoid or ear (Başar,
Rahn, Demiralp, & Sch€ urmann, 1998), ring of electrodes around the neck
(Kuklinski, 1983; Nunez & Srinivasan, 2006), bilateral mastoids or ears
(Gevins & Smith, 2000), or average of all recorded electrodes (Offner,
1950). Unfortunately, none of these points on the body have a true zero
electrical potential, and thus, reference choice has a strong influence on
the signal recorded in each electrode. There is a theoretical advantage to
the average reference in that the surface integral of the potential of a conductor
is zero under most circumstances. This advantage is only well realized in sit-
uations where a large number of electrodes are used (high-density EEG). If the
EEG montage has a small number of electrodes, the average reference can
introduce complex artifacts into the recordings. Furthermore, care must be
Brief Introduction to Electroencephalography 261

taken to remove all of the noisy electrodes before averaging. For this rea-
son, we suggest first recording the data with a single-electrode reference
(e.g., mastoid or vertex). If conversion to an average or other (see below)
reference is desired, this can be done in postprocessing.
While physically recording the signal with respect to a distant reference is
technically not feasible because of electrical noise, recently a mathematical
algorithm has been proposed to simulate this recording configuration
(Chella, Marzetti, Pizzella, Zappasodi, & Nolte, 2014; Yao et al., 2005).
This reference electrode standardization technique (REST) model is based
on an assumption of a “three concentric spheres” model of head conduc-
tance (Nunez & Srinivasan, 2006; Rush & Driscoll, 1969). Some simula-
tion studies (Lei & Liao, 2017) have demonstrated advantages of REST
reference, and this reference configuration has been used in a number of
experimental settings (Chella et al., 2014; Chella, Pizzella, Zappasodi, &
Marzetti, 2016; Liu et al., 2015; Tian & Yao, 2013; Yao et al., 2005).
The advantage of REST reference relative to average is moderate for a
high-density EEG recording in simulation studies (Lei & Liao, 2017).
The algorithm for computing REST reference has been recently published
in an open-source MATLAB® toolbox (Dong et al., 2017). The best choice
of reference may depend on the specifics of the analysis performed, but it
most strongly affects analytic techniques that look at the interrelations
between different EEG signals (e.g., coherence, connectivity, etc.). Thus,
it seems prudent to us to verify that the overall conclusions are not strongly
affected by the choice of reference scheme. The only way to currently
accomplish this is to build in different reference schemes into the analysis
pipeline.

2.2 Filtering and Sampling Frequency

All modern applications of the analysis of electrophysiological signals such as
the EEG are performed digitally. That is signals are sampled with a certain
sampling rate onto a digital computer where analysis is subsequently per-
formed. Thus, an experimenter needs to decide how frequently to sample
the data. If the data are sampled too frequently, the resulting datasets will
be unnecessarily large and this would complicate subsequent analysis. Con-
versely, if the data are sampled too sparsely in time, then information will be
lost. Decisions concerning sampling frequency should be based on Nyquist
theorem which states that if the highest frequency of interest in the data is f
then the sampling interval dt between consecutive samples should be 1/(2f ).
262 Alex Proekt

Most power in the EEG signals is contained in the relatively low frequencies
(Niedermeyer & da Silva, 2005; Nunez & Srinivasan, 2006), thus sampling
of EEG beyond 1 kHz is probably unnecessary. Downsampling from the
original sampling rate can be performed post hoc if desired for speeding
up the analysis code. To avoid introducing sharp discontinuities in the signal
we suggest using a low-pass filter prior to downsampling rather than simply
removing some data points.
Most forms of EEG analysis require some filtering of the data. Filtering
generally refers to selectively attenuating some frequencies for the purposes
of enhancing others. Typically filters fall into several categories, such as high
pass, low pass, band pass, notch, based on the kind of frequencies they select
for. For instance, high-pass filter selectively attenuates lower frequencies,
while low-pass filter attenuates higher frequencies. Filters can differ by their
design such as for instance finite impulse response and infinite impulse
response (IIR) filters. Much like the choice of reference, choice of filter
type and filter settings can have dramatic consequences on the data. For
instance, consider a simple signal that is constructed by adding two sine
waves (Fig. 1). In order to extract just the higher frequency oscillation a
sixth order Butterworth (IIR) high-pass filter was applied to the signal.
In the filtered waveform (dashed red trace in the bottom panel), the slow
oscillation is indeed suppressed well. This is because the two frequencies
were chosen to be quite different from each other for the purposes of illus-
tration. Yet, the filtered waveform does not recover the original sine wave
that went into the construction of the signal (compare the dashed and the
solid line). In this example, the filtered waveform is phase shifted relative to

Fig. 1 Filtering introduces phase shifts. The signal of interest consists of two sine
waves: slow (black) and fast (red) in the top panel. The overall signal is the sum of
the two oscillations (blue) in the middle panel. A sixth order high-pass Butterworth filter
was constructed in MATLAB® using butter command and applied to the data. The cut-
off frequency was set in the middle of the frequency interval separating the fast and
the slow oscillation. The filtered signal (dashed bottom) does indeed recover the fast
oscillation. However, this oscillation is phase shifted with respect to the original signal
(red solid line).
Brief Introduction to Electroencephalography 263

the original sine wave. This is just one of many distortions in the signal that
filters can induce (Percival & Walden, 1993). Thus, it is our opinion that fil-
tering should be minimized. This is especially relevant for applications where
phases of oscillations are important such as phase coupling which constitutes
the basis of the bispectrum—a mathematical technique at the core of the BIS
index of anesthetic depth (Liu, Singh, & White, 1996). Some of the distor-
tions induced by filtering such as phase delays can be minimized by using post
hoc filtering of the data rather than using the filters built into the EEG ampli-
fier. One simple method, for instance, is to run the same filter forward and
backward along the same strip of data. This way the phase delays cancel each
other out, and the overall signal phase is better preserved. This method is
implemented in the MATLAB® filtfilt command that we find rather useful
in analysis of EEG data.
Despite best recording practices it is not always possible to completely
remove the 60 Hz (or 50 Hz outside the US) oscillation from the original
recording. Thus, a common practice is to apply a notch filter. A notch
(also known as band stop) filter selectively attenuates oscillations in a narrow
frequency band. Yet, band stop filters are most problematic. Conventional
notch filters are known to produce band holes in the spectrum, significant
distortions around the notched frequencies, as well as phase distortions and
Gibbs rippling in the time domain. An alternative approach to conventional
notch filter has been proposed in Mitra (2007). The basic ideas are that the
signal is first subjected to a sliding window multitaper spectral estimation
(see below). Then a Thomson spectral F-test is computed for the regression
coefficient fit to each complex-valued element in the spectrum. This detects
statistically significant deviations of a particular narrow-band signal from the
expected spectrum. The complex-valued spectral coefficients identified
using this method can be used to construct a sine wave which is then sub-
tracted from the data in the time domain. This avoids many of the pitfalls of
conventional notch filtering. The code for performing this type of filtering
can be found in the Chronux toolbox for MATLAB® or in CleanLine plu-
gin for EEGLAB (Bokil, Andrews, Kulkarni, Mehta, & Mitra, 2010).

2.3 Artifact Rejection

The importance of proper artifact rejection cannot be overstated for
processing any EEG recording. Because EEG electrodes are placed on the
skull rather than directly onto the brain, the electrical signals recorded by
the electrodes are not solely due to electrical fields generated in the brain.
264 Alex Proekt

Figure Display Settings Help

Stack

Norm
Blink Blink

Scale
23

110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
Muscle Chan. Time Value +
CLOSE << < 110 > >> 23
78 112.4321 –1.8351 –

Fig. 2 Artifacts in EEG. Snapshot of high density (128 channels) of EEG during the awake
state in a human subject. The signals are displayed using the eegplot function. This func-
tion is a part of the EEGLAB toolbox and is quite useful in visually inspecting the signals.
Blink artifacts are much larger in amplitude than the EEG signals. Muscle artifact appears
as characteristic high-frequency oscillations. Removing these artifacts is critical for
further analysis. Using eegplot function, artifacts can be marked in the signal. Regions
that contain artifacts can be excluded from subsequent analysis.

For instance, electrical activity in the muscles can be much larger in ampli-
tude than electrical activity generated by the brain. Other very common
sources of artifacts are blinks and movements (e.g., Fig. 2). There are two
classes of approaches to artifact removal: remove the segments of data con-
taining the artifact, and remove the artifact from the data. Analysis of most
datasets involves some combination of both of these approaches. Obviously
it is desirable to keep as much of the data as possible. Thus, we will briefly
mention here some techniques that can be used to remove the artifacts from
the data. One way to remove the muscle artifact is to simply low-pass filter
the data below 30 Hz or so. Most EMG activity occurs at faster frequencies
and filtering will therefore attenuate activity due to muscle. That being said,
some neuronal activity of interest such as gamma oscillations, for instance,
does occur above 30 Hz, and this way of removing muscle artifact will also
remove the signal of interest.
Another popular approach that has been successfully used to remove arti-
facts from EEG data is based on an assumption that the spatial location of
sources of noise (and neuronal activity) is constant. For instance, the elec-
trical fields generated by eye blinks ought to be consistently stronger in the
frontal electrodes. A class of methods known as the independent compo-
nent analysis (ICA) is a particularly useful tool with multiple publications
in EEG literature (Delorme & Makeig, 2004; Delorme, Sejnowski, &
Makeig, 2007; Jung, Makeig, Bell, & Sejnowski, 1998; Vigário, 1997;
Brief Introduction to Electroencephalography 265

Vigario, Sarela, Jousmiki, Hamalainen, & Oja, 2000, to name a few). ICA
is at the core of a popular open-source MATLAB®-based EEG analysis
toolbox called EEGLAB (Delorme & Makeig, 2004) and is also utilized
in another MATLAB®-based EEG toolbox called Fieldtrip (Oostenveld,
Fries, Maris, & Schoffelen, 2011). There are several fast algorithms that
compute ICA (e.g., Bell & Sejnowski, 1997; Oja & Yuan, 2006), but they
all seek to accomplish the same goal—to breakdown the multivariate
distribution of observed data (multiple channels of EEG recordings in
this case) into multiple independent and non-Gaussian sources that can
be added together to reconstruct the original signal. In the case of ICA,
“independence” is accomplished by attempting to minimize the mutual
information between signals (Bell & Sejnowski, 1997). Deviation from
Gaussian distribution is attained by maximizing kurtosis. At the comple-
tion of an ICA run, one ends up with a mixing and unmixing matrix.
These matrices quantify how each individual channel of EEG contributes
to the activation of each independent component. Rather than studying
statistical properties of the raw signals (e.g., spectrum) recorded in each
EEG channel, one can instead study statistical properties of each indepen-
dent component (e.g., Fig. 3) (Delorme & Makeig, 2004). This figure
illustrates a typical application of ICA to identify and remove blink artifacts
from EEG data. Once an independent component corresponding to blinks
for instance is identified, it can be removed from the unmixing matrix
without totally removing all of the data that occurs during the blink. While
there are many published methods for automatic EEG artifact rejection
(Krishnaveni, Jayaraman, Anitha, & Ramadoss, 2006; LeVan, Urrestarazu,
& Gotman, 2006; Li, Ma, Lu, & Li, 2006; Mognon, Jovicich, Bruzzone,
& Buiatti, 2011; Shao, Shen, Ong, Wilder-Smith, & Li, 2009, to name
a few), we strongly advocate using semiautomated methods where the data
are processed using one of the above-mentioned algorithms but is subse-
quently inspected by an expert observer.

3. BRIEF INTRODUCTION TO SPECTRAL ANALYSIS OF

THE EEG
It is believed that oscillations play a significant role in neuronal
processing (Buzsaki, 2004). Thus, many analytic techniques for EEG and
other forms of neuronal activity recordings focused on identifying and char-
acterizing the different oscillations. One very broad class of methods that has
been used extensively in neuroscience in general and in analysis of EEG in
266 Alex Proekt

Time course of Scalp maps of

Original EEG ICA components Corrected EEG
ICA components

Fp1 1 Fp1
Fp2 2 Fp2
F3 3 F3
F4 4 F4
C3 5 C3
C4 6 C4
A2 7 A2
P3 8 P3
P4 9 P4
O1 10 O1
O2 11 O2
F7 12 F7
F8 13 F8
T3 14 T3
T4 15 T4
T5 16 T5
T6 17 T6
Fz 18 Fz
Cz 19 Cz
Pz 20 Pz
EOG1 21 EOG1
EOG2 22 EOG2

0 1 2 3 0 1 2 3 0 1 2 3
Time (s)
Fig. 3 Artifact rejection using independent component analysis. Image from the
EEGLAB tutorial. Close-up of blink and muscle EEG artifact is shown on the left. Note
the similarity to those shown in Fig. 2. ICA decomposition of the EEG converts the data
from channels  time matrix to component  time matrix. Note that activation of com-
ponents 1 and 2 closely corresponds to the blink artifacts, while the high frequency of
activity in components 14 and 15 appears similar to the muscle activity. This intuition is
corroborated by the fact that the majority of the blink artifact is found in the frontal
leads (third column), while muscle artifacts are most closely related to temporalis activ-
ity. Components that contain artifacts (Berger, 1932; Buzsáki et al., 2015; Gibbs, 1937;
Nunez & Srinivasan, 2010) can be removed from the mixing matrix. This matrix can
be inverted to then recover the EEG data without the artifact.

the context of anesthesia specifically is spectral analysis. The mathematical

basis of spectral analysis is the Fourier transform which transforms a function
in time domain (e.g., EEG) into a complex-valued function of frequency.
This corresponds to expressing a function as a sum of sines and cosines.
Yet, the proper application of spectral analysis to complex and nonstationary
signals such as those recorded in the EEG cannot be accomplished by simply
applying the fast Fourier transform algorithm (Press, 2007) to the recorded
signals. This section will only introduce the very basics of spectral analysis
Brief Introduction to Electroencephalography 267

and make suggestions concerning a good choice of analytic techniques.

A much more detailed account of spectral analysis can be found in Bokil
et al. (2010), Mitra (2007), and Percival and Walden (1993). Several very
useful and pragmatic tutorials can be found at http://chronux.org/.
All recorded signals are evaluated at discrete time points set by the sam-
pling frequency. If we had an infinitely long stretch of data we would be able
to properly evaluate its discrete Fourier transform. Yet, all real signals are
finite. Furthermore, when the signals are nonstationary (see below), even
smaller chunks of data have to be chosen for spectral analysis. Taking a chunk
of data is equivalent to multiplying the signal by a function that is equal to 1
within the desired window and is 0 elsewhere. Thus, taking a chunk of data
for spectral analysis is equivalent to convolving the Fourier transform of the
signal with a Fourier transform of a rectangle. The Fourier transform of the
rectangle, however, distorts the signal by introducing both narrow and
broadband bias and profoundly distorts the spectrum. Other functions can
be convolved with the signal to extract a segment of data (e.g., Hanning,
Tukey, Exponential, B-Spline, Boxcar, Parzen, Welch, etc.). The choice
of the window has significant implications for spectral analysis. Yet, one spe-
cific choice of a spectral window—discrete prolate spheroidal (Slepian)
sequences—deserves special attention. These sequences constitute an opti-
mal choice of a windowing function in a sense that concentrate power in the
desired frequency range while minimizing power outside of the range.
Remarkably, in order to find such windowing function, one needs to solve
an eigenvalue problem where the desired functions emerge as eigenvectors.
Because the matrix is symmetric, the resultant eigenvectors are orthogonal.
As a consequence, the same short segment of data can be convolved with each
one of such eigenvectors in turn to get an independent estimate of the spec-
trum of the signal within the data. This allows for estimation of the confidence
intervals around the spectrum for a given window of data. Confidence inter-
val estimation can be performed using a variety of bootstrapping methods
such as Jacknife for instance. These discrete prolate spheroidal sequences
are also known as tapers and are the basis for what is referred to as multitaper
spectral analysis (Bokil et al., 2010; Mitra, 2007; Percival & Walden, 1993).
While many alternative choices are possible, our opinion is that for most appli-
cations, multitaper spectral analysis is a very good choice of technique for
spectral estimation. Some code for performing multitaper spectral estimation
and computing Slepian sequences is native to MATLAB®. Additional spectral
estimation routines for multitaper spectra and coherences are found in the
Chronux toolbox for MATLAB®. Python packages for multitaper spectral
268 Alex Proekt

estimation are also freely available at http://www.thomas-cokelaer.info/

software/spectrum/html/contents.html.
When performing multitaper spectral estimation, one needs to settle on a
choice of several key parameters that have a strong influence on results. The
first crucial choice is the number of tapers. This choice is related to the degree
of desired smoothing in the frequency domain. Recall that the tapers max-
imize power in the frequency range W. It turns out that the maximum num-
ber of tapers is K ¼ 2NWΔt where N is the number of points in a window of
data and Δt is the inverse of the sampling frequency. Thus, the choice of the
number of tapers and the choice of the window length is closely related.
Choosing an appropriate spectral window plays a very strong role in the
resulting spectral analysis. The choice of an appropriate window length
depends on the specifics of experimental design and the question the analysis
is trying to answer. Choosing a particular window length is equivalent to
assuming that the signal is approximately stationary and linear over the course
of this window (Mitra, 2007; Percival & Walden, 1993). In our experience
some applications (e.g., Alonso et al., 2014; Solovey et al., 2015) require
extremely short windows, while other (e.g., Hudson, Calderon, Pfaff, &
Proekt, 2014) require exceptionally long spectral windows. As is the case
with most other parameters the safest strategy is to explore the range of
parameter values to demonstrate that the specific choice does not have a very
strong influence on the results.

4. MULTIVARIATE SPECTRAL ANALYSIS: COHERENCE

AND PHASE LAG INDEX
Spectrum of a signal is very closely related to its autocorrelation. By
analogy crossspectrum is closely related to the crosscorrelations between sig-
nals. The major advantage of spectra over correlations is that statistically
robust methodologies exist for computing spectral quantities for short seg-
ments of data. Techniques similar to that used for estimation of the spectrum
of a single-time series (i.e., one channel of the EEG) can be generalized to a
pair of signals. One common way of representing the results is to plot the
appropriately normalized absolute value of the crossspectrum between a pair
of signals as a function of frequency. This quantity [0,1] is typically referred
to as coherence and measures frequency-dependent correlations between
the signals. If the coherence value is close to 1 at a particular frequency, this
means that the phase difference between two signals is approximately con-
stant at that frequency. Conversely, if the number is close to 0, then the
Brief Introduction to Electroencephalography 269

phase differences are more uniformly distributed. A version of coherence

analysis referred to as phase lag index (PLI) is essentially coherence where
the phase lag between two signals is nonzero (i.e., the signals are asynchro-
nous) (Stam, Nolte, & Daffertshofer, 2007). In this way, coherence and PLI
are forms of “functional connectivity” (Bastos & Schoffelen, 2015).
There has been very significant interest in quantifying connectivity
between brain regions and brain activity signals. It is not possible to review
the various approaches here in any significant detail. Here, we will only
address a single caveat that complicates all forms of this analysis known as
the common reference problem. Recall that EEG signals measure potential
difference between a pair of points. One such point is called reference. Thus,
if two different electrodes are recording activity with respect to the same
reference, fluctuations in voltage at the reference electrode will be reflected
in both electrodes and would yield spurious connectivity results. One solu-
tion to this problem is to use a different reference for each channel. This can,
for instance, be accomplished by using locally referenced bipolar derivation
(Bosman et al., 2012). This is not assumption free (Bastos & Schoffelen,
2015) and decreases the number of total electrodes by a factor of ½. Closely
related to the common reference problem is the problem of field spread/
volume conduction. As we alluded to earlier for extracellular recordings
such as the EEG, there is a fundamental distinction between the sensors
(i.e., electrodes) and the sources of electric activity. Because electrical fields
travel readily through conductive medium, each sensor will record mixture
of activity from different sources. As discussed earlier, it is not strictly speak-
ing possible to unambiguously extract the different sources from EEG
recordings. Thus, coherence observed in two electrodes may simply reflect
spreading of the electric fields rather than true phase locking between different
sources of neuronal activity (Nunez et al., 1997, 1999; Schoffelen & Gross,
2009; Srinivasan, Winter, Ding, & Nunez, 2007; Winter, Nunez, Ding, &
Srinivasan, 2007). The solution to the problem of volume conduction is con-
ceptually similar to that of the common reference. The choice of referencing
significantly affects the spatial scale on which connectivity is considered
(Nunez et al., 1997). Arguably the most parsimonious solution is to com-
pute a surface Laplacian reference (Srinivasan, Nunez, & Silberstein, 1998).
The simplest form of the Laplacian is to reference each electrode to the
average of all of its nearest neighbors (Cimenser et al., 2011). This tech-
nique has the advantage of mathematical simplicity but at the cost of poten-
tially introducing artifacts into the data. More sophisticated approaches
based on spline approximations of the surface potentials are reviewed in
270 Alex Proekt

Carvalhaes and de Barros (2015). Code for generating Laplacians is readily

available through Fieldtrip (Oostenveld et al., 2011). It has been argued that
PLI is less sensitive to the issues of common reference and field spread
because it only considers non-zero phase lag (Stam et al., 2007). That being
said, synchrony may play an important role in neuronal processing.

5. EFFECTS OF ANESTHETICS ON THE EEG

Changes in the amplitude and frequency of EEG induced with general
anesthesia became readily apparent to early encephalographers (Gibbs, 1937).
Indeed, Gibbs (1937) suggested that “Electroencephalography may be of
value in controlling depth of anesthesia and sedation.” Indeed there are
now many different “depth of anesthesia” monitors in clinical use. Further-
more, there is increasing interest in using the EEG to modify the incidence of
postoperative cognitive disturbances (Whitlock et al., 2014).
Oscillations in the EEG observed during the awake and anesthetized
states are typically grouped into several categories based on their frequencies.

Band Frequency (Hz)

δ <4
θ 4–7
α 8–15
β 16–30
γ >30

It should be noted, however, that by and large the power spectrum of the
EEG is 1/f noise. This means that P ∝ f β where P is the power and f is the
frequency. This conspicuous relationship between power and frequency has
been a subject of vigorous debate (e.g., B
edard, Kr€oger, & Destexhe, 2006;
Beggs & Plenz, 2003; He, Zempel, Snyder, & Raichle, 2010) which we will
not cover here.
In the case of the EEG one significant frequency component that devi-
ates from the usual relationship between power and frequency is the alpha
wave. Indeed, alpha wave was first described by Hans Berger in his original
manuscript on human EEG and it has been referred to as the Berger wave in
the early literature. Alpha wave has been of significant interest in anesthesia
Brief Introduction to Electroencephalography 271

research. In the awake brain alpha wave is most prominent in the occipital
leads when the subject is relaxed and their eyes are closed. It has been noted
by Michenfelder and colleagues (Tinker, Sharbrough, & Michenfelder,
1977) that loss of consciousness is associated with the spatial shift in the
alpha oscillations such that just around the point where responsiveness is lost,
alpha wave assumes an anterior dominance. This anteriorization of the alpha
rhythm has been systematically investigated by John et al. (2001) and was
suggested to be the hallmark of loss of consciousness. Recent robust compu-
tational techniques by Brown and colleagues revealed that frontal alpha oscil-
lations are coherent during loss and recovery of consciousness induced with
propofol (Cimenser et al., 2011). Yet there are some important differences
among the anesthetics with respect to the alpha oscillations. For instance,
dexmedetomidine produces a spindle-like pattern consisting of intermittent
oscillations (12–16 Hz) (Akeju et al., 2016a). Furthermore, Mashour and col-
leagues (Blain-Moraes et al., 2015) found that the anteriorization of alpha
rhythm is not consistently found in the EEG of sevoflurane-anesthetized
human subjects. The same group demonstrated that ketamine—a dissociative
anesthetic—does not elicit an increase in the frontal alpha-band power
(Blain-Moraes, Lee, Ku, Noh, & Mashour, 2014).
Deepening of the anesthetic state is associated with the increase in the
slow oscillations referred to as the delta waves (Gibbs, 1937). These slow
oscillations arise as a result of the interplay between ion channels and syn-
aptic transmission in the thalamocortical circuits (Steriade, McCormick,
& Sejnowski, 1993). In increasing doses many, but not all, anesthetics
elicit delta oscillations (Brown, Lydic, & Schiff, 2010). A notable excep-
tion to this pattern is ketamine. Under ketamine anesthesia delta and
gamma oscillations alternate in a burst-like pattern (Akeju et al., 2016b;
Maksimow et al., 2006). Another interesting exception to the delta wave
pattern associated with deeper stages of anesthesia is nitrous oxide.
Avramov and colleagues first found that delta waves appear transiently
(Avramov, Shingu, & Mori, 1990) during the administration of nitrous
oxide. In a paradoxical reaction Avramov et al. also demonstrated that pain-
ful stimulation, typically associated with desynchronization of the EEG,
under nitrous oxide can be associated with emergence of delta oscillations
(Avramov et al., 1990). There has been a recent resurgence of interest in using
delta oscillations under anesthesia to quantify anesthetic depth. For instance,
Tracey and colleagues observe that loss of consciousness induced with propofol
(Nı́ Mhuircheartaigh, Warnaby, Rogers, Jbabdi, & Tracey, 2013). Warnaby
and colleagues extended these findings to show that slow wave activity
272 Alex Proekt

saturation is a hallmark of loss of consciousness for the volatile anesthetics in

addition to propofol (Warnaby, Sleigh, Hight, Jbabdi, & Tracey, 2017).
If the anesthetic concentration is increased further, one typically
observes a characteristic pattern of the EEG referred to as burst suppression.
Burst suppression, as the name suggests corresponds to alterations between
period of high-amplitude fluctuations in the EEG signals and periods of
suppressed EEG (Steriade, Amzica, & Contreras, 1994). First demonstrated
by Derbyshire, Rempel, Forbes, and Lambert (1936), burst suppression
is a feature of neuronal activity induced by many anesthetics. Interestingly
halothane, in the clinically useful concentration range, does not seem
to reliably elicit burst suppression (Murrell, Waters, & Johnson, 2008;
Yli-Hankala, Eskola, & Kaukinen, 1989).
In addition to the traditional spectral analysis of the EEG, recent studies
turned to the quantification of “functional connectivity” in the anesthetized
states. Functional connectivity can be loosely defined as the statistical inter-
relationship between recordings obtained in different parts of the brain. As
alluded to earlier, care must be taken to ensure that this relationship is not
spuriously observed because of effects of common ground and volume con-
ductance. Several measures of connectivity have been applied to the EEG
under anesthesia. The measure most closely associated with spectral analysis
is coherence. Coherence is the frequency domain counterpart of correla-
tion. The advantage of the coherence measure is that it can be robustly esti-
mated in short segments of data using, for instance, multitaper spectral
estimates. An example of this approach is (Cimenser et al., 2011) that dem-
onstrate that the alpha waves found over anterior leads of the EEG are coher-
ent. A closely related technique is PLI. PLI seeks to minimize effects of
volume conduction by only focusing at coherence observed with a nonzero
phase lag. An example of application of PLI to analyzing the EEG under
ketamine anesthesia is by Blain-Moraes et al. (2014, 2015). Yet, other mea-
sures of functional connectivity rely on information-theoretic quantities
such as transfer entropy a symbolic transfer entropy (Staniek & Lehnertz,
2008). As the mathematics of these measures is rather involved, they would
not be discussed in significant detail here. Some useful code for computing
these quantities can be found in TRENTOOL—a MATLAB®-based toolbox
(Wibral, Vicente, Priesemann, & Lindner, 2011) and MuTE (Montalto,
Faes, & Marinazzo, 2014). An exciting development in the field of functional
connectivity as it relates to the study of anesthesia is the proposal that mech-
anistically distinct anesthetic agents that induce different patterns of oscil-
lations in the EEG may nonetheless uniformly disrupt connectivity
Brief Introduction to Electroencephalography 273

between the frontal and the parietal cortices. This work has been pioneered
by Lee and Mashour (e.g., Ku, Lee, Noh, Jun, & Mashour, 2011; Lee et al.,
2009, 2013).

6. CONCLUDING REMARKS
Here we introduced the very basics of EEG and pointed toward some
useful references that address some of the fundamental issues that compli-
cate analysis of the EEG in much more detail. On the one hand prolifer-
ation of software tools that allow complex processing of EEG signals has
made EEG more accessible to scientists interested in neuronal activity.
Thus, it may no longer be necessary for all researchers involved with
EEG recordings to do very sophisticated mathematical calculations de
novo. On the other hand, this accessibility may give the false impression
that EEG analysis can be performed in an “out of the box” unsupervised
fashion. The major purpose of this chapter was to introduce some caveats
and assumptions involved in nearly all forms of EEG analysis without get-
ting into the mathematical details. The history of the investigation of the
effects of anesthetics on the EEG is almost as long as the history of the
EEG itself. Thus, here we just highlighted some of the classic papers and
some of the recent findings that characterize the effects of anesthetics on
the EEG.

REFERENCES
Akeju, O., et al. (2016a). Spatiotemporal dynamics of dexmedetomidine-induced electroen-
cephalogram oscillations. PLoS One, 11, e0163431.
Akeju, O., et al. (2016b). Electroencephalogram signatures of ketamine anesthesia-induced
unconsciousness. Clinical Neurophysiology, 127, 2414–2422.
Alonso, L. M., et al. (2014). Dynamical criticality during induction of anesthesia in human
ECoG recordings. Frontiers Neural Circuits, 8, 20.
Andrew, C., & Pfurtscheller, G. (1996). Dependence of coherence measurements on EEG
derivation type. Medical & Biological Engineering & Computing, 34, 232–238.
Avramov, M. N., Shingu, K., & Mori, K. (1990). Progressive changes in electroencephalo-
graphic responses to nitrous oxide in humans: A possible acute drug tolerance. Anesthesia
and Analgesia, 70, 369–374.
Başar, E., Rahn, E., Demiralp, T., & Sch€ urmann, M. (1998). Spontaneous EEG theta activity
controls frontal visual evoked potential amplitudes. Electroencephalography and Clinical
Neurophysiology, 108, 101–109.
Bastos, A. M., & Schoffelen, J.-M. (2015). A tutorial review of functional connectivity anal-
ysis methods and their interpretational pitfalls. Frontiers in Systems Neuroscience, 9, 175.
B
edard, C., Kr€oger, H., & Destexhe, A. (2006). Does the 1/f frequency scaling of brain sig-
nals reflect self-organized critical states? Physical Review Letters, 97, 118102.
Beggs, J. M., & Plenz, D. (2003). Neuronal avalanches in neocortical circuits. The Journal of
Neuroscience, 23, 11167–11177.
274 Alex Proekt

Bell, A. J., & Sejnowski, T. J. (1997). The ‘independent components’ of natural scenes are
edge filters. Vision Research, 37, 3327–3338.
€
Berger, H. (1932). Uber das Elektrenkephalogramm des Menschen. Archiv f u€r Psychiatrie und
Nervenkrankheiten, 97, 6–26.
Blain-Moraes, S., Lee, U., Ku, S., Noh, G., & Mashour, G. A. (2014). Electroencephalo-
graphic effects of ketamine on power, cross-frequency coupling, and connectivity in
the alpha bandwidth. Frontiers in Systems Neuroscience, 8, 114.
Blain-Moraes, S., et al. (2015). Neurophysiological correlates of sevoflurane-induced uncon-
sciousness. Anesthesiology, 122, 307–316.
Bokil, H., Andrews, P., Kulkarni, J. E., Mehta, S., & Mitra, P. P. (2010). Chronux:
A platform for analyzing neural signals. Journal of Neuroscience Methods, 192, 146–151.
Bosman, C. A., et al. (2012). Attentional stimulus selection through selective synchronization
between monkey visual areas. Neuron, 75, 875–888.
Brown, E. N., Lydic, R., & Schiff, N. D. (2010). General anesthesia, sleep, and coma. The
New England Journal of Medicine, 363, 2638–2650.
Buzsaki, G. (2004). Neuronal oscillations in cortical networks. Science, 304, 1926–1929.
Buzsáki, G., Anastassiou, C. A., & Koch, C. (2012). The origin of extracellular fields and
currents—EEG, ECoG, LFP and spikes. Nature Reviews. Neuroscience, 13, 407–420.
Buzsáki, G., et al. (2015). Tools for probing local circuits: High-density silicon probes com-
bined with optogenetics. Neuron, 86, 92–105.
Carvalhaes, C., & de Barros, J. A. (2015). The surface Laplacian technique in EEG: Theory
and methods. International Journal of Psychophysiology, 97, 174–188.
Chella, F., Marzetti, L., Pizzella, V., Zappasodi, F., & Nolte, G. (2014). Third order spectral
analysis robust to mixing artifacts for mapping cross-frequency interactions in EEG/
MEG. NeuroImage, 91, 146–161.
Chella, F., Pizzella, V., Zappasodi, F., & Marzetti, L. (2016). Impact of the reference choice
on scalp EEG connectivity estimation. Journal of Neural Engineering, 13, 036016.
Cimenser, A., et al. (2011). Tracking brain states under general anesthesia by using global
coherence analysis. Proceedings of the National Academy of Sciences of the United States of
America, 108, 8832–8837.
Delorme, A., & Makeig, S. (2004). EEGLAB: An open source toolbox for analysis of single-
trial EEG dynamics including independent component analysis. Journal of Neuroscience
Methods, 134, 9–21.
Delorme, A., Sejnowski, T., & Makeig, S. (2007). Enhanced detection of artifacts in EEG
data using higher-order statistics and independent component analysis. NeuroImage,
34, 1443–1449.
Derbyshire, A. J., Rempel, B., Forbes, A., & Lambert, E. F. (1936). The effects of anesthetics
on action potentials in the cerebral cortex of the cat. American Journal of Physiology-Legacy
Content, 116, 577–596.
Destexhe, A., Contreras, D., & Steriade, M. (1999). Spatiotemporal analysis of local field
potentials and unit discharges in cat cerebral cortex during natural wake and sleep states.
The Journal of Neuroscience, 19, 4595–4608.
Dietzel, I., Heinemann, U., & Lux, H. D. (1989). Relations between slow extracellular
potential changes, glial potassium buffering, and electrolyte and cellular volume changes
during neuronal hyperactivity in cat brain. Glia, 2, 25–44.
Dong, L., et al. (2017). MATLAB toolboxes for reference electrode standardization tech-
nique (REST) of scalp EEG. Frontiers in Neuroscience, 11, 601.
Gevins, A., & Smith, M. E. (2000). Neurophysiological measures of working memory
and individual differences in cognitive ability and cognitive style. Cerebral Cortex, 10,
829–839.
Gibbs, F. A. (1937). Effect on the electro-encephalogram of certain drugs which influence
nervous activity. Archives of Internal Medicine, 60, 154.
Brief Introduction to Electroencephalography 275

Grech, R., et al. (2008). Review on solving the inverse problem in EEG source analysis.
Journal of Neuroengineering and Rehabilitation, 5, 25.
He, B. J., Zempel, J. M., Snyder, A. Z., & Raichle, M. E. (2010). The temporal structures
and functional significance of scale-free brain activity. Neuron, 66, 353–369.
Holt, G. R., & Koch, C. (1999). Electrical interactions via the extracellular potential near cell
bodies. Journal of Computational Neuroscience, 6, 169–184.
Hudson, A. E., Calderon, D. P., Pfaff, D. W., & Proekt, A. (2014). Recovery of conscious-
ness is mediated by a network of discrete metastable activity states. Proceedings of the
National Academy of Sciences of the United States of America, 111, 9283–9288.
John, E. R., et al. (2001). Invariant reversible QEEG effects of anesthetics. Consciousness and
Cognition, 10, 165–183.
Jones, E. G. (1998). Viewpoint: The core and matrix of thalamic organization. Neuroscience,
85, 331–345.
Jones, E. G. (2001). The thalamic matrix and thalamocortical synchrony. Trends in Neurosci-
ences, 24, 595–601.
Jung, T.-P., Makeig, S., Bell, A. J., & Sejnowski, T. J. (1998). Independent component anal-
ysis of electroencephalographic and event-related potential data. In Central auditory
processing and neural modeling (pp. 189–197). Springer.
Kirschstein, T., & K€ ohling, R. (2009). What is the source of the EEG? Clinical EEG and
Neuroscience, 40, 146–149.
Krishnaveni, V., Jayaraman, S., Anitha, L., & Ramadoss, K. (2006). Removal of ocular
artifacts from EEG using adaptive thresholding of wavelet coefficients. Journal of Neural
Engineering, 3(4), 338.
Ku, S.-W., Lee, U., Noh, G.-J., Jun, I.-G., & Mashour, G. A. (2011). Preferential inhibition
of frontal-to-parietal feedback connectivity is a neurophysiologic correlate of general
anesthesia in surgical patients. PLoS One, 6, e25155.
Kuklinski, W. S. (1983). Electric fields of the brain. The Neurophysics of EEG by P. L.
Nunez with contributions by R. D. Katznelson. Medical Physics, 10, 119–120.
Lee, U., et al. (2009). The directionality and functional organization of frontoparietal
connectivity during consciousness and anesthesia in humans. Consciousness and Cognition,
18, 1069–1078.
Lee, U., et al. (2013). Disruption of frontal–parietal communication by ketamine, propofol,
and sevoflurane. Anesthesiology, 118, 1264–1275.
Lehmann, D., Strik, W. K., Henggeler, B., Koenig, T., & Koukkou, M. (1998). Brain elec-
tric microstates and momentary conscious mind states as building blocks of spontaneous
thinking: I. Visual imagery and abstract thoughts. International Journal of Psychophysiology,
29, 1–11.
Lei, X., & Liao, K. (2017). Understanding the influences of EEG reference: A large-scale
brain network perspective. Frontiers in Neuroscience, 11, 205.
LeVan, P., Urrestarazu, E., & Gotman, J. (2006). A system for automatic artifact removal in
ictal scalp EEG based on independent component analysis and Bayesian classification.
Clinical Neurophysiology, 117, 912–927.
Li, Y., Ma, Z., Lu, W., & Li, Y. (2006). Automatic removal of the eye blink artifact from
EEG using an ICA-based template matching approach. Physiological Measurement, 27,
425–436.
Lind
en, H., et al. (2013). LFPy: A tool for biophysical simulation of extracellular potentials
generated by detailed model neurons. Frontiers in Neuroinformatics, 7, 41.
Liu, J., Singh, H., & White, P. F. (1996). Electroencephalogram bispectral analysis predicts
the depth of midazolam-induced sedation. Anesthesiology, 84, 64–69.
Liu, Q., et al. (2015). Estimating a neutral reference for electroencephalographic recordings:
The importance of using a high-density montage and a realistic head model. Journal of
Neural Engineering, 12, 056012.
276 Alex Proekt

Maksimow, A., et al. (2006). Increase in high frequency EEG activity explains the poor
performance of EEG spectral entropy monitor during S-ketamine anesthesia. Clinical
Neurophysiology, 117, 1660–1668.
Mitra, P. (2007). Observed brain dynamics. Oxford University Press.
Mognon, A., Jovicich, J., Bruzzone, L., & Buiatti, M. (2011). ADJUST: An automatic EEG
artifact detector based on the joint use of spatial and temporal features. Psychophysiology,
48, 229–240.
Montalto, A., Faes, L., & Marinazzo, D. (2014). MuTE: A MATLAB toolbox to com-
pare established and novel estimators of the multivariate transfer entropy. PLoS One,
9, e109462.
Murrell, J. C., Waters, D., & Johnson, C. B. (2008). Comparative effects of halothane, iso-
flurane, sevoflurane and desflurane on the electroencephalogram of the rat. Laboratory
Animals, 42, 161–170.
Nı́ Mhuircheartaigh, R., Warnaby, C., Rogers, R., Jbabdi, S., & Tracey, I. (2013). Slow-
wave activity saturation and thalamocortical isolation during propofol anesthesia in
humans. Science Translational Medicine, 5, 208ra148.
Niedermeyer, E., & da Silva, F. H. L. (2005). Electroencephalography: Basic principles, clinical
applications, and related fields. Lippincott Williams & Wilkins.
Nunez, P. L., & Srinivasan, R. (2006). Electric fields of the brain: The neurophysics of EEG. USA:
Oxford University Press.
Nunez, P. L., & Srinivasan, R. (2010). Scale and frequency chauvinism in brain dynamics:
Too much emphasis on gamma band oscillations. Brain Structure & Function, 215, 67–71.
Nunez, P. L., et al. (1997). EEG coherency. I: Statistics, reference electrode, volume con-
duction, Laplacians, cortical imaging, and interpretation at multiple scales. Electroenceph-
alography and Clinical Neurophysiology, 103, 499–515.
Nunez, P. L., et al. (1999). EEG coherency II: Experimental comparisons of multiple mea-
sures. Clinical Neurophysiology, 110, 469–486.
Offner, F. F. (1950). The EEG as potential mapping: The value of the average monopolar
reference. Electroencephalography and Clinical Neurophysiology, 2, 213–214.
Oja, E., & Yuan, Z. (2006). The fastICA algorithm revisited: Convergence analysis. IEEE
Transactions on Neural Networks, 17, 1370–1381.
Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J.-M. (2011). FieldTrip: Open source soft-
ware for advanced analysis of MEG, EEG, and invasive electrophysiological data.
Computational Intelligence and Neuroscience, 2011, 156869.
Percival, D. B., & Walden, A. T. (1993). Spectral analysis for physical applications. Cambridge,
UK: Cambridge University Press.
Press, W. H. (2007). Numerical recipes. In The art of scientific computing (3rd ed.). Cambridge
University Press.
Rall, W., & Shepherd, G. M. (1968). Theoretical reconstruction of field potentials and
dendrodendritic synaptic interactions in olfactory bulb. Journal of Neurophysiology, 31,
884–915.
Rush, S., & Driscoll, D. A. (1969). EEG electrode sensitivity—an application of reciprocity.
IEEE Transactions on Biomedical Engineering, BME-16, 204.
Schoffelen, J.-M., & Gross, J. (2009). Source connectivity analysis with MEG and EEG.
Human Brain Mapping, 30, 1857–1865.
Shao, S.-Y., Shen, K.-Q., Ong, C. J., Wilder-Smith, E. P. V., & Li, X.-P. (2009). Automatic
EEG artifact removal: A weighted support vector machine approach with error correc-
tion. IEEE Transactions on Biomedical Engineering, 56, 336–344.
Solovey, G., et al. (2015). Loss of consciousness is associated with stabilization of cortical
activity. The Journal of Neuroscience, 35, 10866–10877.
Srinivasan, R., Nunez, P. L., & Silberstein, R. B. (1998). Spatial filtering and neocortical
dynamics: Estimates of EEG coherence. IEEE Transactions on Biomedical Engineering,
45, 814–826.
Brief Introduction to Electroencephalography 277

Srinivasan, R., Winter, W. R., Ding, J., & Nunez, P. L. (2007). EEG and MEG coherence:
Measures of functional connectivity at distinct spatial scales of neocortical dynamics. Jour-
nal of Neuroscience Methods, 166, 41–52.
Stam, C. J., Nolte, G., & Daffertshofer, A. (2007). Phase lag index: Assessment of functional
connectivity from multi channel EEG and MEG with diminished bias from common
sources. Human Brain Mapping, 28, 1178–1193.
Staniek, M., & Lehnertz, K. (2008). Symbolic transfer entropy. Physical Review Letters, 100,
158101.
Steriade, M., Amzica, F., & Contreras, D. (1994). Cortical and thalamic cellular correlates of
electroencephalographic burst-suppression. Electroencephalography and Clinical Neurophys-
iology, 90, 1–16.
Steriade, M., McCormick, D., & Sejnowski, T. (1993). Thalamocortical oscillations in the
sleeping and aroused brain. Science, 262, 679–685.
Tian, Y., & Yao, D. (2013). Why do we need to use a zero reference? Reference influences
on the ERPs of audiovisual effects. Psychophysiology, 50, 1282–1290.
Tinker, J. H., Sharbrough, F. W., & Michenfelder, J. D. (1977). Anterior shift of the dom-
inant EEG rhythm during anesthesia in the java monkey. Anesthesiology, 46, 252–259.
Vigário, R. N. (1997). Extraction of ocular artefacts from EEG using independent compo-
nent analysis. Electroencephalography and Clinical Neurophysiology, 103, 395–404.
Vigario, R., Sarela, J., Jousmiki, V., Hamalainen, M., & Oja, E. (2000). Independent com-
ponent approach to the analysis of EEG and MEG recordings. IEEE Transactions on
Biomedical Engineering, 47, 589–593.
Warnaby, C. E., Sleigh, J. W., Hight, D., Jbabdi, S., & Tracey, I. (2017). Investigation of
slow-wave activity saturation during surgical anesthesia reveals a signature of neural iner-
tia in humans. Anesthesiology, 127, 645–657.
Whitlock, E. L., et al. (2014). Postoperative delirium in a substudy of cardiothoracic surgical
patients in the BAG-RECALL clinical trial. Anesthesia & Analgesia, 118, 809–817.
Wibral, M., Vicente, R., Priesemann, V., & Lindner, M. (2011). TRENTOOL: An open
source toolbox to estimate neural directed interactions with transfer entropy. BMC
Neuroscience, 12, P200.
Winter, W. R., Nunez, P. L., Ding, J., & Srinivasan, R. (2007). Comparison of the effect
of volume conduction on EEG coherence with the effect of field spread on MEG coher-
ence. Statistics in Medicine, 26, 3946–3957.
Yao, D., et al. (2005). A comparative study of different references for EEG spectral mapping:
The issue of the neutral reference and the use of the infinity reference. Physiological
Measurement, 26, 173–184.
Yli-Hankala, A., Eskola, H., & Kaukinen, S. (1989). EEG spectral power during halothane
anaesthesia. A comparison of spectral bands in the monitoring of anaesthesia level. Acta
Anaesthesiologica Scandinavica, 33, 304–308.