*

Alpha rhythms are predominant during wakefulness

ELECTROENCEPHALOGRAPHY (EEG)
and are most pronounced in the posterior regions of
the head. They are best observed when the eyes are
KATARZYNA BLINOWSKA
closed and the subject is in a relaxed state. They are
PIOTR DURKA
Warsaw University blocked or attenuated by attention (especially visual)
Warszawa, Poland and by mental effort. Mu rhythms have a frequency
band similar to alpha, but their topography and
physiological significance are different. They are
Electroencephalography is a domain concerning recording related to the function of motor cortex and are pre-
and interpretation of the electroencephalogram. Electro- valent in the central part of the head. Mu rhythms
encephalogram (EEG) is a record of the electric signal are blocked by motor functions.
generated by the cooperative action of brain cells, or more *
Beta activity is characteristic for the states of in-
precisely, the time course of extracellular field potentials creased alertness and focused attention, as was
generated by their synchronous action. Electroencephalo- shown in several animal and human studies.
gram derives from the Greek words enkephalo (brain) and *
Gamma activity is connected with information pro-
graphein (to write). EEG can be measured by means of cessing (e.g., recognition of sensory stimuli) (3) and
electrodes placed on the scalp or directly on the cortex. In the onset of voluntary movements. In general, it can
the latter case, it is sometimes called electrocorticogram be summarized that the slowest cortical rhythms are
(ECoG). Electric fields measured intracortically were related to an idle brain and the fastest to information
named Local Fields Potentials (LFP). EEG recorded in processing.
the absence of an external stimulus is called spontaneous
EEG; EEG generated as a response to external or internal The EEG is observed in all mammals, the character-
stimulus is called an event-related potential (ERP). The istics of primate EEG being closest to the human. Cat, dog,
amplitude of EEG of a normal subject in the awake state and rodent EEGs also resemble human EEGs, but have
recorded with the scalp electrodes is 10–100 mV. In case of different spectral content. In lower vertebrates, electric
epilepsy, the EEG amplitudes may increase by almost an brain activity is also observed, but it lacks the rhythmical
order of magnitude. In the cortex, amplitudes are in the behavior found in higher vertebrate recordings.
range 500–1500 mV.

2. SHORT HISTORY OF ELECTROENCEPHALOGRAPHY

1. EEG RHYTHMS
Richard Caton (1842–1926) is regarded as the first scien-
The following rhythms have been distinguished in EEG tist to investigate brain potentials. He worked on the
(Fig. 1): delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), exposed brains of cats and rabbits, measuring electric
beta (13–30 Hz), and gamma (above 30 Hz). Gamma com- currents by means of a galvanometer, where a beam of
ponents are difficult to record by scalp electrodes and their light reflected from its mirror was projected onto a scale
frequency does not exceed 45 Hz; in ECoG components, up placed on a nearby wall. The results (presented in 1875)
to 100 Hz, or even higher, may be registered. The con- showed that ‘‘feeble currents of varying directions pass
tribution of different rhythms to the EEG depends on the through the multiplier when the electrodes are placed at
age and behavioral state of the subject, mainly the level of two points of the external surface, or one electrode on the
alertness. Considerable intersubject differences in EEG gray matter and one on the surface of skull.’’ This ob-
characteristics also exist. EEG pattern is influenced by servation can be regarded as a discovery of electroence-
neuro-pathological conditions, metabolic disorders, and phalographic activity.
drug action (1). Adolf Beck (1863–1939) also investigated spontaneous
activity of the brains of rabbits and dogs. He was the first
to discover (in 1890) the rhythmical oscillations of brain
*
Delta rhythm is a predominant feature in EEGs electrical activity. He also observed the disappearance of
recorded during deep sleep. In this stage, delta waves these oscillations when the eyes were stimulated with
usually have large amplitudes (75–200 mV) and show light, which was the first discovery of so-called ‘‘alpha
strong coherence all over the scalp. blocking.’’ Later, his co-worker Napoleon Cybulski (1854–
*
Theta rhythms rarely occur in adult humans. How- 1919) presented the electroencephalogram in a graphical
ever, they are predominant in rodents; in this case, form by applying a galvanometer with a photographic
the frequency range is broader (4–12 Hz) and waves attachment and was the first to observe epileptic EEG
have a high amplitude and characteristic sawtooth activity in a dog elicited by an electric stimulation (4). In
shape. It is hypothesized that theta rhythms in 1929, the first electroencephalogram was recorded from
rodents serve as a gating mechanism in the informa- the surface of the human scalp by Hans Berger (5).
tion transfer between the brain structures (2). In 1935 witnessed birth of the major fields of today’s
humans, activity in the theta band may occur in clinical electroencephalography. F. Gibbs and H. Davis
emotional or some cognitive states; it can be also showed association of 3/sec spike-wave complexes in EEG
connected with the slowing of alpha rhythms caused with epileptic absences and A. L. Loomis et al. studied
by pathology. human sleep patterns. Also in 1935, the first electroence-
1

Wiley Encyclopedia of Biomedical Engineering, Copyright & 2006 John Wiley & Sons, Inc.
2 ELECTROENCEPHALOGRAPHY (EEG)

50

0

−50
1 5 10 15 s
50

0


−50
1 5 10 15 s
50

0


−50
1 5 10 15 s
50

0


−50
1 5 10 15 s
500
Epilepsy

−500
1 5 10 15 s
Figure 1. Characteristic EEG rhythms, from the top: d (0.5–4 Hz), y (4–8 Hz), a (8–13 Hz), b (13–
30 Hz). The lowest trace–EEG during epileptic seizure, note that the amplitude scale is an order of
magnitude bigger.

phalograph (Grass Model I) started the era of contempor- change of polarity of the inside of the neuron from
ary EEG recording. More information about the history of negative to positive (about þ 30 mV). A subsequent in-
electroencephalography may be found in (1) and (4). crease of membrane permeability to K þ ions (leading to
their outflow from the cell), and a decrease of permeability
for Na þ ions makes the inside of the cell negative again
3. NEUROPHYSIOLOGICAL BASIS OF EEG with respect to the surrounding medium. In this way,
action potential of characteristic spike-like shape (dura-
In the brain, two main classes of cells exist: nervous cells, tion about 1 ms) is created. It obeys the ‘‘all or nothing’’
called neurons (Fig. 2), and glial cells. In both, the resting rule: for supra-threshold stimuli, a pulse of a constant
potential is approximately
80 mV, with the inside of amplitude is generated; for subthreshold excitation, the
cells being negative. The difference of potentials across a neuron doesn’t fire.
cell membrane comes from the difference of concentration PSPs are connected with the phenomena occurring on
of cations: K þ , Na þ , anions Cl
, and large organic the postsynaptic membrane. When action potential ar-
anions. Ca þ þ ions are less abundant, but they have an rives at the synapse, it secretes a chemical substance,
important regulatory role. The potential difference is called mediator or transmitter, which causes a change in
maintained by the active transport of cations K þ to the the permeability of the postsynaptic membrane of the next
inside of the cell and Na þ to the outside, using the energy neuron. As a result, ions traverse the membrane and a
supplied through metabolic processes. difference in potentials across the membrane is created.
Electric activity of neurons is manifested by generation When the negativity inside the neuron is decreased (e.g.,
of action potentials and postsynaptic potentials (PSP). by the influx of Na þ ions), the possibility of firing is higher
Action potentials occur when the electrical excitation of and an excitatory postsynaptic potential (EPSP) is gener-
the membrane exceeds a threshold. Postsynaptic poten- ated. An inhibitory postsynaptic potential (IPSP) is cre-
tials are subthreshold phenomena. The generation of ated when the negativity inside the neuron is increased
action potentials is connected with rapid increase of and the neuron becomes hyperpolarized. Unlike the action
permeability for Na þ ions. Their influx in the cell causes potential, the PSPs are graded potentials, their ampli-
a rapid increase of the potential inside the cell and the
 ELECTROENCEPHALOGRAPHY (EEG) 3

contribution to the electric field of neurons acting syn-

chronously is approximately proportional to their number,
and, for those firing nonsynchronously, as a square root of
their number. For example, if the electrode records action
of 108neurons (which is typical for scalp electrode) and 1%
of them are acting synchronously, their contribution will
be 100 times bigger than the contribution of neurons
acting asynchronously, because 106/sqrt(108) ¼ 100.
The problem of the origins of EEG rhythmical activity
has been approached by electrophysiological studies on
brain nerve cells and by the modeling of electrical activity
of the neural populations (2,3). The question emerges
whether the rhythms are caused by single cells with
pacemaker properties or by the oscillating neural net-
works. It has been shown that some thalamic neurons
display oscillatory behavior, even in the absence of synap-
tic input (7). Evidence exists that the intrinsic oscillatory
properties of some neurons contribute to the shaping of
the rhythmic behavior of networks to which they belong.
However, these properties may not be sufficient to account
for the network rhythmic behavior (2). It is generally
accepted that cooperative properties of networks consist-
ing of excitatory and inhibitory neurons connected by
feedback loops play the crucial role in establishing EEG
rhythms. The frequency of oscillation depends on the
intrinsic membrane properties, on the membrane poten-
tial of the individual neurons, and on the strength of the
synaptic interactions.
In the past, the role of EEG in information processing
has not been fully recognized. However, strong evidence
exists that coherent oscillations in the beta range in a
population of neurons might be the basic mechanism in
feature binding of the visual system (8). It seems that this
observation is not limited to the visual system and that
synchronized oscillatory activity provides an efficient way
Figure 2. From the top: neuron, synapse, postsynaptic potential. to switch the brain system between different behavioral
states and to cause a qualitative transition between modes
of information processing. In this way, neuronal groups
tudes are proportional to the amount of secreted mediator, with a similar dynamic functional state can be formed,
which depends on the excitation of the input neuron. subserving perceptual processes. It has also been postu-
Postsynaptic potentials typically have amplitudes of 5– lated that the role of synchronized oscillatory EEG activ-
10 mV and a time span 10–50 msec. In order to obtain ity in the alpha and theta range is to serve as a gating
supra-threshold excitation, the amplitudes of many post- mechanism to the flow of the information through the
synaptic potentials have to be superimposed in the soma of network. Bursts of oscillatory activity may constitute a
a neuron. A neuron can have very abundant arborizations, mechanism by which the brain can regulate changes of
making up to 10,000 synaptic junctions with other neu- state in selected neuronal networks and change the route
rons (in the human brain, about 1011 neurons exist). of information (2).
The electrical activity of neurons generates currents
along the cell membrane in the intra- and extracellular
spaces, producing an electric field conforming approxi- 4. RECORDING STANDARDS
mately to that of a dipole. Macroscopic observation of
this electric field requires the synchronization of electrical EEG is usually registered by means of electrodes placed
activity of a large number of dipoles oriented in parallel on the scalp. They can be secured by an adhesive (like
(6). Indeed, pyramidal cells of the cortex are, to a large collodion) or embedded in a special snug cap. The resis-
degree, parallel and, moreover, they are synchronized by tance of the connection should be less than 5 KOhms, so
virtue of common feeding by thalamocortical connections the recording site is first cleaned with diluted alcohol, and
(2). The condition of synchrony is fulfilled by the PSPs, conductive electrode paste applied to the electrode cup.
which are relatively long in duration. The contribution Knowledge of exact positions of electrodes is very
from action potentials to the electric field measured ex- important for both interpretation of a single recording as
tracranially is negligible. EEG comes from the summation well as comparison of results, hence the need for standar-
of synchronously generated postsynaptic potentials. The dization. The traditional 10–20 electrode system (9) states
4 ELECTROENCEPHALOGRAPHY (EEG)

positions of 19 EEG electrodes (and two electrodes placed movement, etc. Corresponding signals (EMG, EOG,
on earlobes A1/A2) related to specific anatomic landmarks, ECG, and body movements) registered simultaneously
such that 10–20% of the distance between them is used as with EEG are helpful in the visual rejection of artifact-
the electrode interval (Fig. 3). The first part of derivation’s contaminated epochs.
name indexes the array’s row—from the front of head: Fp, EEG is usually digitized by a 12-bit ADC (analog-
F, C, P, and O. The second part is formed from numbers digital conversion) with the sampling frequency ranging
even on the left and odd on the right side, in the center ‘‘z’’ from 100 Hz for spontaneous EEG and several hundred
or ‘‘0’’. Progress in topographic representation of EEG Hz for ERP to several kHz for recording short latency far-
recordings brought demand for a larger amount of deriva- field ERP. A black diagram of a recording setup is shown
tions. Electrode sites halfway between those defined by in Fig. 4. Prior to sampling, low-pass anti-aliasing filters
the standard 10–20 system were introduced in the ex- are used; high-pass filters are applied in order to eliminate
tended 10–20 system (10). artifacts of lowest frequencies.
EEG is a measure of potential difference; in the refer-
ential (or unipolar) setup, it is measured relative to the
same electrode for all derivations. This reference electrode 5. MATURATION OF EEG
is usually placed on the earlobe, nose, mastoid, chin, neck,
or scalp center. No universal consent exists regarding the EEG evolves with age and achieves its final character at
best position of the reference electrode, because currents 30 years, when it stabilizes and then starts to change
coming from bioelectric activity of muscles, heart, or brain again in the old age. The rate of change is correlated
propagate all over the human body. In the bipolar setup with mental health. EEG development in infancy and
(montage), each channel registers the potential difference adolescence is characterized by a shift of the EEG rhythm
between two particular scalp electrodes. Data recorded in toward higher frequencies. In newborns, slow delta
a referential setup can be transformed into any bipolar rhythms predominate, then the basic frequency shifts
montage, for the sake of display or further processing. The toward theta at the age of 12 months. The posterior slow
common ‘‘average reference’’ montage can be obtained by activity characteristic in young children constantly di-
subtracting from each channel the average activity from minishes during adolescence. Alpha rhythm appears at
all the remaining derivations. The Hjorth transform re- the age of 10 years (1). In young adults (21–30 years), the
ferences each electrode to the four closest neighbors, EEG still shows mild signs of immaturity including con-
which is an approximation of the Laplace transform tribution of 1.5–3 Hz and 4–7 Hz waves during awake
(LT). LT is calculated as a second spatial derivative of a state, normally not seen past the age of 30.
signal, offering information about vertical current density. Physiologically, the maturation process is connected
For best performance, it needs an adequate spatial sam- with the development of dendritic trees and myelination.
pling-interelectrode distance around 20 mm (e.g., 128 Myelin layers produced by glial cells cover the axons of
electrodes on the scalp). The estimates obtained by means neurons and act as an insulator of electrically conductive
of LT for the electrodes lying at the scalp periphery are cells. The propagation of electrical activity is faster and
biased and have to be excluded. less energy-consuming in myelinated fibers.
Contrary to the open question of the reference, the
necessity of artifact rejection is universally acknowledged.
The main problem lies in the lack of a working definition 6. SLEEP EEG
for an EEG artifact—it can stem from muscle or heart
activity (EMG, ECG), eye movement (EOG), external Sleep EEG reveals a characteristic alternating pattern.
electromagnetic field, poor electrode contact, subject’s The classic description of sleep involves division into
stages originally defined by Rechtschaffen and Kales

Figure 3. Electrodes placement in 10–20 system.

 ELECTROENCEPHALOGRAPHY (EEG) 5

Amplifier Paper
recordings

Low-pass
Diff. High-pass
(antialias)
amplifier frequency Amplifier
frequency
filter
filter
Analog
digital Digital
converter storage

Figure 4. Block diagram of recording of a single EEG channel. Differential amplifier measures
potential between two electrodes (one of them is treated as the reference). Analog filters and
adjustable; amplifier prepares the signal for analog-digital conversion (ADC) and storage (lower
path). Before the proliferation of digital media, EEG was stored on folded paper (upper path).

(R&K) (11): stage 1 (drowsiness), stage 2 (light sleep), and low-amplitude rhythms of 15–25 Hz band. Dee-
stage 3 (deep sleep), stage 4 (very deep sleep), and REM pening of drowsiness is connected with enhancement
(dreaming period accompanied by rapid eye movements). of slow activity and occurrence of vertex waves and
The differentiation of the sleep stages involves measure- slow rolling eye movements. According to R&K, the
ment of several signals. Their recording, called a poly- stage 1 is scored when less than 20% of the epoch
somnogram, includes not only EEG, but also contains any alpha activity and EEG consists of
electrooculogram (EOG), electromyogram (muscular ac- medium amplitude mixed frequency (mainly theta)
tivity), and respiration. It may also include measurement activity, sometimes with vertex sharp waves.
of blood flow, electrocardiogram (ECG), and oxygen level *
Stage 2 is characterized by appearance of sleep
in blood. EOG is recorded by means of the electrodes spindles, which are usually considered as a signal of
placed at the canthi of the eyes. As a result of the sleep onset. Slow frequencies ranging from 0.75 Hz to
cornoretinal standing potential (the cornea is positive 4 Hz are usually predominant in stage 2 of sleep;
relative to the fundus), the eye movements produce however, fast frequencies (15–30 Hz) may be present
changes in the potential between electrodes. The EOG too. Summarizing, stage 2 is characterized by spin-
and EMG help to distinguish REM state. The sequence of dles and K complexes, less than 20% of the epoch may
sleep stages is usually illustrated in the form of the contain delta waves.
hypnogram (Fig. 5). Recognition of stages is based on the *
Stage 3 is associated with preponderant slow rhythm
contribution of the different rhythms and the occurrence in the delta frequency (0.75–3 Hz) range; activity of
of characteristic signal transients absent in wake EEG, lower amplitude in 5–9 Hz range is also quite com-
namely sleep spindles, vertex waves, and K complexes. mon. In a sizeable number of healthy subjects, alpha
Sleep spindles are rhythmic waves of frequency 11–15 Hz activity (7–11 Hz) may be intermingled with delta
and duration longer than 0.5s; characteristic increase and rhythm, and, in this alternating pattern, certain
then gradual decrease of amplitude is not always ob- periodicities may occur. K complexes are still present
served. They are most prominent in the central deriva- in sleep stage 3; spindles are less abundant than in
tions; low-frequency spindles (11–12.5 Hz) are more stage 2. Stage 3 is scored when 20–50% of the epoch
pronounced in the frontal and high-frequency spindles contains delta waves of 0.5–2.5 Hz frequency and of
(12.5–15 Hz) in more posterior derivations (12). Vertex 75 mV or greater peak-to-peak amplitude.
wave is a compound potential: a small spike discharge of *
Stage 4 is dominated by slow-wave activity of high
positive polarity preceding a large spike and followed by a
amplitude; K complexes may appear. Stage 4 is
negative wave of latency around 100 ms and often another
scored when more than 50% of the epoch contains
small positive spike. Vertex waves are a kind of auditory-
delta activity conforming to the criteria defined
evoked response (AER), as can be judged from their shape
above.
and place of occurrence. The K complex consists of an
initial sharp component, followed by a slow component
*
REM is characterized by a decrease of EEG ampli-
that fuses with a superimposed fast component. The sharp tude, occurrence of faster rhythms, rapid eye move-
component may be biphasic or multiphasic. Sometimes the ments, and loss of muscular activity. Spectral
K complex is described only as having slow and fast characteristics in REM is polyrhythmic and, on the
components; the initiating sharp component is equated basis of EEG only, it is difficult to distinguish REM
with a vertex wave (1). from stage 1.
Sleep stages may be briefly characterized as follows.
The evolution of slow-wave activity and spindles during
overnight sleep is shown in Fig. 5.
*
Stage 1 (drowsiness) is associated with a decrease of Evidence exists that when the sleep becomes deeper,
alpha rhythm, rhythms in 2–7 Hz frequency band the sources that drive EEG activity move from the poster-
6 ELECTROENCEPHALOGRAPHY (EEG)

m
w
I
REM
II
III
IV

0 1 2 3 4 5 6 7

100
SWA% / epoch

50

20

0
0 1 2 3 4 5 6 7
num.of spindles / epoch

0
0 1 2 3 4 5 6 7
time [h]
Figure 5. Time evolution of the overnight sleep (horizontal scale in hours). Top: – Hypnogram (by
human expert; bottom: SWA and sleep spindles detected automatically from matching pursuit
parameterization of the EEG signal: % SWA denotes the percentage of epoch occupied by
waveforms classified as SWA. Continuous description of the slow-wave sleep is compatible with
the III/IV stages delineation defined R&K (11), as indicated by the 20% an 50% lines. Artifacts not
removed from analysis.

ior regions of the head (prevalent during awake state with intrusion of the alpha rhythm on the slow waves in stages
eyes closed) to the centro-frontal regions (13). 3 and 4 or the appearance of sleep spindles in REM. Also,
During the night, nonREM (NREM) and REM periods normal ultradian NREM/REM cyclicity may be altered or
occur in cycles. Slow-wave sleep is concentrated in the first lost. The normal REM onset in adults appears after about
one-third of the night and is predominant in childhood. one hour or later. Early (about 10 minutes) onset of REM
The sleep pattern changes very much during childhood sleep, called SOREM, may be an effect of previous REM
and adolescence. For newborn babies, REM takes most of deprivation, alcoholism, drug withdrawal, irregular sleep
the sleep time, and in young children, only REM and habits, severe depression, or narcolepxy-catalepsy. In the
nonREM stages can be distinguished. Diminuition of deep latest case, a poor REM cyclicity is observed.
slow-wave sleep and increase in wakefulness continues From the clinical point of view, not only sleep macro-
through entire life span after the age of 1 year. In old age, structure described by hypnogram, but also its micro-
the contribution of stages 3 and 4 decreases markedly and structure, is important. Transient arousals associated
the first REM stage appears later at night. The changes of with unstable sleep conditions are reflected in the EEG.
the sleep pattern may be caused not only by a normal A set of guidelines for arousal scoring has been proposed
aging, but also by degenerative diseases. An atypical by the American Academy of Sleep Medicine (14). An EEG
polysomnogram may be observed in a variety of situations arousal is defined as an abrupt shift in EEG frequency,
(e.g., sleep deprivation, abnormal sleep habits, drug and which may include theta, alpha, or frequencies greater
drug withdrawal effects, sleep pathologies). Dissociated or than 16 Hz, but not spindles. A set of additional conditions
otherwise atypical sleep patterns may be manifested by is given in Guilleminault et al. (14). A certain number of
 ELECTROENCEPHALOGRAPHY (EEG) 7

spontaneous arousals seems to be an intrinsic component An important diagnostic problem is localization of the
of physiological sleep, but their frequent occurrence may epileptic focus, which, in severe cases, can be possibly
be connected with respiratory sleep disorders, nocturnal removed by surgical intervention. Intracranial electrodes
myoclonus, and other clinical conditions. are usually placed in the suspected region, found from the
scalp EEG, in order to better localize the focus. Tests
involving measurement of ERP are performed in order to
7. PATHOLOGICAL CONDITIONS INFLUENCING EEG check if the removal of a given part of the brain will not
impair some vital brain functions. The epileptic focus may
EEG is affected by the CNS disorders (e.g., cerebral not necessarily be detected by imaging techniques such as
anoxia, cerebral inflammatory processes, cerebral palsy, tomography, so the information contained in EEG is
Creutzfeld–Jacob disease, and metabolic and degenerative essential for localization of epileptic foci.
nervous system disorders such as senile and presenile
dementias). It is influenced by brain tumors and cranio-
9. INFLUENCE OF DRUGS
cerebral traumas; in the second case, it can serve as a
measure of patient recovery. EEG is also an important test
EEG is very sensitive to the action of a wide range of
in psychiatric diseases, sleep disorders, and developmen-
pharmacological substances, especially psychotropic
tal disorders. In particular, analysis of evoked potentials is
drugs, anaesthetics, and anticonvulsants. It is also af-
very helpful in diagnosing dyslexia and differentiation
fected by some drugs targeted to organs other than the
between psychogenic and neurogenic disorders. The char-
central nervous system (CNS), such as antihistamines and
acter of EEG changes dramatically in epileptic condition
antihypertensives. Influence of drugs on EEG primarily
and its analysis is a basic tool in this case (1).
include changes in its spectral content and topographic
characteristics. Effects of psychoactive drugs on EEG
could be used to assess their action on the CNS. A
8. EPILEPTIC SEIZURE DISORDERS particular effect of a drug on EEG may be used as an
indication for its potential therapeutic efficiency.
An epileptic seizure is caused by the massive synchroniza-
tion of neuronal electrical activity. During the seizure,
groups of neurons discharge synchronously, creating a 10. EVENT-RELATED POTENTIALS
large amplitude signal and leading to uncontrollable
oscillations. Tumors, infections, trauma, or metabolic Event-related potentials (ERPs) are the changes of spon-
and toxic disorders may be responsible for the synchro- taneous EEG activity related to a specific event. ERPs
nized discharges. Epilepsy is the second most common triggered by particular stimuli, visual (VEP), auditory
neurological disease (15). Its clinical symptoms may in- (AEP), or somatosensory (SEP), are called evoked poten-
volve the loss of awareness, drop attacks, facial muscles tials (EP). It is assumed that ERPs are generated by
and eye movements, aggressive outbursts, prolonged con- activation of specific neural populations, time-locked to
fusional states, and flexor spasms of a whole body. the stimulus, or that they occur as the result of reorgani-
Seizure types can be divided into three main categories zation of ongoing EEG activity. The basic problem in
(15): analysis of ERPs is their detection within the larger
EEG activity. ERP amplitudes are an order of magnitude
smaller than that of the ongoing EEG. Averaging is a
1. Local—the synchronized electrical activity starts in
common technique in ERP analysis; it makes possible the
a well localized part of the brain. The seizure,
reduction of background EEG noise. However, assump-
lasting a few seconds, is accompanied by jerking or
tions underlying the averaging procedure, namely (1) the
spasms, as well as by a loss of consciousness.
background noise is a random process, (2) the ERP is
2. Generalized—the EEG patterns are bilaterally sym-
deterministic and repeatable, and (3) EEG and ERP are
metrical and roughly synchronous; the epileptic
independent, are not well justified.
activity is spread over wide areas of both hemi-
The ERP pattern depends on the nature of the stimula-
spheres simultaneously from the onset of attack.
tion, placement of the recording electrode, and the actual
3. Unclassifiable—different from those described in (1)
state of the brain. ERPs are usually described in terms of
and (2).
the amplitudes and latencies of their characteristic waves
(Fig. 6). The components occurring at different times are
In epileptic discharges, the membrane potential of different in nature; they are named early and late ERP.
cortical and deeper located neurons changes in a dramatic The early ERPs of latency below 10–12 ms (called some-
way, which leads to massive bursts of action potentials times ‘‘far fields’’) are connected with the response of the
and large fluctuations of intra- and extracellular fields. receptors and peripheral nervous system; late ERPs
The seizure initiation is probably connected with the (‘‘near-field’’ potentials) are generated in the brain. In
breakdown of the local inhibitory mechanisms. The crucial late ERPs, exogenous components (primarily dependent
factor in generation of epileptic activity is the synchroni- on characteristics of the external stimulus) and endogen-
zation of neural pools. Mechanisms of this synchronization ous components (which are dependent on internal cogni-
are probably connected with recurrent excitation operat- tive processes) can be distinguished. Endogenous
ing through positive feedback loops. components of latencies above 100–200 ms are influenced
8 ELECTROENCEPHALOGRAPHY (EEG)

EEG power distribution in time and frequency, electrode C1,

right hand finger movement

1
0.5
+3
0

c)
Time (se
0 5 −5
10 15 20 25 30
Frequency (Hz)

Figure 7. EEG power distribution in time and frequency during

voluntary finger movement (movement onset at time 0) for
electrode placed above sensorimotor cortex. The desynchroniza-
tion (decrease of amplitude) is visible for alpha and beta band
before the movement and their synchronization may be observed
after the movement. (EEG filtered in 0.5–100 Hz frequency range.
The power was divided by 1/f in order to suppress the frequency-
dependent background.)

ERPs need not to be time-locked to the stimuli, they can

occur without the fixed phase to the trigger. ERPs caused
by continuous stimuli (e.g., wave-modulated light or am-
plitude-modulated tone) or those that are not time-locked
to the stimulus are preferably analyzed in the frequency
Figure 6. Schematic representations of average auditory ERP domain. As an example, ERPs can serve preceding volun-
(upper picture) and visual ERP (lower picture) in logarithmic tary actions such as speech or movements. They are
time scale, showing the commonly recognized components. Letter usually accompanied by the change in the spectral content
‘‘N’’ denotes negative polarity, ‘‘P’’ positive polarity, usually fol- of the signals. Specifically, before the movement onset and
lowed by the number denoting latency in ms. The components of during the movement, a decrease of activity occurs in the
auditory potentials marked by roman numbers are the brain alpha and beta band in cortical regions connected with
stem-evoked responses (BAEP). They are followed by mid-latency sensorimotor cortex (Fig. 7). This phenomenon is called
exogenous components (MAEP) in the frequency range 10–
event-related desynchronization (16). After the move-
100 ms. The first peak in exogenous visual ERP comes from
ERG (electroretinogram). Exogenous ERP exhibit modality–spe-
ment, an increased synchronization in both these bands
cific features; endogenous ERP are similar in both modalities. occurs. During the action, an increase of activity in the
high frequencies (gamma band) takes place. It is not
always present in scalp recordings, but it is well visible
by the attention to the stimulus. The later components in electrocorticogram.
around 300 ms (P300) reflect recognition and discrimina-
tion between stimuli. P300 amplitude is considered as a
manifestation of CNS activation that reflects attention to
incoming stimulus, when memory representations are 11. EEG ANALYSIS
updated. P300 latency is dependent on the stimulus
classification speed (it is smaller for known stimuli) and The traditional method of EEG analysis is visual inspec-
the latency is connected with individual cognitive cap- tion of the signals plotted on paper. Modern computer
ability. analysis can extend electroencephalographer’s capabil-
ERPs are widely used in clinical practice as tests of the ities by supplying information not directly available
integrity of the sensory pathways or their different dys- from the raw data. However, visual analysis is still a
functions. They are also helpful in the diagnosis of diffused widespread technique, especially for detection of transient
brain diseases (e.g., multiple sclerosis or psychiatric dis- features of signal. In most cases, the agreement of an
orders). automatic method with visual analysis is a basic criterion
for its acceptance.
 ELECTROENCEPHALOGRAPHY (EEG) 9

As a result of its complexity, the EEG time series can be effective for a human observer to look at a map than at a
treated as a realization of a stochastic process, and the table of numbers. A map may help to make a direct
statistical properties can be evaluated by typical methods comparison between the topographic distribution of EEG
based on the theory of stochastic signals. These methods features and an anatomic image given, for example, by the
include probability distributions and their moments tomographic brain scan. Three types of features are most
(means, variances, higher-order moments), correlation commonly mapped for clinical applications (1) direct vari-
functions, and spectra. Estimation of these observables able such as amplitude, (2) transformed variable such as
is usually based on the assumption of stationarity, which total spectral power or relative spectral power in fre-
means that the statistical properties of the signal do not quency band, and (3) the result of statistical test applied
change during the observation time. Although the EEG to given EEG feature.
signals are ever changing, they can be subdivided into The appearance of a map depends very much on the
quasi-stationary epochs when recorded under constant electrode reference system. The recommended representa-
behavioral conditions. On the basis of empirical observa- tion involves surface Laplacians, because this approach
tions and statistical analysis performed by several approximates source current density and cancels a com-
authors, quasi-stationarity can be assumed for EEG mon component caused by volume conduction (6,21).
epochs of approximately 10 seconds in length, measured However, a reliable computation of surface Laplacian
under constant behavioral conditions (1). requires at least 64 electrodes and adequate spatial
EEG signals can be analyzed in the time or frequency sampling is obtained for 128 electrodes. Therefore, quite
domain, and one or several channels can be analyzed at a frequently an approximation of the Laplacian operator by
time. The applied methods involve spectral analysis by Hjorth transform (22) is applied [e.g., it was used as a
Fourier Transform (FT), autoregressive (AR) or autore- preprocessing method improving spatial resolution for
gressive-moving average (ARMA) parametric models, Kal- estimation of synchronization and desynchronisation of
man filters, and time-frequency and time-scale methods EEG activity (16)]. Results obtained by application of
(Wigner distributions, wavelets, matching pursuit). The Laplacian operator may be further ameliorated by deblur-
most common methods used for postprocessing include ring; that is, using a mathematical model of volume
cluster analysis, discriminant analysis, or artificial neural conduction through the skull and scalp to downwardly
networks (ANN). project scalp-recorded potentials, which provides a com-
Estimation of power spectra is one of the most fre- putational estimate of the electrical potentials, that would
quently used methods of EEG analysis. It provides infor- be recorded near the superficial cortical surface (23).
mation about the basic rhythms present in the signal and Interdependence between two EEG signals can be
can be easily and rapidly calculated by means of the Fast found by a cross-correlation function or its analogue in
Fourier Transform (FFT). Maximum entropy power spec- the frequency domain—coherence. Cross-correlation can
trum may be obtained by means of the autoregressive be used for comparison of EEGs from homologous deriva-
model, which can be recommended for the EEG analysis. tions on the scalp. A certain degree of difference between
The AR model represents a filter with a white noise at the these EEGs may be connected with functional differences
input and the EEG series at the output; it is compatible between brain hemispheres, but a low value of cross-
with a physiological model of the alpha rhythm generation correlation may also indicate a pathology. The cross-
(17), but this link is neither specific nor essential. The AR covariance functions have been extensively used in the
model provides a parametric description of the signal and analysis of event-related potentials for the study of the
makes possible its segmentation into stationary epochs. It electrophysiological correlates of cognitive functions (24).
also offers the possibility of detecting nonstationarities by Inter-relationships between EEG time series recorded at
means of the inverse filtering (1). different sites can also be quantified by information mea-
Autoregressive or autoregressive-moving average mod- sures (25) and coherences (26). Usually, an ordinary
els (ARMA), sometimes used for EEG analysis, belong to coherence calculated pair-wise between two signals is
the class of linear models. Some authors use nonlinear used. However, for the ensemble of channels taken from
methods for EEG analysis, based on estimators derived different derivations, the relationship between two signals
from chaos theory, such as attractor dimension or Lyapu- may come from the common driving from another site;
nov coefficients. However, these parameters have a very therefore, partial and multiple coherences should be taken
limited value for EEG, because this signal has a character into account as well (27).
of colored noise and reveals chaotic character only in some If the signals are modeled as a linear mixture of
epochs of epileptic seizures, as was shown by surrogate statistically independent ‘‘sources,’’ their activities can
data tests (18,19) and linear forecasting (20). Moreover, be found by means of the Independent Component Analy-
the above-mentioned chaotic estimators require long sta- sis (ICA) (28). This class of algorithms, usually based on
tionary data epochs, are subject to systematic errors, and the neural networks scheme, is used in general for the
are very sensitive to noise. ‘‘blind source separation’’ problems (BSS). ICA can be seen
The representation of EEG activity for a complete as an extension to the principal component analysis and
ensemble of channels records from scalp electrodes is factor analysis.
usually performed by mapping. The features of EEG can In order to find intrinsic relationships between signals
be extracted from multivariate statistics, so, in this re- from different locations, multivariate autoregressive
spect, graphic representation in the form of maps is model (MVAR) may be applied simultaneously to the
neither necessary nor sufficient. However, it is more whole set of EEG channels. From the MVAR coefficients,
10 ELECTROENCEPHALOGRAPHY (EEG)

partial, multiple, and ordinary (bivariate) coherences can progress was also achieved by introduction of wavelet
be found as well as the transfer function of the system. analysis.
Elements of the transfer matrix of the MVAR model have Wavelet transform (WT) describes signals in terms of
the meaning of Granger causality (19). It relies on predic- coefficients representing their energy content in specified
tion of the future of one channel from the past of the other time-frequency region. This representation is constructed
channels, which allows for the determination of the activ- by means of decomposition of the signal over a set of
ity propagation. Normalized Granger causality is equiva- functions generated by translating and scaling one func-
lent to the directed transfer function (30), which was used, tion called mother wavelet. The basics of wavelet analysis
for example, for the determination of the direction of the can be found in Mallat (36). Some of the applications of
propagation of EEG activity during overnight sleep (13), WT involve ERP component separation and measurement,
during epileptic seizure (31), and for the assessment of time-varying filtering for denoising single trial ERPs,
information flow between brain structures of behaving isolation of specific ERP and EEG rhythms, hearing
animals (32). For the set of mutually dependent signals, as threshold estimation via auditory brain stem-evoked re-
is the case for most of the recorded EEG signals, all sponse measurements, scale-specific topographic analysis,
involved channels have to be processed simultaneously. and data compression [referred in (37)]. WT is especially
Bivariate estimates of Granger causality or DTF can bring useful for evaluation of time-locked phenomena and their
quite misleading results (33). Recently, short-time direc- distinction from the non time-locked events. As the ex-
ted transfer function (Fig. 8) was introduced, which makes amples may indicate, distinction of both kinds of compo-
possible calculation of EEG flows not only as a function of nents in AEP (38) and reconstruction of a single AEP
frequency, but also of time when multiple repetitions of based on discrimination between background and evoked
experiments are available (29,34,35). activity parameterized by means of WT (39). Multiresolu-
The assessment of the time evolution of EEG and ERP tion analysis, offered by WT, provides the measure of EEG
is crucial for understanding the information processing by energy at each decomposition scale. This property was
the brain. Detection of transient EEG features is impor- used to achieve spatial enhancement of ERP to bring out
tant in diagnosis as well; therefore, time-frequency meth- topographic features that might not be seen without
ods operating in a short time scale are needed. The first processing (40). The assessment of complexity of the
method aiming at dynamic analysis is the windowed energy distribution in different frequency subbands may
Fourier transform with a sliding window. Substantial be evaluated in terms of wavelet entropy (41). This
measure was used, for example, to follow EEG evolution
after hypoxic-ischemic injury (42). WT was also used for
automatic detection of arousals during sleep (43). A wide
range of biomedical applications of wavelets together with
basic theory are described, for example, in (44).
Time and frequency resolution in WTs are subject to
certain restrictions that lead to poor frequency resolution
at high frequencies. The representation depends also on
the setting of the time window, which makes WT mostly
suitable to the evaluation of time-locked signals such as
EP, but less appropriate for detecting structures appear-
ing more or less randomly in the signal. This problem has
been approached by application of time-shift and fre-
quency-shift invariant time-frequency distributions from
the Cohen class. However, significant cross terms are
present in these distributions, and sophisticated mathe-
matics has to be applied to diminish their contribution.
Another drawback for EEG applications may stem from
the fact that, as continuous functions of time and fre-
quency, those distributions do not provide direct parame-
terization of signal structures.
These drawbacks are absent in the adaptive time-
frequency approximations, which decompose the signal
into waveforms of well-defined frequency, time occurrence,
Figure 8. Determination of the EEG propagation by means of time span, and amplitude. Such an iterative algorithm—
the SDTF. The arrows represent increase of EEG activity flows in matching pursuit (MP)—was introduced by Mallat and
the beta band in the 1–2 seconds after voluntary movement of the
Zhang in 1993 (45). In Fig. 9 time-frequency energy
right index finger. (Flows calculated as: SDTF functions inte-
grated in the beta band and time 1–2 s after movement in respect
distribution of an epileptic EEG signal, obtained by means
to SDTF in beta band in reference period before movement.) of the MP algorithm, is presented. MP parameterization
Based on the results described in (33). Dynamic propagation of makes possible statistical evaluation of EEG features and
EEG activity in time in the form of movie is available from automatic detection of desired signal structures (46).
Internet at http://eeg.pl. Application of MP to the detection and parameterization
of sleep spindles and slow waves is shown in Fig. 5, where
 ELECTROENCEPHALOGRAPHY (EEG) 11

Figure 9. Time-frequency representations of

EEG energy density during an epileptic sei-
zure, obtained by means of MP method. Upper
plot in 3-D, lower—the same in 2-D, below—
EEG trace (scale in seconds).

certain features of these structures are presented in time 12. LOCALIZATION OF CORTICAL SOURCES OF EEG
together with hypnogram. ACTIVITY
Parameterization by means of time-frequency features
may serve also as an input to the artificial neural net- The determination of geometry and orientation of cortical
works (ANN), usually with the aim of classification, dis- sources of EEG is a complex problem. Electrical activity
crimination, or feature extraction. ANN are constructed propagates along neuronal tracts and by volume conduc-
from artificial neurons (or units), which produce the out- tion. Potentials measured by scalp electrodes are attenu-
put depending on the sum of weighted inputs from other ated by media of different conductivity and complicated
units. Weights are modified in the process of learning. The geometry (cerebrospinal fluid, skull, skin), which results
most popular type of ANN applied for post-processing of in a decrease of their amplitude by over an order of
the EEG signals are networks with one or more hidden magnitude. However, the major problem in localization
layers of neurons (a hidden layer is a layer between input of the sources of EEG activity stems from the fact that
and output layer) and the supervised learning based on different configurations of sources can generate the same
‘‘back-propagation of errors’’ (47). This type of network distribution of potentials on the scalp. Therefore, a unique
was used, for example, for sleep stage scoring (48). Per- solution of the EEG inverse problem can be obtained only
formance of ANN depends heavily on the input para- by introducing extra a priori assumptions. Usually, one or
meters, which was demonstrated in many papers. For several dipole sources are assumed and their positions and
example, in Trejo and Shensa (49), it was found that orientation are estimated by an iterative fit to the mea-
prediction of human performance from EGG parameter- sured field [e.g., (53)]. The number of dipole sources is an
ized by means WT coefficients gives better results than open question; therefore, linear solutions based on dis-
application as input parameters to ANN raw data or tributed source models become more popular. However, in
components found by PCA. The training of the above- this case additional constrains on the solution are also
mentioned ANN with back propagation is based on ‘‘su- needed, like, for example, Laplacian-weighted minimum
pervised learning,’’ which means that the desired output is norm of the solution [LORETA (54)]. In any of these cases,
known. The unsupervised networks are based on competi- the solution space is usually a priori restricted to physio-
tion between units. The procedures of unsupervised learn- logically plausible locations. As a result of the nonunique-
ing minimize the sum of two factors: the cost of code and ness and high sensitivity to noise of the solution, results
the cost of reconstruction (50). Learning vector quantizer must be interpreted with care. For a recent review, see, for
(51) belongs to this type of network. Its version, improved example, (55).
in respect to enhancing informative features, by updating
weights in each step was used, for example, for EEG
classification during externally-paced hand movements 13. SPECIFIC CLINICAL APPLICATIONS
(52).
For particular medical applications, specific methods of
EEG analysis were designed. An example may be the
assessment of ‘‘drug profiles.’’ The method relies on esti-
mation of spectral power in basic frequency bands before
12 ELECTROENCEPHALOGRAPHY (EEG)

and after drug application and finding the significance of 14. COMPUTER-ASSISTED EEG DIAGNOSIS
changes by means of statistical tests. A drug profile is
constructed by plotting the results of a test (commonly in Computer-assisted diagnosis usually consists of two steps:
form of t-value of Student test) for frequency bands d, y, a1, feature extraction and classification. As a primary feature
a2, b1, and b2. The method is based on observation that in the amplitude domain mean, variance and higher-order
drugs of similar profile give similar effects. More recently, moments can be used; in the frequency domain, spectral
nonparametric tests came into use and the shifts of intensities in different frequency bands and their ratios
spectral peaks are considered as well (56). An interesting are mostly applied.
approach is the combination of pharmaco-EEG with phar- One of the first automatic diagnostic methods (67) was
macokinetics, by finding a relation between EEG features based on the observation that an increased amount of slow
and pharmacodynamical models (57). pathological EEG activity might be analogous to the slow
The analysis of overnight sleep is a very tedious and activity seen in the immature EEG. For each electrode,
time-consuming procedure; therefore, many attempts maturity calculated on the basis of spectral features was
were made to automate the procedure of hypnogram compared with the actual maturity. A significant discre-
construction [listed, for example, in (58)] as well as the pancy was considered an abnormality. In another diag-
automatic identification of arousals, for example, (43). nostic system (1), the ratio of slow and fast EEG activity as
Some authors (48) reported quite good performance of well as the degree of asymmetry between homologous
their systems [e.g., the system of sleep scoring based on derivations were taken into account.
assessment of EEG spectral power in frequency bands, The most extended diagnostic system, called Neuro-
their ratios and parameters derived from EMG and EOG metrics (68), is based on standardized data acquisition
signals, followed by classification with the use of ANN techniques and EEG and ERP feature extraction. A
(multilayer perceptron) achieved an average agreement neurometric test battery includes spontaneous EEG spec-
rate of 82%, whereas the interexpert agreement was tral intensities in frequency bands and their ratios and
87.5%]. The current tendencies in sleep EEG analysis similarity of signals from homologous derivations. ERP
are directed toward more continuous description of sleep are quantified in terms of amplitudes and latencies or by
going beyond R&K rules and assessment of sleep micro- means of application of principal component analysis. In
structure. Major progress may be achieved via incorporat- the last approach, the signals from different derivations
ing recognition of transient structures into analysis, such are represented as linear combinations of some basic
as sleep spindles and K complexes, and consideration of waveforms multiplied by weighting factors. In Neuro-
wave amplitude and wave incidence separately, not only in metrics, each parameter is subjected to a transformation,
terms of spectral power (59). The approach that offers such that the difference between individual index and the
possibilities of description of phasic and tonic features of group mean value is divided by the standard deviation of
EEG in the framework of one formalism is matching the whole sample. In this way, the metric is created
pursuit (46). reflecting the relative probability of finding the given
The analysis of epileptic activity serves for localization value within a normal reference group. The next steps of
of epileptic focus, monitoring of interictal activity, predic- diagnostic procedure involve application of multivariate
tion of seizure, and characterization of epileptic dis- statistical methods such as factor analysis, cluster analy-
charges. The analysis of seizures evolution is used for sis, and discriminant analysis. Profiles of neurometric
their classification, which offers the possibility of an features that deviate from age-matched normals have
application of appropriate treatment [e.g., (60)]. An exam- been obtained for patients suffering from cognitive dis-
ple of the evolution of seizure is shown in Fig. 9. For orders, psychiatric illnesses, and neurological dysfunc-
localization of the epileptic focus on the basis of EEG, the tions.
inverse problem has to be solved. Taking into account its The computerized EEG monitoring in the intensive
nonuniqueness and other limitations mentioned above, care neurological units involves measurement and assess-
when the removal of the part of brain is considered, the ment of several signals apart from EEG (e.g., ERP, ECG,
localization requires confirmation by means of imaging heart rate variability, respiration, intracranial pression,
techniques, such as CAT or MRI, and verification by and others depending on the injury). Physiological signal
analysis of EEG from subdural and implanted electrodes analysis can be combined with other diagnostic techniques
(60). A variety of methods for detection of epileptiform [e.g., ultrasound (69)]. Advances in computerized EEG
activity in the EEG have been implemented by means of monitoring in neurological intensive care unit are de-
expert systems [e.g., (61,62)], some of which are available scribed in (70).
commercially (63). A growing interest in the forecasting of Design of contemporary computer-assisted diagnosis
epileptic seizures also exists. Seizure prediction times (CAD) systems is usually based on computation of spectral
from minutes to hours have been reported [for a recent parameters followed by artificial intelligence methods. In
review, see (64)]. However, some of the claimed results (71), the frequency components of EEG were converted
have been severely challenged by appropriate statistical into pseudo-linguistic facts via fuzzification. The EEG
analysis (65,66). features were extracted by expert systems applying sym-
bolic rules formulated by neurologist. The results were
presented as linguistic terms, numerical values, and maps
of temporal extent. In (72), the procedures involved in the
diagnosis consisted of the following steps: (1) real-time
 ELECTROENCEPHALOGRAPHY (EEG) 13

processing and compression of EEG and VEP, (2) brain visual cortex. In: E. Basar and T. H. Bullock, eds., Induced
mapping of spectral powers, (3) classifier design, (4) auto- Rhythms in the Brain. Boston, MA: Birkhauser, 1992.
matic detection of morphologies through ANN, (5) signal 9. H. Jasper, Report of the Committee on Methods of Clinical
analysis through fuzzy modeling, and (6) a knowledge- Examination in Electroencephalography. Electroenceph. Clin.
based approach to classifier design. In the automatic Neurophysiol. 1958; 10:370–375.
system for classification of adult EEG (73), single epochs 10. R. T. Pivik, R. J. Broughton, R. Coppola, R. J. Davidson, N.
of EEG were classified by ANN. Time and space correla- Fox, and M. R. Nuwer, Committee Report: Guidelines for the
tions of outputs from ANN were evaluated by an expert recording and quantitative analysis of electroencephalo-
system, which generated a final report in the form of a graphic activity in research context. Psychophysiology 1993;
30:547–558.
medical diagnosis.
11. A. Rechtschaffen and A. Kales, A Manual: Standardized
Terminology, Technique and Scoring System for Sleep Stages
15. FUTURE OF ELECTROENCEPHALOGRAPHY of Human Subjects. Los Angeles, CA: Brain Information
Service/Brain Research Institute, University of California at
Los Angeles, 1968.
EEG, for many years, has been an important diagnostic
12. J. Zygierewicz, K. J. Blinowska, P. J. Durka, W. Szelenberger,
tool, and more recent investigations have proven its
S. Niemcewicz, and W. Androsiuk, High resolution study of
significance for the understanding of information proces- sleep spindles. Clin. Neurophys. 1999; 110:2136–2147.
sing by the functional brain. At present, the topographical
13. M. Kaminski, K. J. Blinowska, and W. Szelenberger, Topo-
techniques, such as positon emission tomography (PET) or graphic analysis of coherence and propagation of EEG activ-
functional nuclear magnetic resonance (fMRI), are widely ity during sleep and wakefulness. Electroenceph. Clin.
used for the diagnosis of pathologic neurological states Neurophys. 1997; 102:216–227.
and in brain research. However, these methods give 14. Ch. Guilleminault et al., ASDA report EEG arousals: scoring
information about the absorption of certain substances rules and examples. Sleep 1992; 15:173–184.
in specific structures or about the metabolism rate or 15. J. S. Dunkan, S. D. Shorvon, and D. R. Fish, Clinical Epilepsy.
glucose consumption, not directly about the brain electri- New York: Churchill Livingstone, 1995.
cal activity. Although their spatial localization properties 16. G. Pfurtscheller and F. H. Lopes da Silva, Event-Related
are good, their time resolution is much lower than EEG. Desynchronization (ERD) Handbook of Electroencephalogra-
Moreover, in the information processing by brain, EEG phy Clin. Neurophysiol., Revised Series, vol. 6. Amsterdam:
rhythms have a different specific role, which cannot be Elsevier, 1999.
distinguished by imaging techniques. The highest infor- 17. A. van Rotterdam, F. H. Lopes da Silva, J. van Ende, M. A.
mation content is usually connected with high-frequency Viergever, and A. J. Hermans, A model of the spatial-temporal
rhythms (especially gamma) generated by small neuron characteristics of the alpha rhythm. Bull. Math. Biol. 1982;
pools, which makes the information carried by them even 44:283–305.
less accessible to the imaging techniques. Therefore, these 18. P. Acherman, R. Hartman, A. Gunzinger, W. Guggenbuhl, and
techniques are not likely to replace EEG, which is a totally A. A. Borbely, All night sleep EEG and artificial stochastic
noninvasive and low-cost technique capable of providing signals have similar correlation dimension. Electroenceph.
Clin. Neurophysiol. 1994; 90:384–387.
information about relationships between cortical sites and
the time evolution of brain processes. 19. F. H. Lopes da Silva, J. P. Pijn, and D. Velis, Signal processing
of EEG: evidence for chaos or noise. Application to seizure
activity in Epilepsy. In: I. Gath and G. F. Inbar, eds., Advances
BIBLIOGRAPHY in Processing and Pattern Analysis of Biological Signals. New
York: Plenum Press, 1996.
1. E. Niedermayer and F. H. Lopes da Silva, Electroencephalo- 20. K. J. Blinowska and M. Malinowski, Non-linear and linear
graphy. Basic Principles, Clinical Applications, and Related forecasting of the EEG time series. Biol. Cybern. 1991;
Fields, 3rd ed. Baltimore, MD: Williams & Wilkins, 1993. 66:159–165.
2. F. H. Lopes da Silva, The generation of electric and magnetic 21. P. L. Nuez and K. L. Pilgreen, The Spline Laplacian in
signals of the brain by local networks. In: R. Greger and U. clinicalneurophysiology: a method to improve EEG spatial
Windhorst, eds., Comprehensive Human Physiology. Heidel- resolution. J. Clin Neurophys. 1991; 8:397–413.
berg, Germany: Springer-Verlag, 1996. 22. B. Hjorth, An on-line transformation of EEG scalp potentials
3. W. J. Freeman, The physiology of perception. Scientif. Amer. into orthogonal source derivations. Electroenceph. Clin. Neu-
1991; 26:78–85. rophysiol. 1975; 39:526–530.
4. M. A. B. Brazier, A History of the Electrical Activity of the 23. J. Le and A. S. Gevins, Method to reduce blur distorsion from
Brain. The First Half-Century. London: Pitman, 1961. EEGs using a realistic head model. IEEE Trans. Biomed. Eng.
1993; 6:517–528.
5. P. Gloor, Hans Berger on the Electroencephalogram of Man.
Amsterdam: Elsevier, 1969. 24. A. S. Gevins, N. H. Morgan, S. L. Bressler, B. A. Cutillo, R. M.
White, D. Greer, and J. Illes, Event-related covariances dur-
6. P. L. Nunez, Electric Fields of the Brain. New York: Oxford
ing a bimanual visuomotor task. Part I. Methods and analysis
University Press, 1981.
of stimulus and response-locked data. Electroenceph. Clin.
7. H. Jahnsen and R. Linas, Electrophysiological properties of Neurophysiol. 1989; 74:58–75.
guinea-pig thalamic neurons: an in vitro study. J. Physiol.
25. N. J. Mars and F. H. Lopes da Silva, EEG analysis methods
1984; 349:205–226.
based on information theory. In: A. S. Gevins and A. Rémond,
8. C. M. Gray, A. K. Engel, P. Konig, and W. Singer, Mechanisms
underlying the generation of neuronal oscillations in cat
14 ELECTROENCEPHALOGRAPHY (EEG)

eds., Methods of Analysis of Brain Electrical and Magnetic 45. S. Mallat and Z. Zhang, Matching pursuit with time-fre-
Signals, vol. 1. Amsterdam: Elsevier, 1987. quency dictionaries. IEEE Trans. Signal Proc. 1993;
26. B. Schack, A. C. N. Chen, S. Mescha, and H. Witte, Instanta- 41:3397–3415.
neous EEG coherence analysis during the Stroop task. Clin. 46. P. J. Durka and K. J. Blinowska, A unified time-frequency
Nuerophysiol. 1999; 110:1410–1426. parametrization of EEGs. IEEE Eng. Med. Biol. Mag. 2001;
27. P. J. Franaszczuk, K. J. Blinowska, and M. Kowalczyk, The 20:47–53.
application of parametric multichannel spectral estimates in 47. D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning
the study of electrical brain activity. Biol. Cybern. 1985; representations by black-propagating errors. Nature 1986;
51:239–247. 6188:533–536.
28. T.-P. Jung, S. Makeig, M. J. McKeown, A. J. Bell, T.-W. Lee, 48. N. Schaltenbrand, R. Langelle, M. Toussaint, R. Luthringer,
and T. J. Sejnowski, Imaging brain dynamics using indepen- G. Carelli, A. Jacqmin, E. Lainey, A. Muzet, and J. P. Macher,
dent component analysis. Proc. IEEE 2001; 89:1107–1122. Sleep stage scoring using the neural network model: compar-
29. M. Kaminski, M. Ding, W. Truccolo, and S. Bressler, Evaluat- ison between visual and automatic analysis in normal sub-
ing causal relations in neural systems: Granger causality, jects and patients. Sleep 1996; 19:26–35.
directed transfer function and statistical assessment of sig- 49. L. J. Trejo and M. J. Shensa, Feature extraction of event-
nificance. Biol. Cybern. 2001; 85:145–157. related potentials using wavelets: an application to human
30. M. Kaminski and K. J. Blinowska, A new method of the performance monitoring. Brain Lang. 1999; 66:89–107.
description of the information flow in the brain structures. 50. G. E. Hinton, How neural networks learn from experience.
Biol. Cybern. 1991; 65:203–210. Scientif. Amer. 1992; Sept.:104–109.
31. P. J. Franaszczuk and G. K. Bergey, Application of the 51. T. Kohenen, The self-organization map. Proc. IEEE, 1990;
directed transfer function method to mesial and lateral onset 78:1464–1480.
temporal lobe seizures. Brain Topogr. 1998; 11:13–21. 52. G. Pfurtscheller, J. Kalcher, Ch. Neuper, D. Flotzinger, and M.
32. A. Korzeniewska, S. Kasicki, M. Kaminski, and K. J. Bli- Pregenzer, On-line EEG classification during externally-
nowska, Information flow between hippocampus and related paced hand movements using a neural network-based classi-
structures during various types of rat’s behavior. J. Neurosci. fier. Electroenceph. Clin. Neurophysiol. 1996; 99:416–425.
Meth. 1997; 73:49–60. 53. C. Baumgartner, W. W. Southerling, S. Di, and D. S. Barth,
33. R. Kus, M. Kaminski, and K. J. Blinowska, Determination of Investigation of multiple instantaneously active brain
EEG activity propagation: pair-wise versus multichannel sources in electroencephalogram. J. Neurosc. Meth. 1989;
estimate. IEEE Trans. BME 2004; 51:1501–1510. 30:175–184.
34. M. Ding, S. L. Bressler, W. Yang, and H. Liang, Short-window 54. R. D. Pascual Marqui, C. M. Michel, and D. Lehmann, Low
spectral analysis of cortical event-related potentials by adap- resolution electromagnetic tomography: a new method to
tive multivariate autoregressive modeling: data processing, localize electrical activity in the brain. Int. J. Psychophysiol.
model validation and variability assessment. Biol. Cybern. 1994; 18:49–65.
2000; 83:35–45. 55. R. Grave de Peralta Menendez and S. L. Gonzales Andino, A
35. J. Ginter, K. J. Blinowska, M. Kaminski, and P. J. Durka, critical analysis of linear inverse solutions. IEEE Trans.
Phase and amplitude analysis in time-frequency space— Biomed. Eng. 1998; 45:440–448.
application to voluntary finger movement. J. Neurosc. Meth. 56. M. C. Salinsky, B. S. Oken, D. Storzbach, and C. B. Dodrill,
2001; 110:113–124. Assessment of CNS effects of antiepileptic drugs by using
36. S. Mallat, A Wavelet Tour of Signal Processing. San Diego, CA: quantitative EEG measures. Epilepsia 2003; 8:1042–1050.
Academic Press, 1999. 57. M. J. Barbanoj, M. Valle, J. Kulisevsky, V. Perez, and P.
37. J. V. Samar, A. Bopardikar, R. Rao, and K. Swartz, Wavelet Gambus, Uses of pharmaco-EEG and pharmacokinetic-phar-
analysis of nueroelectric waveforms: a conceptual tutorial. macodynamic modeling in the clinical scenario. Meth. Find
Brain Lang. 1999; 66:7–60. Exp. Clin. Pharmacol. 2002; 24:139–144.
38. V. T. Makinen, P. J. May, and H. Tiitinen, Human auditory 58. S. Kubicki and W. M. Herrmann, The future of computer-
event-related processes in the time-frequency plan. Neurore- assisted investigation of the polysomnogram: sleep micro-
port 2004; 15:1767–1771. structure. J. Clin. Neurophysiol. 1996; 13:285–294.
39. E. A. Bartnik, K. J. Blinowska, and P. J. Durka, Single evoked 59. P. Y. Ktonas, Computer-based recognition of EEG patterns.
potential reconstruction by means of wavelet transform. In: R. M. Dasheiff and D. J. Vincent, eds., Frontier Science in
Biologic. Cybernet. 1992; 67:175–181. EEG: Continuous Waveform Analysis (EEG Suppl. 45). New
40. K. Wang, H. Begleiter, and B. Porjesz, Spatial enhancement of York: Elsevier, 1996.
event-related potentials using multiresolution analysis. 60. P. J. Franaszczuk, G. K. Bergey, P. J. Durka, and H. M.
Brain Topogr. 1998; 10:191–200. Eisenberg, Time-frequency analysis using the matching pur-
41. N. V. Thakor and S. Tong, Advances in quantitative electro- suit algorithm applied to seizures originating from the mesial
encephalogram analysis methods. Annu. Rev. Biomed. Eng. temporal lobe. Electroencephalogr. Clin. Neurophysiol. 1998;
2004; 6:453–495. 106:513–521.
42. H. Al-Nashash, J. Paul, W. Ziai, D. Hanley, and N. Thakor, 61. A. J. Gabor, R. R. Leach, and F. U. Dowla, Automated seizure
Wavelet entropy for suband segmentation of EEG during detection using a self-organizing neural network. Electroen-
injury and recovery. Ann. Biomed. Eng. 2003; 31:653–658. cephalogr. Clin. Neurophysiol. 1996; 99:257–266.
43. F. De Carli, L. Nobili, P. Gelcich, and F. Ferrillo, A method for 62. B. L. Davey, W. R. Fright, G. J. Carroll, and R. D. Jones,
the automatic detection of arousals during sleep. Sleep 1999; Expert system approach to detection of epileptiform activity
22:561–572. in the EEG. Med. Biol. Comput. 1989; 27:365–370.
44. M. Akay, Time Frequency and Wavelets in Biomedical Signal 63. J. Gotman, Automatic detection of seizures and spikes. J.
Processing. New York: IEEE Press, 1998. Clin. Neurophysiol. 1999; 16:130–140.
 ELECTROENCEPHALOGRAPHY (EEG) 15

64. B. Litt and K. Lehnertz, Seizure prediction and the presei- cerebrovascular screening system (CERBERUS). Med. Prog.
zure period. Curr. Opin. Neuro. 2002; 15:173–177. Technol. 1995; 21:53–66.
65. R. Aschenbrenner-Scheibe, T. Maiwald, M. Winterhalder, H. 70. B. Rosenblatt and J. Gotman, Computerized EEG monitoring.
U. Voss, J. Timmer, and A. Schulze-Bonhage, How well can Semin. Pediatr. Neurol. 1999; 6:120–127.
epileptic seizures be predicted? An evaluation of a nonlinear 71. C. S. Herrmann, T. Arnold, A. Visbeck, H. P. Hundemer, and
method. Brain 2003; 126:2616–2626. H. C. Hopf, Adaptive frequency decomposition of EEG with
66. Y. C. Lai, M. A. Harrison, M. G. Frei, and I. Osorio, Inability subsequent expert system analysis. Comput. Biol. Med. 2001;
of Lyapunov exponents to predict epileptic seizures. Phys. 31:407–427.
Rev. Lett. 2003; 8:91–96. 72. L. Moreno, J. L. Sanchez, S. Manas, J. D. Pineiro, J. J. Merino,
67. M. Matousek. P. Petersen, and S. Freeberg, Automatic assess- J. Sigut, R. M. Aguilar, J. I. Estevez, and R. Marichial, Tools
ment of randomly selected routine EEG records. In: G. Dolce for acquisition, processing and knowledge-based diagnostic of
and H. Kunkel, eds., CEAN-Computerized EEG Analysis. the electroencephalogram and visual evoked potentials. J.
Stuttgart, Germany: Fischer, 1975. Med. Syst. 2001; 25:177–194.
68. E. R. John, L. Prichep, J. Friedman, and P. Easton, Neuro- 73. C. Castellaro, G. Favaro, A. Castellaro, A. Casagrande, S.
metrics: computer assisted diagnosis of brain dysfunctions. Castellaro, D. V. Puthenparampil, and C. F. Salimbeni, An
Science 1988; 29:162–169. artificial intelligence approach to classify and analyse EEG
69. M. Bodo, G. Thuroczy, I. Nagy, J. Peredi, K. Sipos, P. Harcos, traces. Neurophysiol. Clin. 2002; 32:193–214.
Z. Nagy, J. Voros, L. Zoltay, and L. Ozsvald, A complex