Effects of total sleep-deprivation on waking human

EEG: functional cluster analysis

Seung-Hyun Jina, Sun Hee Nab, Soo Yong Kimb, Dai-Jin Kimc,*

aDepartment of Science Education, Korea National University of Education, Chungbuk 363-791, South KoreabDepartment of Physics, Korea Advanced Institute of Science and Technology, Daejeon 305-701, South Korea

cDepartment of Psychiatry, College of Medicine, Catholic University of Korea, Holy Family Hospital, 420-717 Buchon City, Kyunggi Do, South Korea

Accepted 2 July 2004

Available online 6 August 2004

Abstract

Objective: The purpose of the present study was to investigate the effects of total sleep deprivation (TSD) on brain functions with an

identification procedure for strongly interactive brain regions, relying on functional cluster (FC) analysis in multichannel

electroencephalogram (EEG) data.

Methods: EEGs from 16 electrodes in 18 healthy, right-handed, young male volunteers were recorded before TSD (after normal sleep) and

after 24 h of experimentally induced sleep deprivation. We estimated cluster index to characterize joint interactions among many brain

regions in order to determine if a particular FC is present or not, and if so, its anatomy.

Results: As a result, we confirmed the presence of FC and found different FC patterns in both before and after TSD. The C3 and F7

locations were outside the cluster under the TSD condition, but belonged to the cluster with C4 and F8 before the TSD condition, and the

F3/F4, and O1 locations were new entries to the functional cluster during sustained wakefulness.

Conclusions: These results indicate that the neuronal activities of the C3 and F7 location are functionally unrelated, whereas the F3/F4

locations are functionally involved with the C4, F8, and O1 locations after 24 h TSD. Our results suggest that FC changes with elapsed awake

time and reflects the change of brain function due to TSD.

Significance: This paper shows the existence of FC both before and after TSD, and the anatomy of each FC is different. So FC analysis

would be a potential tool to investigate the simultaneous neuronal activity of human EEGs.

q 2004 Published by Elsevier Ireland Ltd. on behalf of International Federation of Clinical Neurophysiology.

Keywords: Multichannel EEG; Total sleep deprivation; Functional clustering

1. Introduction

TSD has an influence on cognitive functions during

wakefulness, and it is found that a monotonic decline occurs

in memory, complex verbal arithmetic function, reaction

times, and logical reasoning tasks, as sleep deprivation (SD)

increases (Drummond et al., 1999; Kim et al., 2001a;

Lorenzo et al., 1995; Mikulincer et al., 1990). The effects of

accumulating hours of wakefulness are reflected, not only in

the sleeping EEG, but also the waking EEG (Corsi-Cabrera

et al., 1992; Lorenzo et al., 1995). Corsi-Cabrera et al.

1388-2457/$30.00 q 2004 Published by Elsevier Ireland Ltd. on behalf of Intern

doi:10.1016/j.clinph.2004.07.001

* Corresponding author. Tel.: C82-32-340-2140; fax: C82-32-340-

2670.

E-mail address: kdj922@chollian.net (D.-J. Kim).

(1996) reported significant absolute power changes of the

theta and beta bands, with closed eyes after TSD. Lorenzo

et al. (1995) have demonstrated that SD may lead to a linear

increase in the absolute power, which was more prominent

on the theta band with open eyes. SD also yielded an

increase in the EEG spectral amplitude in the 5.00–7.55 and

11.00–13.75 Hz ranges (Dumont et al., 1997), and an

increased power density in the 6.25–9.00 Hz range

(Cajochen et al., 1995).

Quantitative EEG analyses during extended wakefulness

have revealed frequency-specific circadian and homeostatic

influences (Aeschbach et al., 1997; Cajochen et al., 1995,

1999a; Finelli et al., 2000). Both the circadian and

homeostatic processes are oscillatory processes that gen-

erate the human sleep–waking cycle (Dann et al., 1984;

Clinical Neurophysiology 115 (2004) 2825–2833

www.elsevier.com/locate/clinph

ational Federation of Clinical Neurophysiology.

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–28332826

Dijk and Czeisler, 1995). Several researchers have demon-

strated that some aspects of human sleep are primarily

dependent on the circadian process, whereas others are more

susceptible to the influence of the sleep–wake cycle. (Dijk

and Czeisler, 1994, 1995; Shanahan et al., 1997). The

intrusion of low-frequency EEG oscillations during wake-

fulness correlates well with the performance decrements

during wakefulness in the adverse circadian phases

(Cajochen et al., 1999b) as well as to performance lapses

during sustained attention throughout the normal waking

day (Makeig et al., 2000). Low frequency EEG components,

in particular slow-wave activity (SWA), are largely

controlled by mechanisms involved in sleep homeostasis

(Cajochen et al., 2001; Finelli et al., 2000).

More information could be provided from measures that

determine the functional connectivity between different

areas, other than just the level of activity in different areas

(Jausovec, 2000). Several linear methods, for example,

correlation and coherence, have been used to study the

functional connectivity between, and within, brain hemi-

spheres using EEG data (Leocani et al., 1997; Shen et al.,

1999; Weiss and Rappelsberger, 2000). Included among

some new approaches are synchronization likelihood

method, which are sensitive to linear as well as non-linear

interdependencies (Altenburg et al., 2003; Stam and van

Dijck, 2002). The synchronization likelihood can quantify

the interdependency between one EEG channel and all the

other recorded channels (Stam and van Dijck, 2002). This

measure used to investigate the statistical interdependencies

between two time series whether epileptic seizure activity

can be distinguished from non-epileptic activity in the

neonate EEG (Altenburg et al., 2003). In addition, non-

linear asymmetric interdependence measures in multivariate

EEGs was used to assess the interdependencies between two

EEG channels from 10 adult human during different

vigilance states (Pereda et al., 2003). However, these

previous results were obtained without access to investi-

gations of the joint functional interactions between many

brain regions.

Several techniques, including principal component

analysis (PCA) (Corsi-Cabrera et al., 2001), independent

component analysis (ICA) (McKeown et al., 1997) and

multidimensional scaling (MDS) closely related to PCA

(Friston et al., 1996), can be used to find evidence of

functional clustering. PCA is a well-known statistical

technique, which estimates the amplitude of a constant

resonance peak across a series of spectra (Stoyanova and

Brown, 2002), thus the principal components extracted by

PCA can be used to remove coherent activities spread over

the data (Casarotto et al., 2004). PCA was used to

investigate which frequencies are covariant during wakeful-

ness, slow-wave, and paradoxical sleep in the rat (Corsi-

Cabrera et al., 2001), and to reduce the ocular artifacts from

EEGs of normal and dyslexic children (Casarotto et al.,

2004). ICA is an effective algorithm for separating or

estimating waveforms of mutually independent components

from an array of sensors without knowledge of any

characteristics of the transmission channels, and has been

applied to multichannel EEGs (Jin et al., 2002; McKeown

et al., 1997). While PCA, and related approaches, may be

useful in identifying statistical relationships within the data

set (Tononi et al., 1998), the principal or independent

components extracted by PCA or ICA are insufficient to

show the cluster in terms of internal cohesion and external

isolation because these methods have no consideration for

intrinsic statistical dependency within a subset of elements

and extrinsic statistical dependency between that subset and

its complement.

The motivation of using functional cluster (FC) analysis

in this study is to attempt to identify strongly interactive

subsets of elements when analyzing a neural system by

taking intrinsic and extrinsic interactions into consideration

in order to overcome shortcoming raised in other clustering

method like PCA or ICA. Tononi et al. (1998) proposed a

new functional clustering method based on multivariate

measures of statistical dependence, which may involve

more than two regions. Functional clustering results from

the joint interactions between many brain regions, and

functional cluster can be defined as a set of elements that

interact much more strongly between themselves than with

the rest of the system in a specific brain state (Tononi et al.,

1998). The cluster index was defined by the ratio between

the integration and mutual information as a measure of

intrinsic and extrinsic interactions, respectively, for all the

possible subsets of the system, and can serve as a measure

of functional clustering. Integration measures the total

statistical dependence between a subset of elements, with

the mutual information providing the statistical dependence

between that subset of elements and the rest of the system.

The application of this approach was tested on positron

emission tomography (PET) data obtained from schizo-

phrenics and normal controls performing a set of cognitive

tasks, and these results showed distinct differences of

clustering between the two groups (Tononi and Edelman,

2000; Tononi et al., 1998). As previously stated, analyses

that employ one or two EEG signals have been mainly used

to investigate the dynamics or connectivity of EEGs,

whereas FC analysis suggests the possibility to find

simultaneous and strong interactions distributed among

many brain areas.

In the present study, we investigate the FC of 24 h

sleep-deprived human multichannel EEGs. Our aim is to

determine whether there is a FC before and after TSD,

and if so, the particular anatomy of each FC.

2. Methods

2.1. Functional clustering

FC analysis provides a useful tool to characterize the

joint interactions of different elementary subsystems, xi with

Fig. 1. Schematic map of clustering. The diagram shows an example of a

subset composed of 4 components, the CI value was calculated using the

integration between F3/4, C3/4 (cross arrows) and mutual information

between these 4 channels and the other 12 channels (curved arrow).

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–2833 2827

iZ1,.,N, of the whole brain system X, consisting of a

collection of N (the number of EEG electrodes) elementary

subsystems. We computed the cluster index CIðXkj Þ (Tononi

et al., 1998) to identify FC by measuring for many different

subsets as follows:

CIðXkj Þ Z

IðXkj Þ

MIðXkj ;X KXk

j Þ

where Xkj indicates a jth subset composed of k components of

system X, and XKXkj is its complement. MIðXk

j ;XKXkj Þ

indicates mutual information that represents interactions

between Xkj and XKXk

j ; and IðXkj Þ denotes the integration of

Xkj ; which represents a multivariate measure of statistical

independence among the elements Xkj : IðXk

j Þ is a measure for

deviation from statistical independency of the individual

components, that is, integration has a maximum value when

the elements of a subset are completely dependent; and, if the

elements are statistically independent, the integration is zero.

The mutual information MIðXkj ;XKXk

j Þ is calculated as

MIðXkj ;XKXk

j ÞZHðXkj ÞCHðXKXk

j ÞKHðXÞ; where HðXkj Þ;

HðXKXkj Þ and H(X) are entropies of Xk

j and XKXkj

considered independently, and that of the system considered

as a whole, respectively (Papoulis, 1991). IðXkj Þ is calculated

as the difference between the sum of the entropies of all the

individual components xj,i (1%i%k) of Xkj considered

independently, and the entropy of Xkj considered as a

whole (Tononi et al., 1994).

IðXkj Þ Z

Xk

iZ1

Hðxj;iÞKHðXkj Þ

where Xkj is a jth subset composed of k elementary

components xj,i of the system X. The indices j,i (1%i%k)

denote the indices of the elementary components, that is the

index j refer to which subset of k components is considered,

i labels the particular component of that subset:

Xkj Z fxj;i; 1% i%kg

The CI values were calculated for all 16 channels of the

EEG recordings obtained from all subjects. The calculation

of the CI values for each subject was performed for each of

the collection of subsets with size k; that is, the calculation is

composed of k-out-of-16 components over the nZ16!/k!(16Kk)! combinations of the k components for

definition of the CI. For example, if we consider the subsets

composed of 4 components of the whole system, then,

16!/4!(16K4)!(Z1820) combinations exist. If F3, F4, C3

and C4 channels are chosen among 1820 combinations as a

jth subset composed of 4 components ðX4j Þ; as depicted in

Fig. 1, the CI value is calculated according to the above

definition using the integration between F3, F4, C3 and C4

channels (cross arrows) and mutual information between

these 4 channels and the remaining 12 channels (curved

arrow).

Assuming that the multidimensional stationary stochas-

tic process describing the activity of N subsystems, x, is

Gaussian, the entropy is H(Xk)Z0.5 ln(2pe)kjCOV(Xk)j,

with j$j indicating the matrix determinant, and COV(Xk)

the covariance matrix of Xk(Tononi et al., 1998). We

should note that entropy from which integration is

computed might treat the signal both with its linear and

nonlinear properties. In the present study, the integration

and mutual information were calculated using the entropy

according to the aforementioned definition for the mean

removed (zero-mean) EEG data set, and so we consider the

wake EEG signal as a multivariate Gaussian signal, in

agreement with computation of integration in previous

researches (Putten and Stam, 2001; Tononi et al., 1998).

This can be considered as a first approximation, as eyes

closed wake EEG is known to be of deterministic chaotic

nature (Jeong et al., 2001).

In order to confirm the validity of this method, we

tested the integration and mutual information for real

EEGs, simulated non-clustered homogeneous data, and

decomposed signals of EEGs using ICA. We used the

‘infomax’ algorithm proposed by Bell and Sejnowski

(1995), in order to obtain decomposed independent

components from the EEG. Briefly, ICA is a signal

processing method to extract independent sources given

only observed data that are mixtures of unknown sources.

The function of the ICA algorithm is to find a matrix W

that makes the elements u(t)Z[u1(t),.,uN(t)]T of the

linear transform u(t)ZWx(t) of a data vector x(t)Z[x1(t),.,xN(t)]T statistically independent. We separated

EEG data into 16 independent components by using ICA.

We used ICA Matlab codes provided by the Sejnowski

group (http://www.cnl.salk.edu).

Fig. 2 shows a typical example of (a) EEG segments of

4 s from 16 channels; (b) an equivalent non-clustered, or

Fig. 2. (a) EEG time series at 16 channels, (b) the equivalent homogeneous system (total integration IZ5.927) and (c) the resulting ICA-transformed EEG time

series from a subject in the rest condition.

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–28332828

homogeneous system with the same total integration

(here, total integration IZ5.927); and (c) the resulting

ICA-transformed EEG time series from a subject.

2.2. Statistics and cluster map

We assessed the statistical significance of CI values by

computing a Student’s t like parameter, tCI, given by Tononi

et al. (1998).

tCI ZCIðXk

j ÞK hCIðXkHÞi

stdðCIðXkHÞÞ

where CIðXkj Þ is the cluster index of Xk

j ; hCIðXkHÞi is the mean

of the CI distribution for subset size k accumulated from

many sampled equivalent homogeneous systems (100 in this

work), and stdðCIðXHk ÞÞ is the standard deviation of the

distribution. XH (the null hypothesis) denotes a homo-

geneous system with the same overall integration, but

contains no FCs (Tononi et al., 1998). XH is generated via

the linear model xi ZbyiCaP

jsi yj; where yi is Gaussian

signals of zero mean and unit variance with the same size of

input signals, aZ{(NK2)rC2K2[(1KN)r2C(NK2)rC1]1/2}1/2/N and bZ[1K(NK1)a2]1/2. The parameters a and

b are the solutions of the equations

hxi$xjihrij;rij Z r; isj

rij Z 1; i Z j;

(

given the fact that

hyi$yjihdij;dij Z 0; isj

dij Z 1; i Z j:

(

In the present study, the confidence level is P!0.01,

because 100 null systems were used in order to gauge the

likelihood of obtaining a given CI of no functional

clustering is present. Statistical process is designed to test

the null hypothesis that the signal consists of non-clustered

homogeneous system, and the tCI value is higher than zero

when the null hypothesis is rejected.

After all calculations including tCI, the statistically

significant (here, P!0.01) subset, with the largest CI

value among all possible subsets, having results higher than

1, was then chosen. A cluster index much greater than 1

indicates a subset of elements that are strongly interactive

among themselves. A weighted cluster distribution was

constructed using the elements of the statistically significant

subsets for all the subjects. For example, if the T5 channel

was an element of the statistically significant subset for one

subject, then T5 location was weighted as 1 unit. This

process was repeated for all the subjects. A weighted (by the

number of involved locations) cluster distribution could

then be obtained for a given task and group of subjects. A

topographic cluster map was drawn at the scalp location,

using a weighted cluster distribution. Because the extracted

elements from the respective subjects were statistically

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–2833 2829

significant, the weighted distribution constructed by those

elements could show the group properties for a given task.

2.3. Subjects and experimental procedure

EEG recordings were obtained from 18 healthy, right-

handed, young adult males (age 23.4G1.4 years), all of

whom volunteered to participate in the study and signed

informed consent forms. All subjects were free of

neurological or sleep-related disorders, took no medi-

cations, and had normal sleeping habits. Sleep disorders

and habits were assessed by a week of sleep logs and a

questionnaire. In addition, subjects were instructed to

maintain a regular bedtime and rising time, and asked

to avoid alcohol and caffeine intake during the week prior to

the experiment.

All subjects were asked to attend the sleep laboratory

twice, at 1 week intervals, to have their EEGs monitored and

to acclimate to the laboratory and EEG electrodes before

beginning the experimental procedure. The subjects

attended the laboratory in the late afternoon. They slept in

the laboratory from 22:00 h, and their EEGs were recorded

just after awakening, at 07:00 h, after a full night sleep

(22:00–07:00 h), with their eyes closed. These EEGs

correspond to the baseline wake EEG segment, with their

sleep quality maintained as normal, as all subjects had very

similar sleep habits according to our selection, and

acclimated well to the laboratory. The subjects were then

kept awake in the laboratory for 24 h. They were

continuously challenged with tasks and games, and kept

under human supervision in the laboratory, while being

simultaneously monitored to assure the maintenance of

wakefulness. Their EEG recordings were taken at 07:00 h

the following day, with their eyes closed. We recorded the

EEGs after 24 h TSD at the same time as those collected at

baseline considering the circadian time. During the data

collection, an experimenter carefully checked the EEG

recordings to ensure the maintenance of wakefulness. If

there was any indication of a drop in alertness, the subject

was asked to open his eyes and a short break was taken to

allow return to alertness before resuming data collection.

Fig. 3. Mean values of integration (solid line) and mutual information

(dashed line) as a function of a subset size for one EEG; the average

integration (solid line and square) and mutual information (dashed line and

triangle) of the equivalent homogeneous system; the average integration

(solid line and blank square) and mutual information (dashed line and blank

triangle) of independent component decomposed by infomax algorithm.

Integration and mutual information are nearly zero for the independent

EEG sources, however, those of EEG data have appropriate values.

2.4. EEG recordings

The EEGs were recorded from the 16 scalp loci (F7/8,

T3/4, Fp1/2, F3/4, C3/4, P3/4, O1/2 and T5/6) of the

International 10–20 System. The EEGs from 16 channels

against ‘linked earlobes’ were amplified on a Nihon Kohden

EEG-4421 K, with a time constant of 0.1 s. The sampling

frequency was 250 Hz. At each monitoring stage 32.768 s of

data were recorded and digitized by a 12 bit analog-to-

digital converter in an IBM PC. All data were digitally

filtered at 0.5–60 Hz, and each EEG record was judged, by

inspection, to be free from electro-oculographic and move-

ment artifacts and to contain minimal electro-myographic

activity. For our analysis under each condition 4000 data

points (16 s) were used.

3. Results

Fig. 3 demonstrates the average integration and mutual

information as a function of subset size of an EEG data set, a

homogeneous system, and the resulting ICA-transformed

EEG time series. To obtain at each subset size k, a collection

of subsets with size k that is composed of k-out-of-16

components, over 16!/k!(16-k)! combinations of k com-

ponents were calculated and averaged. The average

integration and mutual information are nearly ‘zero’ when

using the independent EEGs, since these data which

decomposed nearly completely independent subsystems

have neither integration nor mutual information. In contrast,

original EEG data have different values much more than

zero for each subset size. Fig. 4, which shows the CI values

of EEG data and its homogeneous system ranked by value of

tCI, supports the validity of the analysis process. CI

values for homogeneous data were almost 1, because

these data have uniform integration and mutual information

among the subsets.

Fig. 5 depicts the 15 most significant subsets, ranked by

tCI value, as indicated by filled circles, with triangles

indicating the CI values for an EEG data set from a subject

before TSD. These CI values were higher than those for

equivalent homogeneous system subsets of the same size in

Fig. 6. Cluster map before the TSD (a), and after 24 h of TSD (b). Before

TSD, a dominant cluster appeared at F7, F8, C3, and C4 and after 24 h TSD,

the cluster was constituted around the C4, F8, F3, F4, and O1 locations. The

C3 and F7 locations were outside the cluster under the TSD condition,

while they belonged to the cluster before the TSD condition, and the F3, F4

and O1 locations were new entries of the FC during sustained wakefulness.

Fig. 4. The CI values of EEG data set (pentagon) and non-clustered

homogeneous system (star). CI values for homogeneous data were almost 1

because these data have uniform integration and mutual information among

the subsets.

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–28332830

Monte Carlo simulations with P!0.01. The elements of the

statistically significant (here, P!0.01) subsets with the

largest CI values were composed of F7, T3, C3, F8, T4 and

C4 in this case.

Fig. 6 (a) and (b) show topographic cluster maps before

TSD and after 24 h of TSD, respectively. A topographic

cluster map was drawn at the scalp location using a

weighted cluster distribution. Following convention, the

nose is at the top of all the figures. The same gray color

Fig. 5. The 15 most significant subsets ranked by the tCI value, as indicated

by filled circles, with triangles indicating the CI values. The statistically

significant (here, P!0.01) subset with the largest CI value was chosen from

all possible subsets with results higher than 1 because a CI much greater

than 1 indicates a subset of elements that are strongly interactive between

themselves, but weakly interactive with the rest of the system. The elements

of the statistically significant (here, P!0.01) subsets with the largest CI

value were composed of F7, T3, C3, F8, T4 and C4 in this case.

scale, from light gray (minimum) to dark gray (maximum),

was used on the appropriate scalp locations to indicate their

weighted cluster values for each figure. The values between

electrodes were linearly interpolated. The locations belonging

to the same cluster were all functionally involved, while the

locations outside the clusters were presumed to be

functionally unrelated (Tononi et al., 1998). Therefore, it

is important to see the main locations that contribute to the

cluster.

Before TSD, a dominant cluster appeared at F7, F8, C3,

and C4, as shown Fig. 6(a). After 24 h TSD, the cluster was

constituted around the C4, F8, F3, F4, and O1 locations, as

illustrated in Fig. 6(b). Notably, the F3/F4, and O1 locations

are new entries to the FC after 24 h TSD, while the C3 and

F7 locations are outside of the cluster under the TSD

condition, but belonged to the cluster before the TSD

condition. These results indicate that the neuronal activities

of the C3 and F7 locations are functionally unrelated,

whereas the F3/F4 and O1 locations are functionally

involved with the C4 and F8 locations after 24 h TSD.

4. Discussion

Results from the present study using a functional

clustering analysis for TSD human EEGs yielded two

major findings. First, the C3 and F7 locations were outside

the cluster under the TSD condition, while belonging to the

cluster before the TSD condition, and the F3/F4 locations

were new entries in the FC during sustained wakefulness.

Second, the occipital area was related to sleep pressure, i.e.

O1 location was a new entry of the cluster under 24 h TSD

condition. That is, before TSD, F7, F8, C3, and C4 locations

formed FC, while C4, F8, F3, F4, and O1 locations were

functionally involved after 24 h TSD.

Many kinds of analyses have been applied to characterize

the relationship between sleep and neural interactions. In

particular, coherence analyses provide a measure of the

functional and structural connectivity between two brain

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–2833 2831

sites on the basis of the cortico-cortical association model

proposed by Thatcher et al. (1987). According to Wright

et al. (1995), coherence between the frontal and occipital

brain sites were decreased in alpha and beta bands,

and the dominant synchronous component of slow wave

activities increased as a function of EEG stages in the

anterior-central areas during the transition from wakeful-

ness to sleep (Tanaka et al., 2000). Using non-linear

interdependency measure, Pereda et al. (2003) showed that

interdependencies among the different brain regions during

quiet sleep were mostly non-linear, asymmetric and greater

than those found during wakeful and active sleep of

neonates. Although these previous researches were infor-

mative, these methods are limited in terms of elucidating the

interactions between different regions at a given time.

Extraction of the functionally interactive subsets of brain

regions is the primary goal of the present study.

Many researchers pointed out the function of the

central cortex in SD studies. Lorenzo et al. (1995) have

demonstrated that SD may lead to deterioration in

performance of vigilance tasks and a linear increase of

power that was more prominent in central deviations than

temporal deviations. Similarly, it was more pronounced on

the left side than the right side with open eyes.

Corsi-Cabrera et al. (1996) reported significant absolute

power changes in the left central cortex during 40 h of

TSD. As a nonlinear study, Jeong et al. (2001) showed

that 24 h sleep-deprived states had lower average

correlation dimension values at the left central (C3)

channel compared to the baseline values. These previous

results showed the linear and nonlinear neural activities of

the left central area, whereas our result have revealed that

a cortical interaction between C3 and C4, and F3 and F4

locations does not occur during sustained wakefulness,

while these regions were involved before TSD. Also, the

result that C3 location was outside the cluster under the

TSD means that the neural activity of this location has no

relation with the cluster under the TSD.

In addition, the neuronal activity of the F7 location was

functionally unrelated, whereas, the F3/F4 locations were

functionally involved. Quantitative EEG analyses during

extended wakefulness have revealed frequency-specific

circadian and homeostatic influences in the frontal regions

(Cajochen et al., 1999b, 2001; Finelli et al., 2000). In

particular, the frontal low-frequency EEG activity (1–7 Hz)

exhibited a prominent increase with waking time, with

little circadian modulation (Cajochen et al., 1999a).

Changes in frontal low-frequency activity during wakeful-

ness are to a large extent determined by the sleep–wake

dependent process; this result reflects the differential levels

of sleep pressure in waking subjects (Cajochen et al.,

2001). Finelli et al. (2000) reported increased theta activity

(5–8 Hz) in the high-sleep pressure protocol in the frontal

areas. Thus, the different frontal activities obtained in our

results reflect the sleep homeostatic process of brain

function due to TSD. Achermann et al. (2001) observed

hemispheric asymmetries in the delta range after 40 h of

SD. They suggested that the anterior predominance of the

delta power, and its selective increase after TSD, was in

accordance with the interpretation that these regional

changes may reflect a higher ‘recovery need’ of the frontal

heteromodal association areas of the cortex. Finelli et al.

(2000) suggested that theta activity (5–8 Hz) in waking,

and slow-wave activity (0.75–4.5 Hz) during sleep, are

markers of a common homeostatic sleep process. Both

effects were largest in the frontal area, as revealed by a

topographic analysis based on 27 deviations. Although

these previous studies have shown that changes in the

frontal low-frequency activities may reflect the homeo-

static sleep process, the results did not address the joint

interactions between the frontal and other regions of the

brain. In this study, the neuronal activity of the F7 location

was shown to be functionally unrelated, whereas the F3/F4

locations were functionally involved with the C4, F8, and

O1 locations after 24 h of TSD.

Occipital regions have been considered in terms of their

relation to sleep events. According to Cantero et al. (2000),

the occipital structures may be the driving force for REM-

alpha bursts generation. Jeong et al. (2001) showed that 24 h

sleep-deprived states had lower average correlation dimen-

sion values at the left occipital channel compared to the

baseline values. Kim et al. (2001b) also reported firing

changes in the underlying cortical neurons of the left

occipital location after 24 h of TSD. The spontaneous eye

blink rate increased following one night of TSD (Barbato

et al., 1995). In addition, Cajochen et al. (2002) commented

that a circadian trough occurred at the melatonin maximum

in the occipital derivation. Our results reveal that occipital

neuronal activity simultaneously occurred within the FC

under high sleep pressure.

In the present study, integration and mutual information

used in the calculation of CI were calculated by means of a

covariance matrix. The entropy of the raw signals, used in

the calculation of the covariance matrix, is related to the

amplitude distribution of the signals (Putten and Stam,

2001), as a typical example if the signals contained in

subsets Xk have the identical amplitude and phases, the

entropy would be zero. The reference montage affects the

relative amplitudes and phases of the EEG signals, so do

integration and mutual information values. Putten and Stam

(2001) reported that source reference montage using the

voltage difference between the recording site and the mean

voltage of the nearest 3w4 neighbor sites would perform

better than the average montage utilizing the voltage

difference between the recording site and the average of

all electrode potentials. Considering these properties, we

used a linked earlobe as a reference electrode suitable for

grasping the relative phase of the signals. The validity of

this method was assessed by examining the average

integration and mutual information for EEGs, homogeneous

data, and decomposed signals using ICA. The average

integration and mutual information are nearly zero when

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–28332832

using the independent EEGs, since these data, which

decomposed nearly completely independent subsystems,

have neither integration nor mutual information. In contrast,

original EEG data have different values much more than

zero for each subset size. More clearly, CI values for

homogeneous data were almost 1, because these systems

uniformly interacted among the subsets. For decomposed

data, CI values were not defined, since all the integration

and mutual information were zero. Therefore, all calculation

processes are considered appropriate.

Although FC does not show the direct relationship

between scalp-recorded electrical and neuronal activities in

the underlying cortical tissue, the simultaneous neuronal

activity could be investigated through the application of FC

analysis. The extension of a FC analysis to an EEG

experiment with higher spatial resolution, along with the use

of various experimental protocols, would lead to a better

understanding of the role of sleep.

Acknowledgements

This research was supported by grant KOSEF-M02-

2004-000-11016-0 and KOSEF-R01-2001-00023-0.

References

Achermann P, Finelli LA, Borbely AA. Unihemispheric enhancement of

delta power in human frontal sleep EEG by prolonged wakefulness.

Brain Res 2001;913:220–3.

Aeschbach D, Dijk DJ, Borbely AA. Dynamics of EEG spindle frequency

activity during extended sleep in human: relationship to slow-wave

activity and time of day. Brain Res 1997;748:131–6.

Altenburg J, Vermeulen RJ, Strijers RL, Fetter WP, Stam CJ. Seizure

detection in the neonatal EEG with synchronization likelihood. Clin

Neurophysiol 2003;114(1):50–5.

Barbato G, Ficca G, Beatrice M, Casiello M, Muscettola G, Rinaldi F.

Effects of sleep deprivation on spontaneous eye blink rate and alpha

EEG power. Biol Psychiatry 1995;38(5):340–1.

Bell AJ, Sejnowski TJ. An information-maximization approach to blind

separation and blind deconvolution. Neural Comput 1995;7:1129–59.

Cajochen C, Brunner DP, Krauchi K, Grew P, Wirz-Justice A. Power

density in theta/alpha frequencies of the waking EEG progressively

increases during sustained wakefulness. Sleep 1995;18(10):890–4.

Cajochen C, Khalsa SBS, Wyatt JK, Caeisler CA, Dijk DJ. EEG and ocular

correlates of circadian melatonin phase human performance decrements

during sleep loss. Am J Physiol 1999a;277:R640–R649.

Cajochen C, Foy R, Dijk DJ. Frontal predominance of a relative increase in

sleep delta and theta EEG activity after sleep loss in humans. Sleep Res

Online 1999b;2:65–9.

Cajochen C, Knoblauch V, Krauchi K, Renz C, Wirz-Justice A. Dynamics

of frontal EEG activity, sleepiness and body temperature under high and

low sleep pressure. NeuroReport 2001;12(10):2277–81.

Cajochen C, Wyatt JK, Czeisler CA, Dijk DJ. Separation of circadian and

wake duration-dependent modulation of EEG activation during

wakefulness. Neuroscience 2002;114(4):1047–60.

Cantero JL, Atienza M, Salas RM. State-modulation of cortico-cortical

connections underlying normal EEG alpha variants. Physiol Behav

2000;71:107–15.

Casarotto S, Bianchi AM, Cerutti S, Chiarenza GA. Principal component

analysis for reduction of ocular artifacts in event-related potentials of

normal and dyslexic children. Clin Neurophysiol 2004;115:609–19.

Corsi-Cabrera M, Ramos J, Arce C, Guevara MA, Ponce-de Leon M,

Lorenzo I. Changes in the waking EEG as a consequence of sleep and

sleep deprivation. Sleep 1992;15(6):550–5.

Corsi-Cabrera M, Arce C, Ramos J, Lorenzo I, Guevara MA. Time course

of reaction time and EEG while performing a vigilance task during total

sleep deprivation. Sleep 1996;19(7):563–9.

Corsi-Cabrera M, Perez-Garci E, Del Rio-Protilla Y, Ugalade E,

Guevara MA. EEG bands during wakefulness, slow-wave, and

paradoxical sleep as a result of principal component analysis in the

rat. Sleep 2001;24(4):374–80.

Dann S, Beersma DGM, Borbely AA. Timing of human sleep: recovery

process gated by a circadian pacemaker. Am J Physiol 1984;246:

R161–R178.

Dijk DJ, Czeisler CA. Paradoxical timing of the circadian rhythm of sleep

propensity serves to consolidate sleep and wakefulness in humans.

Neurosci Lett 1994;166:63–8.

Dijk DJ, Czeisler CA. Contribution of the circadian pacemaker and the

sleep homeostat to sleep propensity, sleep structure, electroencephalo-

graphic slow waves and sleep spindle activity in humans. J Neurosci

1995;15:3526–38.

Drummond SP, Brown GG, Stricker JL, Buxton RB, Wong EC, Gillin JC.

Sleep deprivation-induced reduction in cortical functional response to

serial subtraction. NeuroReport 1999;10(18):3745–8.

Dumont M, Macchi M, Piche M-C, Carrier J, Hebert M. Waking EEG

activity during a constant routine: effects of sleep deprivation. Sleep

Res 1997;26:712–22.

Finelli LA, Baumann H, Borbely AA, Achermann P. Dual electroenceph-

alogram markers of human sleep homeostasis: correlation between

theta activity in waking and slow-wave activity in sleep. Neuroscience

2000;101(3):523–9.

Friston KJ, Frith CD, Fletcher P, Liddle PE, Fracowiak RSJ. Functional

topography: multidimensional scaling and functional connectivity in

the brain. Cereb Cortex 1996;6:156–64.

Jausovec N. Differences in cognitive processes between gifted, intelligent,

creative, and average individuals while solving complex problems: an

EEG study. Intelligence 2000;28:213–37.

Jeong J, Kim D-J, Kim SY, Chae J-H, Go HJ, Kim K-S. Effect of total sleep

deprivation on the dimensional complexity of the waking EEG. Sleep

2001;101(3):523–9.

Jin S-H, Jeong J, Jeong D-G, Kim D-J, Kim SY. Nonlinear dynamics of the

EEG separated by independent component analysis after sound and

light stimulation. Biol Cybern 2002;86(5):395–401.

Kim DJ, Lee HP, Kim MS, Park YJ, Go HJ, Kim KS, Lee SP, Chae JH,

Lee CT. The effect of total sleep deprivation on cognitive functions in

normal adult male subjects. Int J Neurosci 2001a;109(1/2):127–37.

Kim H, Guillerminault C, Hong S, Kim D, Kim S, Go H, Lee S. Pattern

analysis of sleep-deprived human EEG. J Sleep Res 2001b;10(3):

193–201.

Leocani C, Toro C, Manganotti P, Zhuang P, Hallet M. Event-related

coherence and event-related desynchronization/synchronization in the

10 Hz and 20 Hz EEG during self-paced movements. Electroencepha-

logr Clin Neurophysiol 1997;104:199–206.

Lorenzo I, Ramos J, Arce C, Guevara MA, Corsi-Cabrera M. Effect of total

sleep deprivation on reaction time and waking EEG activity in man.

Sleep 1995;18(5):346–54.

Makeig S, Jung T-P, Sejnowski TJ. Awareness during drowsiness:

dynamics and electrophysiological correlates. Can J Exp Psychol

2000;54:266–73.

McKeown MJ, Makeig S, Brown GG, Jung T-P, Kindermann SS, Bell AJ,

Sejnowski TJ. Analysis of fMRI data by decomposition into

independent components. Neurology 1997;48:A417.

Mikulincer M, Babkoff H, Caspy T, Weiss H. The impact of cognitive

interference on performance during prolonged sleep loss. Psychol Res

1990;52(1):80–6.

S.-H. Jin et al. / Clinical Neurophysiology 115 (2004) 2825–2833 2833

Papoulis A. Probability, random variables and stochastic processes, 3rd ed.

New York: McGraw Hill; 1991.

Pereda E, Manas S, De Vera L, Garrido JM, Lopez S, Gonzalez JJ. Non-

linear asymmetric interdependencies in the electroencephalogram of

healthy term neonates during sleep. Neurosci Lett 2003;337(2):101–5.

Putten MJAM, Stam CJ. Application of a neural complexity measure to

multichannel EEG. Phys Lett A 2001;19:131–41.

Shanahan TL, Zeitzer JM, Czeisler CA. Resetting the melatonin rhythm

with light humans. J Biol Rhythms 1997;12:556–67.

Shen B, Nadkarni M, Zappulla RA. Spectral modulation of cortical

connections measured by EEG coherence in human. Clin Neurophysiol

1999;110:115–25.

Stam CJ, van Dijck BW. Synchronization likelihood: an unbiased measure

of generalized synchronization in multivariate data sets. Physica D

2002;163:236–51.

Stoyanova R, Brown TR. NMR spectral quantitation by principal

component analysis: III. A generalized procedure for determination of

lineshape variations. J Magn Reson 2002;154:163–75.

Tanaka H, Hayashi M, Hori T. Topographical characteristics of slow wave

activities during the transition from wakefulness to sleep. Clin

Neurophysiol 2000;111(3):417–27.

Thatcher RW, Walker WA, Giudice S. Human cerebral hemispheres

develop at different rates and ages. Science 1987;236:1110–3.

Tononi G, Edelman GM. Schizophrenia and the mechanisms of conscious

integration. Brain Res Rev 2000;31:391–400.

Tononi G, Sporns O, Edelman GM. A measure for brain complexity:

relating functional segregation and integration in the nervous system.

Proc Natl Acad Sci USA 1994;91:5033–7.

Tononi G, McIntosh AR, Russell DP, Edelman GM. Functional clustering;

identifying strongly interactive brain regions in neuroimaging data.

NeuroImage 1998;7:133–49.

Weiss S, Rappelsberger P. Long-range EEG synchronization during word

encoding correlates with successful memory performance. Cogn Brain

Res 2000;9:299–312.

Wright Jr KP, Badia P, Wauquier A. Topographical and temporal patterns

of brain activity during the transition from wakefulness to sleep. Sleep

1995;18(10):880–9.