0% found this document useful (0 votes)

38 views96 pages

EEG Analysis: Theory and Practice

The document discusses the complexities of electroencephalography (EEG) signal analysis, emphasizing the need for quantification and pattern recognition to support clinical evaluations. It highlights the importance of selecting appropriate analysis methods based on the goals of the study and the characteristics of EEG signals, which can vary based on time and location of recording. Additionally, it covers the statistical properties of EEG signals, including sampling and the significance of maintaining signal integrity during digital conversion.

Uploaded by

k.jawahar Raj

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

38 views96 pages

EEG Analysis: Theory and Practice

Uploaded by

k.jawahar Raj

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 96

Part IX Computer-Assisted EEG Analysis

CHAPTER

EEG Analysis: Theory and Practice

FERNANDO H. LOPES DA SILVA 54
Analysis of electroencephalography (EEG) signals always practical consideration is the number of derivations to be ana-
involves questions of quantification; such questions may con- lyzed and whether the corresponding topographic relations
cern the precise value of the dominant frequency and the simi- have to be determined or the analysis of one or two derivations
larity between two signals recorded from symmetric derivations is enough; the latter may suffice during anesthesia monitoring
at the same time or different times. In these examples, there is a or in sleep research. Whether the analysis of a relatively short
question that can be solved only by taking measures with regard EEG epoch is sufficient or must involve very long records, for
to the EEG signal. Without such measures, EEG appraisal instance, up to 24 hours is another important factor.
remains subjective and can hardly lead to logical systematiza- In short, the method of analysis must be suited to the purpose
tion. Classic EEG evaluation has always involved measuring fre- of the analysis. Among the different purposes are the following:
quency and/or amplitude with the help of simple rulers. The (i) determining whether a relatively short EEG record taken in
limitations of such simple methods are severe, particularly a routine laboratory is normal or abnormal; (ii) classifying an
when large amounts of EEG data must be evaluated and the EEG as abnormal, for example, as epileptiform or hypofunc-
need for data reduction is felt strongly, as well as when rather tional; (iii) evaluating changes occurring in serial EEG; and (iv)
sophisticated questions are being asked, such as whether EEG evaluating trends during many hours of EEG monitoring, such
signal changes occur in relation to internal or external factors, as under intensive care conditions for heart surgery or in long-
and how synchronous are EEG phenomena occurring in differ- term recordings in epileptic patients.
ent derivations.
Clear replies to these questions require some form of EEG GENERAL CHARACTERISTICS
analysis. However, such analysis not only is a problem of quan-
tification, but also involves elements of pattern recognition. The EEG is a complex signal, the statistical properties of which
Every electroencephalographer knows that it is sometimes depend on both time and space. Regarding the temporal char-
extremely difficult to cite exact measures for such EEG phe- acteristics, it is essential to note that EEG signals are ever-
nomena as spikes, sharp waves, or other abnormal patterns; the changing. However, they can be analytically subdivided into
experienced specialist is able to detect them only by “eye- representative epochs (i.e., with more or less constant statistical
balling.” These types of problems may be solved using pattern properties).
recognition analysis techniques, based on the principle that fea- Estimates of the length of such epochs vary considerably
tures characteristic of the EEG phenomena have to be meas- because of dependence on the subject’s behavioral state. When
ured. This phase of feature extraction is followed by classification the latter is kept almost constant, Isaksson and Wennberg (1)
of the phenomena into different groups. EEG analysis thus found that, over relatively short-time intervals, epochs can be
implies not only simple quantification, but also feature extrac- defined that can be considered representative of the subject’s
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

tion and classification. state; in this study, some 90% of the EEG signals investigated had
The primary aim of EEG analysis is to support electroen- time-invariant properties after 20 seconds, whereas less than
cephalographers’ evaluations with objective data in numerical 75% remained time invariable after 60 seconds. Empirical obser-
or graphic form. EEG analysis, however, can go further, actually vations indicate that EEG records obtained under equivalent
extending electroencephalographers’ capabilities by giving behavioral conditions show highly stable characteristics; for
them new tools with which they can perform such difficult example, Dumermuth et al. (2) showed that variations in mean
tasks as quantitative analysis of long-duration EEG in epileptic peak (beta activity) of only 0.8 Hz were obtained in a series of 11
patients and sleep and psychopharmacologic studies. EEGs over 29 weeks. In this respect it is interesting to consider
The choice of analytic method should be determined mainly the studies of Jansen (3) and Grosveld et al. (4); these authors
by the goal of the application, although budget limitations must investigated the possibility of correctly assigning EEG epochs
also be taken into consideration. The development of an appro- (duration 10.24 seconds) to the corresponding subject by means
priate strategy rests on such practical facts as whether analysis of multivariate analysis, using half of the EEG epochs recorded
results must be available in real time and online or may be pre- from 16 subjects as a training set for the classification algorithm.
sented offline. In the past, the former requirement would pose Using 4 to 10 EEG features, it was found that in 80% to 90% of
considerable problems, solvable only by adopting a rather sim- the cases of EEG epochs were assigned correctly to the corre-
ple form of analysis; the development of new computer tech- sponding subject. McEwen and Anderson (5) introduced the
nology has provided more acceptable solutions. Another concept of wide-sense stationarity in EEG analysis; they
1147
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1148 Part IX ■ Computer-Assisted EEG Analysis

proposed a procedure for determining whether a set of signal EEG signals are, of course, time series; they are characterized
samples (e.g., an EEG signal) can be considered to belong to a by a set of values as a function of time. An important problem,
wide-sense stationarity process. Their procedure consisted of however, is whether the general methods for analyzing time
calculating the amplitude distributions and power spectra of series can be applied without restrictions to EEG signals.
sample subsets and showing that they do not differ significantly In Chapter 4, which discusses EEG dynamics, it was men-
using the Kolmogorov–Smirnov statistic. From this study it was tioned that modern mathematical tools are being used to ana-
concluded that, for EEG epochs (awake condition or during lyze EEG signals, assuming that signal generation can be
anesthesia) of less than 32 seconds, the assumption of wide- described using sets of nonlinear differential equations. These
sense stationarity was valid more than 50% of the time. techniques have been developed within the active field of math-
On the basis of this type of empirical observation, it can be ematical research called “deterministic chaos.” In essence, non-
assumed that relatively short EEG epochs (~10 seconds) linear dynamical systems such as the neuronal networks
recorded under constant behavioral conditions are quasi-sta- generating EEG signals can display chaotic behavior; that is,
tionary. Elul (6) remarked that the EEG is related to intermit- their behavior can become unpredictable for relatively long
tent changes in the synchrony of cortical neurons; thus, he periods, and EEG signals may be an expression of chaotic
characterized the EEG as a series of short epochs rather than a behavior. Since new mathematical tools, based on the analysis
continuous process. of complex nonlinear systems such as the correlation dimen-
The fact that EEG signals have different characteristics sion, were introduced in EEG, it became clear that EEG signals
depending on the place over the head where they are recorded may be high-dimensional so that in many cases it is difficult, or
is essential to all EEG recordings. Therefore, in any method of even impossible, to distinguish whether these signals are gener-
EEG analysis, topographic characteristics have to be taken into ated by random or by high-dimensional nonlinear determinis-
account. This means that one should choose EEG montages tic processes (9).
carefully, in view of the objectives of the analysis. The topo-
graphic aspects appear most clearly in the simple case of com- SAMPLING, PROBABILITY DISTRIBUTIONS,
paring EEG records from symmetric derivations; indeed, the CORRELATION FUNCTIONS, AND SPECTRA
use of the subject as his or her own control through right–left
comparisons is a cornerstone of the neurologic examination. EEG signals are continuous variations of potential as a function
Therefore, right–left comparisons are also paramount in any of time. However, in most practical cases where quantitative
practical clinical system of EEG analysis. analysis is applied, signals must be digitized so that they can be
processed by digital computer. This means that the EEG signal
BASIC STATISTICAL PROPERTIES must be processed in such a way that the random variable,
potential as a function of time, will have only one set of discrete
Some of the underlying assumptions of the most common values at a set of discrete time instances. In technical terms, the
methods of EEG analysis will be discussed briefly. Gasser (7) process of analog-to-digital (AD) conversion involves sampling
has provided a more fundamental discussion of this topic; here, combined with the operation of quantizing. According to defi-
general concepts will suffice. nitions commonly used (10), sampling is the “process of
The exact characteristics of EEG signals are, in general obtaining a sequence of instantaneous values of a wave at regu-
terms, unpredictable. This means that one cannot foresee pre- lar or intermittent intervals” and quantization is the “process in
cisely the amplitude of an EEG graphoelement or the duration which the continuous range of values of an input signal is
of an EEG wave. Therefore, it is said that an EEG signal is a divided into nonoverlapping subranges and to each subrange a
realization of a random or stochastic process. Indeed, it is pos- discrete value of the output is uniquely assigned.”
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

sible to determine some statistical measures of EEG signals EEG signal sampling must be performed without changing
that show considerable regularity, such as an average ampli- the statistical properties of the continuous signal. Generally, one
tude or an average frequency. This is a general characteristic of samples an EEG signal at equidistant time intervals (t), thus
random processes, which are characterized by probability dis- transforming the continuous signal into a set of impulses with
tributions and their moments (e.g., mean, variance, skewness, different heights separated by intervals t (Fig. 54.1). An impor-
and kurtosis) or by frequency spectra or correlation func- tant question is the choice of the sampling frequency. This
tions. Such a description of an EEG signal as a realization of a choice is based on the sampling theorem: assuming that a signal
random process implies a mathematical, but not a biophysi- x(t) has a frequency spectrum X(f) such that X(f) = 0 for fN, no
cal, model. information is lost by sampling x(t) at equidistant intervals t
It should be stressed (8) that the biophysical process with fN = 1/(2 t); fN is called the folding or Nyquist frequency.
underlying EEG generation is not necessarily random in The sampling frequency, therefore, must be at least equal to 2fN.
nature, but it may have such a high degree of complexity that A consequence of this theorem is that care has to be taken to
only a description in statistical terms is justified. Gasser (7) ensure that the signal to be sampled has no frequency compo-
has also emphasized this point; even in the case of signals that nents above fN. Therefore, before sampling, all frequency com-
are deterministic (e.g., sinusoids) but very complex (e.g., ponents greater than fN should be eliminated by low-pass
made of many components), a stochastic approach may be filtering. One should keep in mind that sampling at a frequency
the most adequate. below 2fN is not equivalent to filtering; it would produce aliasing

Also definable is a class of statistical functions characteristic

of the random process: mn = E[xn(ti)] with n = 1, 2, 3, …; these
functions are called the nth moments of the discrete random
variable x(ti). The implicit assumption here and in the follow-
ing discussion is that the statistical properties of the signal do
not change in the interval T. Therefore, the moments are inde-
pendent of time ti.
The first moment E[x(ti)] is called the mean of x(ti). It is
often preferable to consider the central moments (i.e., the
moments around the mean); the second central moment is
Figure 54.1 Analog-to-digital (AD) conversion of the continuous
then:
signal is performed at equidistant time intervals, digitizing its amplitude
according to the corresponding quantizing levels. (Adapted from Lopes E[ (x(ti )  E(x(ti ) ) ) 2 ]  m2 (54.2)
da Silva FH, Cooper R, Dumermuth G, et al. Sampling, conversion, and 2
or  or variance of x(ti).
measurement of bioelectrical phenomena. In: Remond A, ed-in-chief;
Similarly, the third central moment E[(x(ti)–(E(x(ti)))3] =
Brazier MA, ed. Handbook of Electroencephalography and Clinical
m3 can be defined; from this can be derived the skewness
Neurophysiology. Vol 4. Part A. Amsterdam: Elsevier; 1976.)
factor 1 = m3/(m2)3/2. The fourth central moment is
E[(x(ti)–(E(x(ti)))4] = m4, from which can be derived the kur-
tosis excess: 2 = m4/(m2)2. In case of a symmetric amplitude,
or signal distortion due to folding of frequency components distribution is 1 = 0; all odd moments are equal to zero. For a
larger than fN onto lower frequencies (11). The analog voltages gaussian distribution the even moments have specific values,
of the signal at the sampling moments are converted to a num- for example, 2 = 3; derivatives from this value indicate the
ber corresponding to the amplitude subrange or level. Most EEG peakedness (2  3) or flatness (2  3) of the distribution
analysis can be performed using 512 to 2048 amplitude levels (Fig. 54.2).
(i.e., 9 to 11 bits). Technical details of AD conversion may be
found in Susskind (12) and, for the special case of EEG signals,
in Lopes da Silva et al. (11) and Steineberg and Paine (13).
The continuous EEG signal is thus replaced by a string of num-
bers x(ti) representing the signal amplitude at sequential sample
moments; the latter are indicated by the index i along the time
axis. The signal is assumed to be a realization of a stationary ran-
dom process x(ti), which is indicated by underlining the letter (x).
In general, a collection of EEG signals of a certain length recorded
under equivalent conditions is available for analysis. The entire
collection of EEG signals is called an ensemble; each member of
the ensemble is called a sample function or a realization.

PROBABILITY DISTRIBUTIONS
The digitized EEG signal values x(ti) can be considered realiza-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

tions of one stochastic variable x(ti) and may be characterized

when stationarity is assumed by a histogram; when in an inter-
val 0  t  T there are na sample points in the interval a
 1/2, na/N is called the relative frequency of occurrence of
the value a, where N is the total number of samples available.
One can define the relative frequencies of all other values simi- Figure 54.2 Examples of EEG signals with corresponding amplitude
larly. When N becomes infinitely large and  infinitely small, distributions. For the EEG signals, the time marks along the horizontal
na/N will tend to a limit value p(x(ti) = a), called the probability axis give the intervals in seconds; the vertical axis is in microvolts; for
of occurrence of x(ti) = a. The set of values of p(x(ti)) is called the amplitude distributions, the horizontal axis is in microvolts, and the
the signal probability distribution, characterized by a mean and vertical axis gives the number (N) of times a certain amplitude class has
a number of moments. Considering that the discrete random been measured in the corresponding EEG epoch. The signals were sam-
variable x can take any of a set of values from 1 to M, the mean pled at 20 Hz. The amplitude distribution of the four EEG signals has
or average of the sample functions is given as follows (E is the the following values of skewness (S) and kurtosis (K): (A) S = 0.17, K
symbol for expectation): = 3.09; (B) S = 0.09, K = 2.41; (C) S = 0.10, K = 3.37; (D) S = 0.07,
K = 2.98. The hypothesis that the amplitude distribution belongs to a
E[x(ti ) ]  c a a # p(x(ti )  a) d  mi
M
(54.1) normal distribution can be rejected at P  0.01 whenever S  0.464
a1 and/or 2.45  K  4.13 with N = 160.

CORRELATION FUNCTIONS AND SPECTRA assuming that the signal x(t) has mean zero. For the value  = 0:

In general terms, successive values of a signal, such as an EEG, T

#
1
which result from a stochastic process are not necessarily inde- RXX (0)  E[x2 (t) ]  Lim x2 (t)dt (54.8)
TSq 2T T
pendent. On the contrary, it is often found that successive discrete
values of an EEG signal have a certain degree of interdependence. which is the signal’s average power or variance . An important
To describe this interdependence, one may compute the signal property of the autocorrelation function is that its Fourier
joint probability distribution. As an example, consider the defini- transform (FT) is:
tion of the joint probability applied to a pair of values at two dis-
crete moments, x(t1) and x(t2); assume that one disposes of N q
realizations of the signal; the number of times that at t1 a value v
and at t2 a value u are encountered is equal to n12. Thus, the joint
Sxx (f)  #q
R()exp(j2f)d  FT(Rxx () ) (54.9)

probability of x(t1) and x(t2) = u may be defined as follows: Sxx(f) is called the power density spectrum, or simply the
n12 power spectrum, a common method of EEG quantification
p(x(ti )  v,x(t2 )  u)  Lim (54.3) (Fig. 54.3). The power spectrum Sxx(f) is a function of frequency
NSq N
(Hz); it gives the distribution of the squared amplitude of differ-
A complete description of the properties of the signal gener- ent frequency components. It should be noted that the word
ated by a random process can be achieved by specifying the power does not have the meaning of dissipated power in an RC
joint probability density function: circuit but is used here in another sense. This discussion deals
 ( (x(t1 ),x(t2 ), p ,x(tn ) ) (54.4)
with a question of time series analysis. In general, a stochastic

for every choice of the discrete time samples t1, t2, …, tn and for
every finite value of n. The computation of this function, how-
ever, is rather complex. A simpler alternative to this form of
description is to compute a number of averages characteristic of
the signal, such as convariance, correlations, and spectra. These
averages do not necessarily describe a stochastic signal com-
pletely, but they may be very useful for a general description of
signals such as EEG.
The convariance between two random variables at two time
samples x(t1) and x(t2) is given by the following expectation:

E[ (x(t1 )  E(x(t1 ) ) ) (x(t2 )  E(x(t2 ) ) ) ] (54.5)

Estimating the covariance between any two variables x(t1)

and x(t2) requires averaging over a umber of realizations of an
ensemble. Another way to estimate the convariance, provided
that the signal is stationary and ergodic (for a discussion of
these concepts see Ref. 14), is by computing a time average, for
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

one realization of the signal, of the product of the signal and a

replica of itself shifted by a certain time k along the time axis.
This time average is called the autocorrelation function:

a x(ti )x(ti  k )
1 N
£ xx (k )  K x(ti )x(ti  k )L  (54.6)
T i1
where k = k•t. Figure 54.3 The power spectrum of an EEG signal is shown in the
lower left plot; vertical axis power indicated here as R(f); horizontal axis
The following description considers continuous random frequency, f, in hertz. On the lower right the corresponding autocorre-
variables x(t), for the sake of simplifying the formulas. Assuming lation function r() is plotted. The power spectrum and autocorrelation
that every sample function, or realization, is representative of the functions are considered to be composed of three components (, , )
whole signal being analyzed, it can be shown that for stationary corresponding to three EEG frequency bands. (Adapted from Zetterberg
and ergodic processes the time average xx() for one realization LH. Experience with analysis and simulation of EEG signals with para-
x(t) is an estimate of the ensemble average Rxx(): metric description of spectra. In: Kellaway P, Petersén I, eds.
Automation of Clinical Electroencephalography. New York, NY: Raven;
Rxx (t)  E[x(t)x(t  t) ] (54.7) 1973:161–201.)

time function may be expressed in one of the several ways, as a where W(fk) is the smoothing window with a duration of
voltage, a length, a velocity, a number of occurrences of a cer- (2p  1) samples or data points. Similarly, one can compute
tain event, and so forth. The power spectrum or simply the a smoothed estimate of the cross-power spectrum (Sxy), which
spectral density of the time function is the FT of the autocorre- might be called Cxy(f).
lation function, the dimension of which is the function’s ampli- The FFT power spectral analysis and its applications are
tude dimension squared. In case the signal dimension is in discussed in more detail below.
volts, the power spectrum is in (V2•sec) or (V2/Hz). Of course, The close relationship between the concepts of variance 2,
if the function’s amplitude is in any other dimension, the inten- autocorrelation (equations 54.6 and 54.7), and power density
sity of the corresponding power spectrum would be yet another spectrum (equation 54.9) has already been made apparent; in
dimension. It is useful to keep a clear distinction between electric fact Rxx(0) = 2 and
power dissipated in an electric circuit (P  1/T #0 V2/Rdt with
T
q
units [V2•sec/]) and power spectrum. s2  #
q
Sxx (f)df (54.14)
A function that represents the average correlation between
two signals x(t) and y(t) may be defined in terms equivalent to The autocorrelation function R() and the power density
expression 54.7: spectrum S(f) correspond thus to the second-order moment of
the probability distribution of the random process.
Rxy (t)  E[x(t)y(t  t) ] (54.10) In case the signals are not gaussian, higher order spectra
moments must be considered. These can be derived as follows.
where the signals x(t) and y(t) are assumed to have means of Assuming that the signal has mean = 0, one can write (as in
zero. Rxy() measures the correlation between the two signals expression 54.7):
and is called the cross-correlation function. Similarly, one can
define the FT of Rxy, which is the cross-power spectrum between Rxx (t1,t2 )  E[x(t)x(t1 )x(t  t2 ) ] (54.15)
signals x and y:
Similar to expression 54.9, the two-dimensional Fourier
Sxy (f)  FT[Rxy (t) ] (54.11) transform FT2 of Rxx(1, 2) can be defined as the bispectrum or
bispectral density:
Fundamental discussions of power spectra and related top-
ics are found in many textbooks on signal analysis, for example,
Bxx (f1,f2 )  FT2 [Rxx (t1,t2 ) ] (54.16)
Refs. 15 and 14. This discussion cannot go into details about
ways of Bendat and Piersol (16) and Otnes and Enochson (17). This discussion cannot go into details about ways of esti-
Application of the frequency analysis principle to EEG signal mating the bispectrum Bxx. For a detailed account of bispec-
analysis has a long history, beginning with the pioneering work tral EEG analysis, refer to Huber et al. (28) and to
of Dietsch (18), Grass and Gibbs (19), Knott and Gibbs (20), Dumermuth et al. (29). It is, however, interesting to note that
Drohocki (21), and Walter (22,23). Brazier and Casby (24) and high Bxx values for a couple of frequencies, f1 and f2, indicate
Barlow and Brazier (25) first computed the autocorrelation phase coupling within the frequency triplet f1, f2, and
functions of EEG signals. The general principles on which this (f1  f2). The third moment of the probability distribution,
work has been based have remained essentially the same since or skewness, is related to the bispectrum. When there exists a
Wiener proposed these signal analysis methods (for a review, sufficiently strong relation between two harmonically related
see Ref. 26). An important advance in computing power spec- frequency components in a signal, there will exist a signifi-
tra has been achieved with the introduction of a new algorithm
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

cant bispectrum and skewness. The process in such a case is

for computing the discrete Fourier transform, known as the fast not gaussian; if it were gaussian with mean zero, the bispec-
Fourier transform (FFT) (27). In this case, it is assumed that one trum would be zero. The bispectrum can be used to deter-
wants to compute the power spectrum of a discrete EEG signal; mine whether the system underlying the EEG generation has
the epoch [x(t1)] is considered as a signal sampled at intervals nonlinear properties. An example of this form of analysis is
t, x(n t) with a total of N samples (n = 1…N). By using the given in Figure 54.4.
discrete FT, the so-called periodogram F(fi) can be computed: This section has demonstrated the progression from the
basic principles of probability distribution and corresponding
` a x(tn )exp(j2 # i ¢f # n ¢t) ` (54.12)
2
¢t N moments to the concepts of autocorrelation, power spectra,
Fxx (fi ) 
N n1 and high-order spectra. It is also of interest to examine the
moments of the spectral density Sxx(f), because this analysis
where fi = i  f with i = 0, 1, 2, …, N. The periodogram can be leads to another set of concepts applicable to EEG analysis, the
smoothed by means of a window W(fk) in order to obtain Pxx(fk), so-called descriptors of Hjorth (30). Thus, one can define the
which is a better estimate of the real power spectrum Sxx(f): nth spectral moment as follows:

Pxx (fi )  a W(fk )Fxx (fik )

p q

kp
(54.13) an  #q
(2f) nSxx (f)df (54.17)

discussion below illustrates how these spectral moments a0, a2,

and a4 are related to the descriptors proposed by Hjorth.

INTERVAL OR PERIOD ANALYSIS

An alternative method of EEG signal analysis is based on meas-
uring the distribution of intervals between zero and other level
crossings, or between maxima and minima.
A level crossing may be defined in general terms as the time
at which a signal x(t) passes a certain amplitude level b; b  0 is
a special case referred to as zero crossing (Fig. 54.5). Knowledge
of the probability density function of the intervals between suc-
cessive zero crossings can be useful in characterizing some statis-
tical properties of the signal x(t) (mean value 0). p0() can be
called the probability distribution density function of the inter-
vals between any two successive zero crossings and p1(), corre-
sponding to the total time  between successive zero crossings at
Figure 54.4 Contour map of the normalized bispectrum (also called
which the signal changes in the same direction (i.e., from posi-
bicoherence) of an EEG signal recorded from a subject who presented
tive to negative or vice versa). In practice, these functions can be
an alpha variant (the corresponding power spectrum is shown in Fig.
approximated by computing histograms of the interval length
54.8). The plot shows three maxima in the value of bicoherence
between two successive zero crossings or between zero crossings
(0.25). One is at the intersection of approximately 5 and 5 Hz (phase
at which the signal has a derivative with the same sign. The
coupling between 5, 5, and 10 Hz); another one is at the intersection of
moments of the distribution function can also be computed; the
about 7 and 7 Hz (phase coupling between 7, 7, and 14 Hz). Still
simplest case is to compute the average number of zero crossings
another is at the intersection of about 10 and 10 Hz (phase coupling
per time unit (N0) (e.g., per second) of the signal x(t):
between 10, 10, and 20 Hz). This means that the two peaks seen in the
power spectrum of Figure 54.8 at 5 and 10 Hz, respectively, are harmon-
ically related, that is, 5 Hz is one-half subharmonic of the dominant
alpha frequency. Moreover, there is another component at 20 Hz, diffi-
cult to see in the power spectrum, of Figure 54.8, which is also harmon-
ically related to the alpha frequency (i.e., a second harmonic of the
alpha component is also present). Another component at about 7 Hz
related to 14 Hz can also be identified. (This component may be distin-
guished as a small notch at the flank of the 10-Hz peak in the power
spectrum of Fig. 54.8.)

The zero-order moment is then:

#
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

a0  Sxx (f)df (54.18)

q

which is equal to the variance 2.

It can be shown (for derivation, see Ref. 31) that the second-
order moment is defined by the following expression:
q
d2Rxx (t) dx(t) 2
a2  #
q
(2f) 2Sxx (f)df  
dt2
`
t0
 E c
dt
d

(54.19)
and the fourth-order moment is: Figure 54.5 A: An EEG signal x(t). B: Plot of the corresponding power
q 4 d x(t) 2 2 spectrum w(f) as function of frequency f in hertz. C: Plot of the autocor-
#
d Rxx ()
a4  (2f) 4Sxx (f)df   4 `  Ec d relation function (). D: Plot of p0() (i.e., the distribution density function
q d t0 dt2
of the intervals between any two successive zero crossing). E: Plot of p1()
(54.20) (i.e., the distribution density function of the time between successive zero
In this way the spectral moments relate to the derivatives of crossings at which the signal x(t) changes in the same direction, from pos-
the autocorrelation function Rxx() and of the signal x(t). The itive to negative or vice versa). (Illustration courtesy of R.A.F. Pronk.)

#
tion function or the power spectrum, or by distribution of
N1
0    p0 ()d (54.21) intervals between level crossings. In all cases, the EEG was
0
treated as a stochastic signal without a specific generation
This means that the number of zero crossings per time unit model. Therefore, all the previously described methods and
N0 equals the reciprocal of the mean interval length . related ones are nonparametric methods. Parametric methods
In some cases it is helpful to determine the probability density may also be used to analyze EEG signals; in such cases one
function of the intervals between two adjacent zero crossings assumes the EEG signal to be generated by a specific model. For
where the sign of x(t) changes from negative to positive or vice example, assuming that the EEG signal is the output of a linear
versa. In this case, it is necessary to compute additionally the zero filter given a white noise input allows characterization of the
crossing interval distribution of the first derivative of x(t), y(t) linear filter by a set of coefficients or parameters (e.g., it may
(i.e., y(t) = dx(t)/dt). If the signal x(t) to be analyzed is quasi-sta- correspond to an autoregressive model as explained below).
tionary and has a gaussian distribution, a mathematical relation Therefore, EEG analysis methods can be divided into two
between Nk, the average rate of zero crossings of the kth deriva- basic categories, parametric and nonparametric. Such a division
tive of x(t), and the power spectrum Sxx(f) can be shown (32–34): is conceptually more correct than the more common differen-
q tiation between frequency and time domain methods because,
#f
1>2
2k2
Sxx (f)df as has been explained, such methods as power spectra in the
0
Nk  4≥ q ¥ (54.22) frequency domain and interval analysis in the time domain are
#f 0
2k
Sxx (f)df closely related; indeed, they represent two different ways of
describing the same phenomena. The methods of EEG analysis
described here are classified as shown in Table 54.1.
In interval analysis, only the values Nk for k = 0, 1, 2 are usu- Not all EEG analysis methods can be assigned to one of the two
ally computed. N0 thus represents the average rate of zero cros- general categories just described. Those having mixed character
sings of dx(t)/dt (i.e., the rate of intervals between extremes of (i.e., methods that have, as a starting point, a nonparametric or
the signal x(t)); N2 represents the average rate of zero crossings parametric approach that is combined with pattern recognition
of d2x(t)/dt2 (i.e., the rate of intervals of the a nonparametric or techniques) must be considered separately. The latter fall into the
a parametric approach that is combined inflection points category of pattern recognition methods. Last, this section shall
of x(t)). discuss topographic analysis methods, in which the emphasis is on
It can be shown (32) that expression 54.22 can separately topographic relations between derivations. Not included here are
also be given in terms of the autocorrelation function: the evoked potentials, which are discussed elsewhere.
N0 Rxx () 1>2 A thorough review of the main techniques currently in use
 c () 2 c 1  d d  fg (54.23) in EEG analysis has been edited by Gevins and Remond (37).
2 Rxx (o)
For more details on methods of analyzing brain electric signals,
where fg is the so-called gyrating frequency (35). the reader is referred to this authoritative handbook.
These relations between the number of zero crossings per
time unit and either spectral moments or the autocorrelation
function for EEG signals have been studied in detail by Table 54.1
Saltzberg and Burch (36), who concluded that, when the pur-
pose is to monitor long-term changes in the statistical proper- EEG Analysis Methods
ties of EEG signals, it is legitimate to use average zero-crossing
rates to calculate moments of the power spectral density. Nonparametric Methods
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Instead of measuring intervals between zero crossings, one can Amplitude distributions
characterize a signal by determining intervals between successive
Interval or period distributions
maxima (or minima), which defines a “wave,” or between a max-
imum and the immediately following minimum or vice versa, Amplitude–interval scatter plots
which defines a “half-wave.” The section “Mimetic Analysis” con- Correlation functions
siders some of the variants of interval analysis as applied to EEG
Auto- and cross-correlation
signals; the straightforward applications of interval analysis are
described in the section “Time–Frequency Analysis.” Complex demodulation
Power spectral analysis
Time-varying spectra
EEG SIGNAL PROCESSING
METHODS IN PRACTICE Cross-spectral functions (coherence and phase)
Bispectra
The previous section considered the statistical properties of
EEG signals as realizations of random processes, explaining Walsh and Haar transforms
how such signals can be characterized by the corresponding Hjorth slope descriptor
probability distribution and its moments, by the autocorrela-

NONPARAMETRIC METHODS kurtosis also deviated significantly from the values expected for a
gaussian distribution depending on sleep stage (48,49). These
Amplitude Distribution observations have led to a study of higher order moments of the
A random signal can be characterized by the distribution of spectral density function using bispectral analysis.
amplitude and its moments. An example of an amplitude distri- The method recommended to test whether EEG amplitude
bution is shown in Figure 54.2. The first question that is asked distributions are gaussian is that proposed by Gasser (7,38); it
regarding the amplitude distribution of an EEG epoch is whether involves calculating skewness and kurtosis after correction in
the distribution is normal or gaussian. The most common tests of view of the possibility that adjacent samples may have a large
normality are the chi-square goodness-of-fit test (17), the (e.g., 0.50) degree of correlation. The allowed kurtosis and
Kolmogorov–Smirnov test, or the values of skewness and kurtosis skewness values can be found in statistical tables. Kurtosis in
(7,38). It has been shown (39) that, for the small EEG samples most cases without paroxysmal activity or artifacts is within the
usually analyzed, the Kolmogorov–Smirnov test is more powerful limits allowed to accept the normality hypothesis; skewness dif-
than the chi-square test. It should be emphasized that in order to ferent from zero is encountered particularly in those cases in
apply these tests of goodness-of-fit, two requirements must be sat- which harmonic components are present in the power spectra.
isfied: stationarity and independence of adjacent samples. The In such instances, the bispectrum exists (see below).
first requirement was considered in the previous sections. The sec- An alternative method of calculating measures of EEG ampli-
ond requirement is a well-known prerequisite for the application tude was developed by Drohocki (50) and is used mainly in psy-
of the statistical tests of the type we consider here. Persson (40) chopharmacologic and psychiatric studies (see review in Ref. 51).
has clearly pointed out the pitfalls of applying goodness-of-fit This method involves measuring the surface of rectified EEG
tests to EEG amplitude distributions. The problem is that the EEG waves. Its usefulness for routine EEG analysis is limited.
signals are usually recorded at such a sampling rate that, depend-
ing on the spectral composition of the signal, adjacent samples are Interval Analysis
more or less correlated. In this way, the second requirement is Interval or period analysis has been used, as described above, to
commonly violated. This has also been shown clearly by McEwen study the statistical properties of EEG signals in general and in
and Anderson’s (5) statistical study of EEG signals. The degree of relation to other analysis methods, such as autocorrelation func-
correlation between adjacent samples can be deduced from the tions and power spectra. This discussion considers a more prac-
autocorrelation function. Persson found that a correlation coeffi- tical aspect, the simplicity of evaluating EEG signals using
cient of 0.50 or larger for adjacent samples introduces a consider- interval analysis. The method, as originally applied by Saltzberg
able error in interpreting a goodness-of-fit test. His experience et al. (52) and Burch et al. (53), has been shown to be useful pri-
with EEG signals led to the conclusion that sampling rates in most marily in quantifying EEG changes induced by psychoactive
cases should be restricted to about 20/sec in order to achieve an drugs (54–57), monitoring long-term EEG changes during anes-
acceptably small degree of correlation between adjacent samples. thesia (58,59), psychiatry (60), and sleep research (see Ref. 35).
It is of general interest to know an EEG sample’s type of ampli- When using interval analysis, it is good practice to compute
tude distribution. Several studies have been carried out, mainly not only the zero crossings of the original EEG signal, but also
investigating whether or not EEG amplitude distributions were those of the signal’s first and second derivatives, to obtain more
gaussian. Saunders (41), using a sample rate of 60/sec, epoch information about the spectral properties of the signal. One
lengths of 8.33 seconds, and the chi-square test, concluded that disadvantage of this method is sensitivity to high-frequency
alpha activity had a gaussian distribution; this confirmed previous noise in the estimation of zero crossings. This problem can be
results from Lion and Winter (42) and Kozhevnikov (43), who avoided by introducing hysteresis, that is, by creating a dead
used analog techniques. On the contrary, Campbell et al. (44), band (e.g., between a and –a V) so that no zero crossing can
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

using a sample rate of 125/sec, epoch lengths of 52.8 seconds, and be detected when the signal has an amplitude between those
the chi-square test, concluded that most EEG signals had non- limits. In this way, Pronk et al. (58) have found that a dead band
gaussian distributions; however, it is likely that in this case the dual between 3 and –3 V is a good practical choice. Another dis-
requirements of stationarity and independence were not met. advantage is that, when examining histograms of zero-crossing
The results obtained by Elul (45) are of special interest counts, it is easy to underestimate the contribution of low-
because he examined EEG time-varying properties using ampli- frequency components, of which there may be very few, and to
tude distributions for epochs of 2 seconds (200 samples/sec, chi- overestimate fast frequency components. These disadvantages
square goodness-of-fit test); this study most certainly failed to are particularly evident when zero-crossing histograms and
meet the requirement of independence. Nevertheless, Elul found power spectra of the same signal are compared as shown in
that a resting EEG signal could be considered to have a gaussian Figure 54.6 (61). Sometimes corrections are made to enhance
distribution 66% of the time, whereas, during performance of a the number of long intervals in relation to the short ones, but
mental arithmetic task, this incidence decreased to 32%. this may complicate the interpretations even more.
Evaluating a small series of waking EEGs in twins, Dumermuth Another approach is to compute zero-crossing intervals only
(46,47) found amplitude distribution deviations from gaussian- within determined frequency bands; this may solve the problem
ity in the majority of the subjects; he tested the normality of missing superimposed waves (62,63).
hypothesis by way of the third- and fourth-order moments, The main advantage of zero-crossing analysis is ease of com-
skewness, and kurtosis. In adult sleep EEG, skewness and putation, which makes this method particularly attractive for

the interval between zero crossings, and in amplitude by the

peak-to-trough amplitudes. This hybrid method had been pro-
posed repeatedly in the past by Marko and Petsche (64), Leader
et al. (65), Legewie and Probst (62), and Pfurtscheller and Koch
(66); it has been applied intensively in a clinical setting by
Harner (67) and Harner and Ostergren (68). The latter called
this method “sequential analysis” because the amplitude and
interval duration of successive half-waves are analyzed, dis-
played, and stored in sequence in real time. The method used by
these authors requires that the sampling rate be at least 250/sec,
the zero level be updated continuously by estimating the run-
ning mean zero level, and, as just discussed, there be a dead
band to avoid the influence of high-frequency noise. The high-
frequency sampling is desirable in order to obtain a relatively
accurate estimate of the peaks and troughs. The amplitude and
the interval duration of a half-wave are defined by the
peak–trough differences in amplitude and time; the amplitude
and the interval duration of a wave are defined by the mean
amplitude and the sum of the interval durations of two consec-
utive half-waves. These data are displayed in a scatter diagram
as illustrated in Figure 54.7.

Figure 54.6 Two examples for comparison of iterative interval analysis

and power spectra of the same EEG signals. The intervals are plotted as
inverse frequencies. The agreement is fairly good in the case presented
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

above; a pronounced rhythmic component (peak at about 11 Hz) is pres-

ent. However, in the plot below, the interval analysis emphasizes in a
marked way the high-frequency components. (Adapted from Matejeck M,
Schenk GK. Die iterative Intervall-Analyse-Ein methodischer Beitrag zur
Quantitativen Beschreibung des Elektroenzephalogramms in Zeitbereich.
Figure 54.7 A display of sequential analysis obtained in real time. The
In: Schenk GK, ed. Die Quantifizierung des Elektroenzephalogramms.
dots represent individual half-waves displayed within 2 msec of their
Konstanz: AEG Telefunken; 1973:293–306.)
occurrence in each of the eight channels. The distribution of dots, for
example, in the 8- to 16-Hz range (frequency equivalents of wavelength
the online quantification of very long EEG records, for exam- are used) gives an indication in amplitude and frequency of the alpha
ple, during sleep or intensive monitoring. To perform interval rhythm. Side-by-side comparison of homologous areas allows assess-
analysis, it is useful to combine it with prefiltering (31) in the ment of symmetry. Marking in y-axis indicates 50 V. (Illustration cour-
analysis of narrow-band signals. tesy of R.N. Harner; also in Lopes da Silva FH, Cooper R, Dumermuth
G, et al. Sampling, conversion, and measurement of bioelectrical phe-
Interval–Amplitude Analysis nomena. In: Remond A, ed-in-chief; Brazier MA, ed. Handbook of
Interval–amplitude analysis is the method by which the EEG is Electroencephalography and Clinical Neurophysiology. Vol 4. Part A.
decomposed in waves or half-waves, defined both in time, by Amsterdam: Elsevier; 1976.)

Correlation Analysis However, correlation analysis has lost much of its attractive-
In practical terms, the computation of correlation functions in ness for EEG analysts since the advent of FT computation of
the 1950s and 1960s constituted the forerunner of contemporary power spectra. The latter technique is less time-consuming and
spectral analysis of EEG signals (25,69,70) and provided an therefore more economical, and, in general terms, more power-
impetus to implement EEG quantification in practice. However, ful. Above all, it is difficult to determine from an autocorrelation
the computations were time-consuming and therefore not widely function EEG components when the signals contain more than
used. A simplified form of correlator was introduced, based on one dominant rhythm, an investigation that can be done simply
the fact that auto- or cross-correlation functions can be approxi- by using the power spectrum (Fig. 54.8). Nevertheless, it should
mated by replacing the signals x(t) and x(t  ) (see equation be noted that the simplified methods of correlation analysis just
54.4) by their signs (sign x(t) and sign x(t  ), where sign x(t) = described and used in the 1960s can still have practical value in
1 for x(t)  0 and sign x(t) = –1 for x(t)  0), as demonstrated simple problems, such as computing an alpha average.
by McFadden (71). The function thus defined is called the polar- The computation of autocorrelation functions has been
ity coincidence correlation function, and it has proved useful in revived due to the introduction of such parametric analysis
EEG analysis (72–74). Another simplified form of EEG analysis methods as the autoregressive model, which, as described
that is akin to correlation has been used by Kamp et al. (75) and below, implies the computation of such functions. Michael and
Lesèvre and Remond (76). It can be called autoaveraging and Houchin (77) have even proposed a method of segmenting
consists of making pulses at a certain phase of the EEG (e.g., zero EEG signals based on the autocorrelation function.
crossing, peak, or trough) that are then used to trigger a device Related to correlation functions is the method of complex
that averages the same signal (autoaveraging) or another signal demodulation (78). With this method, a particular frequency com-
(cross-averaging). In this way, rhythmic EEG phenomena can be ponent (e.g., ~10 Hz) can be detected and followed as a function
detected and some characteristic measures obtained. of time. In this case, a priori knowledge of the component to be

Figure 54.8 Left-hand column: Different ways of plotting

the spectrum of the same EEG epoch, the bicoherence of
which is shown in Figure 54.4. First plot: y-axis, power in
dB, and x-axis, frequency (Hz) along a linear scale; the
90% confidence band of the spectral estimate is indicated.
Second plot: y-axis, power in V2/Hz, and x-axis as above.
Third plot: y-axis, power in dB, and x-axis, frequency (Hz)
logarithmic scale (this way emphasizes somewhat the low-
frequency components). Fourth plot: y-axis, power in
V2/Hz, and x-axis, frequency (Hz) along a logarithmic
scale. Right-hand column: First plot: squared coherence
(Coh or (2)) between two symmetric EEG signals; the
power of one is shown in the plots on the left side. Second
plot: the same function as above; along the vertical axis the
z transformed coherence is plotted, z = (1/2)ln((1  )/
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

(1 – )); the advantage of this form of presentation lies in

the fact that, in this case, the confidence bands are the
same for the whole curve and are not dependent on
the value of 2. Third plot: phase spectrum corresponding to
the coherence spectrum shown above.

analyzed is necessary. Assuming, thus, that in an EEG signal a Length

component at about 10 Hz exists and should be followed, one can The length of the epoch T to be analyzed must be selected. It is
set an “analysis oscillator” at 10 Hz; the oscillator output and the important to take into account that the epoch should be short
signal are then multiplied. The product contains components at enough to avoid nonstationarity segments but long enough to
the sum frequency (~20 Hz) and at the difference frequency (~0 obtain the desired level of frequency resolution f; the maximum
Hz). This product is smoothed so that only the difference compo- f is, of course, f = 1/T. In many clinical applications one uses
nents (at about 0 Hz) are considered. In this way, phase and T = 5 or 10 seconds.
amplitude of EEG frequency components can be detected and
their modulation in time determined. Complex demodulation Frequency Smoothing and Ensemble Averaging
has been used to analyze rhythmic components of visual poten- The estimate of one frequency point of a periodogram Fxx(f1) of
tials (79) and sleep spindles (80). This method is similar to a direct one EEG epoch has a chi-square distribution with only 2 degrees
Fourier analysis in which an EEG signal is multiplied by sines and of freedom. The number of degrees of freedom must be increased
cosines at a particular frequency in the study of evoked potentials and the estimate variance reduced either by averaging for a num-
(81) and also the method of phase-locked loop analysis as used to ber of equivalent epochs or by smoothing over adjacent frequency
detect sleep spindles (82,83). components. Sometimes both ensemble averaging and frequency
smoothing are used. Generally, the spectral estimate Pxx(f1) (equa-
Power Spectra Analysis
tion 54.13) should correspond to at least 60 degrees of freedom
A classical way of describing an EEG signal is in terms of fre- (91), which allows acceptable estimates of spectral values. This
quency as established by the common EEG frequency bands. implies that an ensemble of at least 30 epochs should be used if
It is possible to obtain information on the frequency compo- only ensemble averaging is carried out. The number of degrees of
nents of EEG signals using interval or period analysis. freedom can also be increased, at the expense of frequency resolu-
However, the most appropriate methods in this respect are tion f, by using a spectral window W(fk) (see equation 54.13). A
analog filtering or Fourier analysis, using either expression spectral window is defined by its form and duration. The duration
54.9 (i.e., the FT of the autocorrelation function) or expres- at the base is given by the distance between truncation points.
sions 54.12 and 54.13 (i.e., the periodogram). Several forms of Using a window with a large base reduces the variance but
analog filtering were introduced in the early days of EEG increases the bias of the estimator. An excessively large window
research; that technique reached a technical level appropriate decreases too greatly the equivalent frequency resolution f. In
for clinical application mainly due to the work of Walter practice, therefore, a complex compromise between all the afore-
(22,23). Even in the 1960s banks of active analog filters were mentioned points must be reached. Details about the technicalities
used to decompose EEG signals into frequency components of choosing the appropriate form W(fk) can be found in Jenkins
(84–87). In 1975, Matousek and collaborators compared ana- and Watts (14) and Künkel and EEG Project Group (92). A good
log and digital techniques of EEG spectral analysis and deal of freedom in the choice of the spectral window is tolerable;
demonstrated clearly the superiority of digital techniques. the appropriate choice depends on the practical use of spectral
Digital methods are more accurate and flexible; using digital analysis. In EEG quantification in the clinical routine, it is com-
computers simplifies multichannel analysis. mon to compute average spectra by making averages of ensembles
The crucial landmarks in the development of EEG quantifi- of 10 epochs of 10 seconds (N = 1024) each, using an elliptic win-
cation methods have always followed technical advances: first, dow five sample points wide for smoothing; the equivalent band-
banks of active analog filters as just described; second, large dig- width is thus 0.5 Hz. The resulting estimate corresponds,
ital computers (88,89); and third, a fast algorithm for digital therefore, to less than 100 (more precisely, 93) degrees of freedom,
computation of discrete FTs, Cooley and Tukey’s (27) so-called owing to the fact that for each frequency component the power
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

FFT. The latter has since been used extensively in EEG analysis estimate is based on 2 degrees of freedom; this number must be
(review in Ref. (90)). multiplied by 10 (epochs), by 5 (window width), and by a factor
This chapter cannot discuss the technical aspects of applying 0.93 that corresponds to the fact that the window is elliptic.
FFT spectral analysis to EEG quantification; for these aspects,
the reader is referred to Matousek et al. (35), Dumermuth (90), Calibration
and the books of Jenkins and Watts (14), Otnes and Enochson The dimension of power spectra in EEG analysis is intensity per
(17), and Gevins and Remond (37). It is sufficient to state here bandwidth; the unit of measurement is in V2/Hz (93).
that, when planning to perform FFT spectral analysis, the elec- Calibration can be carried out using sine waves, as proposed by
troencephalographer should consider the basic issues described Abraham et al. (94), Clusin et al. (95), Dumermuth and Flühler
in the following sections. (96), and Matousek et al. (35); Sciarretta G, Erculiani P (97)
proposed a simple method, using a single rectangular pulse,
Digitization and Prefiltering that has practical advantages.
Digitization and prefiltering were discussed in relation to the
sampling process. It is necessary to define beforehand the Graphic Representation
frequency range over which the spectrum should be com- The graphic representation of power spectra merits special
puted, not only to avoid aliasing, but also to minimize compu- attention. In most instances, the EEG analyst needs a plot of
tation time. power spectra, as shown in Figure 54.8. In most cases, the

vertical scale is simply the spectral density as computed by way

of the Fourier coefficients (i.e., in V2/Hz). A preferred alterna-
tive is plotting the spectral intensity along a logarithmic scale.
The advantage of choosing log power intensity instead of sim-
ply power intensity is that the confidence intervals of the for-
mer are independent of the values of the spectral intensity.
Another technique involves computing the square root of the
spectral intensity and plotting it along the vertical axis.
Frequency is usually presented along a linear scale calibrated in
hertz; however, one may prefer, if the most attention is to be
paid to the lower frequencies (delta and theta), to compress the
frequency scale in the higher frequency range by plotting log Hz
(Fig. 54.8) along the horizontal axis or a more compressed scale
for frequencies higher than, for instance, 15 Hz. The presenta-
tion of a power spectrum plotting log spectral intensity verti- Figure 54.9 Display of a compressed spectral array showing the alpha
cally and frequency horizontally, and where the higher rhythm and the effects of various artifact-inducing maneuvers on the
frequency components are plotted in a more compressed way background activity. Note the change in alpha peak frequency from the
than the lower ones, is useful in routine clinical situations. eyes-closed situation when shaking occurs. Note also the influence of
artifacts in the spectra, particularly at tapping the electrodes; the large
Time-Varying Spectra peaks at low frequencies are artifactual. (Adapted from Bickford RG.
Time-varying spectra are often computed in order to analyze Computer analysis of background activity. In: Remond A, ed. EEG
more or less slowly changing EEG records. Such spectra can be Informatics. A Didactic Review of Methods and Applications of EEG Data
plotted simply by using the so-called compressed spectra array Processing. Amsterdam: Elsevier; 1977:215–232.)
(Fig. 54.9) as introduced by Bickford et al. (98,99). This method
is particularly valuable in obtaining an overall view of EEG
spectral changes for intraoperative or sleep monitoring (100). of characterizing EEG changes in relation to specific events,
Another form of plotting time-varying power spectra is by such as eyes closing/opening (101), fists closing/opening
using contour plots (i.e., plots of frequency against time), as (102), word association tests (103), and similar events. The
shown in Figure 54.10; in such plots, points corresponding to problem here is to quantify time-locked changes in EEG
equal values of power spectra computed from successive epochs spectra by way of ensemble averaging, using a particular
are connected by contour lines. These plots provide useful, eas- event as a trigger. Kawabata (101) considered this problem
ily interpretable visual displays of the evolution of power spec- analytically and proposed a formalism to compute time-
tra as a function of time. varying EEG spectra. Using this construct, he could show that
The computation of time-varying power spectra is partic- initially at eye closure, power within the alpha band
ularly important in those studies in which the problem is that increases, with the greatest concentration in the center

Figure 54.10 Contour plot of power spectra: note the fre-

quency shifts and increase in power intensity occurring in
the second part of the registration. (Adapted from
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Dumermuth G. Fundamentals of spectral analysis in elec-

troencephalography. In: Remond A, ed. EEG Informatics. A
Didactic Review of Methods and Applications of EEG Data
Processing. Amsterdam: Elsevier; 1977:83–105.)

frequency; later the center frequency shifts to a lower fre- The fact that EEG baseline values (i.e., pre-event segments)
quency. When the eyes open, the alpha power decreases and can change from trial to trial makes a statistical analysis based
the center frequency increases. on ensemble averages and standard deviations particularly dif-
Pfurtscheller and Aranibar (102) designed a method for ana- ficult. Kamp and Vliegenthart (103) proposed resolving this
lyzing EEG changes related to sets of stimuli such as those used type of difficulty by analyzing EEG epochs immediately before
to study the phenomenon of contingent negative variation and immediately after the event causing the change. In this way
(CNV). According to this method, the 6-second EEG epoch a pre-event epoch of, for instance, 4 seconds and a postevent
occurring before and after the event of interest is subdivided epoch of, for instance, 4 seconds are analyzed. A relatively large
into 1-second overlapping segments. For each segment, a power number of trials are recorded, and the degree to which the spec-
spectrum is computed; in the experiment quoted above, the tral value within a certain frequency band for each postevent
total power (0 to 32 Hz) and the power in the alpha frequency subsegment (e.g., 1-second long) differs from the pre-event
range (7 to 13 Hz) of each segment are averaged over a number epoch evaluated using a nonparametric test (Mann–Whitney
of equivalent segments, and mean values and standard errors test) (106). The end result is given as the number of trials in
calculated. In this manner, these authors demonstrated phasic which a certain frequency band changed significantly at a par-
decreases of power in the alpha band related to sensory stimu- ticular postevent segment. A similar method has been used by
lation and to the interstimulus interval in a CNV paradigm. Arnolds et al. (107), who compared spectral parameters of EEG
Pfurtscheller and Aranibar (104) used the same method to segments occurring after a behavioral event with pre-event val-
study changes in central mu rhythms occurring in relation to ues of the same parameter simply by using the sign test (106).
opening and closing the fists in normal subjects and patients The advantages of using this type of nonparametric method
(Fig. 54.11). A large number of studies where dynamical should be emphasized. Because the baseline values may vary
changes of the ongoing EEG, within different frequency com- dramatically, one runs the risk of failing to detect real EEG
ponents, were detected and characterized have been revised by changes related to a particular event that exist if one compares
Pfurtscheller et al. (105). only mean values. Directly comparing within each trial the
baseline with the postevent values, particularly by means of a
nonparametric statistical test, avoids the difficulty pointed
out above.
Statistical Evaluation
Statistical evaluation of spectra is not done only in the analysis
of time averaging EEG signals as discussed previously.
Frequently, it is helpful to determine whether or not two sets of
EEG power spectra differ significantly. The sets might have
been obtained under two different behavioral conditions or
during administration of two different treatments (e.g., a
placebo or a psychotropic drug); they could have been recorded
from symmetric derivations over the scalp. The question is a
simple one. Given two sets of power spectra, how can one deter-
mine whether they belong to the same population? The answer,
however, is not so simple. To start with, it is necessary to
emphasize that the power spectrum of a certain EEG epoch is
an estimate; thus, it also has a variance. A convenient way of
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

presenting estimate variability is to present the corresponding

confidence bands simultaneously with the average power spec-
trum as indicated previously. The question of testing whether
the two sets of spectra belong to the same population has been
Figure 54.11 Alpha power time course over a 6-second interval calcu- approached using a variety of methods (for references see Ref.
lated during voluntary hand movement (movement onset at 4 seconds). 49). Often analysis of variance (F test) and Student t test are
The scale on the left gives the percentage alpha power. Reference inter- applied (108). The F test in principle should be the first test
val 0 to 2 seconds with an absolute reference power of 6.8 V2 corre- chosen, because power density is a quadratic function. If the
sponding to 100% (this reference power corresponds to 23.8% of the number of degrees of freedom increases, the power density dis-
total power within the frequency band 0 to 32 Hz). The significance lev- tribution tends to normalize so that a t test can be applied.
els for the power decrease (event-related desynchronization [ERD]) are In general, it is advisable to apply to the power spectrum a
indicated on the right scale (10–2 corresponds to P  0.01, etc., sign logarithmic transformation, because it produces a symmetric
test). Note that a decrease of alpha power is indicative for ERD. distribution. Confidence intervals for log Pxx(f) are given
(Adapted from Pfurscheller G, Klimesch W. Functional topography dur- approximately as (35) log Pxx (f)  Za>2 ( 22>N) , where Za/2 is
ing a visuo-verbal judgment task studied with event-related desynchro- the 100 alpha/2 percentage point of the standardized normal
nization mapping. J Clin Neurophysiol. 1992;9:120–131.) distribution and N is the equivalent number of degrees of

freedom (16). Nevertheless, in many applications, especially if applied with good results (117). The important point is to
the number of degrees of freedom is small, it is preferable to define clearly which questions one wishes to answer through
apply nonparametric tests, such as the simple sign test or the coherence functions application. In this context, the most rele-
more powerful Wilcoxon or Mann–Whitney tests. For a vant points are as follows.
detailed analysis of the questions of statistical inference on EEG Is it possible to differentiate spectral components with fre-
data, the reader is referred to Gasser (7,109). This problem has quencies lying close to each other? For example, alpha and mu
been discussed in detail in relation to those psychopharmaco- rhythms may be difficult to differentiate in plots of power spec-
logic investigations in which EEG plays a central role tra but are readily separated using coherence functions com-
(56,57,60,110), but these aspects are too specialized to be con- puted between symmetric transversal derivations because the
sidered here. former show large values of transversal coherences, whereas the
latter have insignificant values (118).
Cross-Spectral Analysis Is it possible to detect the existence of bilateral synchronous
frequency components? Such components may make relatively
Cross-spectral analysis is an important part of EEG spectral
small contributions to power spectra, whereas they may give rise
analysis because it allows quantification of the relationships
to large coherence values. Coherences also may be useful in deter-
between different EEG signals. The section “Basic Statistical
mining the topographic relations of different EEG components.
Properties” mentioned the smoothed estimate of the cross-
The counterpart of coherence is the phase function
power spectrum Cxy(f); this quantity is the product of the
(Fig. 54.8), which provides information on the time relation-
smoothed discrete FT of one signal and the complex conjugate
ships between two EEG signals. An explanation of the use of the
of the other (see for details Ref. 14). Cxy(f) is a complex quan-
term phase is necessary here. Phase is used in the present con-
tity that therefore has a magnitude and phase:
text as a mathematical notion referring to the proportion of the
Cxy (f)  0 Cxy (f) 0 # exp[j£ xy (f) ]
period of a sine wave component of a signal as obtained
(54.24)
through Fourier analysis. The existence of a phase difference
between two EEG signals as obtained from the phase function
where j  11. The function of frequency xy(f) is the phase
can have different meanings. First, assume that the two signals
spectrum. It is useful to define a normalized quantity, the coher-
were recorded from bipolar derivations and that some compo-
ence function, as follows:
nents, for example, between 0 and 3 Hz, show an inverted
0 Cxy (f) 0 2 polarity (phase opposition, in EEG terms); in this case, the
Cohxy (f)  (54.25) phase function computed from the cross-power spectrum
Pxx (f)Pxy (f)
between the two signals will show, for the 0- to 3-Hz compo-
Examples of coherence and phase functions are shown in nents, a phase difference of 180. In the second case, assuming
Figure 54.8. In EEG analysis these functions are computed after that some components of the signal recorded from one deriva-
the application of cross-correlation functions, which was car- tion will be transmitted to the other derivation after a certain
ried out in a way similar to the autocorrelation function as delay time t (in seconds) the phase difference  (in degrees)
described previously (see for details Ref. 35). Coherence func- between the two signals will be linear with frequency in the
tions have been used in several investigations of the EEG signal range f (in Hz) corresponding to those components; in this
generation and their relation to brain functions, including case, the following relationship is valid:
studies of hippocampal theta rhythms (89,111), on limbic
¢£
¢t 
360 # ¢f
structures in humans (112), on thalamic and cortical alpha (54.26)
rhythms (113), on sleep stages in humans (48), on EEG devel-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

opment in babies (91,114), and on the spatial and temporal Until now, phase functions have been little used in routine
structures of dynamical features of local EEG signals (115,116). clinical EEG practice, probably because phase measurements
The latter measured coherence functions between EEG signals are generally difficult to interpret in terms of the two models
recorded using electrodes with 5- to 10-mm spacing from just presented. This is because scalp EEG derivations are a com-
epileptic patients, and found that in both the subdural surface plex representation of underlying cortical activity, so that the
samples and those from temporal lobe depth arrays, coherence potentials recorded at a distance are not easily reduced to clear-
declines with distance between electrodes of the pair, on the cut biophysical processes at the cortical level. Nevertheless,
average quite severely in millimeters. This demonstrates that Gotman’s (119) system of EEG analysis included phase func-
coherence fluctuations are quite local. tion computation in order to detect phase opposition between
The recommended way to evaluate coherence functions sta- the slow frequency components of different bipolar derivations.
tistically is to apply Fisher’s z transformation (14) as used by If the phase difference between the two signals is about 180
Lopes da Silva et al. (113) to analyze EEG signals. Thus, the con- with a significant coherence between the two signals, one can
fidence intervals and bias are dependent on the coherence val- conclude that a phase reversal exists. In Gotman’s system, the
ues (Fig. 54.8). search for phase reversals is performed only in the presence of
The use of coherence functions in routine clinical EEG slow activity. Computing phase functions to determine time
analysis has been rather limited thus far. In one system dedi- delays between EEG signals during epileptic seizures has also
cated to this type of analysis, coherence functions have been been proposed (120). The interpretation of these results,

however, poses a problem. A time delay between two signals can some rhythmic EEG activities have a significant bispectrum;
be concluded with certainty only if there is a linear relationship examples are the mu rhythm, which presents significant rela-
between phase and frequency within a certain frequency band; tions between harmonics of 10 Hz (5, 20, and 30 Hz), and the
if the coherence between the two signals is significant over only psychomotor variant, with relations between 6, 12, 18, and 24
a very narrow frequency band (around a peak), it may be Hz. Lopes da Silva and Storm van Leeuwen (131) found that
impossible to define a best fit line to the phase function. In such alpha rhythms recorded from the cortex also may have a signif-
a case, the result may be impossible to interpret definitively in icant bispectrum with harmonic components at 10 and 20 Hz.
terms of time delay. Instead of using the simple phase function, Moreover, alpha rhythms recorded on the human scalp may
it may be recommended to use a weighted phase function, as also show a significant bispectrum; in a few studied cases (Fig.
proposed by Carter (121), in the sort of problems just dis- 54.4), the so-called alpha variant has been characterized by a
cussed. A fundamental problem, however, is that very often the significant relation between the dominant frequency at 10 Hz
relations between EEG signals cannot be considered linear, so and the one-half subharmonic at 5 Hz. Under such circum-
that the use of coherence is not justified. Alternative methods stances, bispectrum computation disproves the alternative
have been developed (122) in order to overcome this limitation. hypothesis that the two components at 10 and 5 Hz are inde-
Another approach to identifying the source of EEG seizure pendent of each other and thus that low-frequency components
activity is use of a generalized form of coherence analysis, the would correspond to abnormal occipital activity. Furthermore,
so-called spectral regression–amount of information analysis bispectral analysis of some forms of visual evoked potentials
introduced and first applied to EEG analysis by Gersch and (132) has permitted putting in evidence some essential proper-
Goddard (123). This method has been used not only to analyze ties of the visual system.
seizures (124,125), but also to investigate the process underly-
ing the generation of hippocampal theta rhythms (126) and Walsh and Haar Transforms
thalamocortical alpha rhythms (127,128), the organization of
Alternative ways of computing power spectra have been pro-
infantile EEGs (129), and seizure activity in animals (130). This
posed. These include the Walsh and Haar transforms, which
analytic method involves computing first the coherence
can improve computational speed (90). These alternative meth-
between two EEG signals and then the partial coherence based
ods, however, have not yet proved to be of practical interest,
on a third EEG signal. Computing partial coherences implies
particularly because the FFT already provides a satisfactory
eliminating from each of the two EEG signals that part that can
solution.
be regarded as being determined by or predictable on the basis
of the third signal, which constitutes a form of regression analy- Hjorth Slope Descriptors The section “Basic Statistical
sis. If the initial coherence decreases significantly, one can con- Properties” defined the nth spectral moment of the power
clude that the coherence between the two initially chosen spectrum an (equation 54.17). Hjorth (30) and Berglund and
signals is due to the effect of the third one. As indicated in the Hjorth (133) have developed special hardware to compute in
references cited earlier, it is possible to thus determine the pat- real time the spectral moments a0 (equation 54.18), a2 (equa-
tern of interactions between a series of simultaneously recorded tion 54.19), and a4 (equation 54.20). In this way, the spectral
EEG signals and, eventually, to find the more likely source of a moments are not invariant in time as described earlier; rather,
given EEG phenomenon (e.g., seizure or rhythmic activity). spectral moments are allowed to vary as a function of time
(i.e., the statistical properties of the signal can vary in time),
Bispectra meaning that this form of analysis can be applied to nonsta-
Equation 54.16 defines the bispectrum. Although the power tionary signals.
spectrum is sufficient to describe the statistical characteristics Based on these quantities, Hjorth derived the following
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

of signals generated by a stationary gaussian process, deviation parameters, also called descriptors:
of amplitude distribution from normality indicates the need to
examine spectra of higher orders. This is particularly true for Activity, A  a0
M  [ (a2>a0 ) ]
1>2
the spectrum corresponding to the second-order autocovari- Mobility, (54.27)
C  [ (a4 >a2 ) (a2 >a0 ) ]
1>2
ance function R(1, 1): the bispectrum Bxx(f1, f2) can be Complexity,
estimated by smoothing the triple product
Fxx (f1 )Fxx*(f2 )Fxx (f1  f2 ), where Fxx(f) represents the com- Note that a0 = 2 (equation 54.18), that is, the variance of the
plex FT of the signal x(t) and Fxx*(f) represents the complex signal; a2 is the variance of the signal’s first derivative as shown
conjugate (see for details Refs. 28 and 29). Moreover, the bico- in equation 54.19; a4 is the variance of the signal’s second deriv-
herence of signal x(t), which is the normalized bispectrum of ative (equation 54.20). It should be noted that Hjorth’s descrip-
x(t), can be defined. (Do not confuse with coherence, which is tors give a valid description of an EEG signal only if the signals
the normalized magnitude of the cross-spectrum between two have a symmetric probability density function with only one
signals x(t) and y(t).) Until now, few studies have put bispectral maximum (134). This may be true for simple EEG generation
computation to practical use. Nevertheless, the specific infor- models (135,136) but not in general practice. Nevertheless, the
mation yielded by bispectra about the relationship between ease of computing Hjorth’s descriptors makes them attractive
harmonic frequency components in EEG signals can be valu- in real-time EEG analysis. The required calculations involve the
able. For example, Dumermuth et al. (49) have shown that computation of time derivatives only.

It must be noted, however, that computing the descriptor

complexity implies taking the ratio between the second and
first derivatives, so that the possibility of introducing large
errors is considerable. To avoid this, the signal bandwidth must
be rather limited. In the author’s opinion, Hjorth descriptors
can be useful if the EEG patterns to be analyzed have a simple
character, a probability density distribution with only one max-
imum, and change over time is rather gradual. It is, therefore,
not surprising that Hjorth’s descriptors have demonstrated Figure 54.12 Block scheme. A: The filtering process on a time contin-
value in monitoring time-varying EEG signals, for instance, uous signal. B: The autoregressive moving average filter model applied
during sleep (137). This method has also been used to quantify to a time discrete signal. (Adapted from Zetterberg LH. Means and
multichannel EEG recordings obtained under routing condi- methods for processing of physiological signals with emphasis on EEG
tions (138). analysis. In: Lawrence JH, et al. Advances in Biology and Medical
Physics. Vol 16. New York, NY: Academic Press; 1977:41–91.)
PARAMETRIC METHODS
It is reasonable to argue that, in general terms, EEG signals may uous case and the discrete case. According to the continuous
be analyzed by any suitable method regardless of precise knowl- case, the EEG signal x(t) is assumed to result from the operation
edge of their biophysical origins. It may be asked, however, of filtering (with a filter having as transfer function H(s)) on a
whether more appropriate methods of EEG analysis might be noise source with a flat spectrum within the frequency range of
developed if more precise models of the biophysical processes interest. In the discrete case, the EEG signal is given as a set of
underlying the generation of EEG phenomena (e.g., alpha samples x(k) resulting from a filter operation on an input noise
rhythms, delta waves, spike and wave complexes, and so on) signal e(k) with zero mean. The filter, corresponding to the
were available. In the particular case of alpha rhythm genera- autoregressive moving average (ARMA) model, is described by
tion, there exist biophysical models that can help in formulat- a linear difference equation of the following form:
ing a reply to such questions (136,139,140). These alpha
rhythm models have indicated that an EEG with a dominant a0x(k)  a1x(k  1)  p  apx(k  p)  b0e(k)
rhythmic component in the alpha frequency range can be  b1e(k  1)  p  bqe(k  q) (54.28)
described by a filter network with parameters related to physi-
ologically acceptable variables submitted to a noise input. This where q  p. The relation between x(k) and e(k) is given by the
filter network can be analyzed in a first approximation as a lin- sets of coefficients a1, …, ap and bi, …, bq with a0 = 1. In case
ear processor. This processor can be realized in terms of a math- bi = 0 for i = 1, …, q, we are left with the so-called autoregre-
ematical model. A special case of this model is the mixed ssive (AR) model:
autoregressive model as described by Zetterberg (141) and the
autoregressive model used by Gersch (142), Fenwick et al. x(k)  a1x(k  1)  p  apx(k  p)  e(k) (54.29)
(143), and Bohlin (144). Such methods are called parametric,
because in such cases the EEG signals are described in terms of The computation problem, therefore, is to estimate the coef-
a mathematical model characterized by a set of parameters. ficients. An important step in this estimation is defining the
A link may be said to exist between this type of mathemati- minimum number of coefficients to be computed.
cal model and the biophysical model of alpha rhythm genera- Fast algorithms exist to enable computation of those coef-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

tion, but this link is neither specific nor essential. The use of ficients; they are described in detail by Zetterberg (31),
such mathematical models in EEG analysis is yet to be justified Makhoul (154), and Eykhoff (155) among others, and
through pragmatic arguments. These models provide a practi- employ several criteria for estimating the order of the model.
cally useful method for quantifying EEG signals, not only in Using Durbin’s algorithm, it was found in a group of EEG
order to compute spectra (142–150), but also to detect EEG recordings of epileptic patients that the minimal order of the
transient nonstationarities such as epileptiform spikes and model was, in about 70% of the cases, equal to or smaller
sharp waves (146,151) and to subdivide the EEG into quasi-sta- than 5 (156). However, when one wishes a faithful reproduc-
tionary segments (152,153). Parametric methods allow consid- tion of the power spectral density, many coefficients may be
erable EEG data reduction. For instance, using an needed (31). In most applications it is sufficient to compute
autoregressive model, it is possible to describe an EEG signal the AR model of the EEG signal, so this section need not
using a few coefficients; by following the values of these coeffi- consider the special problems regarding ARMA model
cients, the signal’s time-varying properties can be traced. The computation.
coefficients can be used to classify EEG spectra using, for
instance, cluster analysis; moreover, the model can also be used Computation of Power Spectra
to help detect nonstationary events. The basic model can be The computation of power spectra using an AR or ARMA
described following the scheme of Figure 54.12, as proposed by model presents no special difficulties. Using a special algorithm
Zetterberg (31). In this figure, two cases are shown: the contin- (spectral parameter analysis [SPA]) developed at Zetterberg’s

(31) laboratory, both an estimation of the model parameters

and the best spectral representation can be obtained. In the case
of the ARMA model (equation 54.28) spectral density of a sig-
nal sampled with sampling interval is given by:

a bi (exp(  j2fi ¢t) )

q 2

i0
Pxx (f)  2 ∞ ∞
a aj (exp(  j2fi ¢t) )
P (54.30)

i0

In case of the AR model, the spectral density is as follows:

2
Pxx (f) 
` a aiexp(  j2fi ¢t) `
P 2 (54.31)

i0 Figure 54.13 Power spectrum of an EEG signal analyzed with a fifth-
order model. It consists of a low-frequency component () and two res-
In both cases described, spectral density is estimated using
onance peaks ( and ); the components are described by the
the sets of coefficients. In Zetterberg’s original computational
parameters G (power),  (bandwidth), and f (peak frequency). Dashed
procedure, the SPA, the EEG analysis is based on the ARMA
lines denote the individual spectral components; the solid line indicates
model (31). In this form of analysis, Zetterberg not only com-
the total spectrum. (Adapted from Isaksson A, Wennberg A. Visual eval-
putes the EEG power spectrum, but also decomposes the spec-
uation and computer analysis of the EEG—a comparison.
trum into a number of components to achieve a degree of data
Electroencephalogr Clin Neurophysiol. 1975;38:79–86.)
reduction; he usually distinguishes three spectral components
so that Pxx(f) is written as the sum of three components (see
also Fig. 54.3):
erties of the estimated noise, ě(k), resulting from the inverse
autoregressive filtering:
Pxx (f)  P (f)  P (f)  P (f) (54.32)

ê(k)  xk  a aix(k  i)
p
(54.34)
The delta () component is described by a first-order model; i1
both alpha () and beta () components require second-order
The EEG signal is said to be nonstationary for t = nT if the
models. The  component is defined by a power parameter G
null hypothesis can be rejected (i.e., if ě(k) deviates at a certain
and a frequency parameter , which denotes the correspon-
probability level from a noise with a normal distribution).
ding bandwidth; G is defined as:
Thus, nonstationarities in an EEG signal can be detected; this is
q
G  #q
P (f)df (54.33)
particularly interesting in the detection of EEG transients of
epileptic patients, as shown by Lopes da Silva et al.
(146,157,158). A simple test on each sample of the estimated
The rhythmic components ( and ) require two frequency
noise can give an indication of the stationarity of the signal at
parameters, the center or resonance frequency f or f, and two
that moment. However, instead of testing ě(k), a detection
power parameters, G or G, defined as in equation 54.33. An
function d(k) is used in order to obtain a certain degree of
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

example of the spectral decomposition of an EEG signal calcu-

smoothing; d(k) is defined as follows:
lated in this way is shown in Figure 54.13.

a c ̂ d
Spectral analysis can be performed much faster with the AR km
ê(n) 2
than with the ARMA model. It is also easy to use this model to d(k)  (54.35)
nkm
analyze multidimensional processes so that not only auto, but
also cross-power spectra can be computed. Because the square of a normally distributed variable (with
unity variance) follows a chi-squared distribution, the detec-
Inverse Autoregressive Filtering tion function should also have a chi-squared distribution with
The AR model can also be used in an inverted form, which a number of degrees of freedom (2m  1). The null hypothesis
leads to the inverse autoregressive filtering operation. Assuming can then be tested at a certain level, for example, at p  103. An
that an EEG signal results from a stationary process, it is pos- application of this process of inverse filtering for the detection
sible to approximate it as a filtered noise with a normal distri- of transient nonstationarities (epileptiform events) in EEG is
bution. Consequently, passing such an EEG signal through the illustrated in Figure 54.14.
inverse of its estimated autoregressive filter should result in a
normally distributed noise N with mean zero and variance 2. Time-Varying Signals: Kalman Filtering
The null hypothesis is that an EEG signal follows the assump- Parametric models can be extended in order to analyze time-
tion of stationarity and can be expressed in terms of the prop- varying signals. A method of analyzing this type of EEG signal

model of order p with q = p – 1; a good choice appeared to be

a value of p* = 11 or 13.

Segmentation Analysis
The original purpose of segmentation analysis as introduced by
Praetorius et al. (153) and Bodenstein and Praetorius (152) was
to find in an EEG signal those segments that could be consid-
ered to have unvarying statistical properties.
This means that those segments should be considered as
being quasi-stationary, and the segments could have variable
length. This necessitated the development of criteria for estab-
lishing divisions between segments. These authors based their
analysis on a parametric model of the EEG, an autoregressive
model as defined by equation 54.29. Consult the aforemen-
tioned references for details. Duquesnoy (159) proposed an
EEG segmentation method related to that just described. A
Figure 54.14 Scheme of the principle of automatic spike detection problem in applying this analytic method is the difficulty of
(ASD) analysis method using simulated signals. Top: The hypothesis is defining clinical–neurophysiologic boundaries between seg-
that the interictal EEG of an epileptic patient results from filtered noise ments. Therefore, judgment of whether the method produces
to which spikes have been added. For simplification in this scheme, the segments acceptable on clinical–neurophysiological grounds is
spikes are not represented as being the output of a filtering process; this, rather subjective and depends strongly on personal criteria.
however, would have been more realistic. Bottom: The analysis consists Nevertheless, this method may be useful in reducing data in
of computing an autoregressive filter model representing the hypotheti- analyses of long EEG records recorded under variable behav-
cal process, determining the corresponding inverse filter through which ioral conditions.
the EEG signal is passed, and squaring and smoothing the resulting error Michael and Houchin (77) proposed a similar method based
signal. The end result or detection signal is tested using the chi-square simply on computing a running autocorrelation function,
statistic; the time samples lying above a certain level are indicated by which ensures a quicker procedure. Barlow (163) used the
thin lines under the curve. Note that in this example the ASD program method devised by Michael and Houchin (77) to compare the
detected at the correct time samples the spikes that had been added to performance of automatic adaptive segmentation with those of
the filtered noise. (Adapted from Lopes da Silva FH, van Hulten K, selective analog filtering and inverse digital filtering in auto-
Lommen JG, et al. Automatic detection and localization of epileptic foci. matic evaluation of significant EEG changes associated with
Electroencephalogr Clin Neurophysiol. 1977;43:1–13.) carotid clamping. Of the three methods, the former was clearly
the best.
Adaptive segmentation was used to analyze a series of clini-
consists of applying the so-called Kalman estimation method of cal EEGs showing a variety of normal and abnormal patterns
tracking the parameters describing the signals (159–161). The (164); the computer method was used based on the autocorre-
input signal to a hypothetical processor responsible for gener- lation function. By means of this algorithm, EEG segments were
ating the EEG signal is assumed to be a normally distributed defined; similar segments were then clustered without supervi-
noise e(k). A model is assumed in order to represent the sion. The study concluded that minimal supervision of the clus-
observed signal; the process dynamics are represented by an tering process may be necessary. Nevertheless, this adaptive
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

autoregressive model. segmentation method is useful for obtaining significant data

The main objective of this procedure is achieved by means of reduction and has practical value for the clinical neurophysiol-
a recursive algorithm called the Kalman filter to obtain esti- ogist. A review of methods for analyzing nonstationary EEGs
mates of the model coefficients using earlier estimated data. has appeared (165).
This involves updating based on new samples of the time series
(31). The Kalman filtering procedure is not simple to imple- MIMETIC ANALYSIS
ment; for an appropriate procedure, it is necessary to choose,
properly, the order of the model and the initial conditions. An This form of analysis has been developed mainly by Remond
application of this method in the subdivision of EEG signals and collaborators (166–169) and is based on the general con-
(segmentation procedure) is described below. Without entering cept that automatic EEG analysis should mirror the visual
into the details of different procedures of Kalman filtering, it is analysis performed by electroencephalographers in their daily
of interest to note that a measure of EEG signal stationarity can practice. This is why it has been called mimetic analysis (168).
be derived from the application of this method. However, this analytic form uses tools common to other meth-
Isaksson (162) has introduced for EEG analysis an algorithm ods, particularly those nonparametric methods based on signal
called SPARK, which stands for spectral parameter analysis, features characterized in the time domain, namely
based on recursive Kalman filtering. He found that an AR interval–amplitude analysis (68). The peculiar aspect of
model of order p* = p  q gave as good results as an ARMA Remond’s mimetic analysis is that the whole procedure of

extracting EEG features and sets of features follows a syntactic

approach: half-waves and minimal descriptors correspond to
linguistic characters or letters, significant wave series such as K
complexes and spindles to words, segments composed of wave
series such as rhythms to paragraphs, and ensembles of seg-
ments to chapters or sections.
Based on these features it is possible to construct tables or
graphs that demonstrate synoptically the distribution of the
different features in an EEG epoch and determine their statisti-
cal properties for several epochs and derivations. A similar type
of analysis has also been proposed by Schenk (170). These
methods, which have as common background an iterative inter-
val analysis, tend to emphasize the high-frequency components
of the signal, compared directly to spectral analysis (see also
Fig. 54.6) (171). The section “Interval Analysis” stressed the
rather intimate relationship between interval analysis based on
the signal and its derivatives and the spectral moments (see
equation 54.22). Therefore, the methodology used by Remond
and collaborators does not differ essentially from spectral
analysis. The main difference is that mimetic analysis combines
feature extraction with segmentation and logical classification.

MATCHED FILTERING OR Figure 54.15 Continuous correlation coefficient write-out at three dif-
TEMPLATE MATCHING ferent points in time for an EEG signal and a template. In each record
the rectangle indicates the time window for comparisons of template
Matched filtering is a form of pattern analysis in which a certain and EEG. In the first two instances there is no match, but in the third
pattern or template (i.e., a set of values in the EEG signal x(t)) the template and EEG match exactly (they are identical); hence, the cor-
is detected by using cross-correlation (equation 54.10) between relation coefficient reaches a peak at 1.0. Sampling rate, 50 Hz; number
x(t) and a priori defined pattern m(t). (For the sake of simplic- of points in template, 72; duration of template, 1.24 seconds. (Adapted
ity the underscore of the stochastic variable is omitted in the from Barlow JS, Dubinsky J. Some computer approaches to continuous
following.) As in equation 54.10, one may write: automatic clinical EEG monitoring. In: Kellaway, P, Petersén I, eds.

Quantitative Analytic Methods in Epilepsy. New York, NY: Raven;
s(t)  # x(  t)m()d
0
(54.36) 1976:309–327.)

The efficiency of the estimator s(t) is defined as follows:

EEG signals by way of time-varying spectra, namely in the form
[s(t) ] 2 of spectral arrays, and by applying segmentation methods. A
0 t T
1 (54.37)
c #
tT
2
x dt d c # m dt d
0
2 more recent development in this respect is the introduction of
methods that can combine analysis both in time and in fre-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

quency in an optimal way. Some of these methods use a special

The estimator reaches its maximum value (=1) when m(t) is class of basis functions, the so-called wavelets. A function can
identical to x(t) and when both signals are aligned perfectly be accepted as a wavelet if it satisfies the following relation:
along the time axis. In this case, the template m(t) can best be q
extracted from the signal x(t). Various algorithms can be used
to complete this operation efficiently. Saltzberg and Burch (36),
#q
(t)dt  1 (54.38)

Herolf (172), Zetterberg (173), Lopes da Silva et al. (146), This means that wavelets have typically a waveform of a
Barlow and Dubinsky (174), and Pfurtscheller and Fischer damped oscillation. The essential point of this method consists
(175) have all suggested using matched filtering to help epilep- in decomposing the EEG signal in a set of wavelet functions
tiform events (Fig. 54.15). defined as follows:
(t  u) 1
TIME–FREQUENCY ANALYSIS s,u (t)   (54.39)
2s s
Above we have already mentioned that an important problem where s represents the time scale and u is the translation vari-
in EEG analysis is the fact that EEG signals, in general, can only able. Based on this definition a set of orthogonal wavelets can
be considered stationary during relatively short epochs. This be constructed that forms an orthonormal basis. A given func-
has led to the development of several ways of analyzing such tion, such as an EEG signal, may be characterized by the

corresponding wavelet coefficients. For a basic theoretical treat- the parameters p and k are sampled at intervals 2j. The proce-
ment of this issue, the reader is referred to Mallat (176). Wavelet dure is iterative and it is stopped as the set of waveforms is able
analysis was applied to ongoing EEG signals (177,178) and to to explain a given amount of the signal’s variance.
evoked potentials (179,180) with interesting results. The set of The corresponding results can be visualized by means of the
wavelets is limited. To represent nonstationary EEG signals, a so-called Wigner maps, an example of which is given in
wider repertoire of basis functions is desirable. With this aim in Figure 54.16 (183) for the analysis of an epoch of sleep EEG where
view, the method called matching pursuit was developed by the detection of different types of sleep spindles is put in evidence.
Mallat and Zhang (181) and applied to the detection of tran- An alternative way to compute the time evolution of the fre-
sients in EEG signals by Durka and Blinowska (182). A large set quency spectrum is to apply a windowed Fourier transform
of basis functions can be obtained by scaling, translating, and that gives information about gradual changes in frequency
modulating a window function g(t): spectra in the course of time. It was shown that this method can
1(t  u) it be useful in the analysis of ictal activity (184).
g(t)  g e (54.40) EEG signals recorded from several derivations represent large
2s s
data sets that may give related, or even redundant, information.
where in the time domain the function is concentrated around u Multivariate statistical methods can be useful in reducing such
with a width proportional to s, and in the frequency domain its large data sets and in determining a small number of statistically
energy is concentrated around  with a spread proportional to 1/s. independent components. Classically this has been accomplished
The minimal time–frequency variance corresponds to the by factor analysis or principal component analysis. Pioneering
condition that g(t) is gaussian. By means of the matching pursuit studies of Walter et al. (185) showed that alpha rhythms recorded
algorithm, using a dictionary of such basic functions, a conven- from the posterior cerebral regions could be accounted for by
ient expansion of a given signal can be obtained, as explained in two independent orthogonal components. It should be noted,
detail by Mallat and Zhang (181) and by Durka and Blinowska however, that while these methods are useful in data reduction,
(182). In practice, since EEG signals are available as real discrete they do not give information on the nature and location of phys-
time series, the basis function has the following form: iologic generators of EEG signals (186). An alternative approach
k consists in applying independent component analysis (ICA), ini-
g( , ) (n)  K( , )gj (n  p)cos a 2 n   b (54.41) tially proposed by Bell and Sejnowski (187) and Makeig et al.
N
(188–190). ICA is a method that can be used to separate a
where the index  = (j, k, p) is the discrete analog of (, s, u) of number of statistically independent signals, or sources, from an
equation 54.40, N = 2L represents the number of samples, and equal number of linear mixtures of these sources. The basic
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 54.16 Wigner plot of the analysis of an EEG epoch of 20 seconds recorded during light sleep using a match-
ing pursuit algorithm. Here the main objective was to detect sleep spindles automatically and to compare the results
with those obtained by visual inspection by experts. The spindles automatically detected are indicated by the letters
A to F. Spindles indicated by A and B were also detected by the experts. C, D, and E were classified by the experts as
single spindles, but F was outside the section marked by the experts. (Adapted from Durka PJ. Time Frequency
Analyses of EEG [PhD thesis]. Department of Physics, University of Warsaw; 1996.)

assumption is that EEG/MEG signals recorded at a given site at eling and ICA was proposed by Gómez-Herrero et al. (197), to
the level of the scalp result from the sum of the projected activi- estimate the sources of alpha rhythms and their propagation in
ties of a number of multiple brain sources, sometimes contami- the brain. The application of this methodology to 20 subjects
nated by extracerebral (artifacts) sources. ICA method aims at under resting conditions suggested that the major alpha generator
separating these sources in an optimal way. Mathematically the process consists of a strong bidirectional feedback between thala-
process of computing ICA consists of a simple transformation of mus and cuneus, while the precuneus appeared to participate also
one matrix (containing the sources) to another matrix (contain- in the generation of the alpha rhythm, which is in accordance with
ing the mixtures or recorded signals) by multiplication with a results obtained using other methodologies (see Chapter 5).
mixing matrix. The inverse of the latter can be used to decom-
pose the set of mixtures into the original sources. The problem is
that one does not know, a priori, how the sources are combined, SPATIOTEMPORAL DYNAMICAL ANALYSIS
that is, the coefficients of the mixing matrix are not known. To OF EEG: THE QUESTION OF PHASE
estimate these coefficients by way of the ICA method, the set of RELATIONSHIPS, SYNCHRONIZATION,
independent components is estimated by finding the minima of AND CAUSALITY
their mutual information. In case this process is done successfully
the independent components represent the original sources. It In the past editions of this book we present a historical
should be realized that the sources should be non-gaussian and overview of several aspects regarding methodological
that the number of sources should not exceed the number of approaches applied to the issue of spatial analysis of EEG sig-
mixtures. Anemüller et al. (191) presented a generalized method nals. These are now presented and discussed more extensively
that considers the EEG sources as eliciting spatiotemporal activ- in Chapter 55. Here we focus on some general considerations
ity patterns, corresponding, for example, to trajectories of active with respect to the analysis of phase relationships between
processes, propagating across the cortex. This led them to pro- EEG/MEG signals recorded from different sites and the related
pose a model of convolutive signal superposition, in contrast to concepts of phase locking and synchronization.
the commonly used instantaneous mixing model. In this way the We must emphasize that the determination of phase relation-
sources of spatiotemporal dynamics of EEG signals recorded ships between EEG/MEG signals is of great interest to get insight
during a visual attention task could be identified. Further ICA into the dynamics of underlying brain processes. Two main con-
has been applied to short-time Fourier transforms of EEG/MEG ditions have to be fulfilled in order to estimate relevant phase
signals, in order to improve the identification of sources of rhyth- relations. One is that spatial sampling is optimal for the phenom-
mic activity (192). ena of interest; another one is that the influence of volume cur-
There are many kinds of algorithms to compute ICA. Klemm rents is reduced to a minimum, as mentioned above. It should
et al. (193) reported a comparative study of the performance of always be remembered that EEG phenomena with potential fields
22 algorithms, with the objective of applying them in EEG that have spatial frequencies higher than the spatial sampling fre-
analysis. The results of this comparative study may help to select quency may give rise to important errors of interpretation (198).
a task-specific algorithm to analyze a variety of EEG patterns. Synchronization between EEG/MEG signals may be put in
A concern about the interpretation of the results of ICA is evidence in different ways. In EEG literature the terms synchro-
the question of how reliable the estimated components are. To nization and desynchronization are commonly used to refer to
answer this question Groppe et al. (194) presented a new algo- changes in power spectra of EEG activities in a given frequency
rithm for assessing the reliability of ICAs based on applying band, such as desynchronization or synchronization of alpha
ICA separately to split-halves of a data set. These authors rhythms with respect to a given event, and are expressed by the
showed further that ICA reliability is enhanced by removing the terms “event-related desynchronization or ERD,” and “event-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

mean EEG at each channel for each epoch of data rather than related synchronization or ERS” (199). Implicit to this terminol-
the mean EEG in a prestimulus baseline period. ogy is the assumption that when the activities of the neuronal
With the advent of combined fMRI and EEG recording, it elements in a population are not phase-locked to each other, the
has been proposed to apply ICA to decompose EEG and fMRI resulting amplitude—or power—in the EEG decreases (desyn-
signals in order to facilitate the integration of the two modali- chronized state), and the reverse happens when the neuronal
ties. This opens interesting possibilities to obtain multimodal activities are highly phase-locked (synchronized state). These
integrated images (195). An interesting field of application is in EEG phenomena are presented in Chapter 45, and thus will not
the field of epilepsy. In this respect Marques et al. (196) decom- be discussed further in this section.
posed EEG signals using ICA and identified the relevant com- Here we will focus on the notion of phase with respect to
ponents’ time courses to find the regions exhibiting fMRI signal pairs, or multiple, EEG/MEG signals: this means phase locking at
changes related to interictal activity. a distance, and not only locally within a circumscribed neuronal
A problem that one is faced while interpreting correlations population. In particular, considerable attention has been given
between EEG signals recorded from the scalp is the well-known to EEG/MEG oscillations within the gamma frequency range
smearing effect of the volume conductor currents. One method occurring in-phase at a certain time, as possibly constituting a
that may help to circumvent this difficulty is to use ICA as a pre- general mechanism of transient association between neuronal
liminary step in the analysis. With this objective in view a novel assemblies underlying sensory perception (200–202). This
methodology based on multivariate autoregressive (MVAR) mod- would constitute the substrate of the “binding” mechanism by

means of which the neuronal representation of different features the filtered signals is computed by means of the Hilbert trans-
of a complex sensory stimulus would be bound together to form form; finally the degree of phase synchrony, called phase-locking
a unified percept, a “gestalt.” This concept was also generalized to value (PLV) by these authors, is estimated by averaging the phase
brain systems engaged in memory processing, as the (para)hip- differences on the unit circle in the complex plane. In this way the
pocampal system, where synchronized gamma oscillations in the phase can be displayed separately from the amplitude compo-
entorhinal–hippocampal circuits may allow distributed neu- nent for a given frequency range (201,205,206). Using this
ronal populations to form functional assemblies necessary for methodology it was shown, for example, that the scalp
the formation of memory traces (203). EEG/MEG of subjects performing the perceptive task of recog-
The estimation of phase relationships is not trivial. Several nizing human faces induces a long-distance pattern of phase syn-
algorithms are used in order to estimate phase synchrony. The chronization that represents active coupling of the underlying
method used by the Paris group (204) starts by choosing neural populations. This coupling appears to be necessary for the
EEG/MEG epochs and applying a band-pass filter to select the realization of this cognitive task (207). An example is shown in
frequency band of interest; thereafter, the instantaneous phase of Figure 54.17, which illustrates how the perception of a human
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 54.17 Average scalp distribution of gamma EEG activity and phase synchrony. Color coding indicates gamma
power (averaged in a 34- to 40-Hz frequency range) over an electrode and during a 180-msec time window, from stimula-
tion onset (0 msec) to motor response (720 msec). Gamma activity is spatially homogeneous and similar between condi-
tions over time. In contrast, phase synchrony is markedly regional and differs between conditions. Synchrony between
electrode pairs is indicated by lines, which are drawn only if the synchrony is beyond the distribution of shuffled data sets
(P , 0.01). Black and green lines correspond to a significant increase or decrease in synchrony, respectively. (Adapted from
Rodriguez E, George N, Lachaux JP, et al. Perception’s shadow: long-distance synchronization of human brain activity.
Nature. 1999;397(6718):430–433.) (See color insert)

face (shown in the upright position) induces a long-distance sta- (212,213) or variable resolution electromagnetic tomography
ble pattern of phase synchronization that corresponds to the (VARETA) (214).
moment of perception, while this was not the case when the faces Another novel approach based also on a measure of phase
were presented in an inverted position, and thus were difficult to synchronization was introduced for the analysis of visual
identify as faces. Later, this study was replicated by Trujillo et al. evoked responses to intermittent light stimulation in photosen-
(208) who revealed that some other aspects should be taken into sitive epileptic patients (215,216). The method consists in esti-
account with respect to the analysis of phase relationships. These mating the phase dispersion of each frequency component
authors put in evidence the dependency of phase relations on the present in the EEG/MEG. In this way a phase clustering index
frequency range at which the phase relationships were estimated, can be defined that differs from that used by Lachaux et al.
and on the EEG montage used. They were able to replicate (205). Applying the phase clustering index it was found that the
Rodriguez et al.’s results but only at the frequency range that patients who develop epileptiform discharges during the light
exhibited the largest gamma power for the face upright position. stimulation present an enhancement of the phase clustering
Additionally they found, however, substantial phase synchrony in index in the gamma frequency band in comparison to that at
other conditions, but at different frequencies. Furthermore, they the driving frequency, appearing before the occurrence of
showed also that the use of EEG signals recorded using Laplacian paroxysmal activity (Fig. 54.18). Thus, the phase clustering
“reference-free” derivations (209–211) yields more robust results index reflects the degree of excitability of the underlying neural
than using another referential montage. The judicious choice of system and it suggests the existence of nonlinear dynamics.
the EEG montage appears to be essential to obtain reliable EEG In general the interpretation of phase relationships between
phase relationships. It may be noted here that one may use other EEG/MEG signals is not straightforward. Indeed it implies the
methods of source estimation with the same purpose, such as low specification of a biophysical model. The point is to specify how
resolution brain electromagnetic tomography (LORETA) the neuronal sources of the signals communicate one with the
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 54.18 Spatial distribution of the relative phase clustering index (rPCI) changes per magnetic sensor over the magne-
toencephalography (MEG) helmet. Left plot: Average of the means of rPCI from four trials where intermittent light stimulation
at 20 Hz was followed by a photoparoxysmal response (PPR), compared with two trials where light stimulation was not fol-
lowed by PPR (middle plot), and with the average of four trials in an age- and sex-matched control subject (right plot).
(Adapted from Parra J, Kalitzin SN, Iriarte J, et al. Gamma-band phase clustering and photosensitivity: is there an underlying
mechanism common to photosensitive epilepsy and visual perception? Brain. 2003;126:1164–1172.) ((See color insert))

other, that is, what kind of transfer function relates both sources. This nonlinear correlation ratio h2 can be computed for a
A general biophysical model may be defined, as an example, as fol- number of time delays t between x and y; the estimated time
lows: at a given site in the brain there is a source of neuronal activ- delay between the two signals corresponds to the value of t for
ity, for instance, a generator of signals in the gamma frequency which h2 is maximum. Low values of h2 denote that signals x
range; this activity propagates to another site but suffers some and y are independent; high values of h2 mean that signal y may
transformation (amplification or attenuation) in the way; as a first be explained by a transformation (possibly nonlinear) of signal
approximation, we may assume that this transformation is linear, x. Bidirectional associations reflect the invertibility of the non-
and that one signal may be considered as input and the other as linear mapping while highly asymmetric, unidirectional associ-
output of a linear system; in general for any such system there is a ations indicate noninvertible mapping. This implies that high
mathematical relationship between the gain, that is, the relation asymmetric association values may indicate essentially nonlin-
between the amplitudes of the input and output signals, and the ear, irreversible dynamics of the underlying system (225).
phase function. Under certain conditions, the phase function is In addition to the estimation of h2, a derived quantity was
uniquely determined from the gain function. The class of systems proposed by Guye et al. (226) that gives information about the
with this property is called minimum phase shift systems. This causality of the association, and was called the direction index D.
refers to the smallest possible phase lag that can be obtained for This quantity takes into account both the estimated time delay
physically realizable systems with a specified gain function. More between signals x and y (t) and the asymmetric nature of the
precisely, the gain and phase of the system’s transfer function are nonlinear correlation coefficient h2 (values of the h2 coefficient
related by a set of equations called the Hilbert transform pair (cf., are different if the computation is performed from x to y or from
Ref. 217). In the case of EEG signals, one may assume that the y to x). Values of parameter D range from –1.0 (x is driven by y)
phase spectrum between the two signals is composed of two com- to 1.0 (y is driven by x). More recently Kalitzin et al. (225) pre-
ponents: one corresponding to the minimum phase component, sented a general definition of the nonlinear association index h2,
as described above, and the other to a phase component due to the demonstrating rigorously that the index measures the best
existence of a time delay. The former can be estimated from the dynamical range of any nonlinear map between signals. A fur-
gain function. Then this component can be subtracted from the ther refinement of the nonlinear association analysis is the so-
measured phase spectrum and, if a time delay exists, the difference called “partialization” of the association measure between two
should be a phase component that is linear with frequency. From signals. Indeed in cases where a nonlinear association index h2
the slope of the latter a time delay can be computed. This method between two signals may be caused by a third one, acting as a
was, for example, applied to the beta/gamma activities recorded common source, the influence of the third one can be removed
from the prepyriform cortex and the anterior entorhinal cortex of and the residual signals’ correlations may be computed to deter-
a cat (218). mine whether the association between them is caused by the
We should note, however, that in the analysis of EEG/MEG common influence of the third signal, or not.
signals one encounters quite often nonlinear relations, so that In addition to these nonlinear association methods, some
methods based on linear assumptions, such as coherence and others have been proposed and implemented to detect general
cross-correlation (70,24), in those cases are not the most appro- types of interactions between EEG/MEG signals based on the
priate. To estimate the degree of association between two signals theory of nonlinear dynamical systems, such as the concept of
and the corresponding time delay in a more general way, the non- generalized synchronization that can be applied to any kind of
linear correlation coefficient h2 as a function of time shift time series (227), based on the theory of nonlinear dynamical
between the two signals was first introduced in EEG signal analy- systems that was introduced by Rulkov et al. (228). According to
sis by Pijn and colleagues (219,220) and has also been shown to this concept synchronization between two dynamical systems X
give reliable measures for the degree and direction of functional (the driver) and Y (the response) exists when the state of the
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

coupling between neuronal populations in different types of response system Y is a function of the state of the driving sys-
epilepsy (221–224). This method has some major advantages tem, X: Y F(X). Assuming that F is continuous and that two
over other signal analysis methods like coherency, mutual infor- points on the attractor of X, xi and xj, are close to each other, the
mation, and cross-correlation, as it can be applied independently corresponding points on attractor Y, yi and yj, will be also close
of whether the type of relationship between the two signals is lin- to each other. The probability that embedded vectors are closer
ear or nonlinear. The basic assumption is that if the amplitude of to each other than a certain small critical distance is estimated
signal y is considered as a function of the amplitude of signal x, for each discrete time pair (i, j). This is done for each signal, or
the value of y given a certain value of x can be predicted accord- channel k of a set of M channels. Thereafter, the number of
ing to a nonlinear regression curve. The variance of y according channels Hi,j for which the embedded vectors Xk,i and Xk,j are
to this regression curve is called the explained variance, that is, it closer together than the critical distance can be calculated.
is explained or predicted on the basis of x. By subtracting the Inspired by these theoretical concepts, Stam and van Dijk (227)
explained variance from the total variance, one obtains the unex- defined a synchronization likelihood Sk,i,j for each EEG/MEG
plained variance. The nonlinear correlation ratio h2 expresses the channel k and each discrete time pair (i, j) as:
reduction of variance of y that can be obtained by predicting the
y values according to the regression curve as follows: h2 =
If 0 Xk,i  Xk,j 0 6 ek,i:Sk,i,j  (Hi,j  1)> (M  1)
(explained variance – unexplained variance)/total variance,
which takes values in the range [0, 1]. If 0 Xk,i  Xk,j 0 7 ek,i:Sk,i  0 (54.42)

By averaging over all j, the synchronization likelihood Sk,i during motor imagery using EEG signals recorded from
can be estimated. Thus, Sk,i is a measure that describes how electrodes overlying cortical sensorimotor areas (237).
strongly signal k at time i is synchronized to all other M – 1 sig- An interesting development in this field that was mainly
nals. The synchronization likelihood is a measure of the inspired by the need of analyzing multiple images produced
dynamical interdependencies of EEG/MEG signals, both linear in fMRI studies, and further to analyze combined fMRI–EEG
and nonlinear, as a function of time, and may be used to quan- recordings. This domain has been called dynamic causal
tify phase relations of nonstationary time series. A number of modeling (DCM) (238). This methodology goes beyond
applications have been published, as, for example, in a study of functional connectivity analysis as revealed by the statistical
EEG changes related to the performance of a visuosemantic methods described above, but it includes explicit biophysical
task (229), and also in clinical study of EEG synchronization in models of the processes that are responsible for the genera-
mild cognitive impairment and Alzheimer disease (230). tion of the signals and the causal relationships between them.
We may conclude that when brain activities are coupled in This is what has been called effective connectivity, in contrast
time, and when one uses the term “synchrony” in the context with functional connectivity. Accordingly, DCM is based on
of brain signals, one should take into account that synchro- two kinds of models: of neural population dynamics and of
nous activities do not necessarily occur with time delay equal the transfer between the neural activity and the signals being
to zero, but with a consistent time delay compatible with the recorded, either EEG or fMRI signals (see for details of these
mechanism of propagation of neural activity. To account for models Chapter 4; 239,240). This approach is most challeng-
the case where neuronal activities recorded from different cor- ing for the future because it brings together analytic method-
tical areas appear to occur synchronously with zero delay, as in ologies with the necessity of constructing biophysical models
the case of the experiments of Roelfsema et al. (231), where of the phenomena of interest, be these EEG or fMRI signals
zero time-lag synchrony was found between cortical areas in a or combinations of both.
visuomotor task in cats, two mechanisms can be proposed:
one is that both activities depend on a third source that REFERENCES
reaches at the same time both cortical areas where the record-
ings were made; another is that the two cortical areas have 1. Isaksson A, Wennberg A. Spectral properties of nonstationary EEG
strong bidirectionally re-entrant properties. This was, for signals, evaluated by means of Kalman filtering: application exam-
ples from a vigilance test. In: Kellaway P, Petersén I, eds.
example, experimentally demonstrated in the case of epilepti-
Quantitative Analytic Studies in Epilepsy. New York, NY: Raven;
form afterdischarges elicited from a focus in the hippocampus
1976:389–402.
spread to homologous sites in the contralateral hippocampus 2. Dumermuth G, Gasser T, Hecker A, et al. Exploration of EEG com-
following commissural systems that may be strong enough to ponents in the beta frequency range. In: Kellaway P, Petersén I, eds.
ensure the forming of one bilateral oscillating system with Quantitative Analytic Studies in Epilepsy. New York, NY: Raven;
zero interhemispheric delay (232). 1976:533–558.
Above we have discussed a number of issues with respect to 3. Jansen BH. EEG Segmentation and Classification [thesis].
the question of EEG phase synchronization. This question, Amsterdam: Free University; 1979:237 pp.
however, may be considered in the more general context of 4. Grosveld FM, Jansen BH, Hasman A, et al. La reconnaissance des
causal relationships between EEG signals, that is, the question individus à l’intérieur d’un groupe de seize sujets normaux. Rev
of causality. We have already indicated that the nonlinear asso- Electroencephalogr Neurophysiol Clin. 1976;9:295–297.
5. McEwen J, Anderson GB. Modelling the stationarity and
ciation index h2 allows to infer causal relations between two
Gaussianity of spontaneous electroencephalographic activity.
signals and to estimate the corresponding time delays. This IEEE Trans Biomed Eng. 1975;22:361–369.
property is an example of how to infer causal relations between 6. Elul R. The genesis of the EEG. Int Rev Neurobiol. 1972;15:227–272.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

two time series in a statistical sense. Already in 1956 Wiener 7. Gasser T. General characteristics of the EEG as a signal. In:
proposed that given two time series, one can be considered Remond A, ed. EEG Informatics. A Didactic Review of Methods and
causal with respect to the other if the second one can be better Applications of EEG Data Processing. Amsterdam: Elsevier;
predicted by adding knowledge about the first one. Later 1977:37–52.
Granger (233) applied this concept to linear regression models 8. Siebert WM. The description of random processes. In:
particularly in studies of econometrics. These concepts were Communications Biophysics Group, Siebert WM, eds. Processing
adapted to the spectral domain by Geweke (234) using MVAR Neuroelectric Data. Cambridge, MA: MIT Press; 1959:66–87.
models. Besides many applications of these statistical concepts 9. Pijn JP, Van Neerven J, Noest A, et al. Chaos or noise in EEG sig-
nals; dependence on state and brain site. Electroencephalogr Clin
of causality in different scientific fields, analytic methods based
Neurophysiol. 1991;79(5):371–381.
on these concepts have entered the field of neurosciences in the 10. Jay F, ed-in-chief. IEEE Standard Dictionary of Electrical and
course of the 1980s. A comprehensive review of these early Electronics Terms. The Institute of Electrical and Electronics
neuroscientific applications can be found in Kami ń ski et al. Engineers, Inc. New York, NY: Wiley; 1977.
(235). Among many others, one example of such methods is 11. Lopes da Silva FH, ed. Cooper R, Dumermuth G, et al. Sampling,
the “short-time directed transfer function (SDTF)” that allows conversion and measurement of bioelectrical phenomena. In:
estimating the propagation of brain activity as a function of Remond A, ed-in-chief; Brazier MA, ed. Handbook of
frequency and time (236). This was applied, for example, to a Electroencephalography and Clinical Neurophysiology. Vol 4. Part A.
study of the propagation of gamma and beta components Amsterdam: Elsevier; 1976.

12. Susskind AK, ed. Notes on Analog–Digital Conversion Techniques. 38. Gasser T. Goodness-of-fit tests for correlated data. Biometrika.
New York, NY: Wiley; 1957 1975;62:563–570.
13. Steineberg CA, Paine LW. Methods and techniques of data conver- 39. Lilliefors HW. On the Kolmogorov–Smirnov test for normality with
sion. Ann N Y Acad Sci. 1964;115/2:614–626. mean and variance unknown. J Am Stat Assoc. 1967;67: 399–402.
14. Jenkins GM, Watts DG. Spectral Analysis and its Applications. San 40. Persson J. Comments on estimations and tests of EEG amplitude dis-
Francisco, CA: Holden Day; 1968. tributions. Electroencephalogr Clin Neurophysiol. 1974;37: 309–313.
15. Blackman RB, Tukey JW. The Measurement of Power Spectra. New 41. Saunders MG. Amplitude probability density studies on alpha and
York, NY: Dover Press; 1958. alpha-like patterns. Electroencephalogr Clin Neurophysiol. 1963;15:
16. Bendat JS, Piersol AG. Measurements and Analysis of Random 761–767.
Data. New York, NY: Wiley Interscience; 1971. 42. Lion KS, Winter DF. A method for the discrimination between signal
17. Otnes RK, Enochson L. Digital Time Series Analysis. New York, NY: and random noise of electrobiological potentials. Electroencephalogr
Wiley; 1978. Clin Neurophysiol. 1961;5:109–111.
18. Dietsch G. Fourier Analyse von Elektroenzephalgorammen des 43. Kozhevnikov VA. Some methods of automatic measurement of the
Menschen. Pflugers Arch Ges Physiol. 1932;230:106–112. electroencephalogram. Electroencephalogr Clin Neurophysiol. 1958;
19. Grass AM, Gibbs FA. A Fourier transform of the electroencephalo- 10:269–278.
gram. J Neurophysiol. 1938;1:521–526. 44. Campbell J, Bower E, Dwyer SJ, et al. On the sufficiency of auto-
20. Knott JR, Gibbs FA. A Fourier transform of the EEG from one to correlation functions as EEG descriptors. IEEE Trans Biomed Eng.
eighteen years. Psychol Bull. 1939;36:512–513. 1967;BME14:49–52.
21. Drohocki Z. L’Électroencéphalographie du cerveau. C R Soc Biol. 45. Elul R. Gaussian behaviour of the electroencephalogram: changes
1938;129:889–893. during performance of mental task. Science. 1969;164:328–331.
22. Walter WG. Automatic low frequency analyzer. Electron Eng. 46. Dumermuth G. Variance spectra of electroencephalogram in
1943;16:9–13. twins. A contribution to the problem of quantification of EEG
23. Walter WG. An improved low frequency analyzer. Electron Eng. background activity in childhood. In: Kellaway P, Petersén I, eds.
1943;16:236–240. Clinical Electroencephalography in Childhood. Stockholm:
24. Brazier MAB, Casby JU. Cross-correlation and autocorrelation Almqvist & Wiksell; 1968:119–154.
of EEG potentials. Electroencephalogr Clin Neurophysiol. 1952; 47. Dumermuth G. Die Anwendung von Varianzspectra für einen
4:201. quantitativen Vergleich von EEG bei Zwillingen. Helv Paediatr
25. Barlow JS, Brazier MAB. A note on a correlator system for brain Acta. 1969;24:45–54.
potentials. Electroencephalogr Clin Neurophysiol. 1954;6:321–325. 48. Dumermuth G, Walz W, Scollo Lavizzari G, et al. Spectral analysis
26. Wiener N. Cybernetics. Cambridge, MA: MIT Press; 1961. of EEG activity during sleep stages in normal adults. Eur Neurol.
27. Cooley JW, Tukey JW. An algorithm for the machine calculation of 1972;7:265–296.
complex Fourier series. Math Comput. 1965;19:297–301. 49. Dumermuth G, Gasser T, Lange B. Aspects of EEG analysis in the
28. Huber PJ, Kleiner B, Gasser T, et al. Statistical methods for investi- frequency domain. In: Dolce G, Künkel H, eds. Computerized EEG
gating phase relations in stationary stochastic processes. IEEE Analysis. Stuttgart: Fischer; 1975:429–457.
Trans Audio Electroacoust. 1971;19:78–86. 50. Drohocki Z. L’intégrateur de l’Électroproduction cérébrale pour
29. Dumermuth G, Huber PJ, Kleiner B, et al. Analysis of the interre- l’Électroencéphalographie quantitative. Rev Neurol Paris. 1948;80:
lations between frequency bands of the EEG by means of the bis- 617–619.
pectrum. Electroencephalogr Clin Neurophysiol. 1971;31:137–148. 51. Goldstein L. Time domain analysis of the EEG. The integrative
30. Hjorth B. EEG analysis based on time domain properties. method. In: Dolce G, Künkel H, eds. CEAN—Computerized EEG
Electroencephalogr Clin Neurophysiol. 1970;29:306–310. Analysis. Stuttgart: Fischer; 1975:251–270.
31. Zetterberg LH. Means and methods for processing of physiologi- 52. Saltzberg B, Burch NR, McLennan MA, et al. A new approach to
cal signals with emphasis on EEG analysis. In: Lawrence JH. signal analysis in electroencephalography. IRE Trans Med Electron.
Advances in Biology and Medical Physics. Vol 16. New York, NY: 1957;8:24–30.
Academic Press; 1977:41–91. 53. Burch NR, Nettleton WJ, Sweeney J, et al. Period analysis of the
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

32. Cohen BA. Period analysis of the electroencephalogram. Comput electroencephalogram on a general purpose digital computer. Ann
Programs Biomed. 1976;6:269–276. N Y Acad Sci. 1964;115:827–843.
33. Rice SO. Mathematical analysis of random noise. In: Wax N, ed. 54. Fink M. EEG and human psychopharmacology. Annu Rev
Noise and Stochastic Processes. New York, NY: Dover; 1954:133–294. Pharmacol. 1969;9:241–258.
34. Saltzberg B, Edwards RJ, Heath RG, et al. Synoptic analysis of EEG 55. Fink M. EEG profiles and bioavailability measures of psychoactive
signals. In: Enslein K, ed. Data Acquisition and Processing in Biology drugs. In: Itil TM, ed. Psychotropic Drugs and the Human EEG.
and Medicine. Vol 5. Oxford: Pergamon Press; 1968:267–307. Basel: S. Karger; 1974:76–98.
35. Matousek M ed., Barlow JS, Dumermuth G, et al. Frequency and 56. Fink M. Cerebral electrometry—quantitative EEG applied to human
correlation analysis. In: Brazier MAB, Walter DO, eds. Evaluation psychopharmacology. In: Dolce G, Künkel H, eds. Computerized EEG
of Bioelectrical Data from Brain, Nerve and Muscle, II. In: Remond Analysis. Stuttgart: Fischer; 1975:271–288.
A, ed-in-chief. Handbook of Electroencephalography and Clinical 57. Itil TM, Shapiro DM, Herrmann WM, et al. HZI systems for EEG
Neurophysiology. Vol 5. Part A. Amsterdam: Elsevier; 1973. parameterization and classification of psychotropic drugs.
36. Saltzberg B, Burch NR. Period analytic estimates of moments of Pharmakopsychiatrie. 1979;12:4–19.
the power spectrum. A simplified EEG time domain procedure. 58. Pronk RAF, Simons AJR. Processing of the electroencephalogram
Electroencephalogr Clin Neurophysiol. 1971;30:568–570. in cardiac surgery. Comput Programs Biomed. 1984;18(3):181–189.
37. Gevins A, Remond A, eds. Methods of Analysis of Brain Electrical 59. Pronk RAF, de Boer SJ, Cornelissen RCM, et al. Computer assisted
and Magnetic Signals. Handbook of Electroencephalography and patient monitoring during open heart surgery with the aid of the
Clinical Neurophysiology. Vol 1. Amsterdam: Elsevier; 1987: EEG. In: van Eijnsbergen B, Lopes da Silva FH, eds. Progress Report
541–582 [new series]. No. PR6. Utrecht: Institute of Medical Physics TNO; 1976:224–228.
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
Chapter 54 ■ EEG Analysis: Theory and Practice 1173

60. Itil TM. Digital computer period analyzed EEG in psychiatry and 81. Regan D. Evoked potentials in basic and clinical research. In:
psychopharmacology. In: Dolce G, Künkel H, eds. CEAN— Remond A, ed. EEG Informatics. A Didactic Review of Methods
Computerized EEG Analysis. Stuttgart: Fischer; 1975:289–308. and Applications of EEG Data Processing. Amsterdam: Elsevier;
61. Matejcek M, Schenk GK. Die iterative IntervallAnalyse—Ein 1977:319–346.
methodischer Beitrag zur Quantitativen Beschreibung des 82. Broughton R, Healey T, Maru J, et al. A phase locked loop device
Elektroenzephalogramms in Zeitbereich. In: Schenk GK, ed. Die for automatic detection of sleep spindles and stage 2. Electroen-
Quantifizierung des Elektroenzephalogramms. Konstanz: AEG cephalogr Clin Neurophysiol. 1978;44:677–680.
Telefunken; 1973:293–306. 83. Campbell K, Kumar A, Hofman W. Human and automatic vali-
62. Legewie H, Probst W. Online analysis of EEG with a small com- dation of a phase-locked loop spindle detection system. Electro-
puter. Electroencephalogr Clin Neurophysiol. 1969;27:533–535. encephalogr Clin Neurophysiol. 1980;48:602–605.
63. Schwarzer F, Reets H. Machines for EEG analysis. Electroencephalogr 84. Kaiser E, Petersén I, Selldin U, et al. EEG data representation in
Clin Neurophysiol. 1966;20:278. broadband frequency analysis. Electroencephalogr Clin Neurophysiol.
64. Marko H, Petsche H. Ein Gerat zur gleichzettigen Frequenz und 1964;17:76–80.
Amplitudenanalyse von EEGKurven bei frei wählbarer Zeitbasis. 85. Matousek M, Petersén I, Friberg S. Automatic assessment of ran-
Arch Psychiat Nervenkr. 1957;196:191–195. domly selected routine EEG records. In: Dolce G, Künkel H, eds.
65. Leader HS, Cohn R, Wehrer AL, et al. Pattern reading of the clinical CEAN—Computerized EEG Analysis. Stuttgart: Fischer; 1975:
electroencephalogram with a digital computer. Electroencephalogr 421–428.
Clin Neurophysiol. 1967;23:566–570. 86. Storm van Leeuwen W. Comparison of EEG data obtained with
66. Pfurtscheller G, Koch W. Eine maschinelle Methode zur EEG frequency analysis and with correlation methods. Electro-
Klassifikation. Methods Inform Med. 1972;11:233–237. encephalogr Clin Neurophysiol Electroencephalogr Clin Neurophysiol
67. Harner RN. Computer analysis and clinical EEG interpretation— Suppl. 1961;20:37–40.
perspective and application. In: Dolce G, Künkel H, eds. CEAN— 87. Storm van Leeuwen W. Complementary of different analysis
Computerized EEG Analysis. Stuttgart: Fischer; 1975:337–343. methods. Electroencephalogr Clin Neurophysiol. 1964;16:136–139.
68. Harner RN, Ostergren KA. Sequential analysis of quasistable 88. Walter DO. Spectral analysis of electroencephalograms: mathe-
and paroxysmal activity. In: Kellaway P, Petersén I, eds. Quantitative matical determination of neurological relationships from records
Analytic Studies in Epilepsy. New York, NY: Raven; 1976:343–361. of limited duration. Exp Neurol. 1963;8:155–181.
69. Barlow JS, Brown RM. An Analog Correlator System for Brain 89. Walter DO, Adey WR. Analysis of brain wave generators as mul-
Potentials. Research Lab Electronics, Technical Report No. 300. tiple statistical time series. IEEE Trans Biomed Eng. 1965;12:8–13.
Cambridge, MA: Massachusetts Institute of Technology (MIT); 90. Dumermuth G. Fundamentals of spectral analysis in electroen-
1955. cephalography. In: Remond A, ed. EEG Informatics. A Didactic
70. Brazier MAB, Barlow JS. Some applications of correlation analysis Review of Methods and Applications of EEG Data Processing.
to clinical problems in electroencephalography. Electroencephalogr Amsterdam: Elsevier; 1977:83–105.
Clin Neurophysiol. 1956;8:325–331. 91. Vos JE. Representation in the frequency domain of nonstationary
71. McFadden JA. The correlation function of a sine-wave plus noise EEGs. In: Dolce G, Künkel H, eds. CEAN—Computerized EEG
after extreme clipping. IRE Trans Inform Theory. 1956;2(2):82–83. Analysis. Stuttgart: Fischer; 1975:41–50.
72. Kaiser JF, Angell RK. New Techniques and Equipment for Correlation 92. Künkel H, EEG Project Group. Hybrid computing system for
Computation. Technical Memo 7668TM2, Servomechanisms Lab. EEG analysis. In: Dolce G, Künkel H, eds. CEAN—Computerized
Cambridge, MA: MIT Press; 1957. EEG Analysis. Stuttgart: Fischer; 1975:365–385.
73. Lopes da Silva FH. Dynamic Characteristics of Visual Evoked 93. Walter DO. On units and dimensions for reporting spectral inten-
Potentials [thesis]. Utrecht, The Netherlands: University of Utrecht; sities. Electroencephalogr Clin Neurophysiol. 1968;24:486–487.
1970:126 pp. 94. Abraham F, Brown D, Gardiner M. Calibration of EEG power
74. Sologub EB. EEG cross correlation analysis during the formation spectra. Commun Behav Biol Part A. 1968;1:31.
of motor dynamic stereotype in human. Zh Vyssh Nerv Deiat. 95. Clusin W, Trapani G, Roccaforte PA. A numerical approach to
1965;15:32–41 [in Russian]. matching amplification for the spectral analysis of recorded EEG.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

75. Kamp A, Storm van Leeuwen W, Tielen AM. A method for auto- Electroencephalogr Clin Neurophysiol. 1970;28:639–641.
and cross-relation analysis of the EEG. Electroencephalogr Clin 96. Dumermuth G, Flühler H. Some modern aspects in numerical
Neurophysiol. 1965;19:91–95. spectrum analysis of multichannel electroencephalographic data.
76. Lesèvre N, Remond A. Variations in the average visual response in Med Biol Eng. 1967;5:319–331.
relation to the alpha phase “autostimulation.” Electroencephalogr 97. Sciarretta G, Erculiani P. A proposal for calibrating EFG spectro-
Clin Neurophysiol. 1967;23:578–579. grams. Electroencephalogr Clin Neurophysiol. 1978;5:674–676.
77. Michael D, Houchin H. Automatic EEG analysis: a segmentation 98. Bickford RG. Computer analysis of background activity. In:
procedure based on the autocorrelation function. Electro- Remond A, ed. EEG Informatics. A Didactic Review of Methods and
encephalogr Clin Neurophysiol. 1979;46:232–235. Applications of EEG Data Processing. Amsterdam: Elsevier; 1977:
78. Walter DO. The method of complex demodulation. In: Walter DO, 215–232.
Brazier MAB, eds. Advances in EEG Analysis. Electroencephalogr 99. Bickford RG, Brimm J, Berger L, et al. Application of compressed
Clin Neurophysiol. 1968;(suppl 27):51–57. spectral array in clinical EEG. In: Kellaway P, Petersén I, eds.
79. Childers DG, Pao MT. Complex demodulation for transient Automation of Clinical Electroencephalography. New York, NY:
wavelet detection and extraction. IEEE Trans Audio Electroacoust. Raven; 1973:55–64.
1972;AU20:295–308. 100. Johnson LC, Hanson K, Bickford RG. Effect of flurazepam on sleep
80. Kumar A. The complex demodulation method for detection of a spindles and K complexes. Electroencephalogr Clin Neurophysiol.
waves and sleep spindles of the human EEG in real time. 1976;40:67–77.
Proceedings of International Conference on Advanced Signal 101. Kawabata N. A nonstationary analysis of the electroencephalo-
Processing Techniques, Lausanne. 1975:355–361. gram. IEEE Trans Biomed Eng. 1973;20:444–452.
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1174 Part IX ■ Computer-Assisted EEG Analysis

102. Pfurtscheller G, Aranibar A. Event-related cortical desynchro- IEEE Trans Acoustics, Speech Signal Processing. 1981;ASSP29(3):
nization detected by power measurements of scalp EEG. 619–621.
Electroencephalogr Clin Neurophysiol. 1977;42:817–826. 123. Gersch W, Goddard G. Locating the site of epileptic focus by
103. Kamp A, Vliegenthart W. Sequential frequency analysis: a spectral analysis methods. Science. 1970;169:701–702.
method to quantify event related EEG changes. Electro- 124. Gersch W, Tharp BR. Spectral regression–amount of information
encephalogr Clin Neurophysiol. 1977;42:843–846. analysis of seizures in humans. In: Kellaway P, Petersén I, eds.
104. Pfurtscheller G, Aranibar A. Voluntary movement event-related Quantitative Analytical Studies in Epilepsy. New York, NY: Raven;
desynchronization (ERD): normative studies. In: Pfurtscheller G, 1976:509–532.
Buser P, Lopes da Silva FH, et al., eds. EEG Activities and Cortical 125. Tharp BR, Gersch W. Spectral analysis of seizures in humans.
Functioning, Developments in Neuroscience. Vol 10. Amsterdam: Comput Biomed Res. 1975;8:503–521.
Elsevier; 1980:151–177. 126. Etévenon P. Étude méthodologique de l’Electroencéphalographie
105. Pfurtscheller G, Stancak A Jr, Neuper C. Event-related synchro- quantitative. Application à quelques examples [Thèse]. Paris VI:
nization (ERS) in the alpha band—an electrophysiological corre- Université Pierre et Marie Curie; 1977.
late of cortical idling: a review. Int J Psychophysiol. 1996;24:39–46. 127. Lopes da Silva FH, Vos JE, Mooibroek J, et al. Partial coherence
106. Siegel S. Nonparametric Statistics for the Behavioural Sciences. analysis of thalamic and cortical alpha rhythms in dog. A contri-
Tokyo: McGraw-Hill; 1956. bution towards a general model of the cortical organization of
107. Arnolds DEAT, Lopes da Silva FH, Aitink AW, et al. Hippocampal rhythmic activity. In: Pfurtscheller G, Buser P, Lopes da Silva FH,
EEG and behaviour in dog. I. Hippocampal EEG correlates of et al., eds. EEG Activities and Cortical Functioning. Developments
gross motor behaviour. Electroencephalogr Clin Neurophysiol. 1979; in Neuroscience. Vol 10. Amsterdam: Elsevier; 1980:33–59.
46:552–570. 128. Lopes da Silva FH, Vos JE, Mooibroek H, et al. Relative contribu-
108. Etévenon P, Pidoux B. From biparametric to multidimensional tions of intracortical and thalamocortical processes in the gener-
analysis of EEG. In: Remond A, ed. EEG Informatics. A Didactic ation of alpha rhythms revealed by partial coherence analysis.
Review of Methods and Application of EEG Data Processing. Electroencephalogr Clin Neurophysiol. 1980;50:449–456.
Amsterdam: Elsevier; 1977:193–214. 129. Vos JE, Lammerstsma AA, Van Eykeren LA. Ordinary and partial
109. Gasser T. Statistical handling of EEG data. Pharmakopsychiatrie. coherences of bipolar and quasiunipolar derivations of infant
1979;12:210–219. electroencephalograms. In: Random Signal Analysis, IEE Cofn.
110. Abt K. Statistical problems in the analysis of comparative phar- Publ. No. 159. London; 1977:154–160.
maco-EEG trials. Pharmakopsychiatrie. 1979;12:228–236. 130. Rappelsberger P, Petsche H, Vollmer R, et al. Rhythmicity in
111. Walter DO, Adey WR. Spectral analysis of electroencephalograms seizure patterns—intracortical aspects. In: Speckmann EJ,
recorded during learning in the cat. Exp Neurol. 1963;8:155–181. Caspers H, eds. Origin of Cerebral Field Potentials. Stuttgart:
112. Brazier MAB. Studies of the EEG activity of limbic structures in Thieme; 1978:80–97.
man. Electroencephalogr Clin Neurophysiol. 1968;25:309–318. 131. Lopes da Silva FH, Storm van Leeuwen W. The cortical alpha
113. Lopes da Silva FH, van Lierop THMT, Schrijer CF, et al. rhythm in dog: depth and surface profile of phase. In: Brazier
Organization of thalamic and cortical alpha rhythms: spectra and MAB, Petsche H, eds. Architectonics of the Cerebral Cortex. IBRO
coherence. Electroencephalogr Clin Neurophysiol. 1973;35: 627–639. Monograph Series. Vol 3. New York, NY: Raven; 1978:319–333.
114. Prechtl HFR, Vos JE. Verlaufsmuster der Frequenzspektren 132. Reits D. Cortical Potentials in Man Evoked by Noise Modulated Light
and Kohärenzen bei schlafenden normalen und neurologisch [thesis]. Utrecht, The Netherlands: University of Utrecht; 1975.
abnormalen Neugeborenen. In: Schenk GK, ed. Die Quantifizierung 133. Berglund K, Hjorth B. Normierte SteilheitsBeschrei bungspara-
des Elektroenzephalogramms. Konstanz: AEG Telefunken; 1973: meter und der en physikalischer Sinn hinsichtlich der EEG
167–188. Deutung. In: Schenk G, ed. Quantifizierung des EEGs. Konstanz:
115. Bullock TH, McClune MC, Achimowicz JZ, et al. EEG coherence AEG Telefunken; 1973:249–257.
has structure in the millimeter domain: subdural and hippocam- 134. Denoth F. Some general remarks on Hjorth’s parameters used in
pal recordings from epileptic patients. Electroencephalogr Clin EEG analysis. In: Dolce G, Künkel H, eds. CEAN—Computerized
Neurophysiol. 1995;95:161–177. EEG Analysis. Stuttgart: Fischer; 1975:9–18.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

116. Bullock TH, McClune MC, Achimowicz JZ, et al. Temporal fluc- 135. Hjorth B. Time domain descriptors and their relation to a particular
tuations in coherence of brain waves. Proc Natl Acad Sci U S A. model for generation of EEG activity. In: Dolce G, Künkel H, eds.
1995;92:11568–11572. CEAN—Computerized EEG Analysis. Stuttgart: Fischer; 1975:3–8.
117. Storm van Leeuwen W, Arntz A, Spoelstra P, et al. The use of 136. Lopes da Silva FH, Hoeks A, Smits H, et al. Model of brain rhyth-
computer analysis for diagnosis in routine electroencephalogra- mic activity, the alpha rhythm of the thalamus. Kybernetik.
phy. Rev Electroencephalogr Neurophysiol Clin. 1976;62:318–327. 1974;15:27–37.
118. Storm van Leeuwen W, Wieneke GH, Spoelstra P, et al. Lack of 137. Caille EJ, Bassano JL. Value and limits of sleep statistical analysis.
bilateral coherence of mu rhythm. Electroencephalogr Clin Objective parameters and subjective evaluations. In: Dolce G,
Neurophysiol. 1978;44:140–146. Künkel H, eds. CEAN—Computerized EEG Analysis. Stuttgart:
119. Gotman J. Experiments in the Automation and Quantification of Fischer; 1975:227–235.
EEG Interpretation: Localized Brain Lesions and Epilepsy [thesis]. 138. Lütcke A, Mertins L, Masuch A. Die Darstellung von
Montreal: McGill University; 1976. Grundaktivität, Herd, und Verlaufsbefunden sowie von paroxys-
120. Brazier MAB. Spread of seizure discharges in epilepsy: anatomi- malen Ereignissen mit Hilfe der von Hjorth angegebe nen
cal and electrophysiological considerations. Exp Neurol. 1972; normierten Steilheitsparameter (vorläufige Mitteilungen). In:
36:263–272. Matejcek M, Schenk GK, eds. Quantitative Analysis of the EEG.
121. Carter GC. Time Delay Estimation. Report TR5335. New London, Konstanz: AEG Telefunken; 1973:259–280.
CT: Naval Underwater Systems Center; 1976. 139. Zetterberg LH. Stochastic Activity in a Population of Neurons. A
122. Mars NJI, van Arragon GW. Time delay estimation in nonlinear Systems Analysis Approach. Report No. 2.3.153/1. Utrecht:
systems using average mutual amount of information analysis. Institute of Medical Physics TNO; 1973:23 pp.
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
Chapter 54 ■ EEG Analysis: Theory and Practice 1175

140. Zetterberg LH, Kristiansson L, Mossberg K. Performance of a 161. Mathieu M. Analyse de l’electroencéphalogramme par prédiction
model of a local neuron population. Biol Cybern. 1978;31:15–26. linéatre [Thèse]. Paris: Université Pierre et Marie Curie; 1976:
141. Zetterberg LH. Estimation of parameters for a linear difference 173 pp.
equation with application to EEG analysis. Math Biosci. 1969;5: 162. Isaksson A. SPARK—A Sparsely Updated Kalman Filter with
227–275. Application to EEG Signals. Telecommunication Theory,
142. Gersch W. Spectral analysis of EEGs by autoregressive decompo- Technical Report No. 120. Stockholm: Royal Institute of
sition of time series. Math Biosci. 1970;7:205–222. Technology; 1974.
143. Fenwick PBC, Michie P, Dollimore J, et al. Mathematical simula- 163. Barlow JS. Analysis of EEG changes with carotid clamping by selec-
tion of the electroencephalogram using an autoregressive series. tive analog filtering, matched inverse digital filtering and automatic
Int J Biomed Comput. 1971;2:281–307. adaptive segmentation: a comparative study. Electroencephalogr
144. Bohlin T. Analysis of EEG Signals with Changing Spectra. Stockholm: Clin Neurophysiol. 1984;58:193–204.
IBM New Lab Technology; 1971:118 pp. Paper TP 18.212. 164. Creutzfeldt OD, Bodenstein G, Barlow JS. Computerized EEG
145. Isaksson A, Wennberg A. Visual evaluation and computer analy- pattern classification by adaptive segmentation and probability
sis of the EEG—a comparison. Electroencephalogr Clin Neuro- density function classification. Clinical evaluation. Electro-
physiol. 1975;38:79–86. encephalogr Clin Neurophysiol. 1985;60:373–393.
146. Lopes da Silva FH, Dijk A, Smits H. Detection of nonstationari- 165. Barlow JS. Methods of analysis of nonstationary EEGs, with
ties in EEGs using the autoregressive model. An application to emphasis on segmentation techniques: a comparative review.
EEGs of epileptics. In: Dolce G, Künkel H, eds. CEAN— J Clin Neurophysiol. 1985;2:267–304.
Computerized EEG Analysis. Stuttgart: Fischer; 1975:180–199. 166. Baillon JF, Bienenfeld G, Findji F, et al. Lecture, mesure et traite-
147. Pfurtscheller G, Haring G. The use of an EEG autoregressive ment automatique de l’analyse mimétique de l’EEG. Rev
model for the time saving calculation of spectral power density Electroencephalogr Neurophysiol Clin. 1976;6:255–270.
distribution with a digital computer. Electroencephalogr Clin 167. Remond A. An EEGer’s approach to automatic data processing.
Neurophysiol. 1972;33:113–115. In: Dolce G, Künkel H, eds. CEAN—Computerized EEG Analysis.
148. Rappelsberger P, Petsche H. Spectral analysis of the EEG by Stuttgart: Fischer; 1975:128–136.
means of autoregression. In: Dolce G, Künkel H, eds. CEAN— 168. Remond A. From graphoelements to EEG pattern recognition.
Computerized EEG Analysis. Stuttgart: Fischer; 1975:27–40. Electroencephalogr Clin Neurophysiol Suppl. 1978;34:141–145.
149. Wennberg A, Zetterberg LH. Application of a computer-based 169. Remond A, Renault B. La theorie des objets electroencephalo-
model for EEG analysis. Electroencephalogr Clin Neurophysiol. graphiques. Rev Electroencephalogr Clin Neurophysiol. 1972;
1971;31:457–468. 2:241–256.
150. Zetterberg LH. Experience with analysis and simulation of EEG 170. Schenk GK. Die Quantifizierung des EEG mittels vektorieller
signals with parametric description of spectra. In: Kellaway P, Iterationstechnik, einer Simulationsmethode der visuellen
Petersén I, eds. Automation of Clinical Electroencephalography. Analyse. In: Schenk GK, ed. Die Quantifizierung des Elektroence-
New York, NY: Raven; 1973:161–201. phalogramms. Konstanz: AEG Telefunken; 1973:307–343.
151. Herolf M. A Recursive Detector. Telecommunication Theory, 171. Ahlblom G, Zetterberg LH. A Comparative Study of Five Methods
Technical Report No. 99. Stockholm: Royal Institute of Technology; for Analysis of EEG. Technical Report No. 112. Stockholm: Royal
1975. Institute of Technology; 1975:56 pp.
152. Bodenstein G, Praetorius HM. Feature extraction from the 172. Herolf M. Detection of Pulse-shaped Signals in EEG Telecommu-
encephalogram by adaptive segmentation. Proc IEEF. 1977;65: nication Theory, Technical Report No. 41. Stockholm: Royal
642–657. Institute of Technology; 1971.
153. Praetorius HM, Bodenstein G, Creutzfeldt O. Adaptive segmen- 173. Zetterberg LH. Spike detection by computer and by analog
tation of EEG records: a new approach to automatic EEG analy- equipment. In: Kellaway P, Petersén I, eds. Automation of Clinical
sis. Electroencephalogr Clin Neurophysiol. 1977;42:84–94. Electroencephalogr. Clin. Neurophysiol. New York, NY: Raven;
154. Makhoul J. Linear prediction: a tutorial review. Proc IEEE. 1973:227–234.
1975;63:561–580. 174. Barlow JS, Dubinsky J. Some computer approaches to continu-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

155. Eykhoff P. System Identification. Parameter and State Estimation. ous automatic clinical EEG monitoring. In: Kellaway P, Petersén
London: Wiley; 1974. I, eds. Quantitative Analytic Methods in Epilepsy. New York, NY:
156. Lopes da Silva FH. Analysis of EEG nonstationarities. Raven; 1976:309–327.
Electroencephalogr Clin Neurophysiol Suppl. 1978;34:163–179. 175. Pfurtscheller G, Fischer G. A new approach to spike detection using
157. Lopes da Silva FH, Dijk A, Smiths H, et al. Automatic detection a combination of inverse and matched filter techniques.
and pattern recognition of epileptic spikes from surface and Electroencephalogr Clin Neurophysiol. 1978;44: 243–247.
depth recording in man. In: Schenk GK, ed. Die Quantifizierung 176. Mallat SG. A theory of multiresolution signal decomposition: the
des Elektroenzephalogramms. Konstanz: AEG Telefunken; 1973: wavelet representation. IEEE Trans Pattern Anal Machine Intell.
425–436. 1989;11:674–693.
158. Lopes da Silva FH, van Hulten K, Lommen JG, et al. Automatic 177. Blanco S, D’Atellis CA, Isaacson S, et al. Time frequency analysis
detection and localization of epileptic foci. Electroencephalogr of electroencephalogram series (II): Gabor and wavelet trans-
Clin Neurophysiol. 1977;43:1–13. form. Phys Rev E. 1996;54:6661–6672.
159. Duquesnoy AJ. Segmentation of EEGs by Means of Kalman 178. Schiff SJ, Aldrouby A, Unser M, et al. Fast wavelet transformation of
Filtering. Progress Report No. PR5. Utrecht: Institute of Medical EEG. Electroencephalogr Clin Neurophysiol. 1994;91: 442–455.
Physics TNO; 1976:87–92. 179. Bartnik EA, Blinowska KJ. Wavelets: new method of evoked
160. Isaksson A. On Time Variable Properties of EEG Signals Examined potential analysis. Med Biol Eng Comput. 1992;30:125–126.
by Means of a Kalman Filter Method. Telecommunication Theory. 180. Geva AB, Pratt H, Zeevi YY. Multichannel wavelet-type decom-
Technical Report No. 95. Stockholm: Royal Institute of Technology; position of evoked potentials: model-based recognition of gener-
1975. ator activity. Med Biol Eng Comput. 1997;35(1):40–46.
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1176 Part IX ■ Computer-Assisted EEG Analysis

181. Mallat SG, Zhang Z. Matching pursuit with time frequency dic- 203. Fell J, Klaver P, Elfadil H, et al. Rhinal–hippocampal theta
tionaries. IEEE Trans Signal Processing. 1993;41:3397–3415. coherence during declarative memory formation: interaction with
182. Durka PJ, Blinowska KJ. Analysis of EEG transients by means of gamma synchronization? Eur J Neurosci. 2003;17(5): 1082–1088.
matching pursuit. Ann Biomed Eng. 1995;23:608–611. 204. Le Van Quyen M, Foucher J, Lachaux J, et al. Comparison of
183. Durka PJ. Time Frequency Analyses of EEG [PhD thesis]. Hilbert transform and wavelet methods for the analysis of neu-
Department of Physics, Warsaw, Poland: University of Warsaw; ronal synchrony. J Neurosci Methods. 2001;111(2):83–98.
1996. 205. Lachaux JP, Rodriguez E, Martinerie J, et al. Measuring phase
184. Quian Quiroga R, Blanco S, Rosso OA, et al. Searching for hidden synchrony in brain signal. Hum Brain Mapp. 1999;8:194–208.
information with Gabor transform in generalized tonic–clonic 206. Tass P, Rosenblum MG, Weule J, et al. Detection of n:m phase
seizures. Electroencephalogr Clin Neurophysiol. 1997;103: 434–439. locking from noisy data: application to magnetoencephalogra-
185. Walter DO, Rhodes JM, Brown D, et al. Comprehensive spectral phy. Phys Rev Lett. 1998;81:3291–3294.
analysis of human generators in posterior cerebral regions. 207. Rodriguez E, George N, Lachaux JP, et al. Perception’s shadow:
Electroencephalogr Clin Neurophysiol. 1966;20:224–237. long-distance synchronization of human brain activity. Nature.
186. van Rotterdam A. Limitations and difficulties in signal process- 1999;397(6718):430–433.
ing by means of the principal components analysis. IEEE Trans 208. Trujillo LT, Peterson MA, Kaszniak AW, et al. EEG phase syn-
Biomed Eng. 1970;17:268–269. chrony differences across visual perception conditions may
187. Bell AJ, Sejnowski TJ. An information-maximization approach to depend on recording and analysis methods. Clin Neurophysiol.
blind separation and blind deconvolution. Neural Comput. 2005;116(1):172–189.
1995;7(6):1129–1159. 209. Pernier J, Perrin F, Bertrand O. Scalp current density fields: con-
188. Makeig S, Bell AJ, Jung TP, et al. Independent component analy- cept and properties. Electroencephalogr Clin Neurophysiol. 1988;
sis of electroencephalographic data. In: Touretzki D, Mozer M, 69(4):385–389.
Hasselmo M, eds. Advances in Neural Information Processing 210. Nunez PL, Pilgreen KL. The spline-Laplacian in clinical neuro-
Systems. Cambridge, MA: MIT Press; 1996:145–151. physiology: a method to improve EEG spatial resolution. J Clin
189. Makeig S, Jung T-P, Bell A, et al. Blind separation of auditory Neurophysiol. 1991;8(4):397–413.
event-related brain responses into independent components. 211. Nunez PL, Silberstein RB, Cadusch PJ, et al. A theoretical and
Proc Natl Acad Sci U S A. 1997;94(20):10979–10984. experimental study of high resolution EEG based on surface
190. Makeig S, Westerfield M, Jung TP, et al. Dynamic brain sources of Laplacian and cortical imaging. Electroencephalogr Clin
visual evoked responses. Science. 2002;295(5555):690–694. Neurophysiol. 1994;90:40–57.
191. Anemüller J, Sejnowski T, Makeig S. Complex independent com- 212. Pascual-Marqui RD, Michel CM, Lehmann D. Low resolution
ponent analysis of frequency-domain electroencephalographic electromagnetic tomography: a new method for localizing elec-
data. Neural Netw. 2003;16(9):1311–1323. trical activity in the brain. Int J Psychophysiol. 1994;18:49–65.
192. Hyvärinen A, Ramkumar P, Parkkonen L, et al. Independent 213. Pascual-Marqui RD. Standardized low resolution brain electro-
component analysis of short-time Fourier transforms for spon- magnetic tomography (sLORETA): technical details. Methods
taneous EEG/MEG analysis. Neuroimage. 2010;49(1):257–271. Find Exp Clin Pharmacol. 2002;24D:5–12.
193. Klemm M, Haueisen J, Ivanova G. Independent component analy- 214. Bosch-Bayard J, Valdés-Sosa P, Virues-Alba T, et al. 3D statistical
sis: comparison of algorithms for the investigation of surface elec- parametric mapping of EEG source spectra by means of variable
trical brain activity. Med Biol Eng Comput. 2009;47(4): 413–423. resolution electromagnetic tomography (VARETA). Clin
194. Groppe DM, Makeig S, Kutas M. Identifying reliable independent Electroencephalogr. 2001;32(2):47–61.
components via split-half comparisons. Neuroimage. 2009;45(4): 215. Kalitzin SN, Parra J, Velis DN, et al. Enhancement of phase cluster-
1199–1211. ing in the EEG/MEG gamma frequency band anticipates transitions
195. Eichele T, Calhoun VD, Debener S. Mining EEG–fMRI using inde- to paroxysmal epileptiform activity in epileptic patients with known
pendent component analysis. Int J Psychophysiol. 2009;73(1):53–61. visual sensitivity. IEEE Trans Biomed Eng. 2002;49:1279–1286.
196. Marques JP, Rebola J, Figueiredo P, et al. ICA decomposition of 216. Parra J, Kalitzin SN, Iriarte J, et al. Gamma-band phase clustering
EEG signal for fMRI processing in epilepsy. Hum Brain Mapp. and photosensitivity: is there an underlying mechanism com-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

2009;30(9):2986–2996. mon to photosensitive epilepsy and visual perception? Brain.

197. Gómez-Herrero G, Atienza M, Egiazarian K, et al. Measuring 2003;126:1164–1172.
directional coupling between EEG sources. Neuroimage. 2008; 217. Papoulis A. Probability, Random Variables, and Stochastic
43(3):497–508. Processes. 9th ed. Tokyo: McGraw-Hill Kogakusha; 1965.
198. Ryynanen OR, Hyttinen JA, Malmivuo JA. Effect of measurement 218. Boeijinga PH, Lopes da Silva FH. Modulations of EEG activity in
noise and electrode density on the spatial resolution of cortical the entorhinal cortex and forebrain olfactory areas during odour
potential distribution with different resistivity values for the sampling. Brain Res. 1989;478(2):257–268.
skull. IEEE Trans Biomed Eng. 2006;53(9):1851–1858. 219. Lopes da Silva F, Pijn JP, Boeijinga PI. Interdependence of EEG sig-
199. Pfurtscheller G, Lopes da Silva FH. Event-related EEG/MEG syn- nals: linear vs. nonlinear associations and the significance of time
chronization and desynchronization: basic principles. Clin delays and phase shifts. Brain Topogr. 1989; 2(1–2):9–18.
Neurophysiol. 1999;110(11):1842–1857. 220. Pijn JP, Vijn PC, Lopes da Silva FH, et al. Localization of epilep-
200. Engel AK, Singer W. Temporal binding and the neural correlates togenic foci using a new signal analytical approach. Neurophysiol
of sensory awareness. Trends Cogn Sci. 2001;5(1):16–25. Clin. 1990;20(1):1–11.
201. Varela F, Lachaux JP, Rodriguez E, et al. The brainweb: phase syn- 221. Wendling F, Bartolomei F, Bellanger JJ, et al. Interpretation of inter-
chronization and large-scale integration. Nat Rev Neurosci. dependencies in epileptic signals using a macroscopic physiological
2001;2(4):229–239. model of the EEG. Clin Neurophysiol. 2001;112 (7):1201–1218.
202. Fell J, Fernández G, Klaver P, et al. Is synchronized neuronal 222. Meeren HK, Pijn JP, Van Luijtelaar EL, et al. Cortical focus drives
gamma activity relevant for selective attention? Brain Res Brain widespread corticothalamic networks during spontaneous
Res Rev. 2003;42(3):265–272. absence seizures in rats. J Neurosci. 2002;22(4):1480–1495.
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
Chapter 54 ■ EEG Analysis: Theory and Practice 1177

223. Bartolomei F, Wendling F, Régis J, et al. Pre-ictal synchronicity in 232. Granger CWJ. Investigating causal relations by econometric mod-
limbic networks of mesial temporal lobe epilepsy. Epilepsy Res. els and cross-spectral methods. Econometrica. 1969; 3, 74, 24–39.
2004;61(1–3):89–104. 233. Fernandes de Lima VM, Pijn JP, Nunes Filipe C, et al. The role
224. Uva L, Librizzi L, Wendling F, et al. Propagation dynamics of of hippocampal commissures in the interhemispheric transfer
epileptiform activity acutely induced by bicuculline in the hip- of epileptiform afterdischarges in the rat: a study using linear
pocampal–parahippocampal region of the isolated Guinea pig and non-linear regression analysis. Electroencephalogr Clin
brain. Epilepsia. 2005;46(12):1914–1925. Neurophysiol. 1990;76(6):520–539.
225. Kalitzin SN, Parra J, Velis DN, et al. Quantification of unidirec- 234. Geweke J. Measurement of linear dependence and feedback
tional nonlinear associations between multidimensional signals. between multiple time series. J Am Stat Assoc. 1982;77:304–324.
IEEE Trans Biomed Eng. 2007;54(3):454–461. 235. Kamiński M, Ding M, Truccolo WA, et al. Evaluating causal rela-
226. Guye M, Régis J, Tamura M, et al. The role of corticothalamic tions in neural systems: granger causality, directed transfer func-
coupling in human temporal lobe epilepsy. Brain. 2006;129(pt tion and statistical assessment of significance. Biol Cybern. 2001;
7):1917–1928. 85(2):145–157.
227. Stam CJ, van Dijk BW. Synchronization likelihood: an unbiased 236. Korzeniewska A, Crainiceanu CM, Kuś R, et al. Dynamics of
measure of generalized synchronization in multivariate data event-related causality in brain electrical activity. Hum Brain
series. Physica D. 2002;163:236–251. Mapp. 2008;29(10):1170–1192.
228. Rulkov NF, Sushchik MM, Tsimring LS, et al. Generalized syn- 237. Ginter J Jr, Blinowska KJ, Kamiński M, et al. Propagation of EEG
chronization of chaos in directionally coupled chaotic systems. activity in the beta and gamma band during movement imagery
Phys Rev E Stat Phys Plasmas Fluids Relat Interdiscip Top. 1995; in humans. Methods Inf Med. 2005;44(1):106–113.
51(2):980–994. 238. Friston KJ, Tononi G, Reeke GN Jr, et al. Value-dependent selection
229. Micheloyannis S, Vourkas M, Bizas M, et al. Changes in linear in the brain: simulation in a synthetic neural model. Neuroscience.
and nonlinear EEG measures as a function of task complexity: 1994;59(2):229–243.
evidence for local and distant signal synchronization. Brain 239. David O, Harrison L, Friston KJ. Modelling event-related responses
Topogr. 2003;15(4):239–247. in the brain. Neuroimage. 2005;25(3):756–770.
230. Stam CJ, van der Made Y, Pijnenburg YAL, et al. EEG synchro- 240. Stephan KE, Harrison LM, Kiebel SJ, et al. Dynamic causal mod-
nization in mild cognitive impairment and Alzheimer’s disease. els of neural system dynamics: current state and future extensions.
Acta Neurol Scand. 2003;108:90–96. J Biosci. 2007;32(1):129–144.
231. Roelfsema PR, Engel AK, König P, et al. Visuomotor integration
is associated with zero time-lag synchronization among cortical
areas. Nature. 1997;385(6612):157–161.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

EEG Mapping and Source Imaging

CHRISTOPH M. MICHEL AND BIN HE 55
INTRODUCTION any a priori assumption on the number of sources. However, solv-
ing the underdetermined inverse problem requires a priori
This chapter is concerned with the use of EEG as a functional assumptions based on information other than the number of
imaging method. Great advancements have been made in the sources, preferentially incorporating physiologic or biophysical
past several years in recording and analyzing of high-resolution knowledge. The correctness of these assumptions determines the
EEG. Powerful EEG systems have been designed that allow fast correctness of the source estimation. The fourth section will dis-
and easy recording from hundreds of channels simultaneously, cuss the different source reconstruction algorithms that are cur-
even in clinical settings. Sophisticated pattern recognition algo- rently used and show examples of applications.
rithms have been developed to characterize the topography of the The spatial resolution of high-density EEG with sophisti-
scalp electric fields and to detect changes in the topography over cated source localization methods in realistic geometry head
time and between experimental or clinical conditions. New models has become very impressive and the images that are
methods for estimating the sources underlying the recorded scalp produced are as tempting as pictures from other functional
potential maps have been constructed and applied to numerous imaging methods, particularly because they show direct neu-
experimental and clinical data. The incorporation of anatomical ronal signaling rather than indirect metabolic changes. But EEG
information, as obtained from magnetic resonance imaging has a second important attraction: the high temporal resolu-
(MRI) in the individual subject, has boosted the use of electro- tion. This temporal resolution combined with electrophysio-
physiologic neuroimaging and has stimulated clinical interest. logic neuroimaging leads to the possibility to elucidate the
We will here discuss these recent developments and present temporal dynamics of neuronal signaling in large-scale neu-
the current state of the art in electrophysiologic neuroimag- ronal networks and directly estimate network connectivity. The
ing, thereby extending other recent detailed reviews on this last section will discuss such analysis methods.
topic (1–6). The power of EEG as a functional neuroimaging method is
Neuronal activities in the brain generate current flows in the largely underestimated and many impressive experimental and
head volume conductor, reflected as electric potentials over the clinical studies using these tools have not received the attention
head surface, where they give rise to a specific topographical they merit. The reason is manifold. First, functional MRI has
map. The proper recording and analysis of these maps are pre- received a unique status of being able to reduce brain activity to
cursors for source localization. A great deal of localization the underlying sources nonambiguously. Second, misinterpre-
information can already be derived from these maps, but their tations of EEG and evoked potential waveforms due to a lack of
incorrect interpretation can also lead to a misleading conclu- understanding of the properties of electromagnetic fields, of
sion about the putative generators. The second section of this the role of the reference electrode, and of the influence of non-
chapter will deal with the proper recording of the scalp poten- neuronal signals such as myogenic or occulomotor activity
tial fields and the characterization, description, and statistical
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

resulted in a number of claims that later proved to be wrong.

comparison of scalp potential maps. It will also summarize Third, the EEG is somehow harmed by history. The term EEG
analysis methods that are based on spatiotemporal characteris- is still often related to the artistic interpretation of graph ele-
tics of potential maps, thereby leading to data reduction and a ments by some skilled neurophysiologists. The magnetoen-
priori constraints for subsequent source localization. cephalogram (MEG) that basically measures the same neuronal
The propagation of the electric potential in the brain that is activity with the same limitations does not suffer from this his-
generated by the active neuronal populations is modulated by tory and is easily considered as a neuroimaging method by pub-
the conductivity properties of the different tissues and by the lic encyclopedias such as Wikipedia. With this chapter we would
shape of the head. If these parameters are known, the electric like to diminish this incorrect historical view and show that the
potentials that a given current source in the brain produces on EEG has considerably matured and can now be considered as a
the surface electrodes can be calculated. This so-called forward powerful, flexible, and affordable imaging technology.
solution is the basis of every source localization method. The
third section of the chapter will discuss the different head mod-
els and the current knowledge on head conductivities and their
MAPPING OF THE SCALP ELECTRIC FIELD
influence on the scalp potential maps. Electrophysiologic neuroimaging is based on the recording of the
EEG source localization has evolved from single dipole search- electric potential from a multitude of electrodes distributed over
ing methods to distributed source estimation procedures without the surface of the head. From these simultaneous recordings a
1179
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1180 Part IX ■ Computer-Assisted EEG Analysis

potential map can be constructed for any single moment in time,

depicting the momentary configuration of the potential field (7).
The idea to analyze these topographies instead of waveform mor-
phologies has already been formulated some decades ago (8–10)
and has been called EEG topographical mapping. EEG mapping
is a precursor to source imaging (11), and the proper analysis and
interpretation of EEG maps can give a great deal of information
with regard to the putative sources in the brain. Most impor-
tantly, by physical laws, different map topographies must have
been produced by different source configurations in the brain
(12). Thus, statistical methods that allow determining signifi-
cantly different map topographies over time or between condi-
tions or subjects provide important a priori hypotheses about
whether and when differences in the source localization algo-
rithms can be expected. Analysis of topographic maps is there-
fore an important step in electric source imaging (3).
Visualization and proper inspection of topographic maps is Figure 55.1 High-resolution EEG. Example of an EEG system that
also mandatory for source imaging to assure that maps that are allows fast application of 256 electrodes. The electrodes are intercon-
clean of artifact enter the algorithms. The quality of the maps nected by thin rubber bands and each contains a small sponge that
determines the goodness of the source imaging procedures. It is touches the subject’s head directly (14). The nets are soaked in saline
therefore of crucial importance that these scalp potential fields water before put on the subject’s head. The whole net is applied at once
are recorded and preprocessed in a reasonable manner, and that and needs no skin abrasion and no electrode paste. (HydroCel Geodesic
they are visualized and carefully inspected before applying Sensor Net constructed by Electrical Geodesics Inc., Eugene, OR, USA.)
source imaging algorithms to them. This particularly concerns
EEG that is recorded in noisy environments such as in the MRI
scanner. EEG waveforms that look correct after filtering and frequency that is twice as high as the highest frequency in the
denoising do not yet necessarily indicate that the EEG maps will signal is required to avoid aliasing (Nyquist rate) (see also
be correct and usable for source analysis. Chapter 54). Similar rules apply to sampling in space, since the
In the following we discuss some practical issues related to potential distribution is only sampled at discrete measurement
the recording and construction of topographic maps. This con- points (electrodes) (19,20). Spatial frequencies of the potential
cerns the number and the distribution of the electrodes on the field that are higher than the spatial sampling frequency (i.e.,
scalp to provide an adequate spatial sampling of the potential the distance between electrodes) will distort the map topogra-
field. It also concerns the parametric description of the map phy (21–24) and will lead to misinterpretation of maps and
configuration and the comparison of map topographies in a consequently to mislocalization of the sources.
global and reference-independent way. Further details can be Already many years ago, researchers tried to estimate the
found in Ref. 13. maximal spatial frequency of the scalp electric field based on
theoretical considerations and modeling. These works sug-
Spatial Sampling gested that interelectrode distances of ~2 to 3 cm are preferable
It is clear that proper sampling of the electromagnetic field over (25,26) for proper sampling of the field, which would lead to
the whole scalp needs a large number of sensors. The MEG com- around 100 required electrodes. Freeman et al. (27) suggested
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

munity has consequently quickly moved from low- to high-reso- from spatial spectral density calculations that even less than 1-
lution systems, and most of the MEG laboratories are nowadays cm spacing of electrodes is needed. Lantz et al. (28) and Michel
recording from over ~150 channels. Until recently, this was a et al. (3) performed simulations using dipole forward modeling
severe limitation for the EEG, because application of a large num- (see section "EEG Forward Problem") to calculate the dipole
ber of electrodes was time consuming, uncomfortable, and expen- localization error of different source localization algorithms
sive. However, this is not a limiting factor anymore. EEG systems when different number of electrodes were used. Both studies
of up to 256 electrodes are commercially available and are easy showed that the localization precision does not increase lin-
and fast applicable, even in clinical settings (Fig. 55.1) (14–17). early, but reaches a plateau at about 100 electrodes for fully dis-
The question of how many electrodes are needed for proper tributed inverse solution algorithms.
EEG mapping and source imaging is not completely answered. Several experimental studies used subsampling techniques to
It depends on the spatial frequency of the scalp potential field, establish the number of electrodes needed to correctly recon-
which is limited by the blurring caused by volume conductor struct potential maps and localize the sources. Michel et al. (3)
effects, particularly induced by the low conductivity of the skull demonstrated incorrect lateralization of the source estimated for
(18). The maximal spatial frequency has to be correctly sampled the P100 component of the visual-evoked potential when down-
to avoid aliasing, which appears when the frequency of the sampling from 46 to 19 electrodes, and that an incomplete cover-
measured signal is higher than the sampling frequency. In the age of the scalp surface can lead to complete misplacement of the
case of discrete sampling of time-varying signals, a sampling sources (Fig. 55.2). Luu et al. (29) and Lantz et al. (28) used the

Figure 55.2 Spatial sampling. The figure illustrates the importance of proper sampling of the scalp potential field. Visual-
evoked potentials were recorded from 46 electrodes positioned according to the scheme on the left. A distributed source
localization algorithm (LORETA) was applied to the data at the peak of the P100 component, resulting in a medial occip-
ital source maximum. Data were then downsampled to fewer electrodes that were restricted to the posterior part of the
head. The same localization procedure (with the same spherical head model) at the same time point led to incorrect local-
ization, with even a frontal maximum with 19 occipital electrodes. Using again only 19 electrodes, but distributing them
equally over the head as illustrated on the scheme on the right, leads to a complete sampling of the electric field and to a
more correct localization. (From Michel CM, Lantz G, Spinelli L, et al. 128-Channel EEG source imaging in epilepsy: clin-
ical yield and localization precision. J Clin Neurophysiol. 2004;21:71–83.)(See color insert)

downsample method in clinical data to evaluate the correctness and Driscoll (30). These traditional values were used in the
of localization of pathologic activity. Luu et al. (29) studied above-described simulation and downsampling studies, but they
patients with acute focal ischemic stroke recorded with 128 elec- are most probably incorrect as indicated by several recent studies
trodes and downsampled to 64, 32, and 19 channels. Visual com- (31–33). If the conductivity of the skull is lower as it is proposed
parison of the EEG maps with radiographic images led to the by these studies, the spatial blurring is smaller and the spatial fre-
conclusion that more than 64 electrodes were desirable to avoid quency is higher. Ryynänen et al. (23,24) investigated the relation
mislocalizations of the affected region. More objectively was the between the number of electrodes and the resistance values of the
downsample study of Lantz et al. (28) on 123-electrode record- different compartments. Their computer simulation results sug-
ings from patients with partial epilepsy. Fourteen patients with gest that if the 1:80 ratio is considered as correct, 64 to 128 elec-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

different focus localization were recorded before successful resec- trodes are indeed sufficient. However, when more realistic skull
tive surgery; thus, the location of the epileptic focus was known. conductivity values are used, a higher number of electrodes may
Several interictal spikes were manually identified and then down- be needed. More realistic values for the conductivity ratio are
sampled to 63 and 31 electrodes. Source localization was applied suggested to be between ~20 and 50 (34), depending on the skull
to each single spike and the distance of the source maximum to thickness. In an experimental study of pediatric patients under-
the resected area was determined and statistically compared going intracranial recordings, He and colleagues (35) measured
between the different electrode sets. Significant smaller localiza- the scalp and subdural potentials simultaneously during cortical
tion error was found when using 63 instead of only 31 electrodes. current injection, and used them to estimate the brain-to-skull
Accuracy still systematically increased from 63 to 123 electrodes, conductivity ratio. The experimental data suggested that the
but less significantly (Fig. 55.3). averaged brain-to-skull conductivity ratio is about 25 when using
The above-described studies estimate that around 64 or more the three-sphere head model (33), and about 20 when using the
electrodes are desirable for accurate spatial sampling and recon- realistic geometry finite element head model (35). In newborns
struction of the scalp potential field. However, as shown by the skull thickness is approximately seven to eight times lower
Ryynänen et al. (23,24) in computer simulation studies, these than in adults, leading to a ratio of approximately 14:1 (20,36).
estimations are only valid if we assume a conductivity ratio of Ryynänen et al. (24) suggested from their computer simulation
approximately 1:80 between skull and brain, as proposed by Rush studies that with this ratio, spatial resolution still increases with

Figure 55.3 Influence of number of electrodes on source localization. Evaluation of source localization precision of inter-
ictal discharges in 14 epileptic patients, recorded with 123 electrodes. Single spikes were localized with a linear inverse
solution in the individual brain and the distance of the source maximum to the resected area was measured. Mean and
standard deviation of the distance were compared between the original high-resolution recording and with downsampling
of the same data to fewer channels (but still equally distributed). The top row shows the example of one patient. The dia-
monds indicate the source maximum; the blue area marks the resected zone. Diamonds outside the brain are actually on
another level and projected onto the illustrated slide. The bar graph shows the mean distance to the lesion. Stars indicate
significant differences between the different number of electrodes. A clear significant amelioration of the localization pre-
cision was observed when increasing the number of recording channels. (From Michel CM, Brandeis D. Data acquisition
and pre-processing standards for electrical neuroimaging. In: Michel CM, Koenig T, Brandeis D, et al., eds. Electrical
Neuroimaging. Cambridge: Cambridge University Press; 2009. Modified after Lantz G, Spinelli L, Seeck M, et al.
Propagation of interictal epileptiform activity can lead to erroneous source localizations: a 128 channel EEG mapping study.
J Clin Neurophysiol. 2003;20:311–319.)(See color insert)

256 as compared to 128 electrodes in realistic noise levels. Grieve number of electrodes that are needed to recover spatial features
et al. (20) also suggested that a 256-electrode array is needed in of scalp topography and to estimate the underlying brain sources
infants to obtain a spatial sampling error of less than 10%. that generate the scalp EEG.
However, larger electrode arrays are also more influenced by
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

measurement noise, which affects the spatial resolution. As sug- Map Inspection, Artifact Correction, Interpolation
gested by Ryynänen et al. (23,24), the measurement noise is a A mandatory requirement in EEG and ERP analysis is the detec-
critical limiting factor for the spatial resolution of high-density tion and elimination of artifacts, caused by bad electrode con-
EEG systems. Thus, there is an important interplay between tact, muscle or eye movement activity, or other environmental
number of electrodes, measurement noise, and conductivity val- noise. Besides automatic detection of such artifacts using ampli-
ues of the different compartments of the head. Furthermore, it tude windows, careful visual inspection of the EEG traces is
remains for the moment unclear how much an imperfect spatial mandatory. These requirements evidently also hold for EEG
sampling influences the source imaging. Some data did suggest mapping and source localization, but in addition to the need for
that even with ~32 electrodes, one gains important insight about clean EEG traces, the spatial EEG analysis also requires clean
the underlying brain electric sources by performing source local- maps (13). Even if the EEG traces appear correct, they might be
ization and imaging (37–40). Further studies with direct meas- contaminated by noise that destroys the spatial configuration of
urements of the conductivities at different locations on the head the maps. Spatial inhomogeneities due to bad electrodes can
through current injection (41) or with the aid of MRI (42) will drastically influence source localization outcome because they
be needed to definitely answer the question of the head tissue can lead to strong local gradients. Figure 55.4 illustrates this
conductivity. Further experimental and clinical studies shall also problem. In one case, the bad electrode is readily seen in the 256-
be needed to systematically investigate the question of optimal channel EEG traces, and it is also easily seen in the topographic

but is ignored and misinterpreted as occipital gamma activity

when looking at single occipital electrodes that were recorded
against this nose reference (43).
When data are averaged over sweeps or over subjects, bad
electrodes must be interpolated before averaging. Most of the
commonly used interpolation routines belong to the family of
spline interpolations. Spline interpolations can be linear (based
on a polynomial of first degree), or quadratic, cubic, or of higher
order. Using cross-validation tests, it was concluded that higher
order spherical spline interpolations perform reasonably well in
sufficiently dense electrode arrays (44,45). The estimation of
potentials is less reliable when the potentials are located outside
of the electrode array and not between some electrodes.
Extrapolating potentials beyond the electrode array should thus
be avoided.
Efficient algorithms to detect and eliminate artifacts based
on “abnormal” spatial configurations are increasingly used in
EEG mapping and source localization studies. Most efficient are
Figure 55.4 Influence of artifacts on source localization. Example of a algorithms based on independent component analysis (ICA)
256-channel EEG recording and the reconstruction of the scalp potential (46). The ICA is particularly useful for eye movement artifact
map and the source estimation for this map using a distributed linear detection and elimination, because the artifact is largely inde-
inverse solution (LAURA). The first time point (left) includes a clear arti- pendent from the remaining part of the data.
fact on a left temporal electrode, easily identifiable on the EEG traces and
on the EEG map. This leads to a dominant source under this electrode Topographic Analysis
position in the inverse solution. Excluding (or interpolating) this elec- The traditional analysis of EEG and evoked potentials relies on
trode leads to a dominant source on the contralateral temporal lobe. The waveforms. Parameters of interest are thereby changes in ampli-
artifact channel at the second time point (right) is not easily identifiable tude or frequency, or peaks at certain latency time-locked to
on the traces (artifact trace shown in red). However, the map clearly iden- stimulus presentation. These measures are ambiguous because
tifies this bad right frontal electrode with a negative potential surrounded the EEG is by definition a bipolar signal. Changes of the loca-
by otherwise positive electrodes. Elimination of this bad electrode leads tion of one of the two electrodes will change the values of the
to a unique occipital source. Keeping the electrode for the inverse solu- above parameters. This ambiguity is well known and has led to
tion calculation leads to an additional right frontal source underlying this a large discussion on the reference-dependency of the EEG and
electrode. (Modified from Michel CM, Brandeis D. Data acquisition and the question of the correct recording reference for a certain
pre-processing standards for electrical neuroimaging. In: Michel CM, experimental or clinical condition (47–51).
Koenig T, Brandeis D, et al., eds. Electrical Neuroimaging. Cambridge: This reference problem of the EEG is completely resolved
Cambridge University Press; 2009.)(See color insert) when topographic analysis methods are applied. The potential
map topography does not depend on the reference (3,50,52–54).
The reference only changes the zero level, but the topographical
map by a strong and isolated local potential minimum. The sec- features of the map remain unaffected (53). Thus, the reference
ond bad channel is not easily seen in the traces and is not only introduces a DC shift. This shift is eliminated when apply-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

detected by an amplitude window. However, it is readily visible ing topographic analysis methods including source imaging.
on the map by a negative “island” within the otherwise positive Elimination of the DC level can be achieved by calculating the so-
potential. Keeping these bad channels in the construction of the called common average reference at each moment in time (9).
potential field leads to an incorrect estimation of a focal source One way to sharpen the spatial details of the scalp potential
under this electrode, which completely disappears when these maps is to calculate the scalp current source density, or the sur-
electrodes are taken out. face Laplacian of the potential (21,55,56). The surface Laplacian
This example illustrates the importance of inspecting not of the scalp potential is the second spatial derivative of the poten-
only the EEG traces for abnormal graph elements, but also the tial field in the local curvature in mV/cm2. The surface Laplacian
potential maps for abnormal topographies. It is thereby impor- has been mainly derived from unipolar potential recordings on
tant that the recording reference is included in the electrode the scalp, using algorithms such as finite difference algorithm
array for map construction, because also the reference electrode (57), spherical spline algorithm (58), or realistic geometry spline
is an active electrode and has to be inspected for abnormal val- algorithm (59). The surface Laplacian has been widely used in
ues. The most convincing example has been demonstrated by applications when one wishes to enhance the sensitivity to local
Yuval-Greenberg et al. (43), showing a map with a very steep activity. It can be interpreted as an estimation of the current den-
potential maximum at the nose reference. This was due to the sity entering or exiting the scalp. It emphasizes superficial sources
recording of miniature saccades by this “noncephalic” reference because deeper sources produce smaller potentials on the surface.
electrode. This becomes readily visible when looking at the map, Like the other topographic measures that will be described below,

the surface Laplacian estimates are independent of the position that the GMD inversely correlates with the GFP: GMD is high
of the recording reference, because the potential common to all when GFP is low, that is, maps tend to be stable during high GFP
electrodes is automatically removed (60). and change the configuration when GFP is low (Fig. 55.5). The
Scalp potential maps can be characterized by their strength GMD is itself not a statistical measure. However, if two groups
and their topography. A reference-independent measure of of maps are compared, a statistical statement of the significance
map strength is the global field power (GFP) (8). GFP is the of the topographic differences can be made. This is achieved by
standard deviation of the potentials at all electrodes of an average- performing nonparametric randomization tests based on the
reference map. It is defined as GMD values as, for example, described in Ref. 54.

a i51 (ui 2 u)
N 2 Spatiotemporal Decomposition
GFP 5 (55.1) Source localization procedures can be applied to multichannel
R N EEG and ERP data at any instant in time. With the high sam-
where ui is the voltage of the map u at the electrode i, u is the pling rate of modern EEG systems that easily exceed 1000 Hz,
average voltage of all electrodes of the map u, and N is the num- this leads to a large amount of data from which the relevant
ber of electrodes of the map u. Scalp potential fields with pro- information has to be extracted. Consequently, experimenters
nounced peaks and troughs and steep gradients, that is, very typically predetermine relevant events within a continuous
“hilly” maps, will result in high GFP, while GFP is low in maps time series of data to which source analysis will be applied. This
that have a “flat” appearance with shallow gradients. GFP is a particularly concerns ERP research, where peaks in certain time
one-number measure of the map at each moment in time. windows at certain electrodes are identified and spatially ana-
Displaying this measure over time allows to identify moments of lyzed (64). This traditional approach is however less tenable
high signal-to-noise ratio, corresponding to moments of high with high-density EEG and ERP recordings, because different
global neuronal synchronization (61). scalp sites have different peak latencies, and because the wave-
A reference-independent measure of topographic differ- forms (and thus the peaks) at certain electrodes change when
ences of scalp potential maps is the so-called global map dis- changing the position of the reference electrode.
similarity measure (GMD) (8). It is defined as An alternative to the traditional preselection of relevant
events based on ERP waveforms is to define components on the

a •
basis of the topography of the potential field (65). Most com-
1 N ui 2 u
GMD 5 monly, some kind of spatial factor analysis methods are used for

B a i51
N i51 N this purpose. These methods produce a series of factors that
T (ui 2 u) 2>N
represent a weighted sum of all recorded channels across time.
2
vi 2 v The aim of this factor analysis approach is to find a limited
2

C a i51
¶ number of optimal factors that best represent a given data set.
N
(vi 2 v) 2>N The load for each of these factors (i.e., the goodness of fit) then
varies in time. Each factor represents a certain potential topog-
where ui is the voltage of map u at the electrode i, vi is the raphy, that is, a prototypical map. Source localization applied to
voltage of map v at the electrode i, u is the average voltage of all these maps results in a limited number of putative sources in
electrodes of map u, v is the average voltage of all electrodes of the brain that explain a full time series of multichannel EEG
map v, and N is the total number of electrodes. In order to assure data with time-varying strength.
that only topography differences are taken into account, the two The most commonly used variant of spatial factor analyses is
maps that are compared are first normalized by dividing the the principal component analysis (PCA) (66–69). The first factor
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

potential values at each electrode of a given map by its GFP. of the PCA solution accounts for the maximally possible amount
The GMD is 0 when two maps are equal, and maximally of data variance, and each next orthogonal factor accounts for the
reaches 2 for the case where the two maps have the same maximum possible residual variance. Since factors that con-
topography with reversed polarity. It can be shown that the tribute little to the explained variance can be neglected, the PCA
GMD is equivalent to the spatial Pearson’s product–moment is a powerful exploratory tool to reduce complex multichannel
correlation coefficient between the potentials of the two maps EEG data in space and time. It has been repeatedly applied to
to compare (62). ERP studies with the aim to extract ERP components whose vari-
If two maps differ in topography independent of their ance is related to a given experimental condition. It can provide
strength, it directly indicates that the two maps were generated useful information on how a given experimental manipulation
by a different configuration of sources in the brain (4,11,63). affects ERP components without any a priori assumption about
The inverse is not necessarily true: infinite number of source the shape or number of components in the data set (69–73).
configurations may produce the same scalp potential topogra- The PCA does not allow for cross-correlations between activ-
phy (12). The GMD calculation is therefore considered as a first ities corresponding to separate factors and thus excludes linear
step for defining whether different sources were involved in the dependencies between the factor maps. However, it does not
two processes that were compared. When comparing subsequent exclude dependencies based on higher order correlations. The
maps in time using the GMD, periods where source configura- factor analysis method that also removes these higher order rela-
tion changes appeared can be defined. It is interesting to note tions is called independent component analysis (ICA) (74). The

Figure 55.5 Spatial analysis of evoked potentials. Example of 256-channel visual-evoked potential (VEP) and somatosen-
sory-evoked potential (SEP). VEP from full-field checkerboard reversal presented to the right eye only (left eye covered). SEP
from electric stimulation of the right median nerve. Data represent the grand mean of over 20 subjects. Top row: Overlaid
traces of all 256 channels against the average reference. Second row: Global field power curve (GFP) as a measure of field
strength indicates five dominant peaks in both evoked potentials. Third row: Global map dissimilarity curve (GMD) meas-
uring topography differences between successive time points. It shows low values during extended periods and sharp peaks
at moments of low GFP. Fourth row: Potential maps (seen from top, nose up, left ear left) that were derived from a k-means
cluster analysis of the whole data sets. In both EPs, five maps best explained the data. Each one dominated a given period
as determined by spatial correlation analysis. Vertical dashed lines mark these periods. Last row: Distributed linear inverse
solution applied to each of the five maps, revealing activation and propagation of visual and sensor-motor cortex, respec-
tively. Note that the first period represents extracortical activity in both cases (activity in the right retina for the VEP and in
the brainstem for the SEP). Both areas were not included in the solution space, consequently leading to incorrect localiza-
tion in the source estimation. (For more details see Ref. 106.)(See color insert)

objective of the ICA is sometimes illustrated by the so-called seen in the ERP examples in Figure 55.5 by the stable low global
“cocktail party problem,” where the ICA allows decomposing a dissimilarity (GMD) over extended time periods separated by
sound record from a party into the independent contributions sharp dissimilarity peaks indicating periods of topographic
of the individual persons. Like the PCA, the ICA produces a change. The same observation holds for spontaneous EEG, if
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

weight coefficient for each factor. Each factor is supposed to polarity inversion caused by the intrinsic oscillatory activity of
represent a temporally independent component. the generator processes is ignored (Fig. 55.6) (83–86). This fun-
As described earlier, the ICA can be very useful for detecting damental observation led to the proposal to apply cluster analy-
and removing artifacts such as eye blinks (46), or artifacts pro- sis to the data set to identify a set of topographies that explain a
duced by brain-independent sources such as the ballistocardio- maximum amount of the variance of the data (87). The differ-
gram artifact of the EEG recorded in an MRI scanner (75–77), ence to the above-described factor analysis approach is that the
although negative results have been reported as well (78–80). microstate model allows only one single topography to occur at
More critical is the idea of decomposing the brain processes into one moment in time. Evidently, each topography can represent
a number of statistically independent factors (74,81) because it multiple simultaneously active sources, but they are active
implies that there are indeed a similar number of independent together for a certain time period, forming a large-scale neuronal
processes in the brain. network configuration that is expressed as unique stable map
An alternative to the above-described component analysis topography. During the period of stable topography, the strength
approaches is the so-called microstate segmentation method of the field varies, indicating different level of synchronization of
(82). It is based on the highly reproducible observation that the the simultaneously active areas. In contrast to the ICA-based
topography of the EEG or ERP potential maps remains stable for models of independent brain processes that overlap in time, the
several tens of milliseconds and then abruptly switches to a new microstate model proposes one global brain state per time
configuration in which it remains stable again. This can easily be period, consisting of an interdependent and synchronized

Figure 55.6 Microstate segmentation of

spontaneous EEG. Four seconds of eyes-
closed EEG measured from 42 electrodes
are shown on top. The two blue curves rep-
resent the global field power (GFP) and the
global map dissimilarity, respectively. A
series of potential maps illustrate the data
that has been submitted to a k-means clus-
ter analysis with ignoring strength and
polarity. Four maps best explained the
whole 4 seconds of data. The four maps are
illustrated on the bottom. On the GFP curve
below the map series, the time periods dur-
ing which each map was dominant are
marked in different colors. A shorter period
is zoomed in and all maps during this
period are shown. Marking and connecting
the extreme potentials illustrates the stabil-
ity of topography during each period.
(From Michel CM, Brandeis D. Data acqui-
sition and pre-processing standards for
electrical neuroimaging. In: Michel CM,
Koenig T, Brandeis D, et al., eds. Electrical
Neuroimaging. Cambridge, MA: Cambridge
University Press; 2009.)(See color insert)

network (88). This corresponds well to the neuronal workspace global dissimilarity as metric (54,104,105). Such methods can be
model, which suggests that episodes of coherent activity last a used for an objective analysis of clinical evoked potentials, for
certain amount of time (around 100 msec) and are separated by example, in multiple sclerosis (106). They have been used in
sharp transitions (89,90), as well as to the proposal that neu- numerous experimental ERP studies on sensory and cognitive
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

rocognitive networks evolve through a sequence of quasistable information processing and have allowed creating a microstate
coordination states rather than a continuous flow of neuronal dictionary for different brain functions (107–115). Source local-
activity (91–93). Cross-validation methods following cluster ization applied to these microstate maps has proven to be an effi-
analysis have shown that a very limited number of map topogra- cient way to describe those brain areas that are crucially
phies are needed to explain extended periods of spontaneous implicated in the processing of stimuli and that differ depending
EEG, and that these few configurations follow each other accord- on the task demands (Fig. 55.5) (103).
ing to certain rules (85,87). Changes in the succession and dura-
tion of the microstates have been observed in several pathologic EEG FORWARD PROBLEM
conditions such as depression (94), dementia (95,96), schizo-
phrenia (97,98), and epilepsy (99), as well as after drug intake In this section, we introduce the methods for solving the so-
(100,101). In normal subjects, the duration and frequency of called EEG forward problem, which deals with (i) how to model
appearance of the four most dominant microstate configurations the neuronal excitation within the brain volume and (ii) how to
varies with age (86). Concerning ERPs, the cluster analysis is an model the head volume conduction process in order to quanti-
efficient way to determine different ERP components exclusively tatively link neuronal electric sources with the electric poten-
on the basis of their topography (73,87,102,103). Statistical speci- tials over the scalp. Solving the EEG forward problem can help
ficity of these component maps for different experimental condi- understand the relationship between neuronal sources and the
tions can then be assessed by spatial fitting procedures using the recorded EEG signals, and is also an integrative part of the EEG

inverse problem, which will be discussed in the section “EEG All of the above current source models are often referred to
Inverse Problem.” as distributed current density models. Physically, any bioelectric
source activity can be represented by a continuous distribution
Source Models of primary current density. Mathematically, the current density
The primary sources of EEG are considered to be the postsy- and current dipole share the identical form of equations for
naptic currents flowing through the apical dendritic trees of computing the extracellular potential, supporting the use of dis-
cortical pyramidal cells. Such neuronal currents, when viewed tributed current dipoles to approximate the continuous current
from a location on the scalp surface that is relatively remote to density distribution. The dipole distribution may be viewed as
where the neural excitation takes place (far field), can be mod- discretized version of the continuous current density distribu-
eled as an electric current dipole composed of a pair of current tion in the space domain. Each of such current dipoles repre-
source and sink with infinitely small interdistance. When the sents the regional neuronal activity, and the dipole amplitude
brain electric activity is confined to a few focal regions, each of indicates the amount of synchronized neuronal activity in the
these focal areas of neuronal excitation may be modeled as an local region.
equivalent current dipole (ECD) based on the far field theory In addition to current dipole-based source models, the current
(for a theoretical treatment see the appendix in Chapter 5). monopole model (125) was also used to equivalently represent
Such equivalent dipole model has been widely used in source brain electric activity. Mathematically, such current monopole
localization analysis of EEG in an attempt to better interpret the model can also represent the source activities in a sense that it pro-
origins of the scalp-recorded EEG (116–120). While the ECD is duces almost the same electric potentials on the scalp electrodes.
a simplified model and higher order equivalent source models However, the biologic interpretation of such monopole source
such as the quadrupole have also been studied to represent the model remains unclear. Another alternative source model is based
neural electric sources (121,122), the dipole model has been so on the estimation of the three-dimensional (3D) distribution of
far the most commonly used brain electric source model. A the intracranial potentials instead of the three components of a
number of experimental and clinical studies have demonstrated dipole, as proposed in the ELECTRA source imaging technique
its merits in helping interpreting EEG data and localizing (102,126). It is based on the neurophysiologic consideration that
sources generating the scalp-recorded EEG. ohmic currents dominate the EEG measurements (127). While
When the neuronal sources are no longer confined to a few such 3D potential estimation more closely resembles intracranial
focal regions, the ECD model may not well represent the distrib- recordings, their biologic interpretation is as difficult as the esti-
uted brain electric activity. Distributed current source models mation of current monopoles.
can then be used to represent the whole-brain bioelectric activity.
The essence of the distributed current source models is to model Volume Conductor Models
the neuronal activities over a small region by a current dipole When the source model is determined, the EEG forward prob-
located at each region. The brain activity with any distribution of lem consists of obtaining the distribution of electric potential F
neuronal currents can be approximately represented by a source on the scalp surface, given any known distribution of current
model consisting of a distribution of current dipoles that are density JY inside the brain as well as conductivity values through-
evenly placed within the entire brain volume. At each location, out the head volume. Such solution is usually called EEG for-
three orthogonal dipoles are used in that the weighted combina- ward solution for a given head volume conductor model. The
tion of them is capable of representing an averaged current flow head volume conductor model refers to our assumption on the
with an arbitrary direction in the region. Such source model is shape and conductivity properties of the tissues of the head.
usually called volume current density (VCD) model, in which Head volume conductor models include the infinite homoge-
current dipoles distribute over the entire brain volume (123). The neous model, single-sphere model, three concentric–sphere
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

brain anatomical information can also be used to constrain the model, four concentric–sphere model, realistic geometry
current source space to the cortical gray matter due to its domi- homogenous head model, realistic geometry multicompartment
nant presence of large pyramidal cells. Such anatomical con- head model, and realistic geometry inhomogeneous head
straints can be obtained from existing structural neuroimaging model.
modalities, particularly T1-weighted MRI, which provides high The simplest EEG forward solution is that in the infinite
spatial resolution and a great contrast to differentiate the cortical homogeneous model, where the entire space is assumed to be
gray matter from the white matter and cerebrospinal fluid (CSF). occupied by a homogeneous conductive medium (4,127). The
The current source orientations can be further constrained to be electric potential over the scalp electrodes can then be given as
perpendicular to the cortical surface, because the columnar

# § a r b # YJ dv
organization of neurons within the cortical gray matter con- 1 1 i
£5 (55.3)
strains the regional current flow in either outward or inward nor- 4ps V
mal direction with respect to the local cortical patch (124), and
the gray matter thickness (about 2 to 4 mm) is much smaller rel- where the source element JYi dv behaves like a dipole source,
ative to the “source-to-sensor” distance (60). Under such cortical with a field that varies as 1/r 2, and ∇ represents the divergence
constraints, such source model is usually called cortical current operator (see the appendix in Chapter 5). The impressed cur-
density (CCD) model, in which current dipoles distribute over a rent density JYi , representing neuronal currents, may be inter-
surface in parallel to the epicortical surface. preted as an equivalent dipole source density, which behaves as

a fundamental driving force establishing the electric potentials

within the head volume conductor with an electric conductiv-
ity s.
A more reasonable representation of head is the series of
spherical models, including a single-sphere model (128),
three concentric–sphere model (30,55,129), and four
concentric–sphere model (130,131). In these spherical models,
the shape of the head is approximated by spherical surfaces,
including the scalp, the skull surface, the brain surface, etc. In
such cases, the electric potentials over the scalp surface (the
outer sphere) due to a current dipole can be derived analytically
for the single-sphere model, or by use of special function for the Figure 55.7 A: A realistic geometry multishell boundary element head
multisphere models. Since the low-conductivity skull layer model as derived from structure MRI of a human subject (207). The scalp,
smears significantly the electric potential, the three-sphere skull, and brain surfaces are represented by a number of surface triangles.
model (brain, skull, and scalp), in which the skull layer is incor- Brown and blue surfaces refer to the right and left cortical surfaces. Pink
porated, has been used widely and found to be a good approx- circles refer to the scalp electrodes. B: A realistic geometry inhomoge-
imation to the head volume conductor when the shape of the neous finite element head model as derived from structure MRI and CT of
head is ignored. Furthermore, the CSF layer can also be incor- a human subject (139). Green, gray, blue, and brown regions refer to the
porated as in the four-sphere model (130), although there is no scalp, skull, CSF, and brain. Red refers to the subdural pad of ECoG elec-
converging agreement that the CSF layer must be incorporated trode array, which has low electric conductivity. The red layer is not dis-
when modeling the head volume conductor. played continuously in this figure due to the angle of view.(From Liu ZM,
We should note that the influence of the inhomogeneities of He B. FMRI-EEG integrated cortical source imaging by use of time-variant
the tissues of the head as volume conductor affect differently spatial constraints. NeuroImage. 2008;39(3):1198–1214 [Panel A], and
the EEG and the MEG, since the brain and surrounding tissues Zhang Y, van Drongelen W, Kohrman M, et al. Three-dimensional brain
behave as a medium with constant magnetic permeability. The current source reconstruction from Intra-cranial ECoG recordings.
magnetic field, in contrast to the electric field, is not influenced NeuroImage. 2008;42:683–695 [Panel B]). (See color insert)
by the tissue inhomogeneities, for a given primary current den-
sity distribution (the current density distribution solely
induced by neuronal activation). In particular, for concentric example of three-shell BEM head model is illustrated in Figure
spherical head models, MEG is not influenced by the concentric 55.7A, as derived from structural MRI of a human subject.
layers. However, when the realistic geometry of head tissue is In addition to the BEM that provides forward solutions when
taken into consideration, existence of head tissue inhomo- the head can be modeled by multicompartment model of homo-
geneities introduces secondary current sources, which in turn geneous conductivity profiles within each compartment, the finite
affect the magnetic field out of the scalp. See also discussion in element method (FEM) has been further used to handle the inho-
the appendix of Chapter 5 and in Chapter 42. mogeneous conductivity distribution within the head (135–139).
A major disadvantage of the spherical head model is that it Figure 55.7B shows an example of a FEM head model of a patient
does not make any distinction between areas that generate elec- undergoing subdural recordings. The FEM modeling allows
tric activity (gray matter) and those that do not (white matter proper handling of inhomogeneity of head tissues including the
and ventricles). Constraining the source space to the gray mat- surgical alteration such as the very low-conductivity subdural pad
ter of the individual brain is an important anatomical con- of ECoG electrodes. The FEM forward solutions enable incorpo-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

straint that improves the accuracy of the EEG forward solution. rating the realistic geometries and inhomogeneous conductivities
A simple straightforward way is to map the individual seg- of the head, even allowing inclusion of the anisotropic conductiv-
mented MRI to a sphere and use the analytical multishell spher- ity distribution of the white matter (140). The challenge of using
ical model described above, but with the solution space FEM to solve EEG forward problem is not the FEM algorithm,
constrained to the individual gray matter (132). which has been fairly well developed, but the need to build a FEM
More advanced EEG forward solutions have used numerical mesh model of the head from MRI or CT images. Nowadays this
techniques to take both the conductivity inhomogeneity and the is still a labor-consuming effort to build FEM head models from a
geometry of the head into consideration. The most popular EEG subject’s MRI. Alternative efforts, such as the finite difference
forward solution is based on the boundary element method method (FDM), have been reported to solve the EEG forward
(BEM) (118,133,134). He et al. (118) first reported the use of problem in which the finite difference grids can be reasonably eas-
BEM in solving the EEG forward and inverse problems using a ily built from MRI of a subject. However, it is still not a straight-
realistic geometry homogeneous head model. The low-conduc- forward procedure to automatically segment and classify head
tivity skull layer was incorporated later by the studies of tissues into the FDM model. For recent review of image segmen-
Hämäläinen and Sarvas (133) and Meijs et al. (134). In particu- tation for the purpose of solving the EEG forward problem, see
lar, the numerical treatment of the low-conductivity skull layer Ref. 141. All of these numerical techniques have to utilize the
developed by Hämäläinen and Sarvas (133) made the BEM for- anatomical information provided by other structural imaging
ward solution widely used for EEG as well as MEG studies. An modalities, particularly T1- and proton density–weighted MRI, in

order to segment different brain tissues (i.e., the gray matter and neurophysiology and neuroscience. Another challenge existing in
the white matter) and head structures (i.e., CSF, skull, and scalp). the EEG inverse problem is that it is ill-posed from the viewpoint
Sometimes, CT is used together with MRI to obtain structure of numerical treatment. This ill-posedness of the inverse problem
information of the skull. Image segmentation remains to be an is handled by various signal processing strategies, which has led
important task in realistic geometry inhomogeneous head mod- to great success in the field.
eling regardless of the numerical algorithms to be used. In this section we will discuss the EEG inverse problem in the
In addition, recent development of diffusion tensor mag- following aspects: (i) dipole source localization (DSL); (ii) source
netic resonance imaging (DT-MRI) (142) provides a means to scanning; (iii) distributed source imaging; and (iv) multimodal
estimate the anisotropic conductivity of the cerebral white mat- imaging integrating EEG source imaging with functional MRI.
ter (137,143,144), which may further improve the accuracy of
the EEG forward solution. The white matter anisotropy is Dipole Source Localization
caused by the bundled axon fibers that restrict the direction of The most classic EEG inverse problem shall be the so-called DSL.
ionic movements (137). While a computer simulation study Such DSL approach is applicable when the primary generators of
(137) suggested that the white matter conductivity anisotropy scalp-recorded EEG signals are localized to one or few small
may have effects to EEG forward and inverse solutions, a recent regions within the brain. Given a specific dipole source model,
experimental study suggested otherwise. He and colleagues DSL can yield solutions of the EEG inverse problem by using a
(140) conducted a human study to localize the sources in pri- nonlinear multidimensional optimization procedure, to estimate
mary visual cortex using the single dipole solution and com- the dipole parameters that can best explain the observed scalp
pared with the fMRI activation under the same visual stimuli in potential measurements in a least-square sense (116–118,
the same subjects. Their results indicated that use of the 120,145–147). Further improvement of the DSL can be achieved
anisotropic white matter model did not return significantly dif- by combining EEG with MEG data that may increase informa-
ferent solution as that using the isotropic white matter model. tion content and improve the overall signal-to-noise ratio
Further investigation is needed to examine the effects of white (148,149), or integrated with functional MRI (see the section
matter anisotropy in other brain regions. “Multimodel Integration of EEG with fMRI” for details).
The EEG DSL can be solved either from the scalp-recorded
EEG INVERSE PROBLEM EEG at a given time instant or over a time period. The single
time-slice source localization estimates the dipole parameters
The EEG inverse problem shall be considered as to reconstruct based on the single time “snapshots” of the measured scalp EEG
brain electric sources from scalp-recorded EEG signals. In (120). In practice, an initial starting point (also termed seed
experimental and clinical studies, it is desirable to image the point) is selected, then using an iterative procedure, the assumed
neuronal activity that generates recordable scalp EEG signals. dipole sources are moved around inside the brain volume in an
When the neuronal activity is localized in few focal regions attempt to produce the best match between the measured and
within the brain, the problem becomes to localize the locations dipole-produced scalp electric potentials. This involves solving
of such activated focal regions, which is usually referred to as the forward problem repetitively and calculating the difference
source localization. For example, source localization has been between the measured scalp potential vector wY and the model-
found useful in localizing epileptogenic foci in epilepsy patients. predicted scalp potential vector c Y at each step. The most com-
On the other hand, when the brain activity is not localized in a monly used measure is the squared distance between the two
few focal areas, one needs to image the distribution of neural vectors, which is given by
electric sources within the brain. Such problem becomes the so- Y 72
D 5 7 wY 2 c (55.4)
called source imaging, which shall include source localization
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

since focal sources are special cases of distributed sources. where D is the objective function that is to be minimized (equal
As opposed to the uniqueness of the forward solution, the to GMD if normalized). The EEG dipole inverse solution is
EEG inverse solution is nonunique if no constraints are given due obtained when this objective function is minimized. Due to the
to the low frequency nature and the limitation of EEG recordings effect of measurement noise, usually one can only estimate one
over the scalp surface only. This nonuniqueness issue has repre- or two moving equivalent dipoles by use of the DSL approach. In
sented a great challenge to the field. Over the past two decades, a addition, reciprocal approaches have also been explored in solv-
number of efforts have been made to tackle these challenges and ing the DSL, in an attempt to improve the numerical accuracy of
it is encouraging that we have learnt that it is possible to obtain the transfer matrices (150).
valid estimates of solutions of the EEG inverse problem if reason- The EEG DSL can also be solved from spatiotemporal EEG
able constraints are given on the equivalent source distribution. measurements, which is sometimes termed spatiotemporal
For example, if the brain electric sources are assumed to consist source localization (117). In this approach, multiple dipole
of few moving ECDs (117,118), or a current dipole distribution sources are assumed to have fixed locations inside the brain
over the cortex with orientation being perpendicular to the local during a certain time interval, and the variations in scalp
curvature of the cortex (124), or a current dipole distribution potentials are due only to variations in the strengths of these
over the brain volume with smoothness a priori (123), in such sources. The dipole sources SY are coupled to the scalp poten-
cases solutions of the EEG inverse problem can be estimated tials wY by the lead field matrix A as wY 5 ASY . The task of the
yielding results that are consistent with other findings of clinical spatiotemporal DSL is to determine the magnitudes and

orientations of multiple dipoles, whose parameters could best through the 3D brain volume (solution space) to identify sources
account for the spatial distribution as well as the temporal that produce potential patterns that lie within the signal subspace
waveforms of the scalp EEG measurement. With the incorpo- of the EEG measurements (164). Furthermore, a recursive
ration of the EEG temporal information in the model fitting, method (RAP-MUSIC) has been developed in order to overcome
the spatiotemporal DSL is more robust against measurement the “multiple-peak picking” problem of the original MUSIC scan
noise and artifacts than the single time-slice DSL. However, (166). Recently, another source scanning algorithm (first princi-
once the locations of dipoles become unknowns, the problem ple vectors, FINE) (167) has been introduced to localize sources
becomes a nonlinear optimization problem and the number with high resolution, and extended to include realistic geometry
of dipoles that can be estimated reliably is limited. inhomogeneous head model (168). The applicability of FINE
All equivalent dipole algorithms need an a priori knowledge scanning algorithm has been shown in a group of patients of
of the number and class of the underlying dipole sources. If the focal interictal epileptogenic activity (38). The FINE algorithm
number of dipoles is underestimated for a given model, then (167,168) is developed under the framework of the subspace
the DSL inverse solution is biased by the missing dipoles. On source localization approach (164) and solves the spatiotemporal
the other hand, if too many dipoles are specified, then spurious source localization problem using a scanning strategy instead of
dipoles may be introduced. So while efforts have been made to optimizing a multidimensional cost function. In principle, the
estimate the number of dipoles based on information criterion subspace source localization approach scans the entire possible
theory (151), the prior information with regard to the number source space and calculates the subspace correlation of two sub-
of dipoles remains to be the limitation of EEG DSL. Sometimes spaces (166). One subspace is spanned by each scanned point and
such information is provided based on the neurophysiology another one is estimated from the scalp EEG, which is known as
and pathology such as the case when there is a clear focal region the noise-only subspace. If the subspace correlation is approxi-
of epileptiform activity (152–154), or known physiology such mate to zero against the noise-only subspace for one possible
as somatosensory-evoked potentials (118). source point, this point is regarded as a source. Multiple sources
The EEG DSL solution can be further improved when prior could thus be obtained at multiple extreme values. The FINE
information is available with regard to the possible solution space. algorithm calculates the correlation to a particular subset of the
If the observed EEG signals are known to be mainly produced by noise-only subspace instead of the entire noise-only subspace,
cortical sources, then the solution space can be restricted to the which helps to achieve high spatial resolution. Computer simula-
cortex surface while excluding deep source locations such as the tion results suggested the enhanced spatial resolution and robust-
brainstem. In particular situations, when a priori information is ness against source coherence of FINE in comparison to MUSIC
known on the possible source region (such as the sensory EP data, or RAP-MUSIC (167,168). Applications to epilepsy source local-
or based on preliminary diagnosis of the epilepsy), the solution ization demonstrated the ability of FINE in localizing seizure foci
space can be restricted to only half of the brain or even more (40). Further development of source scanning approach has also
focused on a certain lobe. Computer simulation and experimen- been reported (169) in which a higher order source model is used
tal studies have also demonstrated that the inverse solution of the instead of dipole model, enabling imaging of complex activation
DSL is more accurate when using the realistic geometry head patterns in humans induced by visual stimulation.
model than the simplified spherical head model (155–157). In addition, another technique called beamforming, which is
Furthermore, by registration with the magnetic resonance images, based on linear spatial filtering, can be used to estimate the
the coordinates of the estimated dipole sources can be visualized source activity at a region of interest or every individual loca-
relative to the brain anatomy. Therefore, it has great potential to tion in the source space (170,171). The beamformer for a spe-
reveal the electrophysiologically active neural substrate underly- cific region of interest or source location is derived in an
ing the scalp potential measurements and facilitate comparison attempt to minimize the interference from other locations.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

with other functional imaging modalities (158,159), and has clin- A source localization that was specifically developed for
ical significance in detecting the epileptic foci (160,161), presurgi- localization of focal epileptic activity is a method called EPIFO-
cal localization of sensorimotor cortex (162,163), and some other CUS (172). Like the MUSIC method, this technique searches
applications. for focal sources in the solution space (which can be restricted
to the gray matter). EPIFOCUS is a linear inverse solution and
Source Scanning requires no nonlinear optimization procedure to find the best
The source localization problem can also be solved by means of a fit. The result of EPIFOCUS can be interpreted as the probabil-
source scanning procedure (164–168). The source scanning tech- ity of a focal source (with a certain spatial extent) at any given
nique avoids solving difficult multidimensional nonlinear opti- point in a discrete solution space. The reliability of EPIFOCUS
mization problems (118,146,152). The scanning results from the to determine focal epileptic sources has been demonstrated in
subspace source localization approaches will directly provide an several studies (15,173,174).
estimate of the source distribution in the 3D brain volume. The
source scanning approach, which scans each point in the brain Distributed Source Imaging
volume and returns a metric about how much chance there is a Compared with DSL or source scanning approaches, the dis-
source at the point, may also be used for source imaging. A rep- tributed source imaging approach has received considerable
resentative source scanning algorithm is the multiple signal clas- attention in recent years. Such distributed imaging approaches
sification (MUSIC) algorithm (164), which has been used to scan are motivated by the need to image brain electric activity that

may not be a few focal sources, and to minimize the error due known as dynamic statistical parametric mapping (dSPM)
to misspecification of the number of dipole sources. The dis- (178) and standardized LORETA (sLORETA) (179), based on
tributed source imaging may be categorized into two groups: the MNLS and LORETA algorithms, respectively (180).
distributed current density imaging and equivalent potential The EEG linear inverse solutions enjoy the merits of solving a
imaging, both of which have been studied extensively. linear inverse problem but end up with low-resolution images of
current density or its statistics. One way to improve the spatial
Distributed Current Density Imaging resolution is to generate images with focal source distribution by
Unlike the point dipole source models, the distributed source iteratively repeating the linear inverse computation (181,182).
models do not make any ad hoc assumption on the number of For each step during the iteration, the linearly computed inverse
brain electric sources. Instead, the equivalent sources are dis- solution from the previous step is used as the weighting factors to
tributed over the source space. For example, the distributed constrain the linear source estimates in the current step. As such
source model may consist of a large number of current dipoles a recursive process continues till convergence, the estimated
placed over the 3D brain volume (123) or over the gray matter source distribution tends to shrink to be more focalized. Other
of cortex (124,132). Regardless of the current source models than L2 norm estimates that are to minimize the energy function,
being used, the current dipoles are fixed at preset locations, so Lp norm estimates with P = 1 or other values have also been
they are not movable. The unknown parameters are the dipole explored (147,183–186). A study by Ding and He (186) demon-
moments, which are to be estimated by minimizing the differ- strates the merits of L1 norm–based distributed current density
ence between the measure and source model–predicted scalp imaging in imaging focal sources in humans as induced by
electric potentials. somatosensory stimulation (Fig. 55.8).
Assuming quasistatic condition and the linear properties of
the head volume conductor, the brain electric sources and the Cortical Potential Imaging
scalp EEG could be mathematically described by the following The current density distribution has explicit physical interpre-
linear matrix equation: tation of neuronal currents, thus being used widely in solving
Y 1 nY the EEG inverse problem. Another kind of EEG linear inverse
wY 5 AX (55.5)

where wY is the vector of scalp EEG measurements, X Y is the vec-

tor of source distribution, nY is the vector of additive measure-
ment noise, and A is the transfer matrix relating wY and XY . So the
distributed source imaging is to estimate the source distribu-
tion XY from the noninvasive scalp EEG measurements wY .
Mathematically, this is equivalent to design an inverse filter B,
which can project the measured data into the solution space:
Y 5 BwY
X (55.6)

The above equation indicates that the distributed source

imaging is a linear inverse problem, which avoids the difficulty
of nonlinear multidimensional optimization problem as in the
DSL approach. However, the linear inverse approach is intrinsi-
cally underdetermined, because the number of unknown dis-
tributed sources is much larger than the limited number of
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

scalp EEG electrodes. Additional constraints have to be

imposed in order to obtain unique and well-posed linear
inverse solutions. A well-studied solution to this linear inverse
problem is the so-called general inverse, which is also termed
the minimum norm least-squares (MNLS) inverse, minimizing
the least-square error of the estimated inverse solution X Y under
Y
the constraint wY 5 AX in the absence of noise (175). Other Figure 55.8 Sparse source imaging. The image illustrates the concept of
variations of the MNLS include the lead field normalized sparse source imaging (SSI) using L1 norm–based generalized minimum
weighted minimum norm (WMN) (176), low-resolution brain norm estimate. The new SSI algorithm corrects inaccurate source field
electromagnetic tomography (LORETA) (123), local autore- modeling in previously reported L1-norm algorithms and proposes that
gressive average (LAURA) (177), and others. Furthermore, by sparseness a priori should only be applied to the regularization term, not
normalizing the source estimates with respect to the correspon- to the data term in the formulation of the regularized inverse problem.
ding noise sensitivity, one can assess the statistical significance The right lower part of the illustration shows that the new SSI is evaluated
of the inverse solution and obtain a map of source estimate sta- using somatosensory-evoked potential data with subdural electrocortico-
tistics. Along this line, Dale et al. and Pascual-Marqui et al. have graphic (ECoG) recordings in a human subject. (Cover image of Human
developed two statistical functional mapping techniques, Brain Mapping, September 2008 issue for article by Ding and He (186).)

Figure 55.9 At latency of 30 msec after the onset of right median nerve stimuli for a patient: (A) the recorded scalp poten-
tial maps, (B) the estimated subdural grid potentials, and (C) the direct recorded subdural grid potentials. All the maps are
normalized and the color bars are shown on the right. The CC value between the estimated and recorded subdural grid poten-
tials is 84%. (From He B, Zhang X, Lian J, et al. Boundary element method-based cortical potential imaging of somatosen-
sory evoked potentials using subjects’ magnetic resonance images. Neuroimage. 2002;16:564–576.) (See color insert)

solution is to estimate the extracellular potential distribution approach called ELECTRA (102,126). Although the validation
from the scalp EEG. Cortical potential imaging (CPI) has been is difficult since the depth electrodes are normally not covering
well studied that aims to reconstruct the distribution of electric a large volume, this approach theoretically does suggest the pos-
potentials over the epicortical surface that can equivalently rep- sibility of estimating local field potentials in the brain volume.
resent the enclosed brain electric sources (37,56,129,187–195).
In this case, the source vector X Y in Eq. (55.5) refers to electric Multimodel Integration of EEG with fMRI
potentials instead of current density. Imaging of electric poten- Integrating electrophysiologic source imaging with functional
tial over the epicortical surface has a unique feature that it constraint derived from the hemodynamic imaging modalities
returns cortical potentials, which are routinely measured by use has drawn great attention during the past decade (see Ref. 198
of subdural electrode arrays in epilepsy patients. for a detailed review). The motivation for integrating EEG/MEG
The CPI employs a distributed source model, in which the source imaging with fMRI is based on the different strengths
equivalent sources are distributed in 2D cortical surface, and no and limitations of these two modalities that are precisely the
ad hoc assumption on the number of source dipoles is needed obverse of each other, that is, EEG/MEG enjoys high temporal
(37,129,138,139,187–196). Using an explicit biophysical model resolution but limited spatial resolution, and fMRI has high spa-
of the passive conducting properties of a head, the CPI attempts tial resolution but poor temporal resolution. The fundamental
to deconvolve a measured scalp potential distribution into a assumption of the multimodal integration approach is that
distribution of the electric potential over the epicortical surface. regions in the brain that show increased metabolic activity are
Because the cortical potential distribution can be experimen- also on the average more electrically active over time.
tally measured (190,195,197) and compared to the inverse The earliest efforts in the E/MEG–fMRI integrated imaging
imaging results, the CPI approach is of clinical importance. utilized fMRI statistical parametric maps to obtain a priori
Essentially not affected by the low-conductivity skull layer, CPI information on where brain electric sources are likely located.
offers much-enhanced spatial resolution in assessing the under- The spatial information derived from fMRI has been used to
lying brain activity as compared to the blurred scalp potentials. constrain the locations of multiple current dipoles (199,200), or
Figure 55.9 shows an example of cortical potentials estimated to constrain the distributed source distribution (178,201–207).
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

from the scalp EEG and directly measured using subdural grids When neural activity is confined to a few small regions, the fMRI
in a human subject, around 30 msec after the onset of right activation mapping should yield several hotspots, which can be
median nerve stimuli. The scalp potential map (Fig. 55.9A) used to constraint the equivalent dipole locations or as initial
shows dipolar pattern of N/P30, with frontal negativity and locations in the DSL. Once the dipole locations are recon-
parietal positivity over the left scalp. The smearing effect of the structed from the EEG data, the time course of dipole moments
scalp potential map was greatly reduced in the inversely esti- represents the temporal dynamics of the regional neural activity.
mated cortical potential map (Fig. 55.9B), which shows much Such fMRI-seeded dipole fitting technique is usually used to
more localized areas of positivity and negativity in the posterior retrieve the time course of the brain activity at identified fMRI
edge of the electrode grid. The recorded cortical potential map activation foci instead of imaging brain activity. On the other
is shown in Figure 55.9C. Note the estimated and recorded grid hand, the fMRI-constrained distributed source imaging (i.e.,
potentials have similar distribution patterns, with averaged CC fMRI is used as constraint in the distributed source imaging)
value of 0.84. Moreover, the central sulcus was clearly demar- can be applied to brain sources that are either focal or extended.
cated in both the estimated and recorded grid potential maps, The fMRI-constrained current density imaging has been
by the separation of negative and positive potential extrema. explored, in the framework of Wiener filter (124,178,201) or
In addition to cortical potential, the extracellular potentials WMN frameworks (202,203). The major technical limitations
have also been estimated over the entire 3D brain volume in an of the above-mentioned approaches are primarily due to the

fundamental mismatches between fMRI and EEG (or MEG), filtering applies to the spatial domain instead of the time
owing to different temporal scales in which fMRI and EEG tem- domain. This feature is opposed to the temporal regression of
poral dynamics are generated and collected. The fMRI data, which theoretically ends up with high-resolution
fMRI–EEG/MEG mismatches include fMRI extra sources, fMRI spatial maps of brain activations but with little or no temporal
invisible sources, and the fMRI displacement (201,202,206), and information. To integrate the EEG and BOLD signals, the
it is problematic to constrain the temporally variable current BOLD effect size estimated from the fMRI signal in each voxel
source estimates to “time-invariant” fMRI spatial priors, which is set to be proportional to the time integral of the local source
may result in fMRI false positives or false negatives. power underlying the ERP signals (207). The source estimates
Efforts have been made to tackle this challenge caused by the are further fitted to the EEG data by means of an adaptive
different time scales of BOLD fMRI signals and EEG signals, by Wiener filter (207). Figure 55.10 shows an experimental exam-
means of data-driven approaches. For example, the fMRI ple in a human subject exploring the cortical pathway special-
weighting factor may be selected from data by means of the ized in processing unilateral visual stimuli (207). The
expectation maximization (EM) algorithm (205,208), or alter- experiment included two separate sessions with the identical
natively a two-step estimation algorithm, so-called Twomey visual stimuli for the EEG and fMRI data collection. The visual
regularization, can be used to achieve reasonable estimation of stimulation was a rectangular checkerboard within the lower
the fMRI weighting factor (206). Recently, a new framework for left quadrant of the visual field; the checkerboard pattern was
the fMRI–EEG/MEG integrated neuroimaging has been pro- reversed at 2 Hz. The dynamically integrated EEG–fMRI imag-
posed (207). The system assumes a common neuronal source ing algorithm (207) revealed a pathway sequentially activating
(i.e., synaptic activity), from which fMRI and EEG signals are V1/V2, V3/V3a, V5/V7, and intraparietal sulcus, in general
generated via a temporal low-pass filter and a spatial low-pass agreement with the hierarchical organization of the visual sys-
filter, respectively. The EEG inverse problems essentially deal tem (209). This pathway was also observed in the low-resolu-
with the spatial deconvolution—the process of reversing the tion images reconstructed from the VEP alone. In contrast, a
head volume conduction. The EEG inverse solution retains the fMRI-weighted source imaging algorithm (178,201) showed a
temporal source evolution even though it may fail to recon- false-positive source region in and around V1/V2 at the latency
struct the spatial source distribution. In other words, at every of 212 msec, whereas a more likely high-tier EEG source around
source location, the source waveform estimated from EEG is V5, as observed from the EEG data, was missed. This experi-
much less distorted than its absolute magnitude, since the mental result indicates the promises in dynamic neuroimaging

Figure 55.10 fMRI–EEG integration. A: The pattern-

reversal checkerboard visual stimulation, (B) fMRI acti-
vation map with a corrected threshold P , 0.01, and (C)
the global field power of VEP and the dynamic cortical
source distribution at three VEP latencies (76, 112, 212
msec after the visual onset) imaged from EEG alone
(first row), or fMRI–EEG integration using our proposed
adaptive wiener filter (second row) and the conven-
tional 90% fMRI-weighted algorithm (third row). Both
the source images and the fMRI activation map are visu-
alized on an inflated representation of cortical surface.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

(From Liu Z, He B. FMRI–EEG integrated cortical source

imaging by use of time-variant spatial constraints.
Neuroimage. 2008;39:1198–1214.)(See color insert)

Figure 55.11 Combined EEG–fMRI in epilepsy. Comparison of fMRI and EEG source imaging. A patient with focal
epilepsy has been recorded with 32-channel EEG in a 1.5-T MR scanner. Spikes were marked and the significant BOLD
responses related to these spikes were determined. Two foci were found around the large lesion: one right lateral parietal
and the other mesial parietal. The same spikes recorded in the scanner were cleaned and subjected to EEG source imag-
ing using a distributed linear inverse solution constrained to the gray matter determined from the patient’s MRI. A tempo-
ral propagation of the activity was found, with the initial activity in the right parietal lobe, followed by activation in mesial
parietal areas. Thus, the foci found in the fMRI were confirmed and temporally resolved. The patient was seizure free after
surgical resection of the right lateral parietal focus. (Data collected by M. Siniatchkin, University Hospital of Pediatric
Neurology, Kiel, Germany. For details and more examples see Ref. 215.)(See color insert)

by integrating fMRI with EEG using the model-based adaptive large brain lesions, indicating that conductivity changes due to
Wiener filter (198). such lesions are not as important as one might believe. This has
New EEG recording systems together with advanced artifact been recently demonstrated in a systematic study in a series of
correction algorithms (79,80) allow recovering the EEG signal epileptic patients with large brain lesions (217).
that is acquired in the scanner. This opens new possibilities to
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

study the relation between hemodynamic and electric activity,

CONNECTIVITY ANALYSIS
but also to directly combine the temporal resolution of the EEG
and the spatial resolution of the fMRI during the same brain Static images indicating brain regions responsible for the execu-
state. Several recent studies have used this technique to study the tion of particular tasks do not convey sufficient information
relation between different brain rhythms and the BOLD with respect to how these regions communicate with each
response (210,211) and the relation between the so-called brain other. The concept of brain connectivity now plays an impor-
resting state and specific oscillatory activity recorded with the tant role in neuroscience, as a way to understand the organized
EEG (212–214). With the use of high-density EEG in the scan- behavior of brain regions (218). Previously, some investigators
ner, source imaging of this EEG has also become possible. This have computed cortical connectivity patterns based on hemo-
is particularly interesting in epilepsy where the spike-related dynamic or metabolic measurements such as fMRI (219,220),
fMRI activity can be compared with the source imaging result of whereas the sluggishness of the hemodynamic process con-
the very same spikes. Such studies have recently demonstrated founds its interpretation in terms of neuronal interaction
the capability of the EEG source imaging to temporally disen- (221). The use of EEG data to examine the functional connec-
tangle the different activated regions seen in the spike-triggered tivity has a long and rich history (222,223). A variety of tech-
fMRI as illustrated in Figure 55.11 (215,216). This figure also niques have been used, most of which have amounted to
shows that EEG source imaging is also possible in patients with evaluating the cross-correlation or phase synchronization of

Figure 55.12 Connectivity estimation. Left panel: Functional connectivity

patterns estimated in a subject during the performance of finger tapping
movement, after the EMG onset. Each pattern is represented with arrows
moving from one cortical area toward another. The color and size of the
arrows code the level of strength of the functional connectivity observed
between ROIs. The labels indicate the names of the ROIs employed. Right
panel: Outflow patterns in all the ROIs obtained for the same connectivity
pattern depicted in top left. The figure summarizes in red hues the behavior
of a ROI in terms of reception of information flow from other ROIs, by adding
all the value of the links arriving on the particular ROI from all the others.
The information is represented with the size and the color of a sphere, cen-
tered on the particular ROI analyzed. The larger the sphere, the higher the
value of inflow or outflow for any given ROI. The blue hues code the outflow
of information from a single ROI toward all the others. (From Babiloni F,
Cincotti F, Babiloni C, et al. Estimation of the cortical functional connectiv-
ity with the multimodal integration of high-resolution EEG and fMRI data by
directed transfer function. Neuroimage. 2005;24:118–131.)(See color insert)

signals between pairs of scalp electrodes or sensors (224). space is complicated by the dispersion of electromagnetic sig-
Additionally, it has also been exploited that scalp connectivity nals from the cortex to the sensors.
patterns can be estimated by other methods (225,226). Graph EEG source imaging techniques have been used in combi-
theory-based tools from the study of complex network have nation with connectivity estimators to noninvasively assess the
also been developed to describe the connectivity of large-scale brain connectivity using methods such as structural equation
networks (227). However, the relation between the observed modeling (SEM) (228,229) and the directed transfer function
connectivity pattern in the sensor space and that in the source (DTF) (204). Figure 55.12 shows an example of functional
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 55.13 Ictal source imaging. A: Three second–long 31-channel scalp waveforms for subspace source localization
analysis. B: Example of a 3D scanning result by FINE for an ictal activity displayed with gray MRI slices. The pseudocol-
ors show the reciprocal of subspace correlation (SC). Red: low SC; blue: high SC. The extent of pseudocolors indicates the
coverage of the possible solution space. Three identified sources in the 3D scanning are marked with red, blue, and green
dots, respectively. C: Locations (pseudocolors on MRI images), waveforms (green curves), and causality patterns (big
arrows) for identified sources from B. (From Ding L, Worrell GA, Lagerlund TD, et al. Ictal source analysis: localization
and imaging of causal interactions in humans. Neuroimage. 2007;34:575–586.)(See color insert)

connectivity patterns as estimated by means of DTF from EEG described the historical evolution of these models, but also made
and functional MRI of a human subject during a motor task clear that a distinction between simple source localization and
(204). It shows connectivity among several regions of interests comprehensive source imaging should be made. The single
involved in the motor task. equivalent dipole fitting approach in a simple spherical head
DTF-based connectivity analysis methods showed promises model, which was still state of the art in the last edition of this
in localizing epileptogenic foci from interictal spikes (230,231) book, has been largely replaced by imaging of distributed
and from seizures (232) as recorded from ECoG. The functional sources in the realistic geometry head model defined by struc-
connectivity in epilepsy patients has also been estimated by tural MRI. A rapidly growing number of experimental and clin-
applying DTF analysis to 3D current source activity recon- ical studies appeared, demonstrating the promising capability of
structed from ictal EEG using the FINE source scanning these new techniques. Most impressive are the results in epilepsy
method (40). Figure 55.13 shows an example of the clinical where EEG source imaging is used to localize the epileptogenic
application of connectivity imaging. From ictal EEG data, three zone. The fact that the very same methods also allow to localize
putative sources were estimated by means of the FINE source eloquent cortex with impressive precision renders electric neu-
scanning algorithm. The causality among each pair of the three roimaging one of the most promising methods in the frame-
sources, as determined from the source waveforms, was work of presurgical evaluation of patients with functional and
assessed. The causal interactions from the “red” and “green” structural brain lesions.
sources to other sources are deemed significant but not the The temporal resolution is the key advantage of the EEG.
causal interactions from the “blue” source. Based on this con- However, it also increases the complexity and demands addi-
nectivity analysis, it was concluded that the “red” and “green” tional analysis steps for electric neuroimaging compared to the
sources are deemed to be primary seizure sources while the other (static) imaging procedures. We here described different
“blue” secondary source. Comparison with MR-visible lesions methods to deal with the temporal dynamics of the brain elec-
in the patient supported this conclusion derived by connectiv- tric activity. One of them is based on spatiotemporal decompo-
ity analysis (40). sition of the topographic maps. It allows to define the most
dominant scalp topographies during given time periods and
CONCLUSION thus permits a reduction of the complex data in time and space.
Source imaging procedures can then be applied to this reduced
This chapter focused on modern analysis techniques that con- data set only. Since the potential maps represent the real record-
vert the EEG to a functional neuroimaging modality. This ings and do not yet rely on models, a preprocessing of the data
translation from waveforms to images includes several process- based on the maps might be more prudent than directly con-
ing steps that need to be understood and performed properly. It verting the raw data to source images and perform all analysis in
starts with the appropriate spatial sampling of the scalp poten- the source space. Nevertheless, several interesting studies
tial field and ends with proper statistical evaluation of the appeared recently that showed the possibilities of analysis of the
reconstructed time series of electric activity in the brain. source waveforms. Most interesting are the applications of con-
Concerning the spatial sampling the recent literature nectivity analysis techniques to these source waveforms. They
strongly suggests that a high number of recording channels are allow to study causal interactions between different sources in
desirable to avoid undersampling of the spatial frequency of the the brain.
EEG. With modern technology this is not a limiting issue any- In summary, this chapter tried to illustrate that the temporal
more, and 64 to 128 channels of EEG are recorded in many labs. resolution of the EEG, together with the capability and reliabil-
EEG can nowadays easily be sampled from 200 and more elec- ity of modern source imaging algorithms, has converted the
trodes, with electrode nets that allow fast applications even in EEG to a fully fledged functional neuroimaging method that is
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

clinical routine. not secondary to, but instead is a perfect companion to fMRI
Reconstruction and visualization of the scalp potential maps and other neuroimaging methods.
are an important step in EEG source imaging. On the one hand,
it serves to detect map distortions due to artifacts that are invis- REFERENCES
ible on the EEG traces. On the other hand, EEG mapping allows
to statistically assess time points where map topographies 1. Baillet S, Mosher JC, Leahy RM. Electromagnetic brain mapping.
changed over time or between experimental conditions. By IEEE Signal Processing Mag. 2001;18:14–30.
physical laws, different map topographies indicate different 2. He B, Lian J. High-resolution spatio-temporal functional
configurations of the active sources in the brain. Analysis strate- neuroimaging of brain activity. Crit Rev Biomed Eng. 2002;30:
283–306.
gies that are based on the topography of the potential field also
3. Michel CM, Murray MM, Lantz G, et al. EEG source imaging. Clin
have the important advantage of being completely reference-
Neurophysiol. 2004;115:2195–2222.
free, which is not the case for the analysis of peaks and troughs 4. He B, Lian J. Electrophysiological neuroimaging: solving the EEG
of EEG or ERP waveforms, and also not for the analysis of inverse problem. In: He B, ed. Neural Engineering. Norwell, MA:
coherence or correlations between different electrode sites. Kluwer Academic Publishers; 2005:221–261.
Two main models are involved in EEG source imaging: the 5. Sanei S, Chambers J. EEG Source localization. In: Sanei S, Chambers
head model and the source model. In this chapter we not only J, eds. EEG Signal Processing. New Jersey: Wiley; 2007:197–218.

6. Michel CM, Koenig T, Brandeis D, et al., eds. Electrical Neuroimaging. 28. Lantz G, Grave de Peralta R, Spinelli L, et al. Epileptic source local-
Cambridge, MA: Cambridge University Press; 2009. ization with high density EEG: how many electrodes are needed?
7. Lehmann D. Multichannel topography of human alpha EEG fields. Clin Neurophysiol. 2003;114:63–69.
Electroencephalogr Clin Neurophysiol. 1971;31:439–449. 29. Luu P, Tucker DM, Englander R, et al. Localizing acute stroke-
8. Lehmann D, Skrandies W. Reference-free identification of compo- related EEG changes: assessing the effects of spatial undersam-
nents of checkerboard-evoked multichannel potential fields. pling. J Clin Neurophysiol. 2001;18:302–317.
Electroencephalogr Clin Neurophysiol. 1980;48:609–621. 30. Rush S, Driscoll DA. EEG electrode sensitivity—an application of
9. Lehmann D, Skrandies W. Spatial analysis of evoked potentials in reciprocity. IEEE Trans Biomed Eng. 1969;16:15–22.
man—a review. Prog Neurobiol. 1984;23:227–250. 31. Oostendorp TF, Delbeke J, Stegeman DF. The conductivity of the
10. Duffy FH, ed. Topographic Mapping of Brain Electrical Activity. human skull: results of in vivo and in vitro measurements. IEEE
Boston, MA: Butterworth; 1986. Trans Biomed Eng. 2000;47:1487–1492.
11. Fender DH. Source localization of brain electrical activity. In: 32. Hoekema R, Wieneke GH, Leijten FS, et al. Measurement of the
Gevins AS, Remont A, eds. Methods of Analysis of Brain Electrical conductivity of skull, temporarily removed during epilepsy sur-
and Magnetic Signals. Amsterdam: Elsevier; 1987:355–403. gery. Brain Topogr. 2003;16:29–38.
12. Helmholtz HLP. Ueber einige gesetze der vertheilung elektrischer 33. Lai Y, van Drongelen W, Ding L, et al. Estimation of in vivo human
ströme in körperlichen leitern mit anwendung aud die thierisch- brain-to-skull conductivity ratio from simultaneous extra- and
elektrischen versuche. Ann Phys Chem. 1853;9:211–233. intra-cranial electrical potential recordings. Clin Neurophysiol.
13. Michel CM, Brandeis D. Data acquisition and pre-processing stan- 2005;116:456–465.
dards for electrical neuroimaging. In: Michel CM, Koenig T, 34. Goncalves S, de Munck JC, Verbunt JP, et al. In vivo measurement
Brandeis D, et al., eds. Electrical Neuroimaging. Cambridge, MA: of the brain and skull resistivities using an EIT-based method and
Cambridge University Press; 2009;79–92. the combined analysis of SEF/SEP data. IEEE Trans Biomed Eng.
14. Tucker DM. Spatial sampling of head electrical fields: the geodesic 2003;50:1124–1128.
sensor net. Electroencephalogr Clin Neurophysiol. 1993;87:154–163. 35. Zhang Y, van Drongelen W, He B. Estimation of in vivo brain-
15. Michel CM, Lantz G, Spinelli L, et al. 128-Channel EEG source to-skull conductivity ratio in humans. Appl Phys Lett. 2006;89:
imaging in epilepsy: clinical yield and localization precision. J Clin 223903–223905.
Neurophysiol. 2004;21:71–83. 36. Fifer WP, Grieve PG, Grose-Fifer J, et al. High-density electroen-
16. Russell GS, Jeffrey EK, Poolman P, et al. Geodesic photogrammetry cephalogram monitoring in the neonate. Clin Perinatol. 2006;
for localizing sensor positions in dense-array EEG. Clin Neurophysiol. 33:679–691, vii.
2005;116:1130–1140. 37. Zhang X, van Drongelen W, Hecox KE, et al. High-resolution EEG:
17. Holmes MD. Dense array EEG: methodology and new hypothesis cortical potential imaging of interictal spikes. Clin Neurophysiol.
on epilepsy syndromes. Epilepsia. 2008;49:3–14. 2003;114:1963–1973.
18. Malmivuo JA, Suihko VE. Effect of skull resistivity on the spatial 38. Ding L, Worrell GA, Lagerlund TD, et al. 3D source localization of
resolutions of EEG and MEG. IEEE Trans Biomed Eng. 2004;51: interictal spikes in epilepsy patients with MRI lesions. Phys Med
1276–1280. Biol. 2006;51:4047–4062.
19. Li T-H, North G. Aliasing effects and sampling theorems of SRFs 39. Sperli F, Spinelli L, Seeck M, et al. EEG source imaging in paedi-
when sampled on a finite grid. Ann Inst Stat Math. 1996;49:341–354. atric epilepsy surgery: a new perspective in presurgical workup.
20. Grieve PG, Emerson RG, Isler JR, et al. Quantitative analysis of Epilepsia. 2006;47:981–990.
spatial sampling error in the infant and adult electroencephalo- 40. Ding L, Worrell GA, Lagerlund TD, et al. Ictal source analysis:
gram. Neuroimage. 2004;21:1260–1274. localization and imaging of causal interactions in humans.
21. Srinivasan R, Nunez PL, Tucker DM, et al. Spatial sampling and fil- Neuroimage. 2007;34:575–586.
tering of EEG with spline Laplacians to estimate cortical poten- 41. Ferree TC, Eriksen KJ, Tucker DM. Regional head tissue conduc-
tials. Brain Topogr. 1996;8:355–366. tivity estimation for improved EEG analysis. IEEE Trans Biomed
22. Srinivasan R, Tucker DM, Murias M. Estimating the spatial Eng. 2000;47:1584–1592.
Nyquist of the human EEG. Behav Res Methods Instrum Comput. 42. Gao N, Zhu SA, He B. Estimation of electrical conductivity distri-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

1998;30:8–19. bution within the human head from magnetic flux density meas-
23. Ryynänen OR, Hyttinen JA, Laarne PH, et al. Effect of electrode urement. Phys Med Biol. 2005;50:2675–2687.
density and measurement noise on the spatial resolution of cortical 43. Yuval-Greenberg S, Tomer O, Keren AS, et al. Transient induced
potential distribution. IEEE Trans Biomed Eng. 2004;51:1547–1554. gamma-band response in EEG as a manifestation of miniature
24. Ryynänen OR, Hyttinen JA, Malmivuo JA. Effect of measurement saccades. Neuron. 2008;58:429–441.
noise and electrode density on the spatial resolution of cortical 44. Perrin F, Pernier J, Bertrand O, et al. Spherical splines for scalp
potential distribution with different resistivity values for the skull. potential and current density mapping. Electroencephalogr Clin
IEEE Trans Biomed Eng. 2006;53:1851–1858. Neurophysiol. 1989;72:184–187.
25. Spitzer AR, Cohen LG, Fabrikant J, et al. A method for determin- 45. Fletcher EM, Kussmaul CL, Mangun GR. Estimation of interpola-
ing optimal interelectrode spacing for cerebral topographic map- tion errors in scalp topographic mapping. Electroencephalogr Clin
ping. Electroencephalogr Clin Neurophysiol. 1989;72:355–361. Neurophysiol. 1996;98:422–434.
26. Gevins A, Brickett P, Costales B, et al. Beyond topographic map- 46. Jung TP, Makeig S, Humphries C, et al. Removing electroen-
ping: towards functional–anatomical imaging with 124-channel cephalographic artifacts by blind source separation.
EEGs and 3-D MRIs. Brain Topogr. 1990;3:53–64. Psychophysiology. 2000;37:163–178.
27. Freeman WJ, Holmes MD, Burke BC, et al. Spatial spectra of scalp 47. Desmedt JE, Tomberg C, Noel P, et al. Beware of the average refer-
EEG and EMG from awake humans. Clin Neurophysiol. 2003;114: ence in brain mapping. Electroencephalogr Clin Neurophysiol Suppl.
1053–1068. 1990;41:22–27.

48. Tomberg C, Noel P, Ozaki I, et al. Inadequacy of the average ref- 68. Wackermann J. Beyond mapping: estimating complexity of mul-
erence for the topographic mapping of focal enhancements of tichannel EEG recordings. Acta Neurobiol Exp (Wars). 1996;56:
brain potentials. Electroencephalogr Clin Neurophysiol. 1990;77: 197–208.
259–265. 69. Kayser J, Tenke CE. Trusting in or breaking with convention:
49. Pascual-Marqui RD, Lehmann D. Comparison of topographic towards a renaissance of principal components analysis in electro-
maps and the reference electrode: comments on two papers by physiology. Clin Neurophysiol. 2005;116:1747–1753.
Desmedt and collaborators. Electroencephalogr Clin Neurophysiol. 70. Spencer KM, Dien J, Donchin E. Spatiotemporal analysis of the
1993;88:530–531, 534–536. late ERP responses to deviant stimuli. Psychophysiology. 2001;38:
50. Pascual-Marqui RD, Lehmann D. Topographic maps, source local- 343–358.
ization inference, and the reference electrode: comments on a paper 71. Kayser J, Tenke CE. Optimizing PCA methodology for ERP com-
by Desmedt et al. Electroencephalogr Clin Neurophysiol. 1993;88: ponent identification and measurement: theoretical rationale and
532–536. empirical evaluation. Clin Neurophysiol. 2003;114:2307–2325.
51. Junghofer M, Elbert T, Tucker DM, et al. The polar average refer- 72. Kayser J, Tenke CE. Principal components analysis of Laplacian
ence effect: a bias in estimating the head surface integral in EEG waveforms as a generic method for identifying ERP generator pat-
recording. Clin Neurophysiol. 1999;110:1149–1155. terns: II. Adequacy of low-density estimates. Clin Neurophysiol.
52. Lehmann D, Skrandies W. Reference-free identification of com- 2006;117:369–380.
ponents of checkerboard-evoked multichannel potential fields. 73. Pourtois G, Deplanque S, Michel C, et al. Beyond the conventional
Electroencephalogr Clin Neurophysiol. 1980;48:609–621. event-related brain potential (ERP): exploring the time-course of
53. Geselowitz DB. The zero of potential. IEEE Eng Med Biol Mag. visual emotion processing using topographic and principal com-
1998;17:128–132. ponent analyses. Brain Topogr. 2008;20:265–277.
54. Murray MM, Brunet D, Michel CM. Topographic ERP analyses: a 74. Makeig S, Jung TP, Bell AJ, et al. Blind separation of auditory
step-by-step tutorial review. Brain Topogr. 2008;20:249–264. event-related brain responses into independent components. Proc
55. Perrin F, Pernier J, Bertrand O, et al. Mapping of scalp potentials by Natl Acad Sci U S A. 1997;94:10979–10984.
surface spline interpolation. Electroencephalogr Clin Neurophysiol. 75. Benar C, Aghakhani Y, Wang Y, et al. Quality of EEG in simultane-
1987;66:75–81. ous EEG–fMRI for epilepsy. Clin Neurophysiol. 2003;114:569–580.
56. Nunez PL, Silberstein RB, Cadusch PJ, et al. A theoretical and exper- 76. Briselli E, Garreffa G, Bianchi L, et al. An independent component
imental study of high resolution EEG based on surface Laplacians analysis-based approach on ballistocardiogram artifact removing.
and cortical imaging. Electroencephalogr Clin Neurophysiol. 1994; Magn Reson Imaging. 2006;24:393–400.
90:40–57. 77. Mantini D, Perrucci MG, Cugini S, et al. Complete artifact removal
57. Hjorth B. An on-line transformation of EEG scalp potentials into for EEG recorded during continuous fMRI using independent
orthogonal source derivations. Electroencephalogr Clin Neurophysiol. component analysis. Neuroimage. 2007;34:598–607.
1975;39:526–530. 78. Debener S, Strobel A, Sorger B, et al. Improved quality of auditory
58. Perrin F, Bertrand O, Pernier J. Scalp current density mapping: event-related potentials recorded simultaneously with 3-T fMRI:
value and estimation from potential data. IEEE Trans Biomed Eng. removal of the ballistocardiogram artefact. Neuroimage. 2007;34:
1987;34:283–288. 587–597.
59. He B, Lian J, Li G. High-resolution EEG: a new realistic geometry 79. Grouiller F, Vercueil L, Krainik A, et al. A comparative study of dif-
spline Laplacian estimation technique. Clin Neurophysiol. 2001;112: ferent artefact removal algorithms for EEG signals acquired during
845–852. functional MRI. Neuroimage. 2007;38:124–137.
60. Nunez PL, Srinivasan R. Electric Fields of the Brain: The 80. Debener S, Mullinger KJ, Niazy RK, et al. Properties of the ballisto-
Neurophysics of EEG. 2nd ed. New York: Oxford University Press; cardiogram artefact as revealed by EEG recordings at 1.5, 3 and 7 T
2006. static magnetic field strength. Int J Psychophysiol. 2008;67:189–199.
61. Skrandies W. The effect of stimulation frequency and retinal stim- 81. Makeig S, Westerfield M, Townsend J, et al. Functionally independent
ulus location on visual evoked potential topography. Brain Topogr. components of early event-related potentials in a visual spatial atten-
2007;20:15–20. tion task. Philos Trans R Soc Lond B Biol Sci. 1999;354:1135–1144.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

62. Brandeis D, Naylor H, Halliday R, et al. Scopolamine effects on 82. Lehmann D, Ozaki H, Pal I. EEG alpha map series: brain micro-
visual information processing, attention, and event-related poten- states by space-oriented adaptive segmentation. Electroencephalogr
tial map latencies. Psychophysiology. 1992;29:315–336. Clin Neurophysiol. 1987;67:271–288.
63. Vaughan HGJ. The neural origins of human event-related poten- 83. Lehmann D. Brain electric fields and brain functional states. In:
tials. Ann N Y Acad Sci. 1982;388:125–138. Friedrich R, Wunderlin A, eds. Evolution of Dynamical Structures
64. Picton TW, Bentin S, Berg P, et al. Guidelines for using human in Complex Systems. Berlin: Springer; 1992:235–248.
event-related potentials to study cognition: recording standards 84. Strik WK, Lehmann D. Data determined window size and space-
and publication criteria. Psychophysiology. 2000;37:127–152. oriented segmentation of spontaneous EEG map series.
65. Koenig T, Wackermann J. Overview of analytical approaches. Electroencephalogr Clin Neurophysiol. 1993;87:169–174.
In: Michel CM, Koenig T, Brandeis D, et al., eds. Electrical 85. Wackerman J, Lehmann D, Michel CM, et al. Adaptive segmenta-
Neuroimaging. Cambridge, MA: Cambridge University Press; 2009: tion of spontaneous EEG map series into spatially defined
93–109. microstates. Int J Psychophysiol. 1993;14:269–283.
66. Skrandies W. Data reduction of multichannel fields: global field 86. Koenig T, Prichep L, Lehmann D, et al. Millisecond by millisecond,
power and principal component analysis. Brain Topogr. 1989;2: year by year: normative EEG microstates and developmental
73–80. stages. Neuroimage. 2002;16:41–48.
67. John ER, Easton P, Prichep LS, et al. Standardized varimax descrip- 87. Pascual-Marqui RD, Michel CM, Lehmann D. Segmentation of
tors of event related potentials: basic considerations. Brain Topogr. brain electrical activity into microstates: model estimation and
1993;6:143–162. validation. IEEE Trans Biomed Eng. 1995;42:658–665.

88. Lehmann D, Pascual-Marqui R, Michel CM. EEG microstates. 110. Murray MM, Camen C, Gonzalez Andino SL, et al. Rapid brain dis-
Scholarpedia. 2009;4:7632. crimination of sounds of objects. J Neurosci. 2006;26:1293–1302.
89. Baars BJ. The conscious access hypothesis: origins and recent evi- 111. De Santis L, Spierer L, Clarke S, et al. Getting in touch: segregated
dence. Trends Cogn Sci. 2002;6:47–52. somatosensory what and where pathways in humans revealed by
90. Dehaene S, Sergent C, Changeux JP. A neuronal network model link- electrical neuroimaging. Neuroimage. 2007;37:890–903.
ing subjective reports and objective physiological data during con- 112. Schnider A, Mohr C, Morand S, et al. Early cortical response to
scious perception. Proc Natl Acad Sci U S A. 2003;100:8520–8525. behaviorally relevant absence of anticipated outcomes: a human
91. Grossberg S. The complementary brain: unifying brain dynamics event-related potential study. Neuroimage. 2007;35:1348–1355.
and modularity. Trends Cogn Sci. 2000;4:233–246. 113. Thierry G, Martin CD, Downing P, et al. Controlling for inter-
92. Bressler SL, Tognoli E. Operational principles of neurocognitive stimulus perceptual variance abolishes N170 face selectivity. Nat
networks. Int J Psychophysiol. 2006;60:139–148. Neurosci. 2007;10:505–511.
93. Fingelkurts AA. Timing in cognition and EEG brain dynamics: 114. Arzy S, Molnar-Szakacs I, Blanke O. Self in time: imagined self-
discreteness versus continuity. Cogn Process. 2006;7:135–162. location influences neural activity related to mental time travel. J
94. Strik WK, Dierks T, Becker T, et al. Larger topographical variance Neurosci. 2008;28:6502–6507.
and decreased duration of brain electric microstates in depres- 115. Spierer L, Murray MM, Tardif E, et al. The path to success in audi-
sion. J Neural Transm Gen Sect. 1995;99:213–222. tory spatial discrimination: electrical neuroimaging responses
95. Strik WK, Chiaramonti R, Muscas GC, et al. Decreased EEG within the supratemporal plane predict performance outcome.
microstate duration and anteriorisation of the brain electrical Neuroimage. 2008;41:493–503.
fields in mild and moderate dementia of the Alzheimer type. 116. Kavanagh RN, Darcey TM, Lehmann D, et al. Evaluation of
Psychiatry Res. 1997;75:183–191. methods for three-dimensional localization of electrical sources
96. Koenig T, Prichep L, Dierks T, et al. Decreased EEG synchroniza- in the human brain. IEEE Trans Biomed Eng. 1978;25:421–429.
tion in Alzheimer’s disease and mild cognitive impairment. 117. Scherg M, von Cramon D. A new interpretation of the generators
Neurobiol Aging. 2005;26:165–171. of BAEP waves I–V: results of a spatio-temporal dipole model.
97. Koenig T, Lehmann D, Merlo MC, et al. A deviant EEG brain Electroencephalogr Clin Neurophysiol. 1985;62:290–299.
microstate in acute, neuroleptic-naive schizophrenics at rest. Eur 118. He B, Musha T, Okamoto Y, et al. Electric dipole tracing in the
Arch Psychiatry Clin Neurosci. 1999;249:205–211. brain by means of the boundary element method and its accu-
98. Lehmann D, Faber PL, Galderisi S, et al. EEG microstate duration racy. IEEE Trans Biomed Eng. 1987;34:406–414.
and syntax in acute, medication-naive, first-episode schizophre- 119. Cuffin BN, Cohen D, Yunokuchi K, et al. Tests of EEG localiza-
nia: a multi-center study. Psychiatry Res. 2005;138:141–156. tion accuracy using implanted sources in the human brain. Ann
99. Lantz G, Michel CM, Seeck M, et al. Space-oriented segmenta- Neurol. 1991;29:132–138.
tion and 3-dimensional source reconstruction of ictal EEG pat- 120. Cuffin BN. A method for localizing EEG sources in realistic head
terns. Clin Neurophysiol. 2001;112:688–697. models. IEEE Trans Biomed Eng. 1995;42:68–71.
100. Lehmann D, Wackermann J, Michel CM, et al. Space-oriented EEG 121. Nolte G, Curio G. Current multipole expansion to estimate lat-
segmentation reveals changes in brain electric field maps under the eral extent of neuronal activity: a theoretical analysis. IEEE Trans
influence of a nootropic drug. Psychiatry Res. 1993;50:275–282. Biomed Eng. 2000;47:1347–1355.
101. Kinoshita T, Strik WK, Michel CM, et al. Microstate segmenta- 122. Jerbi K, Baillet S, Mosher JC, et al. Localization of realistic corti-
tion of spontaneous multichannel EEG map series under cal activity in MEG using current multipoles. Neuroimage. 2004;22:
diazepam and sulpiride. Pharmacopsychiatry. 1995;28:51–55. 779–793.
102. Michel CM, Grave de Peralta R, Lantz G, et al. Spatio-temporal 123. Pascual-Marqui RD, Michel CM, Lehmann D. Low resolution
EEG analysis and distributed source estimation in presurgical electromagnetic tomography: a new method for localizing elec-
epilepsy evaluation. J Clin Neurophysiol. 1999;16:225–238. trical activity in the brain. Int J Psychophysiol. 1994;18:49–65.
103. Michel CM, Thut G, Morand S, et al. Electric source imaging of 124. Dale AM, Sereno MI. Improved localization of cortical activity by
human brain functions. Brain Res Brain Res Rev. 2001;36: combining EEG and MEG with MRI cortical surface reconstruc-
108–118. tion: a linear approach. J Cogn Neurosci. 1993;5:162–176.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

104. Pegna AJ, Khateb A, Spinelli L, et al. Unravelling the cerebral 125. He B, Yao D, Lian J, et al. An equivalent current source model and
dynamics of mental imagery. Hum Brain Mapp. 1997;5:410–421. Laplacian weighted minimum norm current estimates of brain
105. Arzy S, Mohr C, Michel CM, et al. Duration and not strength of electrical activity. IEEE Trans Biomed Eng. 2002;49:277–288.
activation in temporo-parietal cortex positively correlates with 126. Grave de Peralta Menendez R, Gonzalez Andino SL, Morand S, et
schizotypy. Neuroimage. 2007;35:326–333. al. Imaging the electrical activity of the brain: ELECTRA. Hum
106. Lascano AM, Brodbeck V, Lalive PH, et al. Increasing the diag- Brain Mapp. 2000;9:1–12.
nostic value of evoked potentials in multiple sclerosis by quanti- 127. Plonsey R. Bioelectric Phenomena. New York, NY: McGraw-Hill; 1969.
tative topographic analysis of multichannel recordings. J Clin 128. Wilson FN, Bayley RH. The electric field of an eccentric dipole in
Neurophysiol. 2009;26:316–325. a homogeneous spherical conducting medium. Circulation. 1950;
107. Murray MM, Michel CM, Grave de Peralta R, et al. Rapid dis- 1:84–92.
crimination of visual and multisensory memories revealed by 129. Wang Y, He B. A computer simulation study of cortical imaging
electrical neuroimaging. Neuroimage. 2004;21:125–135. from scalp potentials. IEEE Trans Biomed Eng. 1998;45:724–735.
108. Ortigue S, Michel CM, Murray MM, et al. Electrical neuroimag- 130. Cuffin BN, Cohen D. Comparison of the magnetoencephalogram
ing reveals early generator modulation to emotional words. and electroencephalogram. Electroencephalogr Clin Neurophysiol.
Neuroimage. 2004;21:1242–1251. 1979;47:132–146.
109. Overney LS, Michel CM, Harris IM, et al. Cerebral processes in 131. Sun M. An efficient algorithm for computing multishell spheri-
mental transformations of body parts: recognition prior to rota- cal volume conductor models in EEG dipole source localization.
tion. Brain Res Cogn Brain Res. 2005;25:722–734. IEEE Trans Biomed Eng. 1997;44:1243–1252.

132. Spinelli L, Andino SG, Lantz G, et al. Electromagnetic inverse 151. Bai X, He B. Estimation of number of independent brain electric
solutions in anatomically constrained spherical head models. sources from the scalp EEGs. IEEE Trans Biomed Eng. 2006;53:
Brain Topogr. 2000;13:115–125. 1883–1892.
133. Hämäläinen M, Sarvas J. Realistic conductor geometry model of 152. Ebersole JS. Non-invasive localization of the epileptogenic focus by
the human head for interpretation of neuromagnetic data. IEEE EEG dipole modeling. Acta Neurol Scand Suppl. 1994;152:20–28.
Trans Biomed Eng. 1989;36:165–171. 153. Lantz G, Ryding E, Rosén I. Three dimensional localisation of
134. Meijs JW, Weier OW, Peters MJ, et al. On the numerical accuracy interictal epileptiform activity with dipole analysis: comparison
of the boundary element method. IEEE Trans Biomed Eng. with intracranial recordings and SPECT findings. J Epilepsy.
1989;36:1038–1049. 1994;7:117–129.
135. Yan Y, Nunez PL, Hart RT. Finite-element model of human head: 154. Ebersole JS, Hawes-Ebersole S. Clinical application of dipole
scalp potentials due to dipole sources. Med Biol Eng Comput. models in the localization of epileptiform activity. J Clin
1991;29:475–481. Neurophysiol. 2007;24:120–129.
136. Zhang YC, Zhu SA, He B. A second-order finite element algo- 155. Cuffin BN. EEG localization accuracy improvements using realis-
rithm for solving the three-dimensional EEG forward problem. tically shaped head models. IEEE Trans Biomed Eng. 1996;43:
Phys Med Biol. 2004;49:2975–2987. 299–303.
137. Wolters CH, Anwander A, Tricoche X, et al. Influence of tissue con- 156. Waberski TD, Buchner H, Lehnertz K, et al. Properties of
ductivity anisotropy on EEG/MEG field and return current com- advanced headmodelling and source reconstruction for the local-
putation in a realistic head model: a simulation and visualization ization of epileptiform activity. Brain Topogr. 1998;10:283–290.
study using high-resolution finite element modeling. Neuroimage. 157. Benar CG, Gotman J. Modeling of post-surgical brain and skull
2006;30:813–826. defects in the EEG inverse problem with the boundary element
138. Zhang Y, Ding L, van Drongelen W, et al. A cortical potential imag- method. Clin Neurophysiol. 2002;113:48–56.
ing study from simultaneous extra- and intracranial electrical 158. Martinez A, Anllo-Vento L, Sereno MI, et al. Involvement of stri-
recordings by means of the finite element method. Neuroimage. ate and extrastriate visual cortical areas in spatial attention. Nat
2006;31:1513–1524. Neurosci. 1999;2:364–369.
139. Zhang Y, van Drongelen W, Kohrman M, et al. Three-dimen- 159. Northoff G, Richter A, Gessner M, et al. Functional dissociation
sional brain current source reconstruction from intra-cranial between medial and lateral prefrontal cortical spatiotemporal
ECoG recordings. Neuroimage. 2008;42:683–695. activation in negative and positive emotions: a combined
140. Lee WH, Liu Z, Mueller BA, et al. Influence of white matter fMRI/MEG study. Cereb Cortex. 2000;10:93–107.
anisotropic conductivity on EEG source localization: compari- 160. Lantz G, Holub M, Ryding E, et al. Simultaneous intracranial and
son to fMRI in human primary visual cortex. Clin Neurophysiol. extracranial recording of interictal epileptiform activity in
2009;120:2071–2081. patients with drug resistant partial epilepsy: patterns of conduc-
141. Withey DJ, Koles ZJ. A review of medical image segmentation: tion and results from dipole reconstructions. Electroencephalogr
methods and available software. Int J Bioelectromagnetism. Clin Neurophysiol. 1996;99:69–78.
2008;10:125–148. 161. Huppertz HJ, Hof E, Klisch J, et al. Localization of interictal delta
142. Basser PJ, Mattiello J, LeBihan D. MR diffusion tensor spec- and epileptiform EEG activity associated with focal epileptogenic
troscopy and imaging. Biophys J. 1994;66:259–267. brain lesions. Neuroimage. 2001;13:15–28.
143. Tuch DS, Wedeen VJ, Dale AM, et al. Conductivity tensor map- 162. Cakmur R, Towle VL, Mullan JF, et al. Intra-operative localiza-
ping of the human brain using diffusion tensor MRI. Proc Natl tion of sensorimotor cortex by cortical somatosensory evoked
Acad Sci U S A. 2001;98:11697–11701. potentials: from analysis of waveforms to dipole source model-
144. Wang K, Zhu S, Mueller BA, et al. A new method to derive white ing. Acta Neurochir (Wien). 1997;139:1117–1124 [discussion
matter conductivity from diffusion tensor MRI. IEEE Trans Biomed 1115–1124].
Eng. 2008;55:2481–2486. 163. Mine S, Oka N, Yamaura A, et al. Presurgical functional localiza-
145. Homma S, Musha T, Nakajima Y, et al. Location of electric current tion of primary somatosensory cortex by dipole tracing method of
sources in the human brain estimated by the dipole tracing method scalp–skull–brain head model applied to somatosensory evoked
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

of the scalp–skull–brain (SSB) head model. Electroencephalogr Clin potential. Electroencephalogr Clin Neurophysiol. 1998;108:226–233.
Neurophysiol. 1994;91:374–382. 164. Mosher JC, Lewis PS, Leahy RM. Multiple dipole modeling and
146. Roth BJ, Ko D, von Albertini-Carletti IR, et al. Dipole localization localization from spatio-temporal MEG data. IEEE Trans Biomed
in patients with epilepsy using the realistically shaped head Eng. 1992;39:541–557.
model. Electroencephalogr Clin Neurophysiol. 1997;102:159–166. 165. Sekihara K, Poeppel D, Marantz A, et al. Noise covariance incor-
147. Bai X, Towle VL, He EJ, et al. Evaluation of cortical imaging tech- porated MEG-MUSIC algorithm: a method for multiple-dipole
niques based on somatosensory evoked potentials. Conf Proc estimation tolerant of the influence of background brain activity.
IEEE Eng Med Biol Soc. 2006;1:1000–1001. IEEE Trans Biomed Eng. 1997;44:839–847.
148. Diekmann V, Becker W, Jurgens R, et al. Localisation of epileptic 166. Mosher JC, Leahy RM. Source localization using recursively
foci with electric, magnetic and combined electromagnetic mod- applied and projected (RAP) MUSIC. IEEE Trans Signal Process.
els. Electroencephalogr Clin Neurophysiol. 1998;106:297–313. 1999;47:332–340.
149. Fuchs M, Wagner M, Wischmann HA, et al. Improving source 167. Xu XL, Xu B, He B. An alternative subspace approach to EEG
reconstructions by combining bioelectric and biomagnetic data. dipole source localization. Phys Med Biol. 2004;49:327–343.
Electroencephalogr Clin Neurophysiol. 1998;107:93–111. 168. Ding L, He B. Spatio-temporal EEG source localization using a
150. Fletcher DJ, Amir A, Jewett DL, et al. Improved method for com- three-dimensional subspace FINE approach in a realistic geometry
putation of potentials in a realistic head shape model. IEEE Trans inhomogeneous head model. IEEE Trans Biomed Eng. 2006;
Biomed Eng. 1995;42:1094–1104. 53:1732–1739.

169. Ding L, Zhang N, Chen W, et al. Three-dimensional imaging of 188. Le J, Gevins A. Method to reduce blur distortion from EEG’s using
complex neural activation in humans from EEG. IEEE Trans a realistic head model. IEEE Trans Biomed Eng. 1993;40:517–528.
Biomed Eng. 2009;56:1980–1988. 189. Srebro R, Oguz RM, Hughlett K, et al. Estimating regional brain
170. van Veen BD, van Drongelen W, Yuchtman M, et al. Localization of activity from evoked potential fields on the scalp. IEEE Trans
brain electrical activity via linearly constrained minimum variance Biomed Eng. 1993;40:509–516.
spatial filtering. IEEE Trans Biomed Eng. 1997;44:867–880. 190. Gevins A, Le J, Martin NK, et al. High resolution EEG: 124-chan-
171. Sekihara K, Nagarajan SS, Poeppel D, et al. Reconstructing spatio- nel recording, spatial deblurring and MRI integration methods.
temporal activities of neural sources using an MEG vector beam- Electroencephalogr Clin Neurophysiol. 1994;90:337–358.
former technique. IEEE Trans Biomed Eng. 2001;48:760–771. 191. He B, Wang Y, Pak S, et al. Cortical source imaging from scalp
172. Grave de Peralta R, Gonzalez S, Lantz G, et al. Noninvasive local- electroencephalograms. Med Biol Eng Comput. 1996;34:
ization of electromagnetic epileptic activity. I. Method descrip- 257–258.
tions and simulations. Brain Topogr. 2001;14:131–137. 192. Babiloni F, Babiloni C, Carducci F, et al. High resolution EEG: a new
173. Lantz G, Spinelli L, Menendez RG, et al. Localization of distrib- model-dependent spatial deblurring method using a realistically-
uted sources and comparison with functional MRI. Epileptic shaped MR-constructed subject’s head model. Electroencephalogr
Disord. 2001;(special issue):45–58. Clin Neurophysiol. 1997;102:69–80.
174. Lantz G, Spinelli L, Seeck M, et al. Propagation of interictal 193. He B, Wang Y, Wu D. Estimating cortical potentials from scalp
epileptiform activity can lead to erroneous source localizations: a EEG’s in a realistically shaped inhomogeneous head model by
128 channel EEG mapping study. J Clin Neurophysiol. 2003; means of the boundary element method. IEEE Trans Biomed Eng.
20:311–319. 1999;46:1264–1268.
175. Hämäläinen MS, Ilmoniemi RJ. Interpreting measured magnetic 194. He B, Lian J, Spencer KM, et al. A cortical potential imaging
fields of the brain: estimation of current distributions. Helsinki: analysis of the P300 and novelty P3 components. Hum Brain
Helsinki University of Technology; 1984:28 pp. Mapp. 2001;12:120–130.
176. Wang JZ, Williamson SJ, Kaufman L. Magnetic source images deter- 195. He B, Zhang X, Lian J, et al. Boundary element method-based
mined by a lead-field analysis: the unique minimum-norm least- cortical potential imaging of somatosensory evoked potentials
squares estimation. IEEE Trans Biomed Eng. 1992;39:665–675. using subjects’ magnetic resonance images. Neuroimage. 2002;16:
177. Grave de Peralta Menendez R, Murray MM, Michel CM, et al. 564–576.
Electrical neuroimaging based on biophysical constraints. 196. Ollikainen JO, Vauhkonen M, Karjalainen PA, et al. A new com-
Neuroimage. 2004;21:527–539. putational approach for cortical imaging. IEEE Trans Med
178. Dale AM, Liu AK, Fischl BR, et al. Dynamic statistical parametric Imaging. 2001;20:325–332.
mapping: combining fMRI and MEG for high-resolution imag- 197. Towle VL, Cohen S, Alperin N, et al. Displaying electrocortico-
ing of cortical activity. Neuron. 2000;26:55–67. graphic findings on gyral anatomy. Electroencephalogr Clin
179. Pascual-Marqui RD, Esslen M, Kochi K, et al. Functional imaging Neurophysiol. 1995;94:221–228.
with low-resolution brain electromagnetic tomography 198. He B, Liu Z. Multimodal functional neuroimaging: integrating
(LORETA): a review. Methods Find Exp Clin Pharmacol. 2002; functional MRI and EEG/MEG. IEEE Rev Biomed Eng.
24(suppl C):91–95. 2008:23–40.
180. Pascual Marqui RD, Sekihara K, Brandeis D, et al. Imaging the 199. Ahlfors SP, Simpson GV, Dale AM, et al. Spatiotemporal activity
electrical neuronal generators of EEG/MEG. In: Michel CM, of a cortical network for processing visual motion revealed by
Koenig T, Brandeis D, et al., eds. Electrical Neuroimaging. MEG and fMRI. J Neurophysiol. 1999;82:2545–2555.
Cambridge, MA: Cambridge University Press; 2009;49–78. 200. Fujimaki N, Hayakawa T, Nielsen M, et al. An fMRI-constrained
181. Gorodnitsky IF, George JS, Rao BD. Neuromagnetic source imag- MEG source analysis with procedures for dividing and grouping
ing with FOCUSS: a recursive weighted minimum norm algo- activation. Neuroimage. 2002;17:324–343.
rithm. Electroencephalogr Clin Neurophysiol. 1995;95:231–251. 201. Liu AK, Belliveau JW, Dale AM. Spatiotemporal imaging of
182. Yao D, He B. A self-coherence enhancement algorithm and its human brain activity using functional MRI constrained magne-
application to enhancing three-dimensional source estimation toencephalography data: Monte Carlo simulations. Proc Natl
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

from EEGs. Ann Biomed Eng. 2001;29:1019–1027. Acad Sci U S A. 1998;95:8945–8950.

183. Matsuura K, Okabe Y. Selective minimum-norm solution of the 202. Wagner M, Fuchs M, Kastner J. fMRI-constrained dipole fits
biomagnetic inverse problem. IEEE Trans Biomed Eng. 1995;42: and current density reconstructions. Proceedings of the 12th
608–615. International Conference on Biomagnetism, Helsinki, 2000:
184. Uutela K, Hamalainen M, Somersalo E. Visualization of magne- 785–788.
toencephalographic data using minimum current estimates. 203. Babiloni F, Babiloni C, Carducci F, et al. Multimodal integration
Neuroimage. 1999;10:173–180. of high-resolution EEG and functional magnetic resonance
185. Huang MX, Dale AM, Song T, et al. Vector-based spatial–tempo- imaging data: a simulation study. Neuroimage. 2003;19:1–15.
ral minimum L1-norm solution for MEG. Neuroimage. 204. Babiloni F, Cincotti F, Babiloni C, et al. Estimation of the cortical
2006;31:1025–1037. functional connectivity with the multimodal integration of high-
186. Ding L, He B. Sparse source imaging in electroencephalography resolution EEG and fMRI data by directed transfer function.
with accurate field modeling. Hum Brain Mapp. 2008;29: Neuroimage. 2005;24:118–131.
1053–1067. 205. Phillips C, Mattout J, Rugg MD, et al. An empirical Bayesian solu-
187. Sidman RD, Vincent DJ, Smith DB, et al. Experimental tests of tion to the source reconstruction problem in EEG. Neuroimage.
the cortical imaging technique—applications to the response to 2005;24:997–1011.
median nerve stimulation and the localization of epileptiform 206. Liu Z, Ding L, He B. Integration of EEG/MEG with MRI and
discharges. IEEE Trans Biomed Eng. 1992;39:437–444. fMRI. IEEE Eng Med Biol Mag. 2006;25:46–53.

207. Liu Z, He B. FMRI–EEG integrated cortical source imaging by 221. Otte A, Halsband U. Brain imaging tools in neurosciences. J
use of time-variant spatial constraints. Neuroimage. 2008;39: Physiol Paris. 2006;99:281–292.
1198–1214. 222. Adey WR, Walter DO, Hendrix CE. Computer techniques in cor-
208. Mattout J, Phillips C, Penny WD, et al. MEG source localization relation and spectral analyses of cerebral slow waves during dis-
under multiple constraints: an extended Bayesian framework. criminative behavior. Exp Neurol. 1961;3:501–524.
Neuroimage. 2006;30:753–767. 223. Gevins AS, Cutillo BA, Bressler SL, et al. Event-related covari-
209. Felleman DJ, Van Essen DC. Distributed hierarchical processing ances during a bimanual visuomotor task. II. Preparation and
in the primate cerebral cortex. Cereb Cortex. 1991;1:1–47. feedback. Electroencephalogr Clin Neurophysiol. 1989;74:147–160.
210. Laufs H, Holt JL, Elfont R, et al. Where the BOLD signal goes 224. Lachaux JP, Rodriguez E, Martinerie J, et al. Measuring phase
when alpha EEG leaves. Neuroimage. 2006;31:1408–1418. synchrony in brain signals. Hum Brain Mapp. 1999;8:194–208.
211. Tyvaert L, Levan P, Grova C, et al. Effects of fluctuating physio- 225. Urbano A, Babiloni C, Onorati P, et al. Dynamic functional cou-
logical rhythms during prolonged EEG–fMRI studies. Clin pling of high resolution EEG potentials related to unilateral
Neurophysiol. 2008;119:2762–2774. internally triggered one-digit movements. Electroencephalogr
212. Laufs H, Kleinschmidt A, Beyerle A, et al. EEG-correlated fMRI Clin Neurophysiol. 1998;106:477–487.
of human alpha activity. Neuroimage. 2003;19:1463–1476. 226. Brovelli A, Ding M, Ledberg A, et al. Beta oscillations in a large-
213. Laufs H, Krakow K, Sterzer P, et al. Electroencephalographic sig- scale sensorimotor cortical network: directional influences
natures of attentional and cognitive default modes in sponta- revealed by Granger causality. Proc Natl Acad Sci U S A.
neous brain activity fluctuations at rest. Proc Natl Acad Sci U S A. 2004;101:9849–9854.
2003;100:11053–11058. 227. Strogatz SH. Exploring complex networks. Nature. 2001;410:
214. Laufs H, Hamandi K, Salek-Haddadi A, et al. Temporal lobe 268–276.
interictal epileptic discharges affect cerebral activity in “default 228. Astolfi L, Cincotti F, Mattia D, et al. Estimation of the effective
mode” brain regions. Hum Brain Mapp. 2007;28:1023–1032. and functional human cortical connectivity with structural equa-
215. Groening K, Brodbeck V, Moeller F, et al. Combination of tion modeling and directed transfer function applied to high-res-
EEG–fMRI and EEG source analysis improves interpretation of olution EEG. Magn Reson Imaging. 2004;22:1457–1470.
spike-associated activation networks in paediatric pharmacore- 229. Astolfi L, Cincotti F, Babiloni C, et al. Estimation of the cortical
sistant focal epilepsies. Neuroimage. 2009;46:827–833. connectivity by high-resolution EEG and structural equation
216. Vulliemoz S, Thornton R, Rodionov R, et al. The spatio-tempo- modeling: simulations and application to finger tapping data.
ral mapping of epileptic networks: combination of EEG–fMRI IEEE Trans Biomed Eng. 2005;52:757–768.
and EEG source imaging. Neuroimage. 2009;46:834–843. 230. Wilke C, Ding L, He B. Estimation of time-varying connectivity
217. Brodbeck V, Lascano AM, Spinelli L, et al. Accuracy of EEG patterns through the use of an adaptive directed transfer func-
source imaging of epileptic spikes in patients with large brain tion. IEEE Trans Biomed Eng. 2008;55:2557–2564.
lesions. Clin Neurophysiol. 2009;120:679–685. 231. Wilke C, van Drongelen W, Kohrman M, et al. Identification of
218. Ioannides AA. Dynamic functional connectivity. Curr Opin epileptogenic foci from causal analysis of ECoG interictal spike
Neurobiol. 2007;17:161–170. activity. Clin Neurophysiol. 2009;120:1449–1456.
219. McIntosh AR, Gonzalez-Lima F. Structural equation modeling 232. Wilke C, van Drongelen W, Kohrman M, et al. Neocortical
and its application to network analysis in functional brain imag- seizure foci localization by means of a directed transfer function
ing. Hum Brain Mapp. 1994;2:2–22. method. Epilepsia. 2009;120:1449–1456.
220. Arthurs OJ, Donovan T, Spiegelhalter DJ, et al. Intracortically dis-
tributed neurovascular coupling relationships within and between
human somatosensory cortices. Cereb Cortex. 2007;17:661–668.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
CHAPTER
Computer-Assisted EEG Pattern
Recognition and Diagnostic Systems
FERNANDO H. LOPES DA SILVA
56
I
f someone would compare older editions of this chapter treatment of pattern recognition theories, the reader is referred
with the present one, he or she would notice how the field to the classic books of Duda and Hart (1), Mendel and Fu (2),
of computerized diagnostic systems has rapidly evolved in and Tou and Gonzalez (3), and to the review of Demartini and
the last decade that is reflected in the emergence of new algo- Vincent-Carrefour (4) that deals with the specific field of EEG.
rithms every year, while some older ones have become obsolete.
Nonetheless, much of the older literature continues to be valid FEATURE EXTRACTION:
inasmuch as it illustrates basic concepts and may guide new SPECIFIC PROBLEMS
researchers to find their way in this frontier between clinical
neurophysiology and the technology of applied signal analysis. The usefulness of any EEG analysis method depends, to a large
This kind of review of relatively old literature may help to avoid extent, on the choice of the set of features that is relevant to
each generation trying to invent the wheel again. Here we pay answer the question being investigated. In this section we con-
attention also to recent developments and new trends. We sider, first, the main types of features used in EEG analysis in
should note that some new algorithms are simple variants of general terms; then we examine how these features can be incor-
previous approaches. Very often the performance of the new porated into EEG classification systems in specialized cases.
ones is not assessed with respect to the older versions that leads
to some lack of transparency in this field. Time-Domain Analysis Methods
A common denominator of the field of computer-assisted Methods aiming at extracting time-domain EEG features were
EEG diagnostic systems is the application of pattern recogni- used in the early period of EEG quantification but are less used
tion methods. The latter constitute a general class of procedures recently. The main features used in this context are derived
applicable in a variety of scientific areas. The first operation in from EEG amplitude analysis: the features ordinarily chosen are
EEG analysis is to define a pattern, that is, to choose a set of fea- mean (m), standard deviation (), skewness, kurtosis, and coef-
tures that are potentially important in identifying the phenom- ficient of variation [(/m)  100] of EEG signals; these are
ena of interest. This set of features constitutes a pattern. The computed from the original signal as discussed in Chapter 54.
second operation may be the classification or clustering of the set Furthermore, one can also define similar features for the recti-
of features. According to the former one must assume a priori fied signal: mean, standard deviation, and coefficient of varia-
that there exists a number of classes (e.g., clinical tion. Furthermore, interval analysis of EEG signals yields a
normal/abnormal) to which the objects must be allocated; number of other features: average frequency of zero crossings of
according to the clustering approach, however, it is not neces- the original signal and also of its first and second derivatives.
sary to define a predetermined number of classes. Rather, in this The combination of amplitude and interval analysis (see
case the aim is to find clusters of objects based on a given sta- Chapter 54) yields a set of features that can characterize EEG
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

tistical criterion. In the classification approach one chooses a signals; in addition to those described above, a few others may
group of EEGs, the so-called learning set, to determine the set of be chosen such as the signal half-wave length and its derivatives
features that gives the best possible discrimination between the (mean, standard deviation, and range), peak-to-peak values per
classes, for example, applying Fisher’s linear discriminant analy- wave (mean and standard deviation), and amplitude range (i.e.,
sis. Thereafter, the best set of features can be used to classify any the difference between the largest and the smallest amplitude
other group of EEGs that constitute the test sets. value within a certain time epoch). Other features that have
Clustering requires little or no specific a priori knowledge; been included in this kind of analyses are Hjorth’s parameters:
the objects are grouped in clusters applying a clustering algo- activity, mobility, and complexity, and parameters extracted
rithm. The user must determine, however, the most convenient using time–frequency analyses, for example, those obtained by
level at which clustering must be stopped, according to the spe- wavelet decomposition, as explained in Chapter 54.
cific problem being analyzed. Thereafter, the relevance of the
clusters obtained with respect to the specific clinical applica- Spectral Analysis Using Nonparametric Methods
tion, or any other application of interest, must be evaluated. The most common features extracted from EEG, however, are
Since in EEG analysis, most methods of analysis follow a pat- derived by way of spectral analysis, such as the spectral intensity
tern recognition approach, explicitly or not, we consider here within the classic frequency bands, namely the mean spectral
the most important aspects of such an approach in relation to intensity (power or amplitude) and the average frequency.
general problems of automatic EEG diagnosis. For a thorough Although this may be rather trivial, it is important to consider
1203
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1204 Part IX ■ Computer-Assisted EEG Analysis

Table 56.1

EEG Features Obtained from Spectral Analysis

Frequency Bands Factor Analysis

Classical Matousek Wieneke
Definition and Gotman Hermann (personal
(Hz) Petersén (5) (6) Weighting Factors, K et al. (7) communication)
Frontal Other
Channels Channels
0  1  1.9 0.4–1.2 1.5 2
2  2  3.4 1.5–3.5 1.6–4.0 4 4 1.5–5.5 1.5–6.0
3.5  1  5.4 4.4–6.4 3 5
5.4  2  7.4 3.5–7.5 6.8–7.2 3 1 5.5–8.5 6.0–9.0
7.4  1  9.9 7.5–9.5 8.6–10.5 9.0–10.5
9.9  2  12.4 9.5–12.5 7.6–12.8 1 1.5 10.5–12.0 10.5–12.5
12.5  1  17.9 12.5–17.5 12.0–18.0 12.5–15.5
18.0  2  23.9 13.2–30.0 1 0.5 18.0–21.0 15.5–18.5
24.0  2 17.5–25.0 21.0–30.0 18.5–28.0
 k 1d 1  k 2d 2  k 3 1  k 4 2
Ratios used
  cs1 k 3    k 6
for discrimination

( and  Asymmetry ratios e

in V),
asymmetry Q

the fundamental question of how to define the EEG frequency University Hospital of Utrecht (Wieneke, personal communica-
bands. According to the generally accepted empirical definition), which factorized frequency bands of power spectra (log-
tions, one may use the subdivision indicated in Table 56.1. A arithmic values) obtained from eight symmetrical derivations
question that should be asked is whether the activities in differ- in 89 patients, are shown in Table 56.1; in this case, normaliza-
ent frequency bands are independent, or not. This question can tion was performed using the band from 5 to 20 Hz as refer-
be answered by means of multivariate statistical analysis of EEG ence. The distribution of the most important frequency factor
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

spectral values. This issue has been investigated thoroughly in loadings obtained in this way for different derivations is given
the 1970s when the first computerized systems were started to in Figure 56.1. We should emphasize that this type of analysis is
be developed. Hermann et al. (7) using factor analysis chose 57 sensitive to normalization and scaling. It is remarkable, how-
relative power values in frequency bands between 1.5 and 30 Hz ever, that the different methods presented in Table 56.1 yield
with frequency resolution f = 0.5 Hz, as well as absolute power results that display a considerable degree of overlapping with
values. In this way, it was found that the power spectrum could be respect to the different frequency bands. The frequency bands
broken down into the frequency bands indicated in Table 56.1. calculated in this way are also clearly compatible with those
Dymond et al. (8) also performed factor analysis of power spec- used in classical EEG; therefore, it may be said that a subdivi-
tra (log transformed) of bilateral centro-occipital leads and sion in frequency bands as used by Matousek and Petersén (5)
extracted four main factors having high loadings within the fol- or Gotman et al. (9) is acceptable for routine clinical EEG
lowing frequency bands: 0 to 8, 6 to 12, 12 to 20, and 20 to 30 Hz; analysis. If one deals with a completely defined group of EEGs
they also extracted factors associated with EEG asymmetry. (e.g., in psychopharmacologic studies where one has a group of
It should be noted, however, that applying factor analysis to subjects receiving a drug and a control group), factor analysis
sets of power spectra is not simple. The results depend on (i) may preferably be applied to such a specific group in order to
whether the spectra are expressed in power or in root mean define the optimal frequency bands that should be used in that
square (RMS) values, (ii) the way the spectra have been normal- specific study. Within the defined frequency bands, several pri-
ized, (iii) the derivations that were included, and (iv) the mary spectral features, such as absolute power intensity in V2
subject population. The results of an investigation of the or in dB, relative power, square root of power, and average
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
Chapter 56 ■ Computer-Assisted EEG Pattern Recognition and Diagnostic Systems 1205

Figure 56.1 Factors found using factor analysis of EEG power spectra. Factor analysis of log-
arithmic power spectra for several derivations was done; power spectra were normalized in
relation to the power within the frequency band between 5 and 20 Hz. Each factor is repre-
sented either by a parallelogram or simply by a horizontal line. The latter or the base of the
parallelogram indicates the frequency interval within which the factor accounts for more than
50% of the variance; the top line of the parallelogram indicates the frequency interval within
which more than 70% of the variance is accounted for by the corresponding factor. The fac-
tors are numbered in the order of decreasing eigenvalues from 1 to 6 or 7. A varimax rotation
was used. The data were obtained from EEGs of 243 patients, each consisting of 100-second
epochs recorded with eyes closed. (Courtesy of G. Wieneke.)

frequency within a band, can be computed. Secondary spectral lized alpha 1, x(18) = normalized alpha 2, x(19) = normalized
features can also be derived. Several types of secondary features beta 1, and x(20) = normalized beta 2.
have been proposed based on empirical criteria; clinical appli- Friberg et al.’s (10) model is defined by the following linear
cation has validated those proposed by Matousek and Petersén equation: calculated EEG age = a(0)  a(1)x(1)  … 
(5) and by Gotman et al. (9). Matousek and Petersén (5) inves- a(20)x(20). The coefficients a(i), with i = 0 to 20, were esti-
tigated 20 features extracted from the frequency spectrum of mated by minimizing the sum of squares of the differences
each EEG derivation. This study was based on the authors’ between the subject’s actual age and the calculated EEG age.
claim that an increased amount of slow frequency in the EEG in The correlation coefficients between actual and calculated EEG
abnormal cases might be considered analogous to the relatively age varied between 0.88 for derivations C3-C0 and C0-C4 and
large amount of slow activity seen in the normal but immature 0.86 for derivations F7-T3 and F8-T4. Those authors found
EEG. Initially, the EEG score chosen as being the most clearly that, according to their model, the calculated EEG age tended to
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

age-related was the ratio between theta-band activity (3.5 to be greater than zero when the line was extrapolated down to an
7.5 Hz) and the alpha-band activity (7.5 to 12.5 Hz) added with actual age of zero. To avoid this, they introduced two new vari-
a constant factor. ables: the calculated EEG maturity and the actual EEG maturity;
The same group later reinvestigated this problem (10). They the former is linearly related to the calculated EEG age and the
used as normative data the RMS of the spectral values com- latter to the actual age. The ratio between calculated and actual
puted within the frequency bands, indicated in Table 56.1, for a EEG maturity is called the ratio of EEG normality, because the
number of derivations (FT-T3, C3-C0, T3-T5, P3-01, and the authors found that this ratio is closely related to the degree of
symmetrical ones) of 562 EEG recordings from healthy individ- EEG (ab)normality. To calculate such a ratio, the maximal
uals aged 1 to 21 years. A number of ratios between RMS values actual EEG maturity of any individual is fixed to correspond to
were also computed. In total, 20 spectral features per derivation 22 years (age-related EEG changes are considered to be small
were calculated, as follows: x(1) = delta activity, x(2) = theta, beyond that age). The clinical implications of this form of fea-
x(3) = alpha 1, x(4) = alpha 2, x(5) = beta 1, x(6) = beta 2, x(9) ture extraction and data reduction are discussed below.
= alpha 1/alpha 2, x(10) = beta l/(alpha 1  alpha 2), x(11) = Gotman and his collaborators based their procedure for
beta 2/(alpha 1  alpha 2), x(12) = beta 1/beta 2, x(13) = extracting spectral features on the widely accepted assumption
delta/theta, and x(14) = sum of delta, theta, alpha 1, alpha 2, that some kind of relation between slow and fast EEG activity
beta 1, and beta 2; features from 15 through 20 are normalized should characterize the degree of EEG abnormality. Moreover,
amplitudes in relation to x(14) for the following bands: x(15) = they pointed out that a relative measure of spectral intensity is
normalized delta, x(16) = normalized theta, x(17) = norma- preferable to an absolute measure because the latter depends on
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1206 Part IX ■ Computer-Assisted EEG Analysis

a number of spurious factors (e.g., skull thickness). Therefore, analyses to achieve differential diagnosis between patients’
these investigators analyzed the potential of several ratios, for (sub)populations. In this way, neurometric test batteries were
example, (delta  theta)/(alpha  beta), using different weigh- constructed and applied to several clinical problems. A general
ing factors and frequency band subdivisions, to discriminate battery consists typically of the following features: spectral
the EEG between normal and abnormal subjects (slow-wave composition, coherence, and symmetry indices of the sponta-
type of abnormality). The best weighing factors for different neous resting EEG; in addition, brainstem auditory evoked
frequency bands and areas are given in Table 56.1. The same potential (BAEP) and brainstem somatosensory evoked poten-
investigators introduced still another important spectral fea- tial (BSEP) to unilateral stimuli, checkerboard pattern reversal
ture, a degree of asymmetry. To compute this feature, the scalp or flash visual EPs, and cortical EPs to different modalities both
was subdivided into four symmetrical regions: frontal (Fp1-F3, to predictable and unpredictable stimuli, are also included. This
Fp1-F7, Fp2-F4, Fp2-F8), temporal (F7-T3, T3-T5, F8-T4, T4- approach can be implemented in a personal computer.
T6), central (F3-C3, C3-P3, F4-C4, C4-P4), and occipital (P3- Profiles of neurometric features that deviate from age-
01, T5-01, P4-02, T6-02). For each region, two asymmetry matched normal subjects have been obtained in several
coefficients were calculated, one for the slow frequencies categories of patients suffering from cognitive disorders (e.g.,
(weighted delta and theta values as given in Table 56.1) and one dementias), psychiatric illnesses (e.g., different types of depres-
for the higher frequencies (weighted alpha and beta values). sions and of schizophrenia), and neurologic dysfunctions, for
The value corresponding with the most active hemisphere was example, compromised cerebral blood flow (14), as discussed
always placed in the numerator. Gotman et al. (9) called the dis- by John et al. (12) and Prichep et al. (15). The clinical relevance
play of these ratios extracted from spectral values canonograms of neurometrics is controversial and has led to publications pre-
(canon is Greek for “ratio”) (Fig. 56.2); the clinical validation of senting opposite points of view by John (16) and Fisch and
these features is discussed later in this chapter. Pedley (17). A special effort was made by John and collabora-
Other spectral features of interest are the spectral peak frequen- tors to apply the neurometrics approach of quantitative EEG
cies and corresponding bandwidths. There are several algorithms analysis to distinguish subgroups of patients with psychiatric
used to calculate peak frequencies: these involve computing a local disorders also with the aim of identifying potential responders
maximum of the curve defining the spectral density. A peak is said to pharmacologic treatments (18). This was applied to patients
to exist when it rises significantly above its surroundings. The suffering from obsessive–compulsive disorder (OCD) who
bandwidth is usually calculated as the frequency interval between received treatment with selective serotonin reuptake inhibitors
the 3-dB points at both sides of the peak. (SSRIs) with the interesting result that the responders and non-
A comprehensive analysis methodology that combines responders presented distinct neurometric profiles (19,20).
quantitative EEG and EP features is the approach introduced by
John and collaborators and reviewed extensively in John et al. Spectral Analysis Using Parametric Methods
(11,12) and Prichep and John (13), under the name of neuro- In Chapter 54 we discussed the general theory of spectral analy-
metrics. This approach is based on the use of standardized data sis employing ARMA or AR models. One of the main advan-
acquisition techniques, computerized feature extraction, statis- tages of these parametric methods of computing power spectra,
tical transformations in order to achieve approximately as proposed initially by Zetterberg (21), is precisely the fact that
Gaussian distributions, age regression equations, and multi- the use of spectral parameter analysis (SPA) avoids having
variate statistical methods, namely discriminant and cluster to subdivide the spectrum in distinct frequency bands

Figure 56.2 Canonogram from subject with multiple metas-

tases in right hemisphere. The size of each polygon is propor-

tional to a slow/fast EEG activity ratio, indicator of
abnormality for a channel. They are arranged in a topograph-
ical pattern corresponding to the position of the derivations on
the subject’s head: frontal on top and occipital on bottom.
Arrows under horizontal lines indicate the asymmetry in slow
EEG activity; arrows above indicate the asymmetry in fast EEG
activity. Sixteen channels, anteroposterior bipolar montage
covering parasagittal and temporal regions; EEG epoch, 40 sec-
onds. (Adapted from Gotman J. Problems of presentation of
analytical results. Electroencephalogr Clin Neurophysiol Suppl.
1978;34:191–197.)

correlation). The computation of an ARMA or AR model yields

an important degree of data reduction. The relevant informa-
tion is thus condensed in the coefficients of the model; the num-
ber of coefficients corresponds to the order of the model. As
shown by Mathieu et al. (24) and Jansen (25), the coefficients
can be used to characterize the EEG directly. The importance of
this approach for EEG pattern classification is discussed later.

The Recognition and Elimination of Artifacts:

Eye Movements and Muscle Artifacts
Physiologic and technical artifacts are the outstanding enemies of
automatic EEG analysis. They must be eliminated if computer
EEG analysis is to be used in clinical practice. It is a general
requirement of EEG recording in any clinical laboratory that the
records have a minimum of technical artifacts, a requirement that
is even more critical in automatic analysis. One way to control the
quality of EEG signals while performing analog-to-digital conver-
sion in the clinical laboratory is by simply deleting those epochs
that are below acceptable standards due to technical or to physio-
logic (e.g., ocular or muscular) artifacts. For example, the techni-
Figure 56.3 Block diagram of the EEG simulator; the  filter represents cian responsible for this operation may delete the series of
a first-order active RC network; the , , and  filters are of second digitized samples immediately preceding an identified artifact.
order; potentiometers independently control the parameters f(2/r) Nevertheless, there will always be situations in which artifacts,
(which determine the resonance frequency f0),  (which determine the particularly those of a physiologic nature, are unavoidable. This is
bandwidth), and  (which determine the zero of the transfer function); particularly important during long-lasting EEG monitoring in
the power parameter is G. (Adapted from Zetterberg LH. Experience several clinical (e.g., EEG-video monitoring of epileptic patients)
with analysis and simulation of EEG signals with parametric description and experimental (e.g., sleep studies) conditions and when com-
of spectra. In: Kellaway P, Petersen I, eds. Automation of Clinical puter-assisted quantification is applied (see also Chapter 35).
Electroencephalography. New York, NY: Raven Press; 1973:161–201.) Eye movements and muscle potentials occur in most records
of a few minutes’ duration; they can distort power spectra and
lead to detection of transient nonstationarities that may be dif-
beforehand. The SPA method describes the EEG as resulting ficult to distinguish from epileptiform events. Eye blinks can be
from noise sources passed through a set of parallel first- or sec- reduced by recording with eyes closed; slow eye movements,
ond-order filters, as illustrated in Figure 56.3. As demonstrated however, are more difficult to avoid. These are bilaterally syn-
by Isaksson and Wennberg (22), the relevant spectral features chronous with a maximum in frontal derivations and represent
can be derived simply. The first-order filter describing the low- an important contribution to the power in the delta band in
frequency band is characterized by two features: the total power these derivations. In the early days of EEG quantification,
G and the total bandwidth (interval from zero to the frequency Gotman (6) discussed several methods of avoiding this type of
corresponding to the 3-dB point); each of the second-order fil- artifact at the very first stage, for example, by subtracting the
ters describing theta, alpha, and beta components is character- electro-oculogram (EOG). This matter has been reviewed by
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

ized by the three features: G (power),  (bandwidth), and Jervis et al. (26) and by Brunia et al. (27). However, the tech-
f (resonance frequency). nique of EOG subtraction may give rise to distortion of the
Isaksson and Wennberg (22) concluded that, for most prac- EEG signals, since the EOG recording also contains brain sig-
tical applications, a SPA model of the fifth order, at the highest, nals (28) that may be partially eliminated by filtering first the
is sufficient, although using this order model only the first- EOG with a low pass of about 8 Hz. The transfer of EOG activ-
order delta component and the second-order alpha and beta ity to the EEG can be analyzed using a frequency domain
components can be described. In a few cases, it may be neces- approach. Eye blinks and slow eye movements have different
sary to use a model of the seventh order to include a second- spectral properties and are transferred in different ways to the
order theta component. In a study comparing the degree of skull. Gain functions for transferring both types of eye move-
visually evaluated slow activity in a large number of artifact- ments to the skull were computed by Gasser et al. (29). These
free EEG epochs with the features identified through SPA of the authors obtained average gain functions that they found to be
same epochs, Isaksson and Wennberg (23) concluded that, for of practical use in correcting EOG artifacts. In other studies, a
most derivations, there was a significant linear correlation frequency domain approach to correcting EOG artifacts has
between the degree of slow activity encountered with visual been proposed (30,31). Similarly, Jervis et al. (32) found that a
inspection and the value of the features G (positive correla- computerized correlation technique provides results superior
tion) and  (negative correlation); in a few cases, there was also to analog techniques for removing eye movement artifacts.
correlation with G (negative correlation) and  (positive Elbert et al. (33) also stressed that the best correction for these

types of artifacts is obtained in the frequency domain, but indi- electrophysiologic data using ICA, time/frequency analysis, and
cated that the correction procedure should be based on more other methods including artifact rejection, as indicated below.
than one EOG derivation, and preferably on three. Fortgens
and de Bruin (34) also obtained good results using the method Transient Nonstationarities: Epileptiform Events
of least squares based on four EOG derivations. The detection of epileptiform events is a typical example of the
Other important physiologic artifacts are muscle potentials. application of a pattern recognition approach in EEG analysis.
Here also it is important to note that electromyographic (EMG) In this case, the epileptiform events (spikes, sharp waves, and
signals affect the EEG power spectrum not only at relatively spike-and-waves) are considered to constitute the “signal,”
high frequencies (30 to 60 Hz) but also even down to 14 Hz whereas the background activity constitutes the “noise.” The
(35). Under normal conditions, there is very little EEG power at difficulty here lies in defining the epileptiform transients, that
the scalp in the 30- to 50-Hz band; if the power is significantly is, the “signals” that one wants to identify. In 1949, Jasper and
large, however, one must suspect contamination with EMG sig- Kershman (47) classified these events into spikes (duration 10 to
nals. Gotman (6) proposed dealing with this problem by 50 milliseconds) and sharp waves (duration 50 to 500 millisec-
introducing a reduction factor with which the activity in the onds). The Terminology Committee of the International
beta band should be multiplied; this factor depends on the Federation of EEG Societies defined spikes as waves with a
spectral intensity integrated over the 30- to 50-Hz band. If this duration of 1/12 second (83 milliseconds) or less, and sharp
is below 1.5 V/Hz, the reduction factor is equal to unity; if the waves as waves with a duration of more than 1/12 second and
activity is larger than 1.5 V/Hz, the reduction factor decreased less than 1/5 second (200 milliseconds) (48). Later, this
linearly to 0.1 as the spectral activity increases up to 5.0 V/Hz. Federation Committee gave somewhat different duration limits
An alternative way to deal with artifacts is that used by for these phenomena, with spikes having a duration from 20 to
Gevins et al. (36), who determined thresholds for head and under 70 milliseconds and sharp waves having a duration of 70
body movement artifacts (under 1 Hz), high-frequency arti- to 200 milliseconds (49). A few other characteristics have been
facts mainly caused by EMG (34 to 50 Hz), and eye movements identified. Spikes and sharp waves should be clearly distin-
(below 3 Hz in frontal derivations) based on a short segment guishable from background activity and have a pointed peak;
that includes those artifacts; thereafter, EEG epochs exceeding their main component should be generally negative relative to
the aforementioned thresholds were simply discarded (37). other scalp areas, and their amplitude variable. A distinction
The need to avoid the contamination with artifacts of rele- between spikes and sharp waves has descriptive value only. The
vant EEG features is so pressing that this area of EEG signal parameter characteristics of spikes found in the human EEG
analysis has been, for decades, in constant evolution. Here we have been studied by Celesia and Chen (50).
briefly review the most relevant approaches. One problem is the difficulty of defining a learning set that
Rather elaborate methods are based on decomposing a set of may be unambiguous. A pioneering investigation in this respect
EEG signals into components that should represent the artifact was carried out by Gose et al. (51); this study revealed consider-
and the EEG signals, respectively. One of these is the spatial filter- able intra- and interrater variability. In practical terms several
ing approach (38,39). According to this method, the topography methods have been used to identify the epileptiform events by
of the artifact is first estimated on the basis of a specific record- increasing the signal-to-noise ratio. Most of them are akin to the
ing where the artifact is clearly evident, since this is, in general, classic approach of Carrie (52), who used as a criterion the ratio
easier to model than the EEG. Thus, the artifact can be described between the amplitude of the second derivative of the EEG signal
as the product of the corresponding topography vector and time and the moving average of similar measurements from a number
waveforms. This can be then subtracted from the EEG signals of preceding and consecutive waves; a ratio of 4 or 5 was said to
contaminated with artifact to yield the corrected signals. indicate an epileptiform event. Most other relevant studies have
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Other methods have been proposed that differ in the way of proposed similar types of measures (53–57). All these methods
separating EEG and artifact signals. With this objective involve a preprocessing stage that constitutes a form of high-pass
Lagerlund et al. (40) used principal component analysis (PCA), filtering (e.g., computing the signal’s second derivative).
but this method has the drawback that PCA yields uncorrelated The method used by Lopes da Silva et al. (58–62) is based on
components while the EEG signals and the artifacts may be cor- an essentially more general form of preprocessing. In this
related. An important step forward in this context has been the method, an EEG epoch is described by way of an AR model that
introduction of independent component analysis (ICA) (see provides the best fit to the background activity. The basic oper-
Chapter 54) that is very effective in separating EEG signals from ation to improve the signal-to-noise ratio consists of passing the
artifacts as shown in a number of applications (41–44). The EEG signal through the inverse filter of this estimated AR model;
application of ICA, however, needs some form of postprocess- this inverse filtering operation yields a new signal that ideally
ing to identify the components corresponding to the EEG sig- should have the properties of uncorrelated white noise. The sta-
nals and to the artifacts. Several strategies and combinations of tistical properties of this new signal are then determined; devia-
approaches particularly with respect to their practical imple- tion of the new signal resulting from inverse filtering from a
mentation are discussed by Ille et al. (38) and by Makeig et al. normal distribution at a certain probability level is thought to
(42). Particularly useful is the software package developed by identify a transient nonstationarity (see Fig. 54.14). The essential
Makeig’s group (45,46) that they called EEGLAB. This is an feature of this method is that inverse filtering of the EEG epoch
interactive Matlab toolbox for processing continuous and eliminates in an optimal way the background activity, allowing
event-related EEG, magnetoencephalogram (MEG), and other the transient nonstationarities to emerge clearly.
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
Chapter 56 ■ Computer-Assisted EEG Pattern Recognition and Diagnostic Systems 1209

Not all transient nonstationarities, however, are necessarily

epileptiform events; some may be physiologic artifacts or other
kinds of EEG transients (e.g., lambda waves or sharp bursts of
alpha waves). After detecting transient nonstationarities, one
must apply a form of pattern recognition to select those that can
be accepted as being epileptiform in nature. This constitutes the
“two-stage analysis approach” proposed by Guedes de Oliveira
and Lopes da Silva (63) and Guedes de Oliveira and Lopes da
Silva (64). Two main pattern recognition methods have been
proposed; one is based on a matched filtering approach and the
other on piecewise characterization of the transient. Matched
filtering using as template a spike-and-wave pattern has been
used (65) to detect epileptiform transients even without prepro-
cessing. Barlow and Dubinsky (66) used a comparable method,
computing the running correlation coefficient between the EEG
signal and a template (see Fig. 54.15). However, the variability of
the waveforms characteristic of such transients presents a seri-
ous difficulty in dealing with this problem in practice.
Pfurtscheller and Fischer (67) combined a preprocessing stage
using inverse autoregressive filtering and a template matching
stage for postselection of relevant epileptiform events. An alter-
native method is to apply a piecewise analysis to the transient
nonstationarities, to identify those that belong to the epilepti-
form class. Smith (68) and Ktonas and Smith (69) proposed
such a piecewise analysis method based on five features
(Fig. 56.4): S1 and S2, the maximum slopes, respectively, before
Figure 56.4 Top: An epileptiform spike; bottom: the corresponding
and after reaching the peak of the spike; S3, the time taken by the
first derivative. The parameters proposed by Ktonas and Smith (69) are
spike to reach the peak after it attains maximum slope; and S4,
shown: S1 and S2, the maximal spike slopes, respectively, before and
the time taken by the spike to reach maximum slope after the
after reaching the peak; S3, the time taken by the spike to reach the
peak. The sum (S3  S4) of the time intervals corresponds to the
peak after it attained maximal slope; S4, the time taken by the spike to
duration of the epileptiform spike (S5). The interval between
reach maximal slope after the peak. The sum S3 = S3  S4 is a meas-
two consecutive zero crossings of the same polarity of the first
ure of the duration of the sharp part of the peak. The time interval
derivative (S6) is also a relevant feature.
between two zero crossings of the same polarity of the first derivative
Frost (70) considered the problem in a simpler form, pro-
is S6. The time duration of the signal shown is 1 second. (Adapted from
posing the following characteristic features. Assuming that an
Lopes da Silva FH. Analysis of EEG nonstationarities. Electro-
epileptiform spike is a triangular wave with a point of origin M
encephalogr Clin Neurophysiol Suppl. 1978;34:163–179.)
at the base, an apex S, and a point of termination P, Frost
defined amplitude as the largest value of the segments MS or SP,
and duration as the interval MP. Furthermore, he used as a on their spatial distribution. Using this methodology, Ktonas
measure of sharpness D, an estimate of the signal’s second and Smith (69), Lopes da Silva et al. (1978), and Gotman (71)
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

derivative. The initial processing step involves comparing the observed that most epileptiform spikes, at the scalp, present a
value of D with a threshold, so that, whenever D is larger than second slope that is steeper than the first, contrary to the qual-
a certain value, a candidate spike is detected. Extracting the fea- itative description of Gloor (72). However, Lemieux and Blume
tures described here requires a relatively high rate of EEG sam- (73) found that the spikes recorded directly from the cortex
pling—at least 200/sec. presented a first slope that was equal or steeper than the second.
The next section considers the practical implications of Another interesting analysis that can be realized using these
these methods in assessing the EEGs of epileptiform patients. methods consists of the quantification of the distribution of
According to the method of Gotman and Gloor (55), at the end epileptiform spikes in relation to the occurrence of seizures.
of an analysis session the computer displays all transients Gotman and Marciani (74) found that the level of spiking is not
detected, whether true or false. The distinction between these related to the probability of seizure occurrence, but they
two types is made off-line in an interactive way. This form of reported an increase in spiking in the days following seizures.
analysis represents a considerable data reduction and provides Very much as in the case of the detection of artifacts, the
a reliable account of the main types of epileptiform transients analysis of epileptiform events has attracted the interest of
present in a given record. many researchers and new approaches are often being intro-
The methods of analysis described in this section not only duced. In most cases, new methods are published without a
are useful in detecting the presence of epileptiform events, but comprehensive comparison with older methods, which make it
also provide quantitative information on the morphology of difficult to evaluate the performance of new approaches with
such events, on their distribution over long periods of time, and respect to previous ones. An interesting exception is the study
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1210 Part IX ■ Computer-Assisted EEG Analysis

of Dumpelmann and Elger (75), which we discuss in detail object is assigned to a certain region or class by several types of
below (see the section “CADS and Epileptiform Events”). decision rules; these are described in detail by Demartini and
Very often several categories of spikes can be distinguished in Vincent-Carrefour (4), among others.
the EEG or MEG of patients with epilepsy that may differ con- To develop and test a classifier, it is important to dispose of a
siderably in waveform and may even be associated with different sufficiently large learning set (i.e., a set of N objects that have been
sources in the brain. Therefore, the population of spikes classified a priori using independent criteria); in the case of EEG
recorded in a given patient should preferably be grouped into analysis, the independent criteria ought to be clinically valid. This
distinct categories before being averaged. This is especially implies that the objects must be classified by expert raters (elec-
relevant if source reconstruction is going to be performed. This troencephalographers) using generally accepted criteria, possibly
implies that a form of clustering of spikes has to be carried out. based on visual inspection, and making use of all relevant clinical
In the simplest case, this may be done by visual inspection by an information. The learning set should contain a sufficient number
experienced electroencephalographer. Such an operation of objects (4). One way to develop an automatic method of EEG
becomes rather complex and time consuming if the number of analysis is to divide the experimental set into two parts. Thus, the
spikes and of channels in the EEG/MEG is quite large. This has first part (learning set) is used to develop the classifier and the sec-
led to the development of computer algorithms to automate the ond to test its performance (test set). A useful alternative if the
identification of clusters of spikes (76). We should add that if the experimental set is too small is the “hold-one-out” strategy, which
purpose of the analysis of epileptiform spikes includes the esti- involves removing one object from the learning set and then resyn-
mation of the localization of the corresponding sources in the thesizing the classifier and trying to recognize the selected object.
brain, it is preferable to use MEG than EEG recordings because This operation should be repeated for each object. The resulting
solutions of the inverse problem are more accurate with MEG error rate is a good estimate of the classifier’s performance.
(see also Chapter 5). One of these studies is that of Van’t Ent et The quality of the learning set is of primary importance. To
al. (77) who performed cluster analysis of MEG epileptiform start with, it is necessary to have knowledge about rater repro-
spikes, grouping spikes according to the similarity between mag- ducibility (intrarater agreement) and validity (interrater agree-
netic field characteristics. Thereafter, the spikes within one clus- ment) as regards evaluation of the EEG records constituting the
ter are averaged to improve the signal-to-noise ratio such that learning set. A few studies have addressed electroencephalogra-
the quality of equivalent dipole source estimates is enhanced. phers’ overall classification of EEG records as normal or abnor-
Similarly Abraham-Fuchs et al. (78) showed that averaging of mal; in such cases, the validity of the visual assessment is usually
similar spike events, recorded in the MEG, substantially about 80% to 90%. Although most raters generally agree on the
improves the signal-to-noise ratio (a more general discussion of division of the EEG into two classes globally (normal or abnor-
clustering algorithms is presented in the next section). mal), classification of short segments or of epileptiform tran-
sients is much less consistent. The same applies to intrarater
Classification and Clustering in EEG Analysis agreement. In the assessment of EEG patterns corresponding to
The previous section considers different ways to find sets of fea- different sleep stages, however, a good degree of interrater
tures that can characterize EEG signals. In this section, we con- agreement can be expected; thus, it is not surprising that meth-
sider very briefly the next phase in pattern recognition, ods of automatically classifying sleep stages have been those
classification and/or clustering. For a detailed account of this more often evaluated in a quantitative way. In assessing epilep-
problem, the reader is referred to Duda and Hart (1) as indi- tiform events (spikes, sharp waves, spikes-and-waves), a large
cated above. It is necessary to consider this question here in degree of interrater variability is also encountered. Gose et al.
order to be able to evaluate quantitative EEG analysis methods (51) found considerable variability in the human detection of
in the clinical laboratory. The essential problem is one of diag- spikes; a total of 948 events were marked as spikes by one or
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

nosis; given a set of EEG epochs that have been analyzed and more electroencephalographers, but only 104 events were
characterized by a number of features, it is necessary to deter- marked by five raters. However, disagreement between raters on
mine what is the performance of the algorithm regarding the individual spikes is not very important; a comparison on a
classification of the EEG epochs in a given number of diagnos- patient basis (30 records seen by five raters) is more important;
tic categories (e.g., normal/abnormal, sleep stages) or to label seen from this viewpoint, the average error rate was only 4%.
EEG transients (e.g., spikes) as epileptiform. For the classification of EEG records in the learning set, it is
One way to solve this problem is to use discriminant analy- important to utilize a structural report such as used by Volavka et al.
sis, which is possible only if one knows a priori that the EEG sig- (79), Rose et al. (80), Gotman et al. (81), Gotman and Gloor (55),
nals belong to a defined number of classes. Assuming that the and Gevins (82). In other words, EEG classes should be defined
analysis involves classifying EEG signals into two classes, nor- unambiguously; the abnormal EEG can be classified as paroxysmal
mal and abnormal, and using a set of features, the feature vec- or irritative, hypofunctional (cortical or centrencephalic) or mixed;
tor defines a point in n-dimensional space. the location of the abnormality (focal: frontal, temporal, central,
In discriminant analysis, the space where all objects (e.g., occipital, lateralized, or diffuse) should also be specified.
EEG epochs characterized by a vector set) are contained must Furthermore, one may use a complementary second-order classifi-
be subdivided into a number of regions; the objects within a cation into diagnostic types related to the global medical diagnosis:
region form one class. The functions that generate the surface space-occupying lesions, metabolic disorders, cerebrovascular
separating the regions are called discriminant functions. An insufficiency, seizure disorders, or psychiatric disorders.

EEG SEGMENTATION AND CLUSTERING analysis of long time series. Van’t Ent et al. (77) reported a spike
clustering analysis that yields meaningful results in neocortical
We should emphasize that EEG records are generally nonsta- localization-related epilepsy, in MEG.
tionary. Although in the clinical laboratory it is usually possible This type of analysis may be criticized similarly to the segmen-
to obtain representative EEG epochs by tightly controlling the tation method described earlier. Here also the number and types
subject’s behavioral state, it is often desirable to distinguish in of classes are arbitrary; an advantage of clustering methods, how-
an EEG signal segment, characterized by different properties, ever, is their flexibility; the same data may be clustered in several
that can be separated automatically. This is particularly impor- ways. This is feasible because all information necessary to charac-
tant in the case of EEGs recorded under intensive care condi- terize the EEG segments is stored in the form of a small number
tions, such as during anesthesia, or in other long-duration of coefficients (e.g., five coefficients of an AR model). Some appli-
records. Ideally equivalent segments thereafter could be cations of this method are discussed in the following section.
grouped together, thus defining a number of classes. An early
effort in this direction was made by Bodenstein and Praetorius NEURAL NETWORK-BASED
(83), who proposed a general method of EEG segmentation;
EEG CLASSIFICATION
they assumed that an EEG should be considered as a sequence
of quasi-stationary segments of varying duration. They used an Neural networks have been employed to classify EEG features.
AR model as described in Chapter 54. Several research groups have successfully explored this
By setting appropriate thresholds, Bodenstein and Praetorius approach. This appears particularly interesting for the classifi-
(83) have been able to formulate explicit criteria for EEG seg- cation of single-EEG epochs. Three types of neural
mentation. The problem, however, is that the validity of this seg- network–based classifications of EEG data were reported: clas-
mentation procedure in relation to clinically clearly defined sification of single-EEG trials for selective averaging (86); clas-
states is difficult to demonstrate. Jansen (25) made an interest- sification of averaged and nonaveraged multichannel EEG data;
ing effort along a similar line by using an algorithm akin to that and classification of single-trial, multichannel EEG data (87).
discussed above but based on a Kalman filter and following a For these classifications, different types of neural networks were
different strategy; this method is called Kalman–Bucy (KB) clus- applied. A back-propagation network was used by Gevins and
tering. Defining segments of variable length based on statistical Morgan (86). Self-organizing feature maps followed by a learn-
criteria proved to be too difficult because a good learning set ing vector quantizer (LVQ), both introduced by Kohonen (see
could not be constructed. An alternative method followed by review, Ref. 88), were used by Pfurtscheller et al. (87). The lat-
Jansen (25) was to divide the EEG into a large number of seg- ter authors used a neural network approach to analyze and clas-
ments with a fixed duration of 1.28 seconds each, and classified sify nonaveraged multichannel EEG data from an experiment
them using an unsupervised learning clustering approach. In where the subject had to press a microswitch with either the left
other words, he used a clustering algorithm to group segments or right hand, whereby the side of movement was indicated by
with similar properties into a number of classes that were not a cue stimulus. On the basis of the spatiotemporal alpha event-
defined a priori. Each 1.28-second segment is characterized by a related desynchronization (ERD) prior to movement, this
feature vector consisting of the five coefficients of the correspon- method of automatic classification was able to predict the side
ding AR model estimated using a Kalman filter, often comple- of the hand movement. One part of the data was used for train-
mented by a measure of amplitude. The statistical approach ing the neural network, the other part to test the performance
used in this case is a form of clustering (see review in Ref. 84). of the network as classifier. Peters et al. (89), using autoregres-
Clustering can be partitional or hierarchical; the former is based sive modeling of EEG time series and artificial neural networks
on a priori knowledge of the place occupied by some objects, (ANNs), developed a classifier that can tell which movement is
which are then used as “seed points” around which clusters
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

performed based on information taken from a segment of the

grow. The latter can have two forms, agglomerative or divisive, EEG signal of a single trial. The classifier’s rate of recognition of
depending on whether one starts from an assembly of as many EEGs not seen before was 92% to 99% on the basis of a 1-sec-
clusters as objects or from one cluster encompassing all objects. ond segment per trial. Thus, the classifier was considered suit-
Jansen (25) used the agglomerative hierarchical clustering able for a so-called brain–computer interface, a system that
approach to group EEG segments of a number of types. This allows one to control a computer, or another device, by means
type of clustering involves an iterative process through which of EEG signals (see also Chapter 57).
the two most similar clusters of the previous step are merged
into a new cluster. The user can stop the process at any point, SEGMENTATION AND CLASSIFICATION
depending on the application. Using statistical criteria, it is pos- IN SLEEP EEG ANALYSIS
sible to delimit the number of classes in such a way that the dis-
tance between their centroids does not fall below a certain value. Several attempts have been made to develop an automatic sleep
Hierarchical clustering of epileptiform spike events has also analyzer based on EEG records, in combination with EOG and
been used in the analysis of interictal EEGs. Guess and Wilson EMG or independently (see a review of classic studies by
(85) presented an application by means of which spike events Johnson (90)). The development of an automatic processor has
are separated into groups, based on topology and morphology, been preceded by a thorough quantitative study of the EEG
which yields an efficient method of performing detailed characteristics during different stages of sleep. These studies

have been directed mainly to quantitative analysis of EEG sig- states. These authors validated their automatic sleep analyzer on
nals recorded from C3-A1 and C4-A2, of EOGs recorded from sleep recordings of nine young subjects. Data from four subjects
the outer canthus of each eye referenced to the ipsilateral mas- were used as learning sets and those of the other five as test sets.
toid, and of EMGs from the submental muscle (see also The sleep stages were classified visually by three human raters.
Chapter 10). Three main characteristics of the sleep EEG have For the five subjects, the average agreement between raters
been identified by computer analysis. ranged from 85.8% to 91.4%; the agreement between the pro-
Sleep spindles have been shown (91) to range in frequency gram and raters ranged from 77.7% to 86.2% and the agreement
from 12.4 to 14.6 Hz in young adults. Gondeck and Smith (92), between the program and the consensus of raters (majority
however, found that frequency can vary about 2 Hz between dif- decision) ranged from 78.8% to 86.4%. Using hybrid systems,
ferent spindles; spindle duration varies between 0.5 and 0.8 sec- Smith and Karacan (106), Gaillard et al. (111), and Gaillard and
onds. Based on a model of the generation of sleep spindles, Kemp Tissot (112) reported similar figures. Poppl (113) used as the
et al. (93) have introduced an optimal detector for this type of first feature extraction method a time-domain amplitude and
activity. Delta activity is the primary feature distinguishing wak- interval analysis procedure, which allowed considerable data
ing and sleep stages; Johnson et al. (91) concluded that the most reduction; by mapping the feature space to maximize the vari-
consistent peak in the spectrum during different sleep stages lies ance ratio between classes (in relation to the variance within
between 0.8 and 1.8 Hz. The delta activity increases between stage classes) and using linear discriminant analysis, a very good
1 (and rapid eye movement [REM]) and stages 2 to 4. With age, (91%) recognition rate for a test run was obtained using the
delta activity decreases in amplitude but not in incidence (94). In hold-one-out strategy. Mathieu et al. (24) obtained EEG features
addition, a very slow oscillation, at 0.5 to 1 Hz, during sleep, was using an autoregressive model of order 9, fitted to a large num-
described by Steriade et al. (95) in cats, which differs from delta ber of 30-second epochs from five sleep EEG recordings of three
waves (see also Chapter 3). This very slow sleep oscillation was different subjects. They found for the three subjects a recogni-
recorded during natural sleep in the EEG (96) and MEG (97) in tion rate of 81% in a test run; when applied to a single patient’s
humans. K complexes have been difficult to analyze automatically, EEG, the recognition rate was 91%. The lower recognition rate
probably because of their large variability. Bremer et al. (98) obtained when using different subjects is a consequence of the
developed a hybrid pattern recognition method for detecting K relatively large intersubject variability. Mathieu et al. pointed out
complexes. Rosa et al. (99) proposed a method for the automatic that grouping subjects by age classes might reduce variability. In
detection of K complexes that yields good practical results. The any case, the most difficult operation of the automatic classifier
method of Rosa et al. is based on a simple model of the neuronal was discriminating, on the one hand, between REM sleep and
network that is responsible for background EEG signals accord- wakefulness and, on the other, between sleep stages 3 and 4. This
ing to the proposal of Kemp (100). In the model, the main path- would probably be facilitated if the AR model features had been
way is represented by a frequency-selective feedback loop. The combined with EOG data and with a supplementary method of
central frequency of the network depends on the time constants detecting delta waves, as described above. This technique may be
of the neuronal elements in the network. Rosa et al. constructed improved still further by incorporating a more accurate detec-
a model that represents the delta activity typical of slow-wave tion of EOG data during REM. Regarding the distinction
sleep (SWS). The K complex is represented as the impulse between stages 3 and 4, the difficulty of the automatic methods
response of such a delta model. is shared by the human raters; thus, many observations combine
This section discusses some of the attempts to analyze auto- stages 3 and 4 into one stage called, simply, SWS. The usefulness
matically EEG signals in relation to sleep stages. The learning of computer analysis in sleep analysis is still unconfirmed; data
sets have been classified according to visual inspection, usually reduction is an obvious advantage, but the purpose of the effort
on the basis of criteria proposed by Dement and Kleitman (101) being carried out must be clearly defined. A possible interesting
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

and Rechtschaffen and Kales (102). Künkel (103) summarized application is in psychopharmacologic studies. Probabilistic
the results obtained by several investigators who used as the first models describing the statistical properties of the hypnogram
extraction procedure a form of spectral or hybrid frequency (i.e., the transitions between stages and their duration) have
analysis (104–106) or period analysis alone or combined with been developed (114–116). Kemp and Kamphuisen (117) intro-
analog filtering (107,108). The mean rate of correct recognition duced a model combining probabilistic and deterministic
of sleep stages varied for the different studies between 60% and aspects of sleep. Such models, based on a Markow chain process,
79%, depending largely on the visual classification method used may be useful for computerized analysis of hypnograms. An
and on the learning set. Martin et al. (109) and Viglione and account of several classic computerized methods of sleep analy-
Martin (110) reanalyzed this problem using a comprehensive sis has been published by Hermann and Kubicki (118). We
methodology; they used power spectra combined with a time- should note that sleep staging is a rather fuzzy process. Indeed it
domain technique to detect delta waves (period 0.55 seconds) is not an easy task to perform a computer-assisted analysis since
exceeding 75 V amplitude, and two EOGs to detect horizontal the standards are not well defined. This is a caveat that should be
eye movements. Interval analysis of delta waves was necessary taken into consideration. Most likely this is the reason why many
because the power in the delta frequency band (0 to 2 Hz) was ingenious algorithms developed in the past decades have not
shown not to be proportional to the number of delta waves gained wide acceptance in practice. One general feature of these
counted by human observers. EOG recordings were considered algorithms is that they are rule-based, and in general their per-
necessary in order to help distinguish between REM and waking formance depends very much on the learning population for

which they were developed. When they are tested in other pop- techniques used in continuous EEG monitoring in the intensive
ulations and other laboratories, many problems arise. More care unit were reviewed by Scheuer (131), including conditions
recently Agarwal and Gotman (119,120) attempted to solve such as cerebral ischemia, acute severe head injury, and coma. It
these limitations by developing an automatic sleep staging is interesting to note that the main EEG frequency ranges that
method that is based on the use of evolving schemes that can be represent ischemic changes in the brain, in a clinical setting,
adjusted depending on the type and quality of polysomno- were examined in detail by Visser et al. (132), who determined
graphical recordings. This algorithm adapts the sleep staging EEG spectral changes as a function of time in the course of
rules to the user preferences and the record being analyzed. A brain ischemia caused by short periods of circulatory arrest
method based on advanced adaptive time–frequency analysis during surgery. After onset of circulatory arrest, the log spectral
was developed by Malinowska et al. (121). This method uses changes of three-epoch moving averages were calculated rela-
Gabor functions that provide optimal definition jointly of fre- tive to the baseline spectrum. Factor analysis was carried out; 17
quency and time. This group pursued the development of auto- EEG periods were selected that showed changes progressing to
matic parametric sleep staging and tested the performance of an isoelectrical period. This analysis revealed four factors that
the system in comparison with visual scoring of sleep stages by represented the spectral EEG changes occurring during circula-
experts. The automatic system showed agreement with visual tory arrest and recovery. The frequency intervals of these factors
staging close to the interexpert concordance (122). Systems for were 0 to 0.5, 1.5 to 3, 7.5 to 9.5, and 15 to 20 Hz for all chan-
sleep staging in ambulatory conditions have also been developed nels. The sequence of events was similar for all derivations. The
and tested (123) including EEG, EMG, and EOG signals; ICA first EEG change after circulatory arrest was an initial increase
was applied. Compared to a standard sleep analysis system, this in alpha power and a decrease in beta power. On average, after
system yields a concordance of 67.2%. Methods were also devel- approximately 15 seconds alpha power started to decrease, beta
oped with the specific aim of assisting in the diagnosis of sleep power decreased further, delta-1 power started to increase, and
apnea syndrome but not including EEG recordings, which falls delta-2 power started to decrease. After approximately 25 sec-
outside the scope of this chapter. The fact that traditional sleep onds, the delta-1 power increase appeared to plateau or to
staging involves a subjective evaluation of EEG and other physi- decrease. Thus, to detect intraoperative cerebral ischemia, mon-
ologic signals according to the classic Rechtschaffen and Kales itoring of changes in these four frequency ranges is preferable
rules has led to discussing the possibility of substituting these to monitoring changes in the classically defined frequency
rules by an objective classification based on parameters bands. Some special uses of EEG computer analysis in clinical
extracted using automatic signal analysis of sleep. This discus- environments, with particular emphasis on long-term EEG
sion is spurred by the fact that there is considerable interrater monitoring, are discussed in detail in Chapter 35, including
variability and the scoring process is time consuming and sub- long-term monitoring in intensive care units during cerebral
jective as discussed by Nieuwenhuijs (124). This discussion, ischemia and coma, but also during anesthesia in general, and
however, is still ongoing. particularly when neuromuscular blockade is used.
Many departments of anesthesia investigated the possibility
of using a simplified index of EEG activity for this purpose.
QUANTITATIVE EEG IN INTENSIVE Thus, a quantified EEG measure was sought to assist the anes-
MONITORING DURING SURGERY thesiologist in interpreting the changes in EEG signals. EEG
power spectral analysis was explored; in particular the spectral
Automatic intensive monitoring of the EEG is of great impor- edge frequency and the median power frequency were EEG
tance when the cerebral circulation is in acute danger, such as parameters applied but were not found to be reliable in practice
during open heart or carotid surgery, in states of recovery or (133). The search for EEG variables that are valid indices of
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

worsening of cerebral function after brain damage, in coma, or depth of anesthesia led to finding that EEG bispectra can yield
during hemodialysis. Monitoring of cerebral function during useful results. Bispectra quantify the nonlinear phase coupling
extended anesthesia is also of interest. EEG changes during between various frequency components of the EEG signals as
anesthesia are well known (see review of early literature in Ref. presented in Chapter 54 (Expression 54.16 and Fig. 54.4).
125); in this situation, a complex number of factors may affect Investigations of the behavior of EEG bispectra and other spec-
neuronal function, cerebral circulation, and the general tral variables were performed using EEG signals recorded con-
acid–base equilibrium in blood and tissues. Disturbances of tinuously from a bifrontal montage (FP1-Cz and FP2-Cz). EEG
these physiologic functions are reflected in EEG changes. parameters were correlated with the Observers’ Assessment of
Therefore, compressed spectral arrays (126) have been used in Alertness/Sedation (OAA/S) scale (134). The bispectra pre-
monitoring these conditions. Because significant data reduc- sented the strongest correlation with OAA/S scores. With
tion is desired in order to implement real-time EEG monitor- increasing sedation a progressive decrease in the magnitude of
ing at a reasonable cost, it is not surprising that systems have the bispectrum was found. The EEG bispectral magnitude was
been developed based on drastic simplification of the EEG sig- found to be a reliable measure for several anesthetics (135,136).
nal, for example, in the cerebral function monitor (CFM) devel- Monitoring devices based on quantified EEG in this way
oped by Maynard (127), Prior (128), and Prior et al. (129). appeared in the 1990s. These devices entered the field of EEG
Pronk (130) has published a review of computerized methods monitoring during anesthesia, such as the BIS monitor system
in perioperative monitoring. More recently the digital (137), but the results must be interpreted carefully taking into

account the specific EEG of a patient, the anesthetic drugs, and possible without distortion of information, (iii) the graphical
other coadjuvant drugs used (e.g., a baseline EEG recording potential to convey adequate communication to the user, and
should always be made before administration of any drug). (iv) operating ease and flexibility. In this respect, it is of para-
Indeed a multicenter study of bispectral EEG analysis for mon- mount importance that the CADS allows interactive opera-
itoring anesthetic effects has shown that BIS is a significant pre- tion, in order to avoid an overflow of information and thus
dictor of patient response to a surgery incision, but the utility speed up the computations. At the same time, it should enable
of the BIS depends on the anesthetic technique used (138). users to adopt their own strategies of selecting analytic facili-
One clinical situation where quantitative EEG monitoring ties to give information on the most interesting features in a
can be indicated is in those cases where barbiturates are used to particular EEG. Because a generally accepted CADS does not
lower intracranial pressure or to control epileptic seizures that yet exist, only a few indications of the most relevant points
may occur after traumatic brain injury and during convulsive that one should take into consideration when implementing
status epilepticus. It is not easy to control the desired doses of such a system in the clinical laboratory will be given here. To
barbiturate. Barbiturates are usually administered until EEG evaluate the basic method of EEG analysis and the possible
burst suppression appears. Monitoring of barbiturate effects on degree of data reduction, a comparative statistical study of dif-
EEG is necessary to establish the lowest useful dose in real time. ferent EEG analysis techniques, available at the time, has been
A prospective study performed at a pediatric intensive care unit carried out in the same database (142). The database, how-
(139), including children with barbiturate-induced coma after ever, was limited; it consisted of 57 EEG records obtained
traumatic brain injury or generalized convulsive status epilepti- from patients with renal insufficiency (2), hepatic coma (5),
cus, showed that the BIS monitor can help to monitor barbitu- brain injury (3), and patients without organic disease, but
rate-induced coma, but it should be used with caution. under psychotropic treatment (3); moreover, only EEGs
Quantitative EEG, whether or not along with auditory recorded from derivation T3-T5 were analyzed. The EEG
evoked potentials, should be considered in order to monitor records were visually assessed by two independent raters; a
brain functions during a variety of surgical procedures, partic- structured report was used. The EEG records were sampled at
ularly with respect to cardiac surgery with cardiopulmonary 204.8 Hz. The correlation between a number of EEG features
bypass and hypothermic circulatory arrest to detect awareness and the visually assessed degree of abnormality was computed
and indirect memory function, as well as to assess the adequacy and the following features were extracted: the RMS value as
of anesthesia (140). indicator of mean amplitude; mean frequency number of
delta and theta waves calculated using zero-crossing interval
COMPUTER-ASSISTED DIAGNOSTIC SYSTEM analysis; power content in the delta and theta frequency bands
calculated using fast Fourier transform (FFT) and the subdi-
For descriptive reasons it is useful to distinguish two types of vision of frequency bands indicated in Table 56.1 (5); power
computer-assisted diagnostic systems (CADS), because they are content in the delta and theta frequency bands as percentage
based on rather different designs. One pertains to the diagnosis of total power; the ratio between power in the theta and alpha
of what might be called hypofunctional states of brain function, bands (theta/alpha); the ratio between power in the delta plus
commonly characterized by some slowing of the dominant theta bands and that in the alpha plus beta bands (Power[delta
EEG frequency components or the appearance of extra-slow  theta/alpha  beta]); and the so-called EEG age quotient
components; the other pertains to the diagnosis of so-called (5) mentioned previously. All measures defined in terms of
irritative states or different manifestations of epilepsy. power were recomputed in terms of amplitude because ampli-
tude is the unit used when employing analog frequency analy-
CADS and Hypofunctional States sis. A few conclusions can be drawn from this early study.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Most systems currently used in clinical laboratories include Time-domain features give, in general terms, worse results
subroutines designed to detect and evaluate hypofunctional than features obtained using spectral analysis; the two most
states. In this field, several systems combine a high degree of revealing features that emerged from this study were the rela-
sophistication with considerable clinical practicality and tive power in “delta plus theta band” (normalized to total
include comprehensive data reduction and specific displays power) and the EEG age quotient. Friberg et al. (10) contin-
for clinician’s use. Thanks to the increased availability of rela- ued this research line, using mainly the so-called ratio of EEG
tively simple computer systems, many dedicated CADS algo- normality, mentioned earlier, in order to obtain an automatic
rithms were developed in the 1970s, such as those of EEG assessment in a wider group of subjects. The overall
Matousek et al. (141,142), Friberg (143), Friberg et al. (10), agreement rate between automatic and visual EEG interpreta-
Gotman (6,144), Gotman et al. (9,81), Gevins et al. (145), tion in several groups of patients was about 80%. Two types of
Künkel and EEG Project Group (146), Storm van Leeuwen et EEG were difficult to classify: those with an alpha activity of 7
al. (147), Mauslby et al. (148), McGillivray and Wadbrook to 8 Hz, which the program tended to classify as abnormal
(149), Binnie et al. (150), Harner and Ostergren (151), Ebe et (contrary to the visual assessment), and those with very low
al. (152), and Bickford et al. (153). The practical success of the amplitudes. In terms of both informative display and interac-
available systems depends not only on the exact method of tive operation, the most attractive system proposed at about
EEG analysis, but also on (i) the system’s capacity to avoid the same time is that of Gotman et al. (9,55,81). A typical out-
and/or eliminate artifacts, (ii) the degree of data reduction put of the original system is shown in Figure 56.2. The

striking advantages of this form of display are the comprehen- rejecting artifacts that may have characteristics similar to
sive presentation of topographical information and the degree epileptiform transients, such as EMG, lambda waves, vertex
of information compression achieved. Furthermore, it is pos- waves, or K complexes in sleep, and positive occipital sharp
sible to obtain, using such a system in an interactive way, other transients.
outputs of spectral analysis, such as plots of spectra of the Taking into consideration the above discussion, it seems
EEG channels, plots of coherence and phase functions, as well desirable to develop CADS in which users can choose whether
as an output indicative of the variability of the four main fre- they want to have detected those events classified with a high
quency bands. Many of these early basic systems became more probability by a consensus of electroencephalographers as
sophisticated in the course of time. The relevance of EEG fea- epileptiform or all events that any electroencephalographer
tures giving information about changes in symmetry in CADS would accept as epileptiform (64,155). The strategic choice
dedicated to hypofunctional states should be stressed. In this would depend on the clinical setting in which the analysis takes
context, a study carried out in order to investigate clinical rel- place. For instance, in routine clinical EEGs, one would proba-
evance of quantitative EEG parameters in ischemic cere- bly prefer a stringent criterion in order to minimize the chance
brovascular disease should be mentioned. Sheorajpanday et al. of false positives, whereas the investigator monitoring long EEG
(154) studied EEGs of patients suffering from subacute and plasma levels of antiepileptic drugs in known epileptic
ischemic cerebrovascular disease. They used a pairwise patients might be inclined to follow less stringent criteria. The
derived brain symmetry index (pdBSI) as important EEG fea- rule-based algorithms developed by Gotman (156,157) and col-
ture for this analysis and determined correlations between laborators (158,159), and implemented in software packages,
EEG parameters, clinical status, and volume of ischemia on have practical value, especially when the objective is to detect
diffusion-weighted imaging (DWI). The main conclusion was epileptiform transients in long EEG recordings, such as during
that pdBSI could reliably discriminate between stroke and a whole night (see Chapter 35 for details of long-term monitor-
TIA patients or control subjects, and correlated significantly ing). In this respect, the fact that the performance of these algo-
with clinical and radiologic status. rithms takes into account the well-known influence of the state
of the ongoing EEG, namely of sleep stages on the occurrence
CADS and Epileptiform Events of epileptiform events, is particularly valuable. In general, these
In contrast to the systems used to analyze EEG patterns in automatic methods yield a relatively large number of false pos-
hypofunctional states, those that have been derived for auto- itives, and thus it is always necessary to perform a secondary
matic recognition and display of epileptiform events have visual reevaluation of the detected events. Nevertheless, this
received a good deal of attention in the past decade. Major automatic CADS achieves a very comprehensive data reduction.
interest with respect to these systems is in laboratories directly Other algorithms have been proposed that may yield smaller
involved in the diagnosis and care of a population of epileptic rates of false positives (155,160–163). A practical conclusion of
patients, particularly in those locations where extensive EEG such studies is that visually corrected (a posteriori) automatic
investigations using intracranial electrodes are performed as a analysis of epileptiform events is a cost-effective procedure for
guide for neurosurgery and where routine long-term EEG the presurgical evaluation of epileptic patients associated with
recording is carried out in combination with the determination video-EEG monitoring (164).
of plasma levels of antiepileptic drugs. In the section “Transient No matter which detection method is chosen, it is always
Nonstationarities: Epileptiform Events,” we discussed already necessary to provide a comprehensive display of the results so
the main features characteristic of these transients and the that topographical interpretation may be made, particularly
problem of finding a consensus among raters about the identi- with regard to the existence of an irritative area.
fication of such transients by visual inspection; here we con- To validate the clinical relevance of CADS, it is important to
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

sider this issue at the more general level of the compare different methods according to a comprehensive pro-
computer-assisted systems (CADS) dedicated to the diagnosis tocol. Such a study was carried out by Dumpelmann and Elger
in epilepsy. CADS in epilepsy may have different objectives, (75), who reported the results of a comparison of the perform-
such as detecting interictal epileptiform transients, detecting ance of two specialist reviewers and of three spike-detection
epileptic seizures, or localizing an epileptogenic area in the brain. approaches with respect to the detection of epileptiform spikes
In the last decade the development of methods that may permit in intracranial recordings from subdural and intrahippocampal
the automatic anticipation of epileptic seizures generated wide depth electrodes in seven patients. The systems analyzed were
interest (this specialized issue is presented in Chapter 30). (i) the “rule-based system” of Gotman’s group, (ii) the “two-
The principal aims of developing CADS for detecting inter- stage system” consisting of an inverse autoregressive filter and a
ictal epileptiform transients are to quantify long-term variations second rule-based stage of Lopes da Silva’s group (64), and (iii)
in transient occurrence rates, especially in relation to a “wavelet detector” using wavelet coefficients of the intracra-
antiepileptic drug therapy, and to determine the topographical nial EEG (iEEG) data developed by the authors. The results are
distribution of such events. The basic methodologies used have quite revealing: the agreement between the two human review-
been described in Chapter 54. The criteria mentioned above are ers with respect to spike identification was less than 50%. The
imprecise and open to subjective interpretation. Nevertheless, automatic systems achieved agreements of 24% (“rule-based
experience has shown that they have pragmatic value. system”), 26% (“wavelet detector”), and 32% (“two-stage sys-
Furthermore, it is also important to consider the problem of tem”) with the individual human reviewers. In spite of the small

proportion of agreements, the same anatomical regions were the lesion visualized in the MRI. In an interesting study a com-
identified by human and automatic EEG analysis as generators parison between the performance of MEG and EEG recordings
for the majority of spikes. This led the authors to conclude that of epileptiform spikes was performed using intracranial EEG
the poor agreement between the human EEG reviewers suggests recordings as reference. Tanaka et al. (183) investigated the
that the definition of spikes and spike-like episodes in intracra- accuracy of spatiotemporal source analysis of MEG and scalp
nial electrodes is far from unequivocal, although the localizing EEG with respect to the propagation of frontotemporal spikes
information is highly consistent by either visual or automatic in patients with partial epilepsy, using a cortically constrained
spike detection, independent of the algorithm used for auto- minimum norm estimate. Furthermore, iEEGs were recorded
matic spike detection. These conclusions are not really surpris- from temporal and frontal lobes in the context of presurgical
ing since in our experience (64) there is considerable variability evaluation. These authors concluded that the spatiotemporal
in how experienced reviewers score epileptiform transients, analysis of MEG spikes models the time course of frontotempo-
while the performance of the automatic methods described ral spikes as observed on iEEG more adequately than the EEG.
above does not differ appreciably from a consensus of a panel In clinical practice epileptiform spike detectors are currently
of eight reviewers. Wilson et al. (165) made a larger reliability used in digital acquisition software applied on-line. Although
study of the performance of human experts in detecting epilep- these methods may differ in detail, they all derive from the
tiform spikes and concluded that the average interrater correla- results obtained in the previous studies described above (for
tion was 0.79. These authors proposed that this database could review see Ref. 184). In general terms they are based on the
serve as a “gold standard” for testing computer algorithms or identification of the features described above, namely sharp-
other readers. The same group (166) developed later a neural ness, duration, slopes, and relative amplitudes. Furthermore, a
network approach that performed automatic grouping of measure of EEG background state is very important since
spikes via hierarchical clustering (using topology and morphol- detection is always a process of extracting a signal (the spike)
ogy), the performance of which was close to that of human from the background, and the changes of the latter necessarily
experts. Appropriately noting that in spite of the many algo- affect the performance of the detector, as shown by Gotman
rithms that have been developed in the past decades dedicated and Wang (159). In addition, information about the distribu-
to the detection of epileptiform spikes, an ideal system contin- tion in space of the detected events is commonly used
ues to be illusive, Harner (167) proposes a spike/nonspike data- (63,77,184–186). As pointed out by Flanagan et al. (185), the
base as a tool for assessing parameters and methods for computation of equivalent dipole models, using appropriate
automatic epileptiform spike detection that is available from detection and preprocessing methods, as indicated above, pro-
the author. vides a spatial parameter for each detected epileptiform event,
Since the advent of the MEG, it has been assumed that this and this may constitute valuable information that may be read-
new methodology would be useful for the localization of epilep- ily combined with MRI and other relevant data, for the clinical
tiform events (168). The early literature has been reviewed by assessment of a patient. A critical review of spike-detection
Sutherling and Barth (169), Ricci (170), and Engel and Ojeman methods in scalp EEG has recently been published by Halford
(171). In the 1990s, the development of large arrays of sensors (187).
for MEG recordings combined with advanced realistic models A new development in this area of endeavor is the explo-
of the brain and surrounding tissues based on magnetic reso- ration of the possibilities offered by the combination of EEG
nance imaging (MRI) scans has led to a number of interesting and functional MRI (fMRI) recordings (188) in search for ways
investigations with the aim of improving spike source localiza- of improving the localization of sources of epileptiform events.
tion (172–182). A conclusion that can be derived from these Al-Asmi et al. (189) studied patients with focal epilepsy and fre-
and similar studies is that the localization of epileptiform quent spikes who were subjected to spike-triggered or continu-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

sources based on MEG data, especially if combined with EEG ous fMRI with simultaneous EEG. The activated regions in
data, using realistic models of the head based on MRI scans, can fMRI were concordant with EEG localization in almost all stud-
provide valuable diagnostic information of particular interest ies and confirmed by intracerebral EEG in some patients. Bursts
in the evaluation of candidates for epilepsy surgery. The inves- of spikes were more likely to generate an fMRI response than
tigation of Van’t Ent et al. (77), introduced above with respect were isolated spikes. The authors concluded that combining
to clustering methodologies, is an example; after clustering EEG and fMRI in focal epilepsy yields regions of activation that
epileptiform spikes the spikes within one cluster were averaged are presumably the source of spiking activity. These regions are
and the underlying sources were estimated using a single equiv- highly linked with epileptic foci and epileptogenic lesions in a
alent current dipole model (see Fig. 5.4). A dipole was esti- significant number of patients. This research area is most
mated at each time sample during the same time window as promising since the quality of simultaneously recording of EEG
used in the clustering procedure. For the forward computa- and fMRI in epileptic patients is becoming practically reliable
tions, a segmentation of the brain from MRI of each subject was (190).
used as volume conductor. Dipole solutions were accepted only Automatic detection of seizures presents another kind of
when the residual error was less than 10%. The MEG data were problems, as discussed in more detail in Chapter 35. In the ini-
transformed to the MRI coordinate system by matching fiducial tial phase of computer-assisted detection of seizures, the recog-
markers. The estimated dipole locations are presented on MRI, nition of petit mal absences characterized by 3/sec
as shown in Figure 56.5. Some of these clusters are very close to spike-and-wave complexes was one main area of investigation.

Figure 56.5 Characterization of epileptogenic networks in the temporal lobe during the transition from pre-
ictal to seizure activity. A: Intracerebral EEG recoding performed in a patient with mesial TLE. B: Color-coded
nonlinear correlation matrices obtained from the pairwise computation of nonlinear correlation coefficient h2
over six different 10-second intervals chosen during the preictal period (1,2), the ictal period (3 to 5), and
after seizure termination (6). C: Graphical representation in which the lines indicate “abnormally strong” cou-
plings between the two considered structures (graph nodes). Only significantly high interdependencies are
represented (i.e., h2 values greater than 0.32; this value corresponds to the average h2 value computed over
the interictal period C2 standard deviations). Line thickness is proportional to h2 values. (Adapted from
Wendling F, Bartolomei F, Senhadji L. Spatial analysis of intracerebral electroencephalographic signals in the
time and frequency domain: identification of epileptognic networks in partial epilepsy. Phil Trans R Soc A.
2009;367:297–316.) (See color insert)

This is understandable considering that these seizures are rela- this type of CADS in petit mal epilepsy are already widely
tively simple to detect in the EEG, while they are of clinical appreciated by researchers interested in quantitative clinical
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

interest in order to investigate correlations between such phe- studies.

nomena and behavior. The technique proposed by Ehrenburg Quite another sort of problem is the automatic recognition
and Penry (191) was designed to recognize generalized spike- of other types of epileptic seizures, mainly of partial complex
and-wave patterns whose main component, the absence spike, seizures characteristic of temporal lobe epilepsy. The interest in
should be detected by way of a procedure based on zero-cross- automatic detection of this type of seizures stems from the fact
ing analysis. The EEG records were classified visually by three that the central objective of EEG in epilepsy is recording an
raters; the consensus of all three was employed as a criterion for electroclinical seizure. In the early phases of computer-assisted
assigning the program’s correct responses. In a test population diagnosis in epilepsy, technical and computer facilities enabled
of 12 patients, the program agreed with the consensus in 85% recording EEG continuously for further off-line analysis
of the cases, and it had 1% overrecognitions, which correspond (194,195). Babb et al. (196) proposed an analog device with the
to “false positives.” The program’s agreement with the consen- objective of performing automatic seizure detection, based on
sus improved to 92% when all sleep sections were eliminated the recognition of high-frequency activity occurring over sev-
from the analyzed EEG. This particularly well-designed study eral seconds. In this system, false alarms were quite frequent
led the authors to conclude that clinical applications of this (30%). Currently, there is software available for the detection of
CADS will lead to reduced costs over visual EEG assessment. seizures in clinical settings that provides reliable results,
Other CADS with the same main objective as the one described although it is not perfect. In a few laboratories these systems are
above have been developed (52,70,192,193). The advantages of being used in combination with methods of stimulation to

influence the development of seizures. Peters et al. (197) (coherence and phase) between EEG seizure records from differ-
described an integrated bedside system for real-time seizure ent derivations. By determining the phase ( in degrees)
detection and automated delivery of electrical stimulation between pairs of derivations at a frequency (f in Hz) with pro-
directly to the brains of subjects undergoing invasive epilepsy nounced coherence, estimated time delays were computed.
surgery evaluation. These authors conclude that this network Although interesting results have been obtained, mainly in
system is proof of the concept of a portable or implantable cases of seizures recorded using electrodes implanted in limbic
device that could serve identical functions. Viglione et al. (198) structures, it should be noted that this method can give
and Viglione (199) attempted earlier to develop miniature ambiguous results. To decide that there is a time delay depends
automatic seizure recognition and warning systems that could on finding a linear relation between phase and frequency over a
be carried by patients; this system was successful in some cases sufficiently wide frequency band. A difference in phase ()
but led to too many false alarms. corresponding linearly to a difference in frequency (f) repre-
A related question is how to estimate the localization of an sents a time delay computed as t = (/f  360).
epileptogenic area, that is, where within the brain epileptic Alternative methods were proposed by Gersch and Goddard
seizures originate. There have been some efforts to localize pos- (208), Gersch and Tharp (209), and Tharp and Gersch (210).
sible sources of seizures on the basis of scalp recordings, and The latter were able to interpret the origin and spread of seizure
using dipole fitting methods, however, with much difficulty activity within the brain of a patient carrying chronically
(200). This is not surprising since it is not likely that the neu- indwelling electrodes, a conclusion that was not possible on the
ronal networks involved in the initiation of an epileptic seizure basis of visual inspection of the records. Gotman (211) applied
may be anatomically restricted to an area that might corre- the same principles to the analysis of some iEEG records and
spond to a discrete dipolar configuration. Nevertheless, this was was able to show that at a contralateral site the seizure activity
attempted by Kobayashi et al. (201), who developed a noninva- is delayed by a few milliseconds compared with the seizures at
sive method of analysis to localize the source and visualize the the focal area. In the same way Gotman and Levtova (212) were
time course of seizures, and to provide the location and orien- able to determine the relationships between amygdala and hip-
tation of the equivalent dipole generating this activity. This pocampus in temporal lobe seizures.
method was applied to scalp seizures in three patients with tem- From these studies, it became clear that the coherence of,
poral lobe epilepsy and single-focus seizures confirmed by and time delays between, different EEG channels during an
intracerebral recordings. A realistic head model based on MRI epileptic seizure usually change rapidly in the course of time.
was used for computation of field distributions. When seizure This implies that such seizures must be analyzed using short
activity was still not visually identifiable on the scalp, the EEG segments. An interesting method, also aiming at determin-
method demonstrated in all scalp seizures a source in the tem- ing the time relations between different EEG signals in such a
poral neocortex corresponding to the region of seizure activity way that it is possible to estimate the flow of information
in intracerebral recordings. More experience with this kind of between different brain sites, has been proposed by Kaminski
methodology is needed to validate this approach. More recent and Blinowska (213), based on autoregressive models. This
and elaborated methods aiming at providing an automatic method can yield interesting results with respect to how epilep-
warning system for epileptic seizures using intracerebral elec- tiform seizure activities may spread in the brain from a focal
trodes (202) and at the detection of the onset of seizures in area (214).
scalp recordings (203) have been developed and are being used To circumvent the limitations of linearity of the methods
in clinical practice (see Chapter 35 for details). described above, new approaches were pursued. In this context
The problem of detecting seizures has been also approached Mars and van Arragon (215) proposed to compute a measure of
applying ANNs providing an interesting way to seizure detec- the average amount of mutual information (AAMI), in the
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

tion (163,204). The study of Wilson et al. (205) is particularly sense of Gelfand and Yaglom (216), between pairs of EEG sig-
interesting because these authors applied the “Reveal algo- nals as a function of the delay time introduced between both
rithm” to the detection of a large number of seizures from 426 signals. It should be noted that this method is related to the
epilepsy patients combining three methods, novel in their cross-correlation as defined by Equation 54.10, but is more gen-
application to seizure detection: matching pursuit, small neural eral, since AAMI is not constrained by a linear relation between
network rules, and a new connected object hierarchical cluster- both signals. The AAMI method of analysis was used for focus
ing algorithm. They found a sensitivity of 76% with a false-pos- localization in animals having a kindled epileptogenic focus
itive rate of 0.11/h, what was a better performance than of other (217). In this way, time delays could be found for certain phases
comparable methods. of epileptic seizures and the spread pattern of these seizures
The objective of localizing an epileptogenic area, or better an obtained. The same method was also applied to human seizures
epileptogenic network, is particularly important in patients with (218). The algorithms based on AAMI, however, proved to be
complex partial seizures resistant to pharmacologic therapy rather cumbersome to apply in practice. This led to the creation
who are candidates for temporal lobectomy. On the basis of of a new method of nonlinear regression analysis (the h2
iEEG recordings, useful results have been obtained by comput- method of Refs. 219–221). This consists of computing a general
ing time relations between EEG signals recorded from different coefficient of nonlinear fit between any pair of signals. The
sites. In this respect, the pioneer work of Brazier (206,207) was applications of this nonlinear regression coefficient to EEG sig-
particularly influential. She computed cross-power spectra nals recorded during seizures in animals (222) revealed that a

large number of EEG signals recorded from different, but func- license for noncommercial use and open-source development,
tionally related, brain areas present clear nonlinear relations. together with a user tutorial and extensive documentation.
The same applies to EEG signals recorded from intracranial Besides this package several other open-source software tools
electrodes in patients (220,223) and in rats with absence-like are now available in the public domain that have been devel-
seizures (224). Thus, this method offers perspectives for the oped to analyze multichannel biomedical signals including
determination of the site of an epileptogenic focus, based on a neurophysiologic data, such as low-resolution brain electro-
set of simultaneously recorded EEG signals. In a recent study of magnetic tomography (LORETA; standard and exact) devel-
Wendling et al. (225) several methods to assess functional brain oped by Pascual-Marqui et al. (228,229). Commercially
connectivity based on signals recorded from different brain available since 1996, the ASA software has become rather pop-
areas during partial complex epileptic seizures (Fig. 56.5) were ular among clinical and cognitive researchers (230). BioSig is an
applied, namely linear and nonlinear regression, phase syn- open-source software library for brain–computer interfaces,
chronization, and generalized synchronization, using a model- where an overview of other open-source software tools for bio-
based methodology. This comparison revealed that there was medical signal processing is available (231).
no “ideal” method, that is, none of the methods performed bet-
ter than the other ones in all studied situations. Nevertheless, CONCLUDING REMARKS
regression methods (linear or nonlinear) showed sensitivity to
the coupling parameter in all tested models with average or The field of computer-assisted diagnosis based on EEG/MEG
good performances, which leads to the conclusion that it is signals with the aim of automatic analysis of such signals in
advisable to first apply these regression methods in order to clinical situations has developed considerably in the past
characterize functional brain connectivity, under normal or decades. In the 1970s in the early stages of this field, many indi-
pathologic conditions, before using more sophisticated meth- vidual groups proposed their own algorithms that were almost
ods that require specific assumptions about the underlying always tested only in the own environment. These algorithms
model of relationship. In addition, these authors recommend to became more sophisticated and more powerful with the
use time–frequency methods when it is interesting to determine advances in computer facilities and software possibilities.
functional coupling in specific frequency subbands (“fre- Nonetheless, most of these algorithms did not reach a wide
quency-locking”) as in epilepsy. In this context, it should be acceptance in practice. One weakness is that very few studies
noted that the choice of frequency bands is critical (226). This made comprehensive comparisons of the performance of dif-
was approached by Ansari-Asl et al. (227) using a linear estima- ferent algorithms on the same test set and using identical crite-
tor based on the computation of the Pearson product–moment ria. Two interesting contributions that carried out such studies
correlation between EEG signals filtered in narrow and overlap- are that of Dumpelmann and Elger (75), who compared the
ping frequency bands. performance of two specialist reviewers and three spike-detec-
An exciting novel development is the possibility of detecting tion approaches with respect to the detection of epileptiform
changes in the EEG that may occur before an epileptic seizure is spikes, and that of Wendling et al. (225,226) who compared the
manifest in the EEG, that is, to be able to anticipate a seizure. performances of several methods to assess functional brain
Chapter 30 is dedicated to this issue. connectivity, as described above. More of such studies are nec-
essary. In the last two decades new possibilities emerged with
BRAIN MAPPING: METHODOLOGY the enhanced availability of comprehensive software packages
that can be used in the clinical routine EEG laboratory, as men-
Brain mapping is the current term used for the methodology of tioned already. With respect to the latter we may also stress the
representing the EEG activity, either spontaneous or evoked, in need for comparative performance assessment under well-con-
the spatial domain as a topographical map projected onto the
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

trolled laboratory conditions, particularly regarding their

scalp. This field has developed very much in the last decade so that potentiality as valuable tools to enhance EEG/MEG evaluation
the present edition includes a specialized chapter (Chapter 55) in clinical settings.
concerned with the use of EEG as a functional imaging method.
REFERENCES
COMPUTER SYSTEMS FOR ROUTINE
1. Duda RO, Hart PE. Pattern Classification and Scene Analysis. New
APPLICATION IN EEG LABORATORY
York, NY: John Wiley & Sons; 1973.
Computer technology has advanced very rapidly with the gen- 2. Mendel JM, Fu KS, eds. Adaptive Learning and Pattern Recognition
eralized use of microprocessor technology. Thus, nowadays, Systems. New York/London: Academic Press; 1970.
3. Tou JT, Gonzalez RC. Pattern Recognition Principles. Reading, MA:
there is a wide choice of computer systems that can be used in
Addison-Wesley; 1974.
the clinical routine EEG laboratory. These systems offer the
4. Demartini J, Vincent-Carrefour A. Topics on pattern recognition.
possibility of data acquisition, editing and processing the data, In: Remond A, ed. EEG Informatics. A Didactic Review of Methods
artifact rejection, statistical analysis, and brain mapping. An and Applications of EEG Data Processing. Amsterdam: Elsevier;
example is the very comprehensive software package developed 1977:107–126.
by Makeig and collaborators (45,46): EEGLAB that is freely 5. Matousek M, Petersén I. Frequency analysis of the EEG in normal
available (http://www.sccn.ucsd.edu/eeglab/) under a public children and adolescents. In: Kellaway P, Petersen I, eds.

Automation of Clinical Electroencephalography. New York, NY: 22. Isaksson A, Wennberg A. An EEG simulator—a means of objective
Raven Press; 1973:75–102. clinical interpretation of EEG. Electroencephalogr Clin
6. Gotman J. Experiments in the Automation and Quantification of Neurophysiol. 1975;39:313–320.
EEG Interpretation: Localized Brain Lesions and Epilepsy [PhD the- 23. Isaksson A, Wennberg A. Visual evaluation and computer analysis
sis]. Montreal, Canada: McGill University; 1976. of the EEG—a comparison. Electroencephalogr Clin Neurophysiol.
7. Hermann WM, Fichte K, Kubicki S. Mathematische rationale für 1975;38:79–86.
die klinische EEG-Frequenzbander. I. Faktorenanalyse mit EEG 24. Mathieu M, Tirsch W, Poppl SJ. Multichannel online EEG analysis
Powerspektralschtzungen zur Definition von Frequenzbandern. by means of an autoregressive model with applications. In:
EEG-EMG. 1978;9:146–154. Matejcek M, Schenk GK, eds. Die Quantifizierung des
8. Dymond AM, Coger RW, Serafetinides EA. Preprocessing by factor Elektroenzephalogramms. Konstanz: AEG-Telefunken; 1975:
analysis of centro-occipital EEG power and asymmetry from three 475–486.
subject groups. Ann Biomed Eng. 1978;6:108–116. 25. Jansen BH. EEG Segmentation and Classification [PhD thesis].
9. Gotman J, Skuce DR, Thompson CJ, et al. Clinical applications of Amsterdam: Free University; 1979.
spectral analysis and extraction of features from electroencephalo- 26. Jervis BW, Ifeachor EC, Allen EM. The removal of ocular artefacts
grams with slow waves in adult patients. Electroencephalogr Clin from the electroencephalogram: a review. Med Biol Eng Comp.
Neurophysiol. 1973;35:225–235. 1988;26:2–12.
10. Friberg S, Matousek M, Petersen I. A Mathematical Model for the 27. Brunia CHM, Mšcks J, van den Berg-Lenssen MMC, et al.
Age Development of the Background Activity in the Human Correcting ocular artifacts in the EEG: a comparison of several
Electroencephalogram. Technical Report, 2:80. Gothenburg, methods. J Psychophysiol. 1989;3:1–50.
Sweden: Research Laboratory of Medical Electronics, Chalmers 28. Berg P, Scherg M. A multiple sources approach to the correction of
University of Technology; 1980. eye artifacts. Electroencephalgr Clin Neurophysiol. 1994;90:229–241.
11. John ER, Prichep LS, Easton P. Normative data banks and neuro- 29. Gasser T, Sroka L, Mocks J. The transfer of EOG activity into the
metrics: basic concepts, methods and results of norm construc- EEG for eyes open and closed. Electroencephalogr Clin
tion. In: Gevins AS, Remond A, eds. Handbook of Neurophysiol. 1985;61:181–193.
Electroencephalography and Clinical Neurophysiology (Revised 30. Whitton JL, Lue F, Moldofsky H. A spectra method for removing
Series). Vol I: Methods of Analysis of Electrical and Magnetic Signals. eye movement artifacts from the EEG. Electroencephalogr Clin
Amsterdam: Elsevier; 1987. Neurophysiol. 1978;44:735–741.
12. John ER, Prichep LS, Fridman J, et al. Neurometrics: computer- 31. Woestenburg JC, Verbaten MN, Slangen JL. The removal of the
assisted differential diagnosis of brain dysfunctions. Science. eye-movement artifact from the EEG by regression analysis in the
1988;239:162–169. frequency domain. Biol Psychol. 1983;16:127–147.
13. Prichep LS, John ER. Neurometrics: clinical applications. In: Lopes 32. Jervis BW, Nichols MJ, Allen EM, et al. The assessment of two
da Silva FH, Storm van Leeuwen W, Remond A, eds. Clinical methods for removing eye movement artifact from the EEG.
Applications of Computer Analysis of EEG and Other Electroencephalogr Clin Neurophysiol. 1985;61:444–452.
Neurophysiological Signals. Handbook of Electroencephalography 33. Elbert T, Lutzenberger W, Rockstroh B, et al. Removal of ocular
and Clinical Neurophysiology. Vol 2. Amsterdam: Elsevier; artifacts from the EEG—a biophysical approach to the EOG.
1986:153–170. Electroencephalogr Clin Neurophysiol. 1985;60:455–463.
14. Jonkman EJ, Poortvliet DCJ, Veering HM, et al. The use of neuro- 34. Fortgens C, de Bruin MP. Removal of eye movement and EOG
metrics in the study of patients with cerebral ischemia. artifacts from the non-cephalic reference EEG. Electroencephalogr
Electroencephalogr Clin Neurophysiol. 1985;61:333–341. Clin Neurophysiol. 1983;56:90–96.
15. Prichep LS, John ER, Mas F. Neurometric functional imaging: 1. 35. O’Donnell RD, Berkhout J, Adey WR. Contamination of scalp
Subtyping of schizophrenia. In: John ER, ed. Machinery of the EEG spectrum during contraction of craniofacial muscles.
Mind. Boston, MA: Birkhauser; 1990:460–471. Electroencephalogr Clin Neurophysiol. 1974;37:145–151.
16. John ER. The role of quantitative EEG topographic mapping or 36. Gevins AS, Yeager CL, Zeitlin GM, et al. On-line computer rejec-
‘neurometrics’ in the diagnosis of psychiatric and neurological dis- tion of EEG artifact. Electroencephalogr Clin Neurophysiol.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

orders: the pros. Electroencephalogr Clin Neurophysiol. 1977;42:267–274.

1989;73(1):2–4. 37. Barlow JS. Artifact processing in EEG data processing. In: Lopes da
17. Fisch BJ, Pedley TA. The role of quantitative topographic mapping or Silva FH, Storm van Leeuwen W, Remond A, eds. Clinical
‘neurometrics’ in the diagnosis of psychiatric and neurological disor- Application of Computer Analysis of EEG and Other
ders: the cons. Electroencephalogr Clin Neurophysiol. 1989;73:5–9. Neurophysiological Signals. Handbook of Electroencephalography
18. Prichep LS, John ER. QEEG profiles of psychiatric disorders. Brain and Clinical Neurophysiology (New Series) Vol 2. Amsterdam:
Topogr. 1992;4(4):249–257. Elsevier; 1986:15–64.
19. Hansen ES, Prichep LS, Bolwig TG, et al. Quantitative electroen- 38. Ille N, Berg P, Scherg M. Artifact correction of the ongoing EEG
cephalography in OCD patients treated with paroxetine. Clin using spatial filters based on artifact and brain signal topogra-
Electroencephalogr. 2003;34(2):70–74. phies. J Clin Neurophysiol. 2003;19:113–124.
20. Prichep LS, Mas F, Hollander E, et al. Quantitative electroen- 39. Scherg M, Ille N, Bornfleth H, et al. Advanced tools for digital EEG
cephalographic subtyping of obsessive–compulsive disorder. review: virtual source montages, whole head mapping, correlation
Psychiatry Res. 1993;50(1):25–32. and phase analysis. J Clin Neurophysiol. 2002;19:91–112.
21. Zetterberg LH. Experience with analysis and simulation of EEG 40. Lagerlund TD, Sharbrough FW, Busacker NE. Spatial filtering of
signals with parametric description of spectra. In: Kellaway P, multichannel electroencephalographic recordings through princi-
Petersen I, eds. Automation of Clinical Electroencephalography. New pal component analysis by single value decomposition. J Clin
York, NY: Raven Press; 1973:161–201. Neurophysiol. 1997;14:73–82.

41. Jung TP, Makeig S, Westerfield M, et al. Removal of eye activity 59. Lopes da Silva FH, Dijk A, Smits H. Detection of nonstationarities
artifacts from visual event-related potentials in normal and clini- in EEGs using the autoregressive model—an application of EEGs
cal subjects. Clin Neurophysiol. 2000;111:1745–1758. of epileptics. In: Dolce G, Kunkel H, eds. CEAN—Computerized
42. Makeig S, Bell AJ, Jung TP, et al. Independent component analysis EEG Analysis. Stuttgart: Fischer; 1975:180–199.
of electroencephalographic data. In: Touretzky D, Mozer M, 60. Lopes da Silva FH, ten Broeke W, van Hulten K, et al. EEG nonsta-
Hasselmo M, eds. Advances in Neural Information Processing. tionarities detected by inverse filtering in scalp and cortical record-
Cambridge, MA: MIT Press; 1996:145–151. ings of epileptics: statistical analysis and spatial display. In:
43. Vigário RN. Extraction of ocular artefacts from EEG using inde- Kellaway P, Petersen I, eds. Quantitative Analytic Studies in
pendent component analysis. Electroencephalogr Clin Epilepsy. New York, NY: Raven Press; 1976:375–387.
Neurophysiol. 1997;103:395–404. 61. Lopes da Silva FH, van Hulten K, Lommen JG, et al. Automatic
44. Vigário RN, JousmŠki V, Hämäläinen M, et al. Independent com- detection and localization of epileptic foci. Electroencephalogr Clin
ponent analysis for the identification of artefact in magnetoen- Neurophysiol. 1977;43:1–13.
cephalographic recordings. In: Jordan MI, Kearns MJ, Solla SA, 62. Lopes da Silva FH. Analysis of EEG nonstationarities.
eds. Advances in Neural Information Processing Systems. Vol 10. Electroencephalogr Clin Neurophysiol Suppl. 1978;34:163–179.
Cambridge, MA: MIT Press; 1998:229–235. 63. Guedes de Oliveira PHH, Lopes da Silva FH. A topographical dis-
45. Delorme A, Makeig S. EEGLAB: an open source toolbox for analy- play of epileptiform transients based on a statistical approach.
sis of single-trial EEG dynamics including independent compo- Electroencephalogr Clin Neurophysiol. 1980;48:710–714.
nent analysis. J Neurosci Methods. 2004;134(1):9–21. 64. Guedes de Oliveira Q, Lopes da Silva FH. Spike detection based on
46. Delorme A, Fernsler T, Serby H, et al. EEGLAB Tutorial. San Diego, a pattern recognition approach using a microcomputer.
CA: University of San Diego; 2006. Electroencephalogr Clin Neurophysiol. 1983;56(1):97–103.
47. Jasper H, Kershman J. Classification of the EEG in epilepsy. 65. Zetterberg LH. Spike detection by computer and by analog equip-
Electroencephalogr Clin Neurophysiol Suppl. 1949;2:123–131. ment. In: Kellaway P, Petersen I, eds. Automation of Clinical
48. Storm van Leeuwen W, Bickford R, Brazier MAB, et al. Proposal Electroencephalography. New York, NY: Raven Press; 1973:227–242.
for an EEG terminology by the Terminology Committee of the 66. Barlow JS, Dubinsky J. Some computer approaches to continuous
International Federation for Electroencephalography and Clinical automatic clinical EEG monitoring. In: Kellaway P, Petersen I, eds.
Neurophysiology. Electroencephalogr Clin Neurophysiol. Quantitative Analytical Methods in Epilepsy. New York, NY: Raven
1966;20:293–320. Press; 1976:309–327.
49. Chatrian GE, Bergamini L, Dondey M, et al. A glossary of terms 67. Pfurtscheller G, Fischer G. A new approach to spike detection
most commonly used by electroencephalographers. Electro- using a combination of inverse and matched filter techniques.
encephalogr Clin Neurophysiol. 1974;31:538–548. Electroencephalogr Clin Neurophysiol. 1978;44:243–247.
50. Celesia GG, Chen R. Parameters of spikes in human epilepsy. Dis 68. Smith JR. Automatic analysis and detection of EEG spikes. IEEE
Nerv Syst. 1976;37:277–281. Trans Biomed Eng. 1972;BME-21:1–7.
51. Gose EE, Werner S, Bickford RG. Computerized spike detection. 69. Ktonas PY, Smith JR. Quantification of abnormal EEG spike char-
Proc San Diego Biomed Symp. 1974;13:193–198. acteristics. Comput Biol Med. 1974;4:157–163.
52. Carrie JRG. Computer-assisted EEG sharp transient detection and 70. Frost JD Jr. Microprocessor-based EEG spike detection and quan-
quantification during overnight recording in an epileptic patient. tification. Int J Biomed Comput. 1979;10:357–373.
In: Kellaway P, Petersen I, eds. Quantitative Analytic Studies in 71. Gotman J. Quantitative measurements of epileptic spike morphol-
Epilepsy. New York, NY: Raven Press; 1976:225–235. ogy in the human EEG. Electroencephalogr Clin Neurophysiol.
53. Gevins AS, Yeager CL, Diamond SL, et al. Sharp-transient analysis 1980;48:551–557.
and thresholded linear coherence spectra of paroxysmal EEGs. In: 72. Gloor P. The EEG and differential diagnosis of epilepsy. In: van
Kellaway P, Petersen I, eds. Quantitative Analytical Studies in Duijn H, Donker DN, van Huffelen AC, eds. Current Concepts in
Epilepsy. New York, NY: Raven Press; 1976:463–482. Clinical Neurophysiology. The Hague: Trio; 1977:9–21.
54. Goldberg P, Samson-Dollfus D, Gremy F. An approach to an auto- 73. Lemieux JF, Blume WT. Automated morphological analysis of
matic pattern recognition of the electroencephalogram: back- spikes and sharp waves in human electrocorticograms.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

ground rhythm and paroxysmal elements. Methods Inf Med. Electroencephalogr Clin Neurophysiol. 1983;55:45–50.
1973;12:155–163. 74. Gotman J, Marciani MG. EEG spiking activity, drug levels and
55. Gotman J, Gloor P. Automatic recognition and quantification of seizure occurrence in epileptic patients. Ann Neurol. 1985;17:
interictal epileptic activity in the human scalp EEG. 597–603.
Electroencephalogr Clin Neurophysiol. 1976;41:513–529. 75. Dumpelmann M, Elger CE. Visual and automatic investigation of
56. Harner RN, Ostergren KA. Sequential analysis of quasi-stable and epileptiform spikes in intracranial EEG recordings. Epilepsia.
paroxysmal activity. In: Kellaway P, Petersen I, eds. Quantitative 1999;40(3):275–285.
Analytic Studies in Epilepsy. New York, NY: Raven Press; 76. Wahlberg P, Lantz G. Methods for robust clustering of epileptic
1976:343–353. EEG spikes. IEEE Trans Biomed Eng. 2000;47:857–868.
57. Saltzberg B, Lustick LS, Heath RG. Detection of focal spiking in the 77. Van’t Ent D, Manshanden I, Ossenblok P, et al. Spike cluster analysis
scalp EEG of monkeys. Electroencephalogr Clin Neurophysiol. in neocortical localization related epilepsy yields clinically signifi-
1971;31:327–333. cant equivalent source localization results in magnetoencephalo-
58. Lopes da Silva FH, Dijk A, Smits H, et al. Automatic detection and gram (MEG). Clin Neurophysiol. 2003;114(10): 1948–1962.
pattern recognition of epileptic spikes from surface and depth 78. Abraham-Fuchs K, Harer W, Schneider S, et al. Pattern recognition
recording in man. In: Schenk GK, ed. Die Quantifizierung des in biomagnetic signals by spatio-temporal correlation and appli-
Elektroenzephalogramms. Konstanz: AEG-Telefunken; 1973: cation to the localisation of propagating neuronal activity. Med
425–436. Biol Eng Comput. 1990.

79. Volavka J, Matousek M, Feldstein S, et al. The reliability of EEG 100. Kemp B. Model-based Monitoring of Human Sleep Stage [PhD
assessment. EEG-EMG. 1973;4:123–130. thesis]. Enschede, the Netherlands: University of Twente; 1987.
80. Rose SW, Penry JK, White BG, et al. Reliability and validity of 101. Dement W, Kleitman N. Cyclic variations in EEG during sleep
visual EEG assessment in third grade children. Clin Electro- and their relation to eye movements, body motility dreaming.
encephalogr. 1973;4:197–205. Electroencephalogr Clin Neurophysiol. 1957;9:673–690.
81. Gotman J, Gloor P, Schaul N. Comparison of traditional reading of 102. Rechtschaffen A, Kales A. Manual of Standardized Terminology,
the EEG and automatic recognition of interictal epileptic activity. Techniques and Scoring System for Sleep Stages of Human Subjects.
Electroencephalogr Clin Neurophysiol. 1978;44:48–60. National Institutes of Health Publication No. 204. Washington,
82. Gevins AS. Quantitative aspects of electroencephalography. In: DC: U.S. Government Printing Office; 1968.
Aminoff MJ, ed. Electrophysiological Approaches to Neurological 103. Künkel H. Die Spektraldarstellung des EEG. EEG-EMG.
Diagnosis. London: Churchill Livingstone; 1979. 1972;3:15–24.
83. Bodenstein G, Praetorius HM. Feature extraction from the 104. Larsen LE, Walter DO. On automatic methods of sleep staging by
encephalogram by adaptive segmentation. Proc IEEE. EEG spectra. Electroencephalogr Clin Neurophysiol.
1977;65:642–652. 1970;28:459–467.
84. Anderberg MR. Cluster Analysis for Application. New York, NY: 105. Lubin A, Johnson LC, Austin NT. Discriminations among states
Academic Press; 1973. of consciousness using EEG spectra. Psychophysiology.
85. Guess MJ, Wilson SB. Introduction to hierarchical clustering. 1969;6:122–132.
J Clin Neurophysiol. 2002;19:144–151. 106. Smith JR, Karacan I. EEG sleep stage scoring by an automatic
86. Gevins AS, Morgan NH. Classifier-directed signal processing in hybrid system. Electroencephalogr Clin Neurophysiol.
brain research. IEEE Trans Biomed Eng. 1986;BME- 1971;31:231–237.
33(12):1054–1068. 107. Itil TM, Shapiro DM, Fink M, et al. Digital computer classifica-
87. Pfurtscheller G, Flozinger G, Mohl W. Prediction of the Side of tions of EEG sleep stages. Electroencephalogr Clin Neurophysiol.
Hand Movements from Single-trial Multi-channel EEG Data using 1969;27:76–83.
a Neural Network; Preliminary Results. Report 299. Graz: Institutes 108. Rossler R, Collins R, Rostman A. A period analysis, classification
for Information Processing; 1991:1–11. of sleep stages. Electroencephalogr Clin Neurophysiol.
88. Kohonen T. Self-organizing Maps. Springer Series in Information 1970;29:358–362.
Sciences. Vol 30. 3rd ed. Berlin/Heidelberg/New York: Springer; 109. Martin WB, Johnson LC, Viglione SS, et al. Pattern recognition of
2001:501 pp. EEG–EOG as a technique for all-night sleep stage scoring.
89. Peters BO, Pfurtscheller G, Flyvbjerg H. Mining multi-channel Electroencephalogr Clin Neurophysiol. 1972;32:417–427.
EEG for its information content: an ANN-based method for a 110. Viglione SS, Martin WB. Automatic analysis of the EEG for sleep
brain–computer interface. Neural Netw. 1998;11(7–8):1429–1433. staging. In: Kellaway P, Petersen I, eds. Automation of Clinical
90. Johnson LC. The EEG during sleep as viewed by a computer. In: Electroencephalography. New York, NY: Raven Press; 1973:
Remond A, ed. EEG Informatics. A Didactic Review of Methods and 269–285.
Applications of EEG Data Processing. Amsterdam: Elsevier; 111. Gaillard JM, Krassvievitch M, Tissot R. Analyse automatique du
1977:385–406. sommeil par un systeme hybride. Electroencephalogr Clin
91. Johnson L, Lubin A, Nautoh P, et al. Spectral analysis of the EEG Neurophysiol. 1972;33:403–410.
of dominant and non-dominant alpha subjects during waking 112. Gaillard JM, Tissot R. Principles of automatic analysis of sleep
and sleeping. Electroencephalogr Clin Neurophysiol. 1969;26: records with a hybrid system. Comput Biomed Res. 1973;6:1–13.
361–370. 113. Poppl SJ. Computer allocation rules for automatic EEG classifi-
92. Gondeck AR, Smith JR. Dynamics of human sleep sigma spindles. cation. In: Dolce G, Kunkel H, eds. CEAN—Computerized EEG
Electroencephalogr Clin Neurophysiol. 1974;37:293–297. Analysis. Stuttgart: Fischer; 1975:202–215.
93. Kemp B, Jaspers P, Franzen JM, et al. An optimal monitor of the 114. Bowe TR, Anders TF. The use of the semi-Markow model in the
electroencephalographic sigma sleep state. Biol Cybern. 1985;51: study of the development of sleep–wake states in infants.
263–270. Psychophysiology. 1979;16:41–48.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

94. Smith JR, Karacan I, Yang M. Ontogeny of delta activity during 115. Yang MCK, Hursch CJ. The use of a semi-Markow model for
human sleep. Electroencephalogr Clin Neurophysiol. 1977;43: describing sleep pattern. Biometrics. 1973;29:667–676.
229–237. 116. Zung WWK, Naylor TH, Gianturco D, et al. A Markov chain
95. Steriade M, Contreras D, Curro Dossi R, et al. The slow model of sleep EEG pattern. Electroencephalogr Clin
(1 Hz) oscillation in reticular thalamic and thalamocortical Neurophysiol. 1965;19:105.
neurons: scenario of sleep rhythm generation in interacting 117. Kemp B, Kamphuisen HAC. Simulation of human hypnograms
thalamic and neocortical networks. J Neurosci. 1993;13(8): using a Markov chain model. Sleep. 1986;9:405–414.
3284–3299. 118. Hermann WM, Kubicki S. Various techniques of computer analy-
96. Achermann PA, Borbély A. Low-frequency (1 Hz) oscillations in sis in nocturnal sleep. In: Degen R, Niedermeyer E, eds. Epilepsy,
the human sleep. Neuroscience. 1997;81:213–222. Sleep and Sleep Deprivation. Amsterdam: Elsevier; 1984:207–229.
97. Simon NR, Manshanden I, Lopes da Silva FH. A MEG study of 119. Agarwal R, Gotman J. Computer-assisted sleep staging. IEEE
sleep. Brain Res. 2000;860(1–2):64–76. Trans Biomed Eng. 2001;48:1412–1423.
98. Bremer G, Smith JR, Karacan I. Automatic detection of the K com- 120. Agarwal R, Gotman J. Digital tools in polysomnography. J Clin
plex in sleep electroencephalograms. IEEE Trans Biomed Eng. Neurophysiol. 2002;19:136–143.
1970;BME-17:314–323. 121. Malinowska U, Durka PJ, Blinowska KJ, et al. Micro- and
99. Rosa AC, da Kemp B, Paiva T, et al. A model-based detector of ver- macrostructure of sleep EEG: a universal adaptive time–
tex waves and K complexes in sleep electroencephalogram. frequency parameterization. IEEE Eng Med Biol Mag.
Electroencephalogr Clin Neurophysiol. 1991;78:71–79. 2006;25(4):26–31.

122. Malinowska U, Klekowicz H, Wakarow A, et al. Fully parametric Technical Report, 3:80. Gothenburg, Sweden: Research
sleep staging compatible with the classical criteria. Laboratory of Medical Electronics, Chalmers University of
Neuroinformatics. 2009;7(4):245–253. Technology; 1980.
123. Porée F, Kachenoura A, Gauvrit H, et al. Blind source separation 144. Gotman J. Problems of presentation of analytical results.
for ambulatory sleep recording. IEEE Trans Inf Technol Biomed. Electroencephalogr Clin Neurophysiol Suppl. 1978;34:191–197.
2006;10(2):293–301. 145. Gevins AS, Yeager CL, Diamond SL, et al. Automated analysis of
124. Nieuwenhuijs DJ. Processed EEG in natural sleep. Best Pract Res the electrical activity of the human brain (EEG): a progress
Clin Anaesthesiol. 2006;20(1):49–56. report. Proc IEEE. 1975;63:1382–1399.
125. Brechner VL, Walter RD, Dillon JB. Practical 146. Künkel H, EEG Project Group. Hybrid computing system for
Electroencephalography for the Anesthesiologist. Springfield, IL: EEG analysis. In: Dolce G, Kunkel H, eds. CEAN—Computerized
Charles C. Thomas; 1962. EEG Analysis. Stuttgart: Fischer; 1975:365–383.
126. Bickford RG, Billinger TW, Fleming NI, et al. The compressed 147. Storm van Leeuwen W, Arntz A, Spoclstra P, et al. The use of
spectral array (CSA). A pictorial EEG. Proc San Diego Biomed computer analysis for diagnosis in routine electroencephalogra-
Symp. 1972;11:365–370. phy. Rev EEG Neurophysiol. 1976;6(2):318–327.
127. Maynard DE. The cerebral function analysis monitor (CFAM). 148. Mauslby RL, Saltzberg B, Lustick LS. Toward an EEG screening
Electroencephalogr Clin Neurophysiol. 1977;43:479. test: a simple system for analysis and display of clinical EEG data.
128. Prior P. Monitoring Cerebral Function: Long-term Recordings of In: Kellaway P, Petersen I, eds. Automation of Clinical
Cerebral Electrical Activity. Amsterdam: Elsevier; 1979. Electroencephalography. New York, NY: Raven Press; 1973:45–53.
129. Prior PF, Maynard DE, Sheaff PC, et al. Monitoring cerebral 149. McGillivray BB, Wadbrook DG. A system for extracting a diagno-
function: clinical experience with new device for continuous sis from the clinical EEG. In: Dolce G, Kunkel H, eds. CEAN—
recording of electrical activity of brain. Br Med J. 1971;2:736. Computerized EEG Analysis. Stuttgart: Fischer; 1975:344–364.
130. Pronk RAF. Peri-operative monitoring. In: Lopes da Silva FH, 150. Binnie CD, Batchelor BG, Bawring PA, et al. Computer-assisted
Storm van Leeuwen W, Remond A, eds. Clinical Applications of interpretation of clinical EEGs. Electroencephalogr Clin
Computer Analysis of EEG and Other Neurophysiological Signals. Neurophysiol. 1978;44:575–585.
Handbook of Electroencephalography and Clinical 151. Harner RN, Ostergren KA. Computed EEG topography.
Neurophysiology. Vol 2. Amsterdam: Elsevier; 1986:93–130. Electroencephalogr Clin Neurophysiol Suppl. 1978;34:151–161.
131. Scheuer ML. Continuous EEG monitoring in the intensive care 152. Ebe M, Homma I, Ishiyama Y, et al. Automatic analysis of clinical
unit. Epilepsia. 2002;43(suppl 3):114–127. information in EEG. Electroencephalogr Clin Neurophysiol.
132. Visser GH, Wieneke GH, Van Huffelen AC, et al. The develop- 1973;34:706.
ment of spectral EEG changes during short periods of circulatory 153. Bickford RG, Brimm J, Berger L, et al. Application of compressed
arrest. J Clin Neurophysiol. 2001;18(2):169–177. spectral array in clinical EEG. In: Kellaway P, Petersen I, eds.
133. Rampil IJ. A primer for EEG signal processing in anesthesia. Automation of Clinical Electroencephalography. New York, NY:
Anesthesiology. 1998;89:980–1002. Raven Press; 1973:55–64.
134. Liu J, Singh H, White PF. Electroencephalogram bispectral analy- 154. Sheorajpanday RV, Nagels G, Weeren AJ, et al. Reproducibility and
sis predicts the depth of midazolam-induced sedation. clinical relevance of quantitative EEG parameters in cerebral
Anesthesiology. 1996;84(1):64–69. ischemia: a basic approach. Clin Neurophysiol. 2009;120(5):845–855.
135. Kearse LA Jr, Manberg P, Chamoun N, et al. Bispectral analysis of 155. Webber WRS, Harner RN, Duffy FH, et al. Spike detection. I.
the electroencephalogram correlates with patient movement to Correlation and reliability of human experts. Electroencephalogr
skin incision during propofol/nitrous oxide anesthesia. Clin Neurophysiol. 1996;98:186–198.
Anesthesiology. 1994;81(6):1365–1370. 156. Gotman J. Automatic seizure detection: improvements and eval-
136. Sigl JC, Chamoun NG. An introduction to bispectral analysis for uation. Electroencephalogr Clin Neurophysiol. 1990;76:317–324.
the electroencephalogram. J Clin Monit. 1994;10:392–404. 157. Gotman J. Automatic detection of seizures and spikes in the EEG.
137. Bard JW. The BIS monitor: a review and technology assessment. In: Lÿders H, ed. Epilepsy Surgery. New York, NY: Raven Press;
AANA J. 2001;69(6):477–483. 1991:307–316.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

138. Sebel PS, Lang E, Rampil IJ, et al. A multicenter study of bispec- 158. Gotman J, Wang LY. State-dependent spike detection: concepts
tral electroencephalogram analysis for monitoring anesthetic and preliminary results. Electroencephalogr Clin Neurophysiol.
effect. Anesth Analg. 1997;84(4):891–899. 1991;79:11–19.
139. Prins SA, de Hoog M, Blok JH, et al. Continuous noninvasive 159. Gotman J, Wang LY. State-dependent spike detection: validation.
monitoring of barbiturate coma in critically ill children using the Electroencephalogr Clin Neurophysiol. 1992;83:12–18.
bispectral index monitor. Crit Care. 2007;11(5):R108. 160. Hosteler W, Doller HJ, Homan RW. Assessment of a computer
140. Sebel PS. Central nervous system monitoring during open heart program to detect epileptiform spikes. Electroencephalogr Clin
surgery: an update. J Cardiothorac Vasc Anesth. 1998;12(2 suppl Neurophysiol. 1992;83:1–11.
1):3–8. 161. Webber WRS, Litt B, Lesser RP, et al. Automatic EEG spike detec-
141. Matousek M, Petersén I, Friberg S. Automatic assessment of ran- tion: what should the computer imitate? Electroencephalogr Clin
domly selected routine EEG records. In: Dolce G, Kunkel H, eds. Neurophysiol. 1993;87:364–373.
CEAN—Computerized EEG Analysis. Stuttgart: Fischer; 162. Webber WRS, Litt B, Wilson K, et al. Practical detection of
1975:421–428. epileptiform discharges in the EEG using an artificial neural net-
142. Matousek M, Arvidsson A, Friberg S. Implementation of analyt- work: a comparison of raw and parameterized EEG data.
ical methods in daily clinical EEG routine. Electroencephalogr Electroencephalogr Clin Neurophysiol. 1994;91:194–204.
Clin Neurophysiol Suppl. 1978;34:199–204. 163. Webber WRS, Lesser RP, Richardson RT, et al. An approach to
143. Friberg S. A Program System for the Automatic Evaluation of the seizure detection using an artificial neural network.
Background Activity in the Human Electroencephalogram. Electroencephalogr Clin Neurophysiol. 1996;98:250–272.

164. Spatt J, Pelzi G, Mamoli B. Reliability of automatic and visual cephalographic and electroencephalographic source analysis.
analysis of interictal spikes in lateralizing an epileptic focus dur- Neuroimage. 2010;50(1):217–222.
ing video-EEG monitoring. Electroencephalogr Clin Neurophysiol. 184. Gotman J. Automatic detection of seizures and spikes. J Clin
1997;103:421–425. Neurophysiol. 1999;16:130–140.
165. Wilson SB, Harner RN, Dufy FH, et al. Spike detection. I. 185. Flanagan D, Agarwal R, Gotman J. Computer-aided spatial clas-
Correlation and reliability of human experts. Electroencephalogr sification of epileptic spikes. J Clin Neurophysiol. 2002;19:
Clin Neurophysiol. 1996;98:186–198. 125–135.
166. Wilson SB, Turner CA, Emerson RG, et al. Spike detection II: 186. Glover JR Jr, Raghavan N, Ktonas PY, et al. Context-based auto-
automatic, perception-based detection and clustering. Clin mated detection of epileptogenic sharp transients in the EEG:
Neurophysiol. 1999;110(3):404–411. elimination of false positives. IEEE Trans Biomed Eng.
167. Harner R. Automatic EEG spike detection. Clin EEG Neurosci. 1989;36:519–527.
2009;40(4):262–270. 187. Halford JJ. Computerized epileptiform transient detection in the
168. Rose DF, Smith PD, Sato S. Magnetoencephalography and scalp electroencephalogram: obstacles to progress and the exam-
epilepsy research. Science. 1987;238:329–335. ple of computerized ECG interpretation. Clin Neurophysiol.
169. Sutherling WW, Barth DS. Magnetoencephalography in clinical 2009;120(11):1909–1915.
epilepsy studies: the UCLA experience. Magnetoencephalogr Adv 188. Ives JR, Warach S, Schmitt F, et al. Monitoring the patient’s EEG
Neurol. 1990;54:231–246. during echo planar MRI. Electroencephalogr Clin Neurophysiol.
170. Ricci GB. Italian contributions to magnetoencephalographic 1993;87(6):417–420.
studies on the epilepsies. Magnetoencephalogr Adv Neurol. 189. Al-Asmi A, Benar CG, Gross DW, et al. fMRI activation in con-
1990;54:247–260. tinuous and spike-triggered EEG–fMRI studies of epileptic
171. Engel J, Ojeman GA. The next step. In: Engel J, ed. Surgical spikes. Epilepsia. 2003;44(10):1328–1339.
Treatment of the Epilepsies. 2nd ed. New York, NY: Raven Press; 190. Salek-Haddadi A, Lemieux L, Merschhemke M, et al. EEG
1993. quality during simultaneous functional MRI of interictal
172. Ebersole JS, Squires KC, Eliashiv SD, et al. Applications of mag- epileptiform discharges. Magn Reson Imaging. 2003;21(10):
netic source imaging in evaluation of candidates for epilepsy sur- 1159–1166.
gery. Funct Imaging. 1995;5:267–287. 191. Ehrenburg BL, Penry JK. Computer recognition of generalized
173. Hari R, Ahonen A, Forss N, et al. Parietal epileptic mirror focus spike-wave discharges. Electroencephalogr Clin Neurophysiol.
detected with a whole-head neuromagnetometer. Neuroreport. 1976;41:25–36.
1993;5:45–48. 192. Carrie JRG. A hybrid computer technique for detecting sharp
174. Kettenmann B, Hummel C, Stefan H, et al. Multichannel magne- EEG transients. Electroencephalogr Clin Neurophysiol.
toencephalographical recordings: separation of cortical 1972;33:336–338.
responses to different chemical stimulation in man. 193. Carrie JRG, Frost JD. Clinical evaluation of a method for quan-
Electroencephalogr Clin Neurophysiol Suppl. 1996;46:271–274. tification of generalized spike-wave EEG patterns by computer
175. Knowlton RC, Laxer KD, Aminoff MJ, et al. Magneto- during prolonged recordings. Comput Biomed Res.
encephalography in partial epilepsy: clinical yield and localiza- 1977;10:449–457.
tion accuracy. Ann Neurol. 1997;42(4):622–631. 194. Ives JR, Thompson CJ, Gloor P. Seizure monitoring: a new tool in
176. Lewine JD, Orrison WW Jr. Magnetoencephalography and mag- electroencephalography. Electroencephalogr Clin Neurophysiol.
netic source imaging. In: Functional Brain Mapping. St. Louis: 1976;41:422–427.
Mosby-Year Book; 1995:369–417. 195. Kamp A, Mars NJI, Wisman T. Long term monitoring of the elec-
177. Merlet I, Garcia-Larrea L, Grégoire MC, et al. Source propagation troencephalogram in epileptic patients. In: Amlaner CJ Jr,
of interictal spikes in temporal lob epilepsy. Correlation between MacDonald DW, eds. Handbook on Biotelemetry and Radio
spike dipole modelling and [18F]fluorodeoxyglucose PET data. Tracking. Oxford: Pergamon Press; 1979:499–503.
Brain. 1996;119:377–392. 196. Babb TL, Mariani E, Crandall PH. An electronic circuit for detec-
178. Paetau R, Kajola M, Karhu J, et al. Magnetoencephalographic tion of EEG seizures recorded with implanted electrodes.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

localization of epileptic cortex—impact on surgical treatment. Electroencephalogr Clin Neurophysiol. 1974;37:305–308.

Ann Neurol. 1992;32:106–109. 197. Peters TE, Bhavaraju NC, Frei MG, et al. Network system for
179. Paetau R, Hamalainen M, Hari R, et al. Magnetoen- automated seizure detection and contingent delivery of therapy.
cephalographic evaluation of children and adolescents with J Clin Neurophysiol. 2001;18(6):545–549.
intractable epilepsy. Epilepsia. 1994;35:275–284. 198. Viglione SS, Ordon VA, Risch F. A Methodology for Detecting
180. Roth BJ, Ko D, von Albertini-Carletti IR, et al. Dipole localization Ongoing Changes in the EEG Prior to Clinical Seizures. West
in patients with epilepsy using the realistically shaped head Huntington Beach, CA: McDonnell Douglas Astronautics Co;
model. Electroencephalogr Clin Neurophysiol. 1997;102:159–166. 1970. MDAC Paper WD 1399.
181. Smith JR, Schwartz BJ, Gallen Ch, et al. Utilization of multichan- 199. Viglione SS. Validation of Epilepsy Seizure Warning System. West
nel magnetoencephalography in the guidance of ablative seizure Huntington Beach, CA: McDonnell Douglas Astronautics Co;
surgery. J Epilepsy. 1995;8:119–130. 1974. MDAC Paper APA 74133.
182. Stephan H, Schuler P, Abraham-Fuchs K, et al. Magnetic source 200. Gotman J. Noninvasive methods for evaluating the localization
localization and morphological changes in temporal lobe and propagation of epileptic activity. Epilepsia. 2003;44(suppl
epilepsy: comparison of MEG/EEG, ECoG and volumetric MRI 12):21–29.
in presurgical evaluation of operated patients. Acta Neurol Scand 201. Kobayashi K, James CJ, Yoshinaga H, et al. The electroencephalo-
Suppl. 1994;152:83–88. gram through a software microscope: non-invasive localization
183. Tanaka N, Hämäläinen MS, Ahlfors SP, et al. Propagation of and visualization of epileptic seizure activity from inside the
epileptic spikes reconstructed from spatiotemporal magnetoen- brain. Clin Neurophysiol. 2000;111(1):134–149.

202. Grewal S, Gotman J. An automatic warning system for epileptic 221. Pijn JPM, Vijn PCM, Lopes da Silva FH, et al. The use of signal
seizures recorded on intracerebral EEGs. Clin Neurophysiol. analysis for the localization of an epileptic focus: a new approach.
2005;116(10):2460–2472. Adv Epileptol. 1989;17:272–276.
203. Saab ME, Gotman J. A system to detect the onset of epileptic 222. Fernandes de Lima VM, Pijn JPM, Filipe CN, et al. The role of
seizures in scalp EEG. Clin Neurophysiol. 2005;116(2):427–442. hippocampal commissures in the interhemispheric transfer of
204. Gabor AJ, Leach RR, Dowla FU. Automated seizure detection epileptiform after-discharges in the rat: a study using linear and
using a self-organising neural network. Electroencephalogr Clin nonlinear regression analysis. Electroencephalogr Clin
Neurophysiol. 1996;99:257–266. Neurophysiol. 1990;76:520–540.
205. Wilson SB, Scheuer ML, Emerson RG, et al. Seizure detection: 223. Pijn JP, Lopes da Silva FH. Propagation of electrical activity: non-
evaluation of the Reveal algorithm. Clin Neurophysiol. linear associations and time delays between EEG signals. In:
2004;115(10):2280–2291. Zschocke S, Speckmann E-J, eds. Basic Mechanisms of the EEG.
206. Brazier MAB. Spread of seizure discharges in epilepsy: anatomi- Boston, MA: Birkhauser; 1993:41–61.
cal and electrophysiological considerations. Exp Neurol. 224. Meeren HK, Pijn JP, Van Luijtelaar EL, et al. Cortical focus drives
1972;36:263–272. widespread corticothalamic networks during spontaneous
207. Brazier MAB. Electrical seizure discharges within the human absence seizures in rats. J Neurosci. 2002;22(4):1480–1495.
brain: the problem of spread. In: Brazier MAB, ed. Epilepsy, Its 225. Wendling F, Ansari-Asl K, Bartolomei F, et al. From EEG signals
Phenomena in Man. New York, NY: Academic Press; to brain connectivity: a model-based evaluation of interdepend-
1973:155–171. ence measures. J Neurosci Methods. 2009;183(1):9–18.
208. Gersch W, Goddard G. Locating the site of epileptic focus by 226. Wendling F, Bartolomei F, Senhadji L. Spatial analysis of intrac-
spectral analysis methods. Science. 1970;169:701–702. erebral electroencephalographic signals in the time and fre-
209. Gersch W, Tharp BR. Spectral regression—amount of informa- quency domain: identification of epileptognic networks in
tion analysis of seizures in humans. In: Kellaway P, Petersen I, eds. partial epilepsy. Phil Trans R Soc A. 2009;367:297–316.
Quantitative Analytic Studies in Epilepsy. New York, NY: Raven 227. Ansari-Asl K, Bellanger JJ, Bartolomei F, et al. Time–frequency
Press; 1976:509–532. characterization of interdependencies in nonstationary signals:
210. Tharp BR, Gersch W. Spectral analysis of seizures in humans. application to epileptic EEG. IEEE Trans Biomed Eng.
Comput Biomed Res. 1976;8:503–521. 2005;52:1218–1226.
211. Gotman J. Interhemispheric relations during bilateral spike-and- 228. Pascual-Marqui RD, Esslen M, Kochi K, et al. Functional imaging
wave activity. Epilepsia. 1981;22:453–466. with low-resolution brain electromagnetic tomography
212. Gotman J, Levtova V. Amygdala–hippocampus relationships in (LORETA): a review. Methods Find Exp Clin Pharmacol.
temporal lobe seizures: a phase-coherence study. Epilepsy Res. 2002;24(suppl C):91–95.
1996;25:51–57. 229. Pascual-Marqui RD. Discrete, 3D Distributed, Linear Imaging
213. Kaminski MJ, Blinowska KJ. A new method of the description of Methods of Electric Neuronal Activity. Part 1: Exact, Zero Error
the information flow in the brain structures. Biol Cybern. Localization. arXiv:0710.3341 [math-ph], http://arxiv.org/pdf/
1991;65:203–210. 0710.3341. Accessed October 17, 2007.
214. Franaszczuk PJ, Bergey GK, Kaminski MJ. Analysis of mesial 230. Zanow F, Knösche TR. ASA—advanced source analysis of contin-
temporal seizure onset and propagation using the directed trans- uous and event-related EEG/MEG signals. Brain Topogr.
fer function method. Electroencephalogr Clin Neurophysiol. 2004;16(4):287–290.
1994;91(6):413–427. 231. Schlögl A, Keinrath C, Zimmermann D, et al. A fully automated
215. Mars NJI, van Arragon GW. Time delay estimation in nonlinear correction method of EOG artifacts in EEG recordings. Clin
systems using average mutual amount of information analysis. Neurophysiol. 2007;118(1):98–104.
IEEE Acoustics Speech Signal Processing. 1981;29(3):619–621. 232. Harner RN. Computer analysis and clinical EEG interpretation—
216. Gelfand IM, Yaglom AM. Calculation of the amount of informa- perspective and application. In: Dolce G, Kunkel H, eds. CEAN—
tion about a random function contained in another such func- Computerized EEG Analysis. Stuttgart: Fischer; 1975:337–343.
tion. Am Math Soc Transact. 1959;12:199–246. 233. Rappelsberger P. The reference problem and mapping of coher-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

217. Mars NJI, Lopes da Silva FH. Propagation of seizure activity in ence: a stimulation study. Brain Topogr. 1989;2:63–72.
kindled dogs. Electroencephalogr Clin Neurophysiol. 234. Teschan PE. Electroencephalographic and other neurophysiolog-
1983;56:194–209. ical abnormalities in uremia. In: Proceedings of the Conference on
218. Mars NJI, Thompson PM, Wilkus RJ. The spread of epileptic Adequacy of Dialysis, Monterey, CA, 1974.
seizures activity in humans. Epilepsia. 1985;26:85–94. 235. Ursin R, Moses J, Naitoh P, et al. REM–NREM cycle in the cat
219. Lopes da Silva FH, Pijn JPM, Boeijinga PH. Interdependence of may be sleep-dependent. Sleep. 1983;6:1–9.
EEG signals: linear versus nonlinear association and the signifi- 236. Walter WG, Shipton HW. A new toposcopic display system.
cance of time delays and phase shifts. Brain Topogr. 1989;2:9–18. Electroencephalogr Clin Neurophysiol. 1951;3:281–292.
220. Pijn JPM. Quantitative Evaluation of EEG Signals in Epilepsy. 237. Williamson SJ, Kaufman L. Advances in neuromagnetic instru-
Nonlinear Associations, Time Delays and Nonlinear Dynamics mentation and studies of spontaneous brain activity. Brain
[PhD thesis]. University of Amsterdam; 1990. Topogr. 1989;2:129–140.

Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
CHAPTER
EEG-Based Brain–Computer
Interfaces
GERT PFURTSCHELLER AND CHRISTA NEUPER
57
INTRODUCTION AND BASIC PRINCIPLES 2. Intentional control: At least one directly recordable brain
signal, which can be intentionally modulated, must provide
A relatively recent development in applied neurophysiology is input to the BCI (electrical potentials, magnetic fields, or
an approach called EEG-based brain–computer interface (BCI). hemodynamic changes).
A BCI translates specific features, automatically extracted from 3. Real-time processing: The signal processing must occur
EEG signals, into signals able to operate computer-controlled online and yield a communication or control signal.
devices in order to assist patients who have highly compro- 4. Feedback: The user must obtain feedback about the success
mised motor functions, such as tetrapalegic patients. This novel or failure of his/her efforts to communicate or control.
approach became possible due to advances both in methods of
EEG analysis and in information technology, along with a bet- It follows from these definitions that each BCI is a closed-
ter understanding of the psychophysiological correlates of cer- loop system with two adaptive controllers: the user’s brain,
tain EEG features. Therefore, it is interesting to take notice of which produces the signals and provides the input to the BCI;
the emerging field of direct brain–computer communication. and the BCI itself, which analyses the brain signals and trans-
A BCI provides the brain with a new nonmuscular commu- forms them to a control signal as the BCI output (Fig. 57.1).
nication channel that can be used to convey messages and com- Any BCI contains components to extract features and classify
mands directly from the brain to the external world without (detect) EEG events. The goal of the feature extraction compo-
using any muscle activity (1). Here, we expand this definition to nent is to find a suitable representation of the EEG signal that
emphasize that any BCI must have the following four compo- simplifies the subsequent classification or detection of specific
nents: patterns of electrical brain activity. That is, the signal features
should encode the commands sent by the user but should not
1. Direct: The signals must be recorded directly from the brain. contain noise and other signal components that can impede the
If a device records signals after they pass through peripheral classification process. There is a variety of feature extraction
nerves or muscles, it is not a BCI. methods used in current BCI systems. A nonexhaustive list of
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 57.1 Principle of a BCI system with the major internal processing steps (A), visual attention-based BCI used
to control a hand orthosis by focusing on one of two flickering lights (LEDs) (B), and motor imagery-based BCI used
to control a virtual reality (VR) hand (C). Motor imagery is the most common mental strategy in BCIs and does not
rely on external stimulation to generate the necessary brain activity (C). BCIs that do rely on external stimulation to
elicit brain activity typically involve spatial visual attention (B).

1227
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1228 Part IX ■ Computer-Assisted EEG Analysis

these methods includes amplitude and band power measures, it yields a larger P300 than the other possible choices. P300
Hjorth parameters, autoregressive parameters, and wavelet coef- BCIs are typically used to spell (14–16) but have been vali-
ficients (2–4). dated with other tasks such as control of a mobile robot (17)
The task of the classifier is to use the signal features provided or a smart home (18).
by the feature extractor to assign the recorded samples of the The SSVEP BCI: Steady-state evoked potentials (SSEPs) occur
signal to a given category of EEG patterns. In the simplest form, when sensory stimuli are repetitively delivered rapidly
detection of an EEG pattern may be made, for instance, by enough that the relevant neuronal structures do not return to
means of a threshold method (5,6). More sophisticated classifi- their resting states. In a BCI application, the user focuses on
cation algorithms of different EEG patterns depend on the use one of several stimuli, each of which flickers at a different rate
of linear or nonlinear classifiers (2,7,8). and/or or phase. Gao et al. (19) described a BCI with 48 flick-
The classifier output, which can be a simple on-off signal or ering lights and a high information transfer rate (ITR) of 68
a signal that encodes a number of different classes, is trans- bits/min. Like P300 BCIs, SSVEP BCIs require no training
formed into an appropriate signal that can then be used to con- and can facilitate rapid communication (9,20,21). SSVEP
trol a variety of devices. For most current BCI systems, the BCIs have also recently expanded to tasks beyond spelling,
output device is a computer screen and the desired output con- such as controlling an avatar in a computer game (22–24) or
sists of the selection of certain targets. Advanced applications controlling an orthosis (25). Some BCI articles argued that
include controlling of spelling systems or other external appa- the SSVEP can only be used for communication when users
ratuses such as prosthetic devices and multimedia applications. have some conscious control of eye muscles and is therefore
Feedback of performance is usually obtained by visualiza- not applicable for patients in the late stages of amyotrophic
tion of the classifier output on a computer screen or by presen- lateral sclerosis (ALS) (1,19). Later work showed that this
tation of an auditory, tactile, or visual feedback signal. Feedback assumption is incorrect; in some cases, SSVEP BCIs can func-
is an integral part of the BCI system because the users observe, tion even when users do not shift gaze (9,26).
for example, selected letters or certain movements simultane- The ERD BCI: Brain rhythms can either display an event-
ously with the brain responses they produce. related amplitude decrease or desynchronization or an
event-related amplitude increase or synchronization (10).
EEG PATTERNS USED AS INPUT FOR A BCI The term ERD BCI describes any BCI system that relies on
the detection of amplitude changes in sensorimotor (mu
The EEG is the most widely used brain signal in BCIs. Two types and central beta rhythms) and/or other brain oscillations,
of changes can be extracted from the ongoing EEG signals: also including short-lasting postimagery beta bursts (beta
ERS, beta rebound) (8,27–29).
1. Event-related potentials (ERPs) display time and phase-
locked changes (evoked) to an externally or internally paced One of the first papers reporting on online classification of
event. Evoked signals include slow cortical potential (SCP) different motor-imagery-induced ERD/ERS patterns were pub-
changes, P300 components, and steady-state-evoked poten- lished by Pfurtscheller et al. (30) and Kalcher et al. (31). At this
tials (SSVEPs) (9). time, beside others, the Wadsworth BCI (1,32), the Berlin BCI
2. Event-related changes in ongoing EEG activity in specific (8), the Graz BCI (33), and variants of the Tübingen BCI (34)
frequency bands. These changes are also time-locked but not use the ERD/ERS as features for single trial EEG classification.
phase-locked (induced). Event-related desynchronization The bit rates reported are between approximately 2 and 17
(ERD) defines an amplitude (power) decrease of a rhythmic bit/min (35,36) up to 35 bits/min (8).
component, whereas event-related synchronization (ERS) The ERD BCI can be operated in two different modes which
characterizes an amplitude (power) increase (10). determine when the user performs a mental task and, there-
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

with, intends to transmit a message. The first mode is externally

Depending on the phenomena analyzed and classified, the
paced (cue-based, computer-driven synchronous BCI) and the
following EEG-based BCI systems can be differentiated:
second mode is internally paced (noncue-based, uncued, user-
The SCP BCI: Beginning in 1979, Birbaumer and coworkers driven asynchronous BCI). In the case of a synchronous BCI, a
published a series of experiments demonstrating operant con- fixed, predefined time window is used. After a visual or auditory
trol of SCPs (see Ref. 11 for review). Operant conditioning is cue stimulus, the subject has to act and produce a specific brain
a learning process with the goal of the self-regulation of brain pattern. Nearly all known BCI systems work in such a cue-based
potentials (e.g., SCP shifts) or brain waves (e.g., sensorimotor mode (1,2,37). An asynchronous protocol requires a continu-
rhythms) with the help of suitable feedback. This process does ous analysis and feature extraction of the recorded brain signal.
not require continuous feedback, but a reward for achieving Thus, such BCIs are generally even more demanding and more
the desired brain potential (wave) change is necessary. complex than BCIs operating with a fixed timing scheme.
Operant conditioning was used in communication systems
for completely paralyzed (locked-in) patients (12,13). MOTOR IMAGERY AS CONTROL STRATEGY
The P300 BCI: The P300 is the positive component of the
evoked potential that may develop about 300 msec after an Several EEG studies indicate that primary sensorimotor areas
item is flashed. The user focuses on one flashing item while are activated when subjects imagine the execution of a hand
ignoring other stimuli. Whenever the target stimulus flashes, movement. Klass and Bickford (38) and Chatrian et al. (39)
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
Chapter 57 ■ EEG-Based Brain–Computer Interfaces 1229

observed blocking or desynchronization of the central mu- The enhancement of oscillatory EEG activity (ERS) during
rhythm with motor imagery. By means of quantification of the motor imagery is a very important aspect in BCI research. For
temporal-spatial ERD pattern, it was clearly shown that one- example, foot motor imagery can induce long-lasting beta
sided hand motor imagery can result in a lateralized activation oscillations during imagery (peri-imagery ERS; Fig. 57.3A)
of sensorimotor areas, similar to that found in the preparatory and/or short-lasting beta bursts after the end of the imagery
phase of a self-paced hand/finger movement (40,41). Such a process (postimagery ERS; Fig. 57.3B) over the foot representa-
pattern of sensorimotor EEG activity related to motor imagery tion area close to the vertex (29,51). The post-imagery ERS is
can also be found in patients with impaired motor function dominant in the beta band with a maximum ~2.5 seconds after
(42,43). To date, a number of more recent electrophysiological brisk cue-paced imagery, can be detected with great accuracy
studies support motor cortex participation in motor imagery (high rate of true positives, TP; see Fig. 57.3C) in the ongoing
(e.g., EEG: 44–48; MEG: 49). EEG and is therefore a good candidate to realize a one-channel
An example is shown in Figure 57.2 in the form of band EEG-based BCI (29,51).
power time courses of 11- to 13-Hz EEG activity. The ERD/ERS Summarizing, it can be stated that motor imagery can mod-
curves show different reactivity patterns during right and left ify sensorimotor rhythms in a similar way to that observed in
motor imagery, displaying a significant band power decrease the preparatory phase of an executed movement. Since motor
(ERD) over the contralateral hand area. It is of interest to note, imagery results in a somatotopically organized activation pat-
first, that contralateral to the side of motor imagery an ERD terns, mental imagination of different movements (e.g., left vs.
and ipsilaterally an ERS were present and, second, that feedback right hand; hand vs. foot) can be an efficient strategy to operate
enhanced the difference between both patterns, and therewith a BCI based on oscillatory EEG activity. The challenge is to
the classification accuracy (see also Ref. 50). detect the imagery-related changes in ongoing EEG recordings.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 57.2 Event-related desynchronization (ERD)/event-related synchronization (ERS) curves (11 to 13 Hz; 95%
confidence intervals indicated) of one representative subject during imagined movements of the left versus right hand
in sessions without feedback (A) and in sessions with continuously present feedback (C). Data were recorded from the
sensorimotor cortex (C3, C4). The time period of cue presentation is indicated by a gray vertical bar. Examples of clas-
sification results of single trials (based on linear discriminant analysis, LDA) of two selected sessions: one without (B)
and one with feedback (D). The x-axis denotes the average size of the distance function (resulting from LDA) for all
left and right trials of one session (for details, see Ref. 69). In the session with feedback, the average distance corre-
sponds to the average length of the feedback bar presented on the screen. Black bars indicate bar movements to the
left side of the screen, white bars indicate bar movements to the right side. The y-axis displays the time points used for
classification. The best classification accuracy for each session is indicated.

Figure 57.4 Schema for the workflow of an imagery-based BCI with

EEG recording (A), feature selection (B), experiment with feedback (C)
and classifier update if necessary (D).

Figure 57.3 Examples of single EEG trials during foot motor imagery patterns, and classification of the individual’s brain activity in
recorded at electrode position Cz and topographic map indicating the different imagery conditions. By applying, for example, the dis-
localized 10-Hz ERS at Cz (A), examples of EEG trials during cue-paced tinction-sensitive learning vector quantization (DSLVQ) (52) to
(second 2) brisk foot motor imagery with postimagery beta ERS the screening data, the most important frequency components
recorded at Cz (B, right panel ), time-frequency map displaying peri- and electrode locations that best discriminate between different
imagery ERD and postimagery ERS at Cz (B, left panel ), and single trial imagery tasks may be identified for each participant, as well as
classification of peri-imagery ERD and postimagery ERS with indicated the accuracy of classification (Fig. 57.4B). After setting up the
true positives (C). initial classifier, subsequent training sessions can start, where
the user receives online feedback of motor imagery–related
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

changes in the EEG (Fig. 57.4C). Depending on the classifica-

TRAINING PARADIGM AND IMPACT tion accuracy, an update of the classifier and further feedback
OF FEEDBACK experiments may be recommended (Fig. 57.4D). This adapta-
tion process between brain and computer can last for many
Before such a motor-imagery-based BCI can be efficiently used, days or weeks in patients. To keep the training period as short
the participants have to undergo training in order to obtain as possible, an efficient training strategy is necessary. One
some control of their brain signals and to maximize the classi- example for this could be the so-called basket game.
fication accuracy of different brain states. Prior to starting In the “basket-game” paradigm, for example, the user has
online feedback sessions with an individual, their brain patterns to mentally move a falling ball into the correct goal (“basket”)
(e.g., related to different types of motor imagery) must be marked on the screen (Fig. 57.5A). If the ball hits the correct
known. To this end, in the first session of an imagery-based BCI basket, it becomes highlighted and points are earned. The
standard protocol, users have to imagine repeatedly different horizontal position of the ball is controlled via the BCI output
kinds of movement (e.g., hand, feet, or tongue movement) in a signal and the velocity can be adjusted by the investigator. The
cue-based mode while their EEG is being recorded (Fig. 57.4A). speed of the ball can be increased run by run until the person
Optimally, this would entail a full-head recording of their EEG, considers it too fast. This approach can find the optimal speed
with topographical and time-frequency analyses of ERD/ERS for a maximum ITR. Experiments with two bipolar EEG

Figure 57.5 Graphical display of the “basket-paradigm” (A). The subject has to direct the ball to the indicated goal
(“basket”). The trial length varies across the different runs. Information transfer rate (ITR) for one subject in relation to
trial length. The black line represents the maximum possible ITR for an error-free classification (B) Modified from Krausz
G, Scherer R, Korisek G, et al. Critical decision-speed and information transfer in the “Graz Brain-Computer Interface.”
Appl Psychophysiol Biofeedback. 2003;28:233–240.

channels and two motor imagery tasks performed by volun- vibrotactile feedback affect the modulation of the mu rhythm,
teers with spinal cord injuries revealed a maximal ITR of 17 haptic information may become a critical component of BCIs,
bits/min with a trial length (falling time of the ball) of 2.5 sec- especially if they are designed to control an advanced neuro-
onds (36, see also Fig. 57.5B). prosthetic device (60).
To keep the training period as short as possible, a well- Despite successful approaches to the development of nonvi-
thought-out training protocol and helpful feedback signals are sual BCI systems, visual presentation of stimuli is the most fre-
essential. The feedback provides the user with information quently used feedback modality in BCI research (1). Typical
about the efficiency of his/her strategy and enables learning. In visual feedback stimuli comprise cursor movement (8,61), a
this context, two aspects are crucial: (i) the exact manner of moving bar of varying size (50,59), and the trajectory of a mov-
how the brain signal is translated into the feedback signal (i.e., ing object like in the basket game (8,36). Other interesting vari-
information content of the feedback; for advantages of provid- ants include color signaling (62) and complex virtual reality
ing continuous or discrete feedback, see Refs. 41 and 50) and environments (28,63).
(ii) the type of feedback presentation. In any case, the influence There is some evidence that a rich visual representation of the
of the feedback on the capacity for attention, concentration, feedback signal, for example, in the form of a three-dimensional
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

and motivation of the user, all aspects which are closely related video game or virtual reality environment, may enhance the
to the learning process, should be considered (see also Ref. 53). learning progress in a BCI task (53,64). Combining BCI and
BCI studies can use different feedback modalities. In the virtual reality technologies could lead to highly realistic and
auditory modality, Hinterberger et al. (54) and Pham et al. (55) immersive feedback scenarios. As an important step in this
coded SCP amplitude shifts in the ascending and descending direction, Pfurtscheller et al. (28) could show that EEG record-
pitches on a major tone scale. Further studies showed that P300 ing and single-trial processing with sufficiently good classifica-
BCIs could also be implemented with auditory rather than tion results are possible in an immersive multi-projection-based
visual feedback (56,57). A BCI using only auditory (rather than stereo and head-tracked virtual reality system (CAVE), and that
visual) stimuli would be of importance providing communica- the obtained signals are even suitable to control events within a
tion support for severely paralyzed patients with visual impair- virtual environment in real time.
ment (58). Although the mentioned studies could show that
BCI communication using auditory stimuli only is possible, SOME EEG-BASED BCI APPLICATIONS
visual feedback turned out to be superior to auditory feedback.
Recently, Chatterjee et al. (59) presented an ERD BCI using a Currently, the most important applications of a BCI include the
motor imagery paradigm and haptic feedback provided by restoration of communication for patients with a locked-in-
vibrotactile stimuli to the upper limb. Although further work syndrome and the control of neuroprostheses in patients with
will be needed to determine how the neural correlates of spinal cord injuries (1,12,65). In addition to these, there is the

important field of neurofeedback to support feedback training introduced recently by Müller and Blankertz (76). The Hex-
in people suffering from epilepsy or stroke (66). Today, the O-Spell application combines asynchronous 2-class ERD BCI
world of BCI application is expanding and new fields are open- control and divides the alphabet into six parts. This raised
ing. One emerging opportunity is to use the BCI to control the average spelling rate up to ~6 letter/min.
multimedia applications (67,68), another is to use the BCI for
user authentication (69,70). Some applications are explained in Control of Neuroprosthesis to Restore
more detail below. Grasp Function
In general, systems for functional electrical stimulation
Control of Spelling Systems in Severely (FES)—the so-called neuroprostheses—are able to restore lost
Paralyzed Patients control/motor functions of body parts after spinal cord injury
Spelling systems are communication aids that allow users to (SCI) with the use of electrical impulses. The control signal
express themselves by selecting letters or items of an alphabet generated by the BCI switches the FES on/off. In a project
and thus form words and sentences. The simplest case involves with a tetraplegic patient, FES resulting in hand grasp was
a binary control signal requiring two distinctive mental activi- controlled by ongoing EEG activity based on an asynchronous
ties. Patients suffering from ALS can learn to control their own BCI. The patient underwent a large number of BCI training
SCP BCI and so to operate the thought translation device sessions with varying types of motor imagery over a period of
spelling device (5,12). By using the same dichotomous selection several months (77). At the end, he was able to induce trains
strategy an ALS patient and a patient suffering from severe cere- of 17-Hz beta oscillations focused on the electrode position
bral palsy (71) learned to operate the virtual keyboard spelling near the vertex (Cz) by foot motor imagery. These mentally
application (72). The spelling rates in these studies varied from induced 17-Hz oscillations were used as a simple asynchro-
0.15 to approximately 1 letter/min. An example for such a nous brain switch to generate a control signal for the opera-
spelling system with an ERD BCI is displayed in Figure 57.6B. tion of the FES using surface electrodes (Fig. 57.6D). With this
The Wadsworth speller, based on mu and beta activity for method the patient was able to grasp a glass at “free will” (for
example, divides the alphabet into four parts (73). In Millán a detailed description of the procedure see Ref. 78). Müller-
and Mouriño (74) an average spelling rate of ~3 letters/min was Putz et al. (79) reported on an implantable neuroprostheses,
reported by using a 3-class BCI. A novel spelling concept, pos- known as the Freehand system (80), coupled with an ERD BCI
sible due to the asynchronous protocol, was introduced by and used in a tetraplegic patient to perform a grasp sequence
Scherer et al. (75). Another efficient selection strategy was (Fig. 57.6C).
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 57.6 Two students playing table tennis (A), ALS patient operating a spelling system at home (B), motor-imagery-
based control of a neuroprosthesis through implanted functional electrical stimulation (FES) of hand muscles (C), motor-
imagery-based control of a neuroprosthesis through FES of hand muscles using surface electrodes (D), wheel chair control
in a virtual environment (E), and control of virtual hand movement (F).(See color insert)

Rehabilitation after Stroke (iv) automatically detect artefacts. It can be expected that mak-
Typically, physical therapy aimed at poststroke motor recovery ing BCIs useful to a wider group of users will open up new fields
focuses on active movement training. Some patients, however, of applications such as entertainment in the next several years.
are so severely disabled that they cannot engage in movement We recently reported on a one-channel EEG-based BCI that
without assistance. Newly developed protocols based on men- detects the postmotor imagery beta rebound (29,51). This beta
tally rehearsing movements (motor imagery) represent an rebound is a relatively stable EEG phenomenon (47) that can be
intriguing backdoor approach to accessing the motor system, detected in single trial EEG during foot motor imagery and
because they can induce an activation of sensorimotor net- used to realize a “brain switch.” Such a brain switch is a BCI sys-
works that the lesions affected (81). tem designed to detect only one brain state (brain pattern) in
It is known that unilateral hand motor imagery can result in the ongoing EEG signal (83). One important feature of such a
a contralateral ERD and simultaneously in an ipsilateral ERS brain switch is that unintended activations should not occur in
after some training sessions (40). Hence, an ERD BCI based on the output signal. That is, the false-positive rate should be zero
movement imagery can provide some measure of attempted or close to zero. An imagery-based brain switch can be used to
activity in the motor regions and reinforce a patient’s sensori- turn on/off the flickering lights of an SSVEP-controlled hand
motor experience during poststroke motor recovery. Feedback orthosis (25). This type of BCI, composed of two sequentially
from the BCI can be solely visual, as in the movement of a vir- operating BCIs, can be seen as an example of a “hybrid BCI.”
tual hand (Figs. 57.1C and 57.6F) or it can occur through a A clear challenge is to develop more effective BCI control par-
prosthetic device like an orthotic hand attached to the patient’s adigms, offering, for instance, three-dimensional control over a
own (64,82). In both cases, not only can positive feedback rein- neuroprostesis or the operation of a spelling device with a speed
force the motor imagery process, but the act of observing the of at least 5 to 10 characters/minute. Such an improvement of
hand movement can itself activate sensorimotor areas. speed and accuracy is possible by analyzing cortical potential
changes recorded with subdural electrode strips or grids or the
detection of firing patterns in intracortical recordings. Figure
PERSPECTIVES FOR THE FUTURE
57.7 shows the three types of potential recording with EEG, sub-
BCI technology is a fast growing field of research and applica- dural electrodes (ECoG), and microelectrode array. The advan-
tions with the potential to improve the quality of life in severely tage of the ECoG over the EEG is the better signal-to-noise ratio,
disabled people. To date, several BCI prototypes exist and most which includes easier detection of gamma activity. Recently,
of them work in a laboratory environment only. Before a BCI bursts of gamma activity between 60 and 90 Hz in ECoG record-
can be used for communication and control at home or work, ings during self-paced limb and tongue movements were
several problems have to be solved, such as how to (i) automat- reported (84,85). These gamma bursts are short lasting, display a
ically select electrode positions and frequency components in a high somatotopic specificity, and are embedded in the alpha and
motor imagery task, (ii) use the fewest number of recording beta ERD lasting for some seconds. Patient-oriented work with
electrodes possible (the optimum is one EEG channel), (iii) subdural electrodes and ECoG single-trial classification have
minimize the training time through game-like feedback, and shown first promising results (3,86).
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 57.7 Schema of electrical potential recording in BCI research with non-invasive EEG elec-
trodes, invasive subdural electrode strip (array), or highly invasive microelectrode array in the cortex.

Studies in monkeys have shown that three-dimensional con- 16. Farwell LA, Donchin E. Talking off the top of your head: toward a
trol is possible when multiunit activity is recorded in cortical mental prosthesis utilizing event-related brain potentials.
areas. Between 32 and 96 microwires were implanted in differ- Electroencephalogr Clin Neurophysiol. 1988;70:510–523.
ent cortical motor areas. After a training period with distinct 17. Bell CJ, Shenoy P, Chalodhorn R, et al. Control of a humanoid
robot by a noninvasive brain-computer interface in humans.
motor tasks, the monkeys were able to achieve three-dimen-
J Neural Eng. 2008;5:214–220.
sional control over the movement of a robotic device by real-
18. Guger C, Daban S, Sellers E, et al. How many people are able to
time transformation of neuronal multiunit neuronal activity control a P300-based brain-computer interface (BCI)? Neurosci
(87). The feasibility of a prosthetic device control in a Lett. 2009;462:94–98.
tetraplegic patient based on recording of neural ensemble activ- 19. Gao X, Xu D, Cheng M, et al. A BCI-based environmental con-
ity through a 96-microelectrode array (Fig. 57.7) was reported troller for the motion-disabled. IEEE Trans Neural Syst Rehabil
by Hochberg et al. (88). These early results suggest that record- Eng. 2003;11:137–140.
ing of intracortical neuronal multiunit activity could provide a 20. Allison BZ, Wolpaw EW, Wolpaw JR. Brain-computer interface
new neurotechnology to restore independence for humans with systems: progress and prospects. Expert Rev Med Devices.
paralysis. 2007;4:463–474.
21. Krusienski D J, Sellers EW, Cabestaing F, et al. A comparison
of classification techniques for the P300 speller. J Neural Eng.
REFERENCES 2006;3:299–305.
22. Faller J, Müller-Putz G, Schmalstieg D, et al. An application frame-
1. Wolpaw JR, Birbaumer N, McFarland DJ, et al. Brain-computer work for controlling an avatar in a desktop based virtual environ-
interfaces for communication and control. Clin Neurophysiol. ment via a software SSVEP brain-computer interface. Special Issue
2002;113:767–791. of Presence: Teleoperators and Virtual Environments. 2010;
2. Pfurtscheller G, Neuper C. Motor imagery and direct brain-com- 19(1):25–34.
puter communication. Proc IEEE. 2001;89:1123–1134. 23. Lalor EC, Kelly SP, Finucane C, et al. Steady-state VEP-based brain-
3. Graimann B, Huggins JE, Schlögl A, et al. Detection of movement- computer interface control in an immersive 3D gaming environ-
related desynchronization patterns in ongoing single-channel ment. EURASIP J Appl Signal Processing. 2005; 19:3156–3164.
electrocorticogram. IEEE Trans Neural Syst Rehabil Eng. 24. Martinez P, Bakardjian H, Cichocki A. Fully online multicom-
2003;11:276–281. mand brain-computer interface with visual neurofeedback
4. Obermaier B, Neuper C, Guger C, et al. Information transfer rate using SSVEP paradigm. Comput Intell Neurosci. 2007; Article ID
in a five-classes brain-computer interface. IEEE Trans Neural Syst 94561.
Rehabil Eng. 2001;9:283–288. 25. Pfurtscheller G, Solis-Escalante T, Ortner R, et al. Self-paced oper-
5. Birbaumer N, Kübler A, Ghanayim N, et al. The thought transla- ation of an SSVEP-based orthosis with and without an imagery-
tion device (TTD) for completely paralyzed patients. IEEE Trans based “brain switch”: a feasibility study towards a hybrid BCI. IEEE
Rehabil Eng. 2000;8:190–193. Trans Neural Syst Rehabil Eng. in press.
6. Levine SP, Huggins JE, BeMent SL, et al. A direct brain interface 26. Kelly SP, Lalor EC, Finucane C, et al. Visual spatial attention con-
based on event-related potentials. IEEE Trans Rehabil Eng. trol in an independent brain-computer interface. IEEE Trans
2000;8:180–185. Neural Syst Rehabil Eng. 2005;52:1588–1596.
7. Millán J, Mouriño J, Franzé M, et al. A local neural classifier for the 27. Pfurtscheller G, Brunner C, Schlögl A, et al. Mu rhythm (de)syn-
recognition of EEG patterns associated to mental tasks. IEEE Trans chronization and EEG single-trial classification of different motor
Neural Netw. 2002;13:678–686. imagery tasks. Neuroimage. 2006;31:153–159.
8. Blankertz B, Dornhege G, Krauledat M, et al. The non-invasive 28. Pfurtscheller G, Leeb R, Keinrath C, et al. Walking from thought.
Berlin brain-computer interface: fast acquisition of effective per- Brain Res. 2006;1071:145–152.
formance in untrained subjects. Neuroimage. 2007;37:539–550. 29. Pfurtscheller G, Solis-Escalante T. Could the beta rebound in the
9. Allison BZ, McFarland DJ, Schalk G, et al. Towards an independ- EEG be suitable to realize a “brain switch”? Clin Neurophysiol.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

ent brain-computer interface using steady state visual evoked 2009;120:24–29.

potentials. Clin Neurophysiol. 2008;119:399–408. 30. Pfurtscheller G, Kalcher J, Flotzinger D. A new communication
10. Pfurtscheller G, Lopes da Silva FH. Event-related EEG/MEG syn- device for handicapped persons The Brain-Computer Interface.
chronization and desynchronization: basic principles. Clin In: Ballabion E, Placencia-Porrero I, Puig de la Bellacasa R, eds.
Neurophysiol. 1999;110:1842–1857. Rehabilitation Technology. Vol 9. Amsterdam: IOS Press; 1993:
11. Birbaumer N, Elbert T, Canavan AG, et al. Slow potentials of the 123–127.
cerebral cortex and behavior. Physiol Rev. 1990;70:1–41. 31. Kalcher J, Flotzinger D, Neuper C, et al. Graz brain-computer
12. Birbaumer N, Ghanayim N, Hinterberger T, et al. A spelling device interface II: towards communication between humans and com-
for the paralyzed. Nature. 1999;398:297–298. puters based on online classification of three different EEG pat-
13. Kübler A, Birbaumer N. Brain-computer interfaces and communica- terns. Med Biol Eng Comput. 1996;34:382–388.
tion in paralysis: extinction of goal directed thinking in completely 32. Sellers EW, Krusienski DJ, McFarland DJ, et al. Noninvasive brain-
paralyzed patients? Clin Neurophysiol. 2008; 119:2658–2666. computer interface research at the Wadsworth Center. In: Dornhege
14. Allison BZ, Pineda JA. Effects of SOA and flash pattern manipula- G, Millán J del R, Hinterberger T, et al., eds. Toward Brain-Computer
tions on ERPs, performance, and preference: implications for a Interfacing. Cambridge, Massachusetts: MIT Press; 2007:31–42.
BCI system. Int J Psychophysiol. 2006;59:127–140. 33. Pfurtscheller G, Müller-Putz GR, Schlögl A, et al. Graz-brain-
15. Donchin E, Spencer KM, Wijesinghe R. The mental prosthesis: computer interface: state of research. In: Dornhege G, Millán J del
assessing the speed of a P300-based brain-computer interface. R, Hinterberger T, et al., eds. Toward Brain-Computer Interfacing.
IEEE Trans Rehabil Eng. 2000;8:174–179. Cambridge, Massachusetts: MIT Press; 2007:65–84.

34. Hinterberger T, Nijboer F, Kübler A, et al. Brain-computer interfaces 53. Pineda JA, Silverman DS, Vankov A, et al. Learning to control brain
for communication in paralysis: a clinical experimental approach. rhythms: making a brain-computer interface possible. IEEE Trans
In: Dornhege G, Millán J del R, Hinterberger T, et al., eds. Toward Neural Syst Rehabil Eng. 2003;11:181–184.
Brain-Computer Interfacing. Cambridge, Massachusetts: MIT Press; 54. Hinterberger T, Neumann N, Pham M, et al. A multimodal brain-
2007:43–64. based feedback and communication system. Exp Brain Res.
35. McFarland DJ, Sarnacki WA, Wolpaw JR. Brain-computer inter- 2004;154:521–526.
face (BCI) operation: optimizing information transfer rates. Biol 55. Pham M, Hinterberger T, Neumann N, et al. An auditory brain-
Psychol. 2003;63:237–251. computer interface based on the self-regulation of slow cortical
36. Krausz G, Scherer R, Korisek G, et al. Critical decision-speed and potentials. Neurorehabil Neural Repair. 2005;19:206–218.
information transfer in the “Graz Brain-Computer Interface”. Appl 56. Furdea A, Halder S, Krusienski DJ, et al. An auditory oddball (P300)
Psychophysiol Biofeedback. 2003;28:233–240. spelling system for brain-computer interfaces. Psychophysiology.
37. Kübler A, Neumann N, Kaiser J, et al. Brain-computer communi- 2009;46:1–9.
cation: self-regulation of slow cortical potentials for verbal com- 57. Klobassa DS, Vaughan TM, Brunner P, et al. Toward a high-
munication. Arch Phys Med Rehabil. 2001;82:1533–1539. throughput auditory P300-based brain-computer interface. Clin
38. Klass SG, Bickford RG. Observations on the rolandic arceau Neurophysiol. 2009;120:1252–1261.
rhythm. Electroencephalogr Clin Neurophysiol. 1957;9:570. 58. Kübler A, Furdea A, Halder S, et al. A brain-computer interface
39. Chatrian GE, Petersen MC, Lazarte JA. The blocking of the controlled auditory event-related potential (P300) spelling system
rolandic wicket rhythm and some central changes related to move- for locked-in patients. Ann N Y Acad Sci. 2009;1157:90–100.
ment. Electroencephalogr Clin Neurophysiol. 1959;11:497–510. 59. Lebedev MA, Nicolelis MAL. Brain-machine interfaces: past, pres-
40. Pfurtscheller G, Neuper C. Motor imagery activates primary sen- ent and future. Trends Neurosci. 2006;29:536–546.
sorimotor area in humans. Neurosci Lett. 1997;239:65–68. 60. Chatterjee A, Aggarwal V, Ramos A, et al. A brain-computer inter-
41. Neuper C, Scherer R, Wriessnegger S, et al. Motor imagery and face with vibrotactile biofeedback for haptic information.
action observation: modulation of sensorimotor brain rhythms J Neuroeng Rehabil. 2007;4:40.
during mental control of a brain-computer interface. Clin 61. McFarland DJ, McCane LM, Wolpaw JR. EEG-based communica-
Neurophysiol. 2009;120:239–247. tion and control: short-term role of feedback. IEEE Trans Rehabil
42. Neuper C, Pfurtscheller G. Motor imagery and ERD. In: Pfurtscheller Eng. 1998;6:7–11.
G, Lopes da Silva FH, eds. Event-Related Desynchronization: 62. Kaplan AY, Lim JJ, Jin KS, et al. Unconscious operant conditioning
Handbook of Electroencephalography and Clinical Neurophysiology. in the paradigm of brain-computer interface based on color per-
Revised Edition, Vol 6. Amsterdam: Elsevier; 1999:303–325. ception. Int J Dev Neurosci. 2005;115:781–802.
43. Neuper C, Müller-Putz GR, Scherer R, et al. Motor imagery and 63. Leeb R, Keinrath C, Friedman D, et al. Walking by thinking: the
EEG-based control of spelling devices and neuroprostheses. In: brainwaves are crucial, not the muscles! Presence 2006;15:500–514.
Neuper C, Klimesch W, eds. Event-Related Dynamics of Brain 64. Ron-Angevin R, Diaz-Estrella A, Reyes-Lecuona A. Development
Oscillations. Amsterdam: Elsevier; 2006:393–409. of a brain-computer interface based on virtual reality to improve
44. McFarland DJ, Miner LA, Vaughan TM, et al. Mu and beta rhythm training techniques. Cyberpsychol Behav. 2005;8:353–354.
topographies during motor imagery and actual movements. Brain 65. Pfurtscheller, G, Müller-Putz G, Scherer R, et al. Rehabilitation
Topogr. 2000;12:177–186. with brain-computer interface systems. IEEE Comput Mag.
45. Caldara R, Deiber MP, Andrey C, et al. Actual and mental motor 2008;41:58–65.
preparation and execution: a spatiotemporal ERP study. Exp Brain 66. Birbaumer N, Cohen LG. Brain-computer interfaces: communica-
Res. 2004;159:389–399. tion and restoration of movement in paralysis. J Physiol. 2007;
46. Neuper C, Scherer R, Reiner M, et al. Imagery of motor actions: 579:621–636.
differential effects of kinaesthetic versus visual-motor mode of 67. Scherer R, Lee F, Schlögl A, et al. Toward self-paced brain-com-
imagery on single-trial EEG. Brain Res Cogn Brain Res. 2005;25: puter communication: navigation through virtual worlds. IEEE
668–677. Trans Biomed Eng. 2008;55:675–682.
47. Pfurtscheller G, Neuper C, Brunner C, et al. Beta rebound after 68. Nijholt A, Tan D, Pfurtscheller G, et al. Brain-computer interfacing
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

different types of motor imagery in man. Neurosci Lett. 2005;378: for intelligent systems. IEEE Intell Syst. 2008;23:72–79.
156–159. 69. Pfurtscheller G, Neuper C. Future prospects of ERD/ERS in the con-
48. Carrillo-de-la-Pena MT, Lastra-Barreira C, Galdo-Alvarez S. Limb text of brain-computer interface (BCI) developments. In: Neuper C,
(hand vs. foot) and response conflict have similar effects on event- Klimesch W, eds. Event-Related Dynamics of Brain Oscillations,
related potentials (ERPs) recorded during motor imagery and Progress in Brain Research. Vol 159. Amsterdam: Elsevier; 2006:19–27.
overt execution. Eur J Neurosci. 2006;24:635–643. 70. Marcel S, Millán J del R. Person authentication using brainwaves
49. Mellinger J, Schalk G, Braun C, et al. An MEG-based brain-com- (EEG) and maximum a posteriori model adaptation. IEEE Trans
puter interface (BCI). Neuroimage. 2007;36:581–593. Pattern Anal Mach Intell. 2007;29:743–748.
50. Neuper C, Schlögl A, Pfurtscheller G. Enhancement of left-right 71. Neuper C, Müller GR, Kübler A, et al. Clinical application of an
sensorimotor EEG differences during feedback-regulated motor EEG-based brain-computer interface: a case study in a patient
imagery. J Clin Neurophysiol. 1999;16:373–382. with severe motor impairment. Clin Neurophysiol. 2003;114:
51. Solis-Escalante T, Müller-Putz GR, Brunner C, et al. Analysis of 399–409.
sensorimotor rhythms for the implementation of a brain switch 72. Obermaier B, Müller GR, Pfurtscheller G. “Virtual keyboard'” con-
for healthy subjects. Biomed Signal Process Control. 2010;5: trolled by spontaneous EEG activity. IEEE Trans Neural Syst
15–20. Rehabil Eng. 2003;11:422–426.
52. Pregenzer M, Pfurtscheller G, Flotzinger D. Automated feature 73. Wolpaw JR, McFarland DJ, Vaughan TM, et al. The Wadsworth
selection with a distinction sensitive learning vector quantizer. Center brain-computer interface (BCI) research and development
Neurocomputing. 1996;11:19–29. program. IEEE Trans Neural Syst Rehabil Eng. 2003;11:204–207.

74. Millán J, Mouriño J. Asynchronous BCI and local neural classifiers: 82. Birbaumer N. Brain-computer interface research: coming of age
an overview of the adaptive brain interface project. IEEE Trans clinical. Neurophysiology. 2006;117:479–483.
Neural Syst Rehabil Eng. 2003;11:159–161. 83. Mason SG, Birch GE. A brain-controlled switch for asynchronous
75. Scherer R, Müller GR, Neuper C, et al. An asynchronously con- control applications. IEEE Trans Biomed Eng. 2000;47:1297–1307.
trolled EEG-based virtual keyboard: improvement of the spelling 84. Crone NE, Miglioretti DL, Gordon B, et al. Functional mapping of
rate. IEEE Trans Neural Syst Rehabil Eng. 2004;51:979–984. human sensorimotor cortex with electrocorticographic spectral
76. Müller K-R, Blankertz B. Toward noninvasive brain computer analysis. II. Event-related synchronization in the gamma band.
interfaces. IEEE Signal Process Mag. 2006;23:125–128. Brain. 1998;121:2301–2315.
77. Pfurtscheller G, Guger C, Müller G, et al. Brain oscillations control 85. Pfurtscheller G, Graimann B, Huggins JE, et al. Spatiotemporal pat-
hand orthosis in a tetraplegic. Neurosci Lett. 2000;292:211–214. terns of beta desynchronization and gamma synchronization in cor-
78. Pfurtscheller G, Müller GR, Rupp R, et al. “Thought”-control of ticographic data during self-paced movement. Clin Neurophysiol.
functional electrical stimulation to restore hand grasp in a 2003;114:1226–1236.
tetraplegic. Neurosci Lett. 2003;351(1):33–36. 86. Schalk G, Miller KJ, Anderson NR, et al. Two-dimensional move-
79. Müller-Putz GR, Scherer R, Pfurtscheller G, et al. EEG-based neu- ment control using electrocorticographic signals in humans.
roprosthesis control: a step into clinical practice. Neurosci Lett. J Neural Eng. 2008;5:75–84.
2005;382:169–174. 87. Wessberg J, Stambaugh CR, Kralik JD, et al. Real-time prediction
80. Keith MW, Peckham PH, Thrope GB, et al. Implantable functional of hand trajectory by ensembles of cortical neurons in primates.
neuromuscular stimulation in the tetraplegic hand. J Hand Surg Nature. 2000;408:361–365.
Am. 1989;14:524–530. 88. Hochberg LR, Serruya MD, Friehs GM, et al. Neuronal ensemble
81. Sharma N, Pomeroy VM, Baron J-C. Motor imagery: a backdoor control of prosthetic devices by a human with tetraplegia. Nature.
to the motor system after stroke? Stroke. 2006;37:1941–1952. 2006;442:164–171.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
CHAPTER
Multimodal Monitoring of EEG
and Evoked Potentials
GERT PFURTSCHELLER
58
GENERAL ASPECTS LONG-TERM MONITORING IN
THE INTENSIVE CARE UNIT
The most comprehensive form of monitoring of cerebral func-
tioning is obtained when electroencephalograph (EEG) and dif- A great number of patients in the ICU are comatose or recover-
ferent modality-evoked potentials (EPs) are recorded and ing from coma, since coma can be caused by many disorders
analyzed continuously and simultaneously. This means that EEG including severe head injury, vascular lesions, encephalitis,
spectra, brainstem auditory-evoked potentials (BAEPs), and cer- posthypoxic or postischemic events, and drug intoxication,
vical and cortical somatosensory-evoked potentials (SEPs) are among others. The monitoring of ICU patients is important to
computed at intervals of seconds to minutes and displayed in a detect critical situations as fast as possible, to facilitate therapeu-
compressed form. In addition to EEG and EPs, other physiologi- tic decisions, to monitor the level of impaired consciousness,
cal signals, such as heart rate (HR), ventilation, temperature, oxy- and to predict the clinical outcome. Another important feature
gen saturation, and blood pressure, may be useful, depending on in the ICU is the determination and documentation of brain
the particular condition. This monitoring can be performed in the death, which is a prerequisite for most organ transplantations.
operating room, in the intensive care unit (ICU), in the sleep lab- The following case report of a patient with unfavorable outcome
oratory, or in the outpatient clinic. For extensive literature on this underlines the importance of monitoring of comatose patients:
subject, see Nuwer (1) and Chapters 39, 40, and 41 of this book. A 57-year-old woman with hypernephroma suffered from
hypertension for several years. She was admitted with a left hemi-
TECHNICAL ASPECTS OF CONTINUOUS sphere hemorrhage and was soporous. She had a right hemiplegia
MONITORING OF EEG SPECTRA and responded appropriately to painful stimuli on the left side.
The dominant EEG frequency was 5 to 6/sec, and there was a
COMBINED WITH EPs
delta-wave focus over the left hemisphere. On the first day, a ven-
The techniques necessary for long-term monitoring of EEG and tricular drainage of cerebrospinal fluid was performed. In the fol-
BAEPs were established in normal subjects by Maresch et al. lowing 6 days, the patient’s condition deteriorated. After 2 days,
(2). They used two EEG channels (Cz-A1 and Cz-A2) with she developed anisocoria. One day later, respiration was insuffi-
broadband amplifiers (1.5 Hz to 1.5 kHz); after amplification cient. The patient showed extensor responses to painful stimuli.
the signals were divided into low-frequency (EEG; lower cut-off After 5 days, pupillary light reaction and ciliospinal, cornea, and
frequency 30 Hz) and high-frequency (BAEP; upper cut-off fre- oculocephalic reflexes were negative, and there was no longer any
quency 250 Hz) branches. In other studies these responses were reaction to painful stimulation. The cranial computed tomogra-
complemented with recordings of SEPs (3,4). The main techni- phy (CT) scan showed an enlargement of the hemorrhage and
cal considerations in this type of analysis are that (a) different signs of an increased intracerebral pressure (ICP).
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

sampling frequencies must be used because of the rather differ- Figure 58.2 shows the various parameters registered from
ent frequency contents of the signal, for example, EEG and the time of admission, throughout the critical rise of ICP, and
BAEP; and (b) the number of samples should be minimized to until brain death was clinically determined. Five days after
save memory storage, since the analysis must be done over long admission, at 6:20 AM, the intracranial pressure rose to a critical
periods of time. A way to solve these technical problems is level of 56 mm Hg, in spite of the ventricular drainage.
introduced in Figure 58.1 by the sampling schema. Amplitudes of the cortical SEPs decreased and finally disap-
Measurements in the operating room or ICU may be easily peared 30 minutes later. The peak of the EEG spectra shifted to
affected by noise. To avoid main artifacts (50 or 60 Hz), the lower frequencies. Pathological HR and heart rate variability
interstimulus interval between auditory and somatosensory (HRV) were observed. Seven hours later, BAEP components
stimulation should be variable. More specifically, intervals IV/V vanished; another 30 minutes later, wave III disappeared,
should be alternatively small or large (e.g., to cancel 50 Hz: audi- and 60 minutes later, all BAEP components were abolished.
tory stimulation 110/90 msec; somatosensory stimulation This example provides evidence that monitoring of only one
210/190 msec). In addition, the stimuli should be applied either signal gives incomplete information. From the SEPs alone, it can
on the positive or negative peak of the main frequency EEG be seen that the cortical component disappeared at 6:50 AM,
component. With this technique, fairly good results are obtained whereas the cervical component remained unchanged for about 8
in unshielded rooms and with electrical equipment nearby. additional hours. The cortical component N20 is very sensitive to

1237
Niedermeyer's Electroencephalography : Basic Principles, Clinical Applications, and Related Fields, edited by Donald L. Schomer, and da Silva, Fernando Lopes, Wolters Kluwer Health, 2010.
ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/unc/detail.action?docID=2032003.
Created from unc on 2023-05-27 00:05:16.
1238 Part IX ■ Computer-Assisted EEG Analysis

Figure 58.1 Sampling scheme of EEG, somatosensory-

evoked potential (SEP), and brainstem auditory-evoked
potential (BAEP) during computer-controlled click (C) and
electrical (E) median nerve stimulation as used for the data
displayed in Figures 60.2 and 60.3B; sampling and electri-
cal and auditory stimulation are indicated. BAEP sampling
lasts for 10 msec, SEP sampling for 50 msec, the SEP sam-
pling window is variable in a range of 90 msec. Computer-
controlled SEP stimulation in intervals of 210/190 msec and
BAEP stimulation in intervals of 111/89 msec.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 58.2 Protocol from a 57-year-old patient with a left hemispheric hemorrhage during deterioration of the comatose
state and ending with brain death. From left to right: compressed EEG spectra from left and right hemispheres; brainstem
auditory-evoked potential (BAEP) to ipsi- and contralateral ear stimulation; BAEP interpeak latency (I–V); cervical (N14)
and cortical (N20) SEPs; central conduction time (CCT); heart rate (HR); heart rate variability (HRV); intracranial pressure
(ICP); systolic (Bpsys) and diastolic (Bpdia) blood pressure; and rectal body temperature (TEMP). Important events are
indicated by arrows. For further explanation, see text. Modified from Hilz MJ, Litscher G, Weis M, et al. Continuous mul-
tivariable monitoring in neurological intensive care patients—preliminary reports on four cases. Intensive Care Med.
1991;17:857–893.

ischemia (6), and its disappearance signals that cortical damage anesthesia is demonstrated in Figure 58.3. The recordings were
occurred early in the morning. BAEPs show gradual deterioration taken from a 31-year-old patient during the initial phase of an
of the waves V, IV, III, and II in the time between the disappear- orthopedic surgical intervention without neurological complica-
ance of the cortical and cervical SEP components. This can be tions. Collection of the biological data started before anesthesia.
interpreted to mean that the brainstem function deteriorated dur- Approximately 1 minute after administration of etomidate
ing that time (starting from the midbrain level). In summary, it is (Hypnomidate) (23 mg), the expected EEG changes occurred
clear that the most comprehensive picture of the cerebral state in (Fig. 60.3; marked with A). The cardiovascular effects intensified
comatose patients can be obtained only by monitoring different after intubation (B) and after application of pancuronium (4 mg)
neuronal systems and signals with use of EEG, SEPs, and BAEPs. (C). With increasing concentration of halothane, changes in EEG
The importance of multimodality EP measurements in spectra and cortical SEPs were accentuated (E). The patient was
patients with severe head injury was documented by Anderson monitored for more than 90 minutes. Despite the massive changes
et al. (7), Greenberg et al. (8), and others. EPs can be even more in the EEG as demonstrated in the band power trend curves due
reliable than intracranial pressure measurements in predicting to the individual concentration, the BAEPs remained nearly
the clinical outcome (7). Combined EEG and EP monitoring unchanged. The cervical SEP showed an increase in latency of
can also be used to differentiate comas due to structural lesions 1 msec, and the cortical N20 component was not identifiable with
from those of metabolic origin (9). deeper levels of anesthesia. Synchronous systolic blood pressure
Besides EEG and EPs, the HR and HRV are displayed in measurements displayed a decrease from 135 to 85 mm Hg.
Figure 58.2. The HRV was initially high, dropped at 1:00 p.m., The protocol of Figure 58.3 is a good example of the influ-
and remained depressed, with an exceptional increase during ence of the level of anesthesia on different neuronal systems.
the disappearance of the cervical component of the SEP. The EEG and cortical SEPs are heavily changed, while BAEPs and
HRV indicates the spontaneous physiological variations in the cervical SEPs are only slightly modified during high levels of
HR modulated by the parasympathetic and sympathetic activ- halothane. It is of interest to see the different behavior of power
ity of the cardiac nerves. This activity originates mainly in the trend curves dependent on the frequency band used. Blood
medullary circulatory center but is also influenced by higher pressure and HR measurements give additional information on
centers (10). HRV in normal subjects and in brain death was the cardiovascular system and should be also recorded.
reported by Schwarz et al. (11) and in newborns by Mehta et al.
(12). Strong correlations exist between EEG and HRV during SLEEP MONITORING IN INFANTS AT RISK
sleep (13). The measurement of HRV, therefore, is a sensitive
parameter to monitor brainstem functions in parallel with or In polysomnography, a great number of different signals must be
instead of BAEPs (BAEP measurements are not always possible recorded and analyzed. One goal of the sleep polygraphy in
in patients with severe head injury). A decrease of HRV, there- infants is to study babies at risk for sudden infant death syndrome
fore, can also indicate deterioration of brainstem function. (SIDS), that is, babies with sleep apnea and near-miss SIDS
(18,19). In this type of monitoring, in addition to the EEG, differ-
MONITORING IN THE OPERATING ROOM ent physiological signals such as HR electro-oculogram (EOG),
electromyogram (EMG), pO2, and pCO2 must be recorded and
In the operating room, there are two different applications for analyzed to reliably classify sleep stages and to differentiate
cerebral monitoring systems. One is to provide the surgeon with between rapid eye movement (REM) (active sleep) and non-REM
continuous information and warnings about potential damage to (NREM) (quiet sleep) state. Additional recording of BAEPs is
the spinal cord and other neuronal structures. Monitoring, there- technically possible and allows for monitoring of the brainstem at
fore, is indicated in spinal cord operations involving risk of the same time. Monitoring of BAEPs during nocturnal sleep
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

ischemia from compression of feeding blood vessels, aneurysm cycles was reported by Bastuji et al. (20). They found that the
surgery, cerebrovascular procedures, and posterior fossa opera- latency of wave V was increased and related to physiological
tions (1,14). The prognostic significance of SEP, BAEP, and serum hypothermia during the night. This gives evidence of how sensi-
S-100B monitoring after aneurysm injury was documented by tive BAEP measurements are and that brainstem signals are
Schick et al. (15). The second reason for cerebral monitoring is to affected not only by pathophysiological but also by physiological
assist the anesthetist in avoiding brain damage caused by hypoxic variations.
and/or ischemic events (16), to control the depth of anesthesia, An example demonstrating the simultaneous recordings of
and to avoid intraoperative wakefulness (17). Further details on EEG and BAEPs in a 4-month-old baby is shown in Figure 58.4.
intraoperative monitoring are presented in Chapter 39. EEG power trend curves, HR, pO2, and pCO2 show characteris-
Because of the different neuronal systems involved in the tic patterns in active sleep and quiet sleep. Slow EEG waves,
generation of EEG, SEPs, and BAEPs and their different sensi- spindle activity (10–12 Hz), and pCO2 increased during quiet
tivity to the effect of drugs, particularly anesthetics, it is very sleep, and HR and HRV decreased. In active sleep, HR and HRV
understandable that combined monitoring of all brain signals increased, and slow-wave activity and spindles decreased. A
together gives more information on neuronal signals than the period of arousal is indicated by HR increase. This example
monitoring of one signal alone. again demonstrates the usefulness of simultaneous monitoring
An example of multiparametric monitoring of EEG spectra of various cardiovascular and cerebral parameters, as, for exam-
combined using different modality EPs under halothane ple, HR, oxygen saturation, EEG spectra, SEPs, and BAEPs.

Figure 58.3 Protocol from a case under halothane anesthesia. From left to right: compressed EEG spectra, power trend
curves (0 to 4, 4 to 8, and 12 to 24 Hz), mean frequency (7 to 14 Hz), BAEPs, cervical and cortical SEPs, CCT, HR, HRV,
and diastolic and systolic blood pressure (BP). On the right side, the level of anesthesia is marked. See text for further
explanation.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Figure 58.4 Monitoring protocol from a 4 month-old baby during active and quiet sleep. From left to right:
Compressed power spectra from left and right hemispheres; BAEPs ipsi- and contralateral to auditory stimulation;
BAEP interpeak latency (I–V); 10- to 14-Hz power trend; HRV; pO2; pCO2; and EEG power trend curves. For further
explanation see text.

REFERENCES 10. Sayers B. Analysis of heart rate variability. Ergonomics. 1973;

16:17–32.
1. Nuwer MR. Evoked Potential Monitoring in the Operating Room. 11. Schwarz G, Pfurtscheller G, Litscher G, et al. Quantification of
New York, NY: Raven Press; 1986. autonomic activity in the brainstem in normal, comatose and
2. Maresch H, Pfurtscheller G, Schuy S. Brain function monitoring: a brain dead subjects using heart rate variability. Funct Neurol.
new method for simultaneous recording and processing of EEG 1987;2:149–154.
power spectrum and brainstem potentials. Biomed Tech. 1983;5: 12. Mehta SK, Super DM, Connuck D, et al. Hear rate variability in
117–122. healthy newborn infants. Am J Cardiol. 2002;89:50–53.
3. Pfurtscheller G, Schwarz G, Schroettner O, et al. Continuous and 13. Ako M, Kawara T, Uchida S, et al. Correlation between electroen-
simultaneous monitoring of EEG-spectra and brainstem auditory cephalography and heart rate variability during sleep. Psychiatr
and somatosensory evoked potentials in the intensive care unit Clin Neurosci. 2003;57(1):59.
and the operating room. J Clin Neurophysiol. 1987;4(4):389–396. 14. Grundy BL. Intraoperative monitoring of sensory-evoked poten-
4. Steller E, Litscher G, Maresch H, et al. Multivariables tials. Anesthesiology. 1983;58:72–87.
Langzeitmonitoring von zerebralen und kardiovaskulären Gröen 15. Schick U, Dohnert J, Meyer JJ, et al. Prognostic significance of
mit Hilfe eines Personal-Computers. Biomed Tech. 1990;35:90–97. SSEP, BAEP and serum S-100B monitoring after aneurysm sur-
5. Hilz MJ, Litscher G, Weis M, et al. Continuous multivariable mon- gery. Acta Neurol Scand. 2003;108(3):161–169.
itoring in neurological intensive care patients—preliminary 16. Prior PF, Maynard DE. Monitoring Cerebral Function. Amsterdam:
reports on four cases. Intensive Care Med. 1991;17:857–893. Elsevier; 1986.
6. Branston NM, Ladds A, Lindsay S, et al. Somatosensory evoked 17. Pichlmayr I, Lips K, Künkel H. The Electroencephalogram in
potentials in experimental brain ischemia. In: Pfurtscheller G, Anesthesia. Fundamentals, Practical Applications, Examples. Berlin:
Jonkman EJ, Lopes da Silva FH, et al., eds. Brain Ischemia: Springer-Verlag; 1984.
Quantitative EEG and Imaging Techniques, Progress in Brain 18. Brass M, Kravath R, Glass L. Death-scene investigation in sudden
Research. Vol. 62. Amsterdam: Elsevier; 1984:185–199. infant death. N Engl J Med 1986;315:100–105.
7. Anderson DC, Bundlie S, Rockswold GL. Multimodality evoked 19. Steinschneider, A. Prolonged apnea and the sudden infant death
potentials in closed head trauma. Arch Neurol. 1984;41:369–374. syndrome: clinical and laboratory observations. Pediatrics. 1972;
8. Greenberg RP, Newlon PG, Hyatt MS, et al. Prognostic implica- 50:646–654.
tions of early multimodality evoked potentials in severely head- 20. Bastuji H, Larrea LG, Bertrand O, et al. BAEP latency changes dur-
injured patients. J Neurosurg. 1981;55:227–236. ing nocturnal sleep are not correlated with sleep stages but with
9. Guerit JM. Medical technology assessment EEG and evoked body temperature variations. Electroencephalogr Clin Neurophysiol.
potentials in the intensive care unit. Clin Neurophysiol. 1999; 1988;70:9–15.
29(4):301–317.
Copyright © 2010. Wolters Kluwer Health. All rights reserved.

Principles of Digital Communication A Top-Down Approach Bixio Rimoldi
75% (4)
Principles of Digital Communication A Top-Down Approach Bixio Rimoldi
311 pages
Analog Communications
No ratings yet
Analog Communications
204 pages
6 - 23 April 2010: GPS Receivers, Receiver Signals and Principals of Operation
No ratings yet
6 - 23 April 2010: GPS Receivers, Receiver Signals and Principals of Operation
198 pages
EEG Stress
No ratings yet
EEG Stress
16 pages
Matching Pursuit and Unification in EEG Analysis - 1st Edition PDF
100% (11)
Matching Pursuit and Unification in EEG Analysis - 1st Edition PDF
16 pages
Digital Signal Processing DSP
No ratings yet
Digital Signal Processing DSP
1,080 pages
Eeg Circuit Design: NSF Project
No ratings yet
Eeg Circuit Design: NSF Project
136 pages
Basic Principles of QEEG
No ratings yet
Basic Principles of QEEG
7 pages
ETE All Chapter MCQ
No ratings yet
ETE All Chapter MCQ
28 pages
Topic20 Sampling Theorem
No ratings yet
Topic20 Sampling Theorem
6 pages
Worksheet 240520 155503
No ratings yet
Worksheet 240520 155503
9 pages
C-T To D-T Conversion Sampling of C-T Signals
100% (1)
C-T To D-T Conversion Sampling of C-T Signals
23 pages
Thesis Signal Sampling Techniques For Data Sampling
No ratings yet
Thesis Signal Sampling Techniques For Data Sampling
201 pages
FFT Vs Wavelet Transforms
No ratings yet
FFT Vs Wavelet Transforms
7 pages
DSP PPT Mod 1
No ratings yet
DSP PPT Mod 1
54 pages
Cours Mansoor Youssefi
No ratings yet
Cours Mansoor Youssefi
95 pages
Diogo Sato T Ese
No ratings yet
Diogo Sato T Ese
70 pages
Lecture 1&2
No ratings yet
Lecture 1&2
36 pages
Instructor: MR Nasolwa Edson Mob: 0765278090 Office: AB14 (Administration Block)
No ratings yet
Instructor: MR Nasolwa Edson Mob: 0765278090 Office: AB14 (Administration Block)
72 pages
HCI M3 Lect24-09.05.25
No ratings yet
HCI M3 Lect24-09.05.25
21 pages
05 Csoverview PDF
No ratings yet
05 Csoverview PDF
87 pages
Unit 3
No ratings yet
Unit 3
39 pages
Analog Design Magazine 2024-3
No ratings yet
Analog Design Magazine 2024-3
31 pages
Multi-Dimensional Framework For EEG Signal Processing and Denoising Through Tensor-Based Architecture
No ratings yet
Multi-Dimensional Framework For EEG Signal Processing and Denoising Through Tensor-Based Architecture
19 pages
Sensors 23 06434 v2
No ratings yet
Sensors 23 06434 v2
27 pages
Ch6 Sampling PDF
No ratings yet
Ch6 Sampling PDF
70 pages
Signal Processing of EEG
No ratings yet
Signal Processing of EEG
27 pages
Accepted Manuscript: Neuroscience Letters
No ratings yet
Accepted Manuscript: Neuroscience Letters
20 pages
Paper On Ewt
No ratings yet
Paper On Ewt
23 pages
A Primer On EEG Spectrograms
No ratings yet
A Primer On EEG Spectrograms
19 pages
Kim Et Al., 2018. EEG Spectral Analysis
No ratings yet
Kim Et Al., 2018. EEG Spectral Analysis
19 pages
Kim 2018
No ratings yet
Kim 2018
19 pages
Samiee 2015
No ratings yet
Samiee 2015
13 pages
Challenges and New Perspectives of Developmental Cognitive EEG - 2022 - NeuroIma
No ratings yet
Challenges and New Perspectives of Developmental Cognitive EEG - 2022 - NeuroIma
15 pages
Li 2021 - Demystifying Signal Processing Techniques To Extract Resting-State EEG Features For Psychologists
No ratings yet
Li 2021 - Demystifying Signal Processing Techniques To Extract Resting-State EEG Features For Psychologists
21 pages
Speech Recognition
No ratings yet
Speech Recognition
14 pages
10.1016@B978 0 444 64032 1.00009 6
No ratings yet
10.1016@B978 0 444 64032 1.00009 6
18 pages
Depression Detection Model Based On Discrete Wavelet Transform Associated With Genetic Algorithm
No ratings yet
Depression Detection Model Based On Discrete Wavelet Transform Associated With Genetic Algorithm
19 pages
Tonner 2006
No ratings yet
Tonner 2006
13 pages
Wavelets For EEG Analysis
No ratings yet
Wavelets For EEG Analysis
17 pages
Anna University: Chennai 600 025 MAY 2012 Department of Electronics and Communication Engineering Project Vivavoce
No ratings yet
Anna University: Chennai 600 025 MAY 2012 Department of Electronics and Communication Engineering Project Vivavoce
42 pages
Module - 5 Wearable Devices Notes
No ratings yet
Module - 5 Wearable Devices Notes
16 pages
Conf - Analysis of Eeg Sig For The Estimation of Concentration Level of Humans
No ratings yet
Conf - Analysis of Eeg Sig For The Estimation of Concentration Level of Humans
8 pages
Chapter 5
No ratings yet
Chapter 5
64 pages
PCS Module 5a
No ratings yet
PCS Module 5a
10 pages
Research Paper 2017-3
No ratings yet
Research Paper 2017-3
11 pages
A Combined Model For A Bucket Brigade Device and Its Input and Output Filters
No ratings yet
A Combined Model For A Bucket Brigade Device and Its Input and Output Filters
8 pages
Extracting Frequency Bands EEG
No ratings yet
Extracting Frequency Bands EEG
9 pages
EPIC: Annotated Epileptic EEG Independent Components For Artifact Reduction
No ratings yet
EPIC: Annotated Epileptic EEG Independent Components For Artifact Reduction
9 pages
5.american Journal
No ratings yet
5.american Journal
11 pages
ERAD04 - P - 67 - Gaussian Model Adaptive Processing (GMAP) For Improved Ground Clutter Cancellation and Moment Calculation
No ratings yet
ERAD04 - P - 67 - Gaussian Model Adaptive Processing (GMAP) For Improved Ground Clutter Cancellation and Moment Calculation
7 pages
JETIR2405449 Published
No ratings yet
JETIR2405449 Published
6 pages
Early Seizure
No ratings yet
Early Seizure
10 pages
A Novel Signal Modeling Approach For Classification of Seizure and Seizure-Free EEG Signals
No ratings yet
A Novel Signal Modeling Approach For Classification of Seizure and Seizure-Free EEG Signals
11 pages
Lab Report 6: Digital System and Processing
No ratings yet
Lab Report 6: Digital System and Processing
8 pages
Analysis of EEG Signals
No ratings yet
Analysis of EEG Signals
16 pages
1848 Article p287
No ratings yet
1848 Article p287
11 pages
(6-12) A Novel Approach For EEG
No ratings yet
(6-12) A Novel Approach For EEG
7 pages
Jms 2
No ratings yet
Jms 2
8 pages
Epileptic Seizure Prediction Based On Features Extracted Using Wavelet Decomposition and Linear Prediction Filter
No ratings yet
Epileptic Seizure Prediction Based On Features Extracted Using Wavelet Decomposition and Linear Prediction Filter
6 pages
Digital Image Processing
No ratings yet
Digital Image Processing
26 pages
Eeg Feature Extraction Methods
No ratings yet
Eeg Feature Extraction Methods
7 pages
18EC62 Module 2 FINAL
No ratings yet
18EC62 Module 2 FINAL
8 pages
EEG Signal Analysis: A Survey: Original Paper
No ratings yet
EEG Signal Analysis: A Survey: Original Paper
18 pages
A Review On Analysis of EEG Signals PDF
No ratings yet
A Review On Analysis of EEG Signals PDF
4 pages
Problem Sheet 5 - Digitisation: (Sampling and Quantization)
No ratings yet
Problem Sheet 5 - Digitisation: (Sampling and Quantization)
6 pages
Wavelet Preprocessing For Automated Neural Network Detection of EEG Spikes
No ratings yet
Wavelet Preprocessing For Automated Neural Network Detection of EEG Spikes
7 pages
Final Paper
No ratings yet
Final Paper
4 pages
Sampling With Automatic Gain Control 4
No ratings yet
Sampling With Automatic Gain Control 4
31 pages
assaIJETT V43P227
No ratings yet
assaIJETT V43P227
5 pages
A New Feature Extraction Method To Improve Emotion Detection Using EEG Signals
No ratings yet
A New Feature Extraction Method To Improve Emotion Detection Using EEG Signals
15 pages
Time Analysis
No ratings yet
Time Analysis
8 pages
Early Detection of Epilepsy Using EEG Signals: Selvin Pradeep Kumar.S Ajitha.L
No ratings yet
Early Detection of Epilepsy Using EEG Signals: Selvin Pradeep Kumar.S Ajitha.L
6 pages
Extraction of Statistical Features For Pattern Analysis During Meditation
No ratings yet
Extraction of Statistical Features For Pattern Analysis During Meditation
5 pages
Electroencephalograph (EEG) Signal Feature Extraction For Unspoken-Speech Identification Using EEGLAB
No ratings yet
Electroencephalograph (EEG) Signal Feature Extraction For Unspoken-Speech Identification Using EEGLAB
5 pages
Feature Extraction Techniques and Classification Algorithms For EEG Signals To Detect Human Stress - A Review
No ratings yet
Feature Extraction Techniques and Classification Algorithms For EEG Signals To Detect Human Stress - A Review
7 pages
Analysis of EEG Signals For Diagnosis of Epileptic Seizures Based On Features Extracted Using Efficient Wavelet Transform
No ratings yet
Analysis of EEG Signals For Diagnosis of Epileptic Seizures Based On Features Extracted Using Efficient Wavelet Transform
5 pages
Successful VSA Analysis Using An Oscilloscope
No ratings yet
Successful VSA Analysis Using An Oscilloscope
18 pages
Chapter Three Source Coding: 1-Sampling Theorem
No ratings yet
Chapter Three Source Coding: 1-Sampling Theorem
19 pages
Study On EEG Signal Processing and Feature Extraction Techniques
No ratings yet
Study On EEG Signal Processing and Feature Extraction Techniques
5 pages
A Pathway To Implement Security System Using EEG Signals of Digit Recognition Activity
No ratings yet
A Pathway To Implement Security System Using EEG Signals of Digit Recognition Activity
8 pages
Nyquist-Shannon Sampling Theorem
No ratings yet
Nyquist-Shannon Sampling Theorem
10 pages
Signals Sampling Theorem
No ratings yet
Signals Sampling Theorem
8 pages
Matlab FE - Toolbox - An Universal Utility For Feature Extraction of EEG Signals For BCI Realization PDF
No ratings yet
Matlab FE - Toolbox - An Universal Utility For Feature Extraction of EEG Signals For BCI Realization PDF
3 pages
Compressive Wideband Power Spectrum Analysis For Eeg Signals Using Fastica and Neural Network
No ratings yet
Compressive Wideband Power Spectrum Analysis For Eeg Signals Using Fastica and Neural Network
7 pages
A New Approach For Diagnosing Epilepsy by Using Wavelet Transform and Neural Networks
No ratings yet
A New Approach For Diagnosing Epilepsy by Using Wavelet Transform and Neural Networks
4 pages
Basic Principles of QEEG
No ratings yet
Basic Principles of QEEG
7 pages
Atlas of the Electrical Generators of Sleep
From Everand
Atlas of the Electrical Generators of Sleep
Dr. Mark Doidge
No ratings yet
Neurofeedback - The Neurofeedback Book for Patients and Therapists : A Symphony of the Mind
From Everand
Neurofeedback - The Neurofeedback Book for Patients and Therapists : A Symphony of the Mind
Jean-Maurice Cecilia-Menzel
No ratings yet
Neuroevolution: Fundamentals and Applications for Surpassing Human Intelligence with Neuroevolution
From Everand
Neuroevolution: Fundamentals and Applications for Surpassing Human Intelligence with Neuroevolution
Fouad Sabry
No ratings yet

EEG Analysis: Theory and Practice

Uploaded by

EEG Analysis: Theory and Practice

Uploaded by

Part IX Computer-Assisted EEG Analysis

EEG Analysis: Theory and Practice

Also definable is a class of statistical functions characteristic

tions of one stochastic variable x(ti) and may be characterized

In general terms, successive values of a signal, such as an EEG, T

E[ (x(t1 )  E(x(t1 ) ) ) (x(t2 )  E(x(t2 ) ) ) ] (54.5)

Estimating the covariance between any two variables x(t1)

one realization of the signal, of the product of the signal and a

cant bispectrum and skewness. The process in such a case is

Pxx (fi )  a W(fk )Fxx (fik )

discussion below illustrates how these spectral moments a0, a2,

INTERVAL OR PERIOD ANALYSIS

The zero-order moment is then:

a0  Sxx (f)df (54.18)

which is equal to the variance 2.

the interval between zero crossings, and in amplitude by the

Figure 54.6 Two examples for comparison of iterative interval analysis

above; a pronounced rhythmic component (peak at about 11 Hz) is pres-

Figure 54.8 Left-hand column: Different ways of plotting

(1 – )); the advantage of this form of presentation lies in

analyzed is necessary. Assuming, thus, that in an EEG signal a Length

vertical scale is simply the spectral density as computed by way

Figure 54.10 Contour plot of power spectra: note the fre-

Dumermuth G. Fundamentals of spectral analysis in elec-

presenting estimate variability is to present the corresponding

It must be noted, however, that computing the descriptor

(31) laboratory, both an estimation of the model parameters

a bi (exp(  j2fi ¢t) )

In case of the AR model, the spectral density is as follows:

example of the spectral decomposition of an EEG signal calcu-

model of order p with q = p – 1; a good choice appeared to be

autoregressive model. segmentation method is useful for obtaining significant data

extracting EEG features and sets of features follows a syntactic

The efficiency of the estimator s(t) is defined as follows:

quency in an optimal way. Some of these methods use a special

2009;30(9):2986–2996. mon to photosensitive epilepsy and visual perception? Brain.

EEG Mapping and Source Imaging

resulted in a number of claims that later proved to be wrong.

potential map can be constructed for any single moment in time,

but is ignored and misinterpreted as occipital gamma activity

Figure 55.6 Microstate segmentation of

a fundamental driving force establishing the electric potentials

where wY is the vector of scalp EEG measurements, X Y is the vec-

The above equation indicates that the distributed source

scalp EEG electrodes. Additional constraints have to be

Figure 55.10 fMRI–EEG integration. A: The pattern-

(From Liu Z, He B. FMRI–EEG integrated cortical source

study the relation between hemodynamic and electric activity,

Figure 55.12 Connectivity estimation. Left panel: Functional connectivity

from EEGs. Ann Biomed Eng. 2001;29:1019–1027. Acad Sci U S A. 1998;95:8945–8950.

EEG Features Obtained from Spectral Analysis

Frequency Bands Factor Analysis

Figure 56.2 Canonogram from subject with multiple metas-

tases in right hemisphere. The size of each polygon is propor-

correlation). The computation of an ARMA or AR model yields

The Recognition and Elimination of Artifacts:

Not all transient nonstationarities, however, are necessarily

performed based on information taken from a segment of the

interest in order to investigate correlations between such phe- studies.

trolled laboratory conditions, particularly regarding their

orders: the pros. Electroencephalogr Clin Neurophysiol. 1977;42:267–274.

localization of epileptic cortex—impact on surgical treatment. Electroencephalogr Clin Neurophysiol. 1974;37:305–308.

with, intends to transmit a message. The first mode is externally

Figure 57.4 Schema for the workflow of an imagery-based BCI with

changes in the EEG (Fig. 57.4C). Depending on the classifica-

ent brain-computer interface using steady state visual evoked 2009;120:24–29.

Figure 58.1 Sampling scheme of EEG, somatosensory-

REFERENCES 10. Sayers B. Analysis of heart rate variability. Ergonomics. 1973;

You might also like