Home Search Collections Journals About Contact us My IOPscience

Neurodevelopment in newborns: a sample entropy analysis of electroencephalogram

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2009 Physiol. Meas. 30 491

(http://iopscience.iop.org/0967-3334/30/5/006)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 141.211.4.224
This content was downloaded on 08/09/2015 at 15:47

Please note that terms and conditions apply.

IOP PUBLISHING PHYSIOLOGICAL MEASUREMENT

Physiol. Meas. 30 (2009) 491–504 doi:10.1088/0967-3334/30/5/006

Neurodevelopment in newborns: a sample entropy

analysis of electroencephalogram

Dandan Zhang1, Haiyan Ding1, Yunfeng Liu1, Congle Zhou2,

Haishu Ding1 and Datian Ye3
1 Department of Biomedical Engineering, Tsinghua University, Beijing 100084,

People’s Republic of China

2 Department of Pediatrics, First Hospital of Peking University, Beijing 100034,

People’s Republic of China

3 Biomedical Engineering Research Centre, Graduate School at Shenzhen, Tsinghua University,

Shenzhen 518055, People’s Republic of China

E-mail: zhangdd05@gmail.com

Received 14 December 2008 accepted for publication 23 March 2009

Published 15 April 2009
Online at stacks.iop.org/PM/30/491

Abstract
The present paper investigates the neural ontogeny of newborns in view of
electroencephalogram (EEG) complexity during active sleep (AS) and quiet
sleep (QS). Sample entropy (SampEn) is applied to EEG recordings from 168
newborns with postmenstrual age (PMA) ranging from 25 to 60 weeks. The
relationship between neurodevelopment and PMA is then explored according
to the statistical analysis of the median and interquartile range of SampEn
curves. It is found that SampEn of EEG during AS is higher than that during
QS. SampEn increases during both AS and QS before about 42 weeks in PMA
while it ceases its increase in QS and even decreases in AS after newborns
reaching term age. A distinct decrease in the interquartile range of SampEn
is found with increasing PMA (from 25 to about 50 weeks), followed by
maintenance of low fluctuation in SampEn curves. The study in this paper
sets the stage for exhaustive investigation of the SampEn of EEG during brain
maturation in newborns. And it could be hoped that SampEn in sleep EEG
might be a useful parameter against which delays and aberrations in brain
maturation might be tested. The SampEn changes during brain maturation
also offer functional clues about neurodevelopment, based on which further
explorations could be done. The significance of this paper is the discovery of
the decrease in EEG complexity after newborns reaching term. Although some
potential neurophysiologic reasons are given, this new discovery might require
more study to investigate. In addition, the fluctuation of EEG complexity is
analyzed for the first time, which helps to understand the EEG maturation in
neurodevelopment.

Keywords: neurodevelopment, EEG, nonlinear analysis, sample entropy

0967-3334/09/050491+14$30.00 © 2009 Institute of Physics and Engineering in Medicine Printed in the UK 491
492 D Zhang et al

1. Background

The neonatal brain is a developmental brain. Sleep takes up most of a newborn’s time during
the early days of their lives. Functional brain ontogeny studies of sleep contribute to the
understanding of neural maturation and may provide answers to many unsettled questions.
The presence of sleep states and a sleep state cycle is an important phenomenon of normality or
abnormality, which helps pediatricians evaluate the risk of developmental delay in newborns.
Electroencephalography (EEG), which reflects the integrated brain activity, constitutes a
valuable tool for evaluation of brain maturity. In the first months of life, EEG could present
major changes (Mizrahi et al 2004) while the clinical symptom is not distinct. So it is able
to provide useful information about brain function prior to clinical manifestation—a window
of opportunity for appropriate interventions (Jordan 1995). In addition, EEG monitoring is
non-invasive and continuous.
There are generally two distinct sleep states in newborns. Active sleep (AS) is defined
as sleep presenting a low-to-moderate voltage, continuous EEG pattern, with rapid eye
movements (REMs) and irregular heart rate. Conversely, quiet sleep (QS) is associated
with the context of a continuous high-voltage slow or discontinuous EEG pattern and less
variability in heart rate, absence of REMs (Swartjes et al 1990, Peirano et al 2003, Scher
2008). Since EEG patterns are different between AS and QS, many researchers (Nunes et al
1997, Ferri et al 2003, Paul et al 2003, 2006, Scher et al 2003, 2005, Scher 2008, Carrozzi
et al 2004, Jenni et al 2004, Pereda et al 2006, Janjarasjitt et al 2008b) have investigated the
EEG data in these two sleep states, respectively.
Early studies of newborn EEG explored specific measures, such as REM count, percentage
of QS, sleep cycle length and spectral energy, all of which were proved as significant
representations of functional brain organization and maturation in healthy preterm and full-
term cohorts (Scher 1997a, 1997b, Scher et al 2003). Jenni et al (2004) investigated the
spectral power of newborn sleep EEG in a longitudinal study and concluded that the nocturnal
patterns of delta, theta and sigma activity could reflect different functional roles in the sleep
process during development. Biagioni et al (2007) examined the maximum interval duration
and minimum burst duration of EEG in very preterm infants and proved that postmenstrual age
(PMA = gestational age at delivery + postnatal age) had a strong influence on both anatomical
and electrophysiological maturation.
In recent years, the application of nonlinear methods (especially nonlinear complexity
methods) to the analysis of EEG has been of great interest to many researchers (Stam 2005).
These nonlinear parameters are essentially independent of classical EEG analysis and can give
complementary knowledge of newborn brain function. Scher et al (2005) used correlation
dimension to study the dynamics of EEG in term and preterm neonates as a function of
gestational age and found that this index was different between sleep states. Meyer-Lindenberg
(1996) explored the correlation dimension of EEG from newborns to adults and the results
showed that EEG complexity was correlated with brain maturation. A more recent study by
Janjarasjitt et al (2008a) investigated the complexity of neonatal EEG in a more detailed frame.
50 healthy neonates were subjected to EEG monitoring and totally 86 recordings with PMA
ranging from 28 to 43 weeks were collected (some subjects with more than one recording).
Then the relationship between neurodevelopment and EEG complexity was explored. They
concluded that the dimensional complexity of neonatal EEG increased with brain maturation.
The data analysis was precise throughout and the results were convincing. Yet we noticed
from their results that the increased tendency of EEG complexity was slightly depressed after
newborns reaching term, i.e. PMA  37 weeks (refer to figures 2 and 3 in that paper). Is this
disturbance of complexity increase in EEG due to data collection or neurodevelopment? How
Neurodevelopment in newborns 493

would EEG complexity develop in older neonates and infants? To find out the answers, the
present paper investigates the trend of EEG nonlinear complexity from early preterm neonates
to infants in a wider PMA range.
The early phase of nonlinear EEG analysis was characterized by the search for low-
dimensional chaotic dynamics in EEG series (Röschke and Aldenhoff 1991, Lutzenberger
et al 1995, Pritchard and Duke 1995, Shen et al 2003, Carrozzi et al 2004). Brain maturation
in EEG was investigated using correlation dimension (Anokhin et al 1996, 2000, Meyer-
Lindenberg 1996, Scher et al 2005, Pereda et al 2006, Janjarasjitt et al 2008a, 2008b), fractal
parameters (Choi et al 2000) and Kolmogorov entropy (Hecox et al 2003), all of which were
considered as a group of useful indexes, providing quantitative assessments of the structure
in EEG (Rapp et al 2001). Around 1990, limitations of aforementioned algorithms became
clear (Stam 2005). Since then, researchers of nonlinear EEG analysis began to focus on
the development of new measures which were more suitable to be applied to noisy,
nonstationary and high-dimensional EEG data (Stam 2005). Then approximate entropy
(ApEn) came forth and became one of the most commonly used measures to quantify the
complexity of EEG. ApEn was introduced by Pincus (1991) and has been proved as a powerful
tool to reveal the cerebral states both in animal models (Al-Nashash et al 2003, Bezerianos et
al 2003) and in human (Papadelis et al 2007). It examines time series for similar epochs: more
frequent and more similar epochs lead to lower values of ApEn. Thus, a low ApEn reflects a
high degree of regularity (Richman and Moorman 2000). Since the ApEn algorithm counts
self-matches when it calculates the pairs of similar epochs, a bias to some extent is introduced,
which reduces the performance of this statistical measure. Soon the limits of the ApEn
method were noticed by many researchers including Pincus himself (Pincus and Goldberger
1994, Pincus 1995). In 2000, Richman and Moorman (2000) first introduced sample entropy
(SampEn) as a new family of statistics, measuring irregularity of clinical and experimental time
series. It improved the performance of ApEn in two aspects (Richman and Moorman 2000):
(1) it agrees much better than ApEn for random numbers with known probabilistic character
over a wide range of operating conditions and (2) it maintains relative consistency where ApEn
does not. Besides statistical improvements, the algorithm of SampEn remarkably increases
the computing speed. All of these advantages make SampEn more promising than ApEn in
practice (Richman and Moorman 2000, Lake et al 2002, Richman et al 2004, Casaseca-de-
la-Higuera et al 2006). SampEn has been used in the investigation of EEG (Ramanand et al
2004, Abásolo et al 2006, Chouvarda et al 2007, Yum et al 2008) and other physiological time
series (Lake et al 2002, Fele-Žorž et al 2008).
This study selects SampEn as the nonlinear measure to investigate EEG recordings in a
large population of newborns. In order to avoid the correlation between EEG samples from
the same subject, only one EEG sample, instead of several ones (Meyer-Lindenberg 1996,
Janjarasjitt et al 2008a, 2008b), is selected from each sleep state in each subject. Single-channel
EEG is performed in this study since it could greatly simplify the process of electrode fixing
and guarantee a prolonged bedside monitoring. The relationship between neurodevelopment
and neonatal/infant age is then explored according to the statistical analysis of median and
interquartile range (i.e. 75th percentile – 25th percentile) of SampEn curves.

2. Methods

2.1. Data and subjects

A human newborn less than 1 month old is defined as neonate. Then it is called an infant until
it reaches 1 year of age. In this study, subjects are divided into six age groups according to their
494 D Zhang et al

Table 1. EEG samples selected from 168 recordings. In total, 281 (127 + 144 + 10) 5 min samples
are selected from 168 neonates.
Age group Subject AS sample QS sample Indeterminate
Early preterm neonates 10 10
Preterm neonates 30 26 30
Term neonates 49 44 46
Old neonates 27 21 24
Infants 34 27 27
Old infants 18 9 17
Total 168 127 144 10

PMAs: early preterm neonates (25 weeks  PMA < 30 weeks); preterm neonates (30 weeks 
PMA < 37 weeks); term neonates (37 weeks  PMA < 42 weeks); old neonates (42 weeks 
PMA < 46 weeks); infants (46 weeks  PMA < 53 weeks); old infants (53 weeks  PMA  60
weeks).
Totally, 168 newborns with PMA ranging from 25 to 60 weeks (42.14 ± 8.25 weeks) are
included in the present study (refer to table 1). Among 168 newborns, EEG monitoring was
performed in 115 in the Department of Pediatrics, First Hospital of Peking University, People’s
Republic of China. All subjects were neurologically normal at the time of EEG recording and
showed normal development at follow-up. Two detecting electrodes were fixed in P3 and P4
in the 10–20 international system (bipolar recording). The sampling rate was 1000 Hz. 16
bit A/D conversion was used. The bioelectric amplifier was provided by Symtop Instrument
Co., People’s Republic of China. The average recording time was about 60 min. Parents of
the infants were informed of methods and purposes of the examination and they gave their
consent. This research had been approved by the Ethics Committee of Peking University and
put on records.
The EEG recordings of 53 other newborns, including 8 early preterm neonates, 23 preterm
neonates and 22 term neonates, were from the Australian EEG Database, which was jointly
maintained by the University of Newcastle and the John Hunter Hospital. All subjects
were neurologically normal at the time of EEG recording. Their EEGs were proved as
normal recordings for the age and stages of sleep by experienced electroencephalographic
technologists. All the 53 EEGs were recorded on a Walter Graphtek Paperless EEG system
with a sampling rate of 167 Hz. 9 bit A/D conversion was used. The average recording time
was 50 min. About 16 electrodes were used during EEG monitoring. The electrode numbers
were not exactly the same in each subject. Only channel P3–P4 (or C3–C4, if no P3–P4) of
the EEG data is selected and analyzed in this study.
We choose channel P3–P4 to record and analyze EEG data for the following reasons
(Prior and Maynard 1986): (1) EEG from parietal regions is less affected by muscle potentials
from facial, jaw and eye movements; (2) this is the triple arterial boundary zone where the
territories of anterior, middle and posterior arteries meet so the EEG here is very sensitive to
ischemia and other related cerebral dysfunctions; (3) EEG amplitudes are relatively large in
the parietal regions; (4) it causes least interference to nursing procedures and least discomfort
to patients. Channel C3–C4 is the nearest bipolar channel to P3–P4 in the standard neonatal
16-lead system so it is selected in this study if there is no P3–P4 channel.
Two 5 min samples free of artifacts (segments contaminated by artifacts are eliminated
by visual inspection) are selected from the EEG recording of each newborn. One sample is
chosen from the middle part of AS, the other from the middle of QS. In addition to EEG
Neurodevelopment in newborns 495

Figure 1. Neonatal sleep EEG and heart rate. HR: heart rate. PMA = 44 weeks.

patterns, subject states (e.g. eye movements and limb movements) during monitoring and
heart rates (refer to figure 1) are also considered when 5 min samples are selected. A total
of 281 artifact-free samples are chosen during AS (127 samples) and QS (144 samples).
Another ten samples are selected from ten early preterm neonates; see table 1. Although it has
been reported that crude sleep-state organization could present as early as 25 week gestation
(Vecchierini et al 2003), we think AS and QS are difficult to differentiate in EEG before PMA
reaching 30 weeks. In order to compare the SampEn results between the early preterm group
and the other five groups, the ten samples of ‘indeterminate’ sleep state are considered both
in AS analysis and QS analysis in this study.

2.2. Data processing

EEGs recorded in First Hospital of Peking University are resampled digitally with the sampling
rate changing from 1000 Hz to 167 Hz. Therefore, all the EEG data have the same sampling
rate. Then band-pass filtering (0.5–30 Hz, 5-order Butterworth filter) is used to minimize
artifacts from sweating, muscle activity and electrical interference since newborn EEGs contain
most of the neuron activity information below 30 Hz (or even lower). The detailed algorithm
and parameter choice of SampEn can be referred to in the literature (Richman and Moorman
2000, Lake et al 2002, Ramanand et al 2004, Papadelis et al 2007). In the present study, we
let N = 2004, m = 2, r = 0.2 SD according to Lake et al (2002) (N: data length; m: embedding
dimension of the vector to be formed; r: tolerance that functions as a noise filter; SD: standard
deviation of the signal). SampEn is calculated from every 2004 data points (12 s), with 1002
points (6 s) overlapping. Therefore, 49 points of SampEn can be obtained from every 5 min
EEG data (sampling rate = 167 Hz). Then the median (denoting SampEn tendency) and the
interquartile range (denoting SampEn fluctuation) of each 49-point SampEn time series are
calculated and used in the statistical analysis in this paper.

2.3. Statistical analysis

Statistical analysis is based on Matlab (The MathWorks). Medians of SampEn are compared
between QS and AS using the Wilcoxon rank sum test. To check the significant differences in
medians of SampEn among age groups, a nonparametric test (Kruskal–Wallis analysis of the
variance) is used. For interquartile ranges of SampEn, scatter diagrams are plotted to reveal
the relationship between SampEn fluctuation and PMA. Methods such as linear regression
496 D Zhang et al

Figure 2. Development of EEG and SampEn in normal sleeping newborns.

are not adopted here because a linear relationship does not exist when PMA is more than
50 weeks.

3. Results

3.1. EEG and SampEn during neurodevelopment

Examples of EEG waveforms and SampEn curves from each age group are shown in
figure 2, which displays a clear developmental trend of newborn sleeping EEG and its
SampEn. Six examples in figure 2 demonstrate that EEG waveforms and the SampEn statistics
undergo a series of distinct changes along with increasing PMA. Early preterm neonate has a
highly discontinuous EEG, which presents as a low and rather flat EEG background added by
occasional electrical impulses. Accordingly, SampEn shows low value and large fluctuations
(see the first example in figure 2 (PMA = 25 weeks)). Then, EEG transforms gradually and
begins to exhibit some continuity during AS after PMA reaching 31 weeks (Mizrahi et al
2004). So the SampEn curve rises during AS (PMA = 32 weeks in figure 2). From then
on, EEG becomes more and more continuous and evolves into its maturation form in term
neonates (PMA  37 weeks). At the same time, SampEn differences between two sleep states
are diminished gradually. In addition, the last four examples (PMA = 39, 44, 52, 60 weeks)
reveal that the values of SampEn decrease slightly while the amplitudes of EEG increase
during this period of neurodevelopment.

3.2. SampEn tendency during maturation: median analysis

The medians of SampEn statistics in every 5 min EEG recording from each subject are
calculated and displayed in figure 3. It is difficult to differentiate AS from QS in EEG until
Neurodevelopment in newborns 497

Figure 3. Relationship between PMA and median of SampEn statistics in QS and AS. Each
median value of SampEn is calculated from the 5 min EEG selected from each subject. The box
of the early preterm group in QS is the same as that in AS to help investigation. ∗P < 0.05.

Table 2. Paired ANOVA of the medians of SampEn statistics during QS between age groups.
Group labels: 1—early preterm; 2—preterm; 3—term; 4—old neonate; 5—infant; 6—old infant.

Pair Difference Confidence interval P

1 and 2 −23.2500 −69.6591 to 23.1591 P > 0.05
2 and 3 −71.1522 −100.9786 to −41.3257 P < 0.05
3 and 4 3.0272 −28.9764 to 35.0308 P > 0.05
4 and 5 33.8194 −1.8365 to 69.4754 P > 0.05
5 and 6 −0.7974 −40.1482 to 38.5535 P > 0.05

PMA reaches 30 weeks. In order to compare the results between the early preterm group and
the other five age groups, medians of SampEn from early preterm subjects are considered both
in AS analysis and in QS analysis.
Shown is the boxplot of medians of SampEn in different age groups. Three horizontal
lines in every box reflect the lower quartile, median and upper quartile of the statistics (i.e.
median of SampEn), respectively. The whisker is equal to one interquartile range of the
medians of SampEn in each group. The ‘+’ marks represent outliers. The notches represent
a robust estimate of the uncertainty for box-to-box comparison. Boxes whose notches do not
overlap indicate that the medians of the two groups differ at the 5% significance level.
It can be observed from figure 3 that SampEn of newborn EEG during AS is higher than
that during QS. The result of the Wilcoxon rank sum test shows that the SampEn differences
in all age groups (except the early preterm group) between QS and AS are significant (P <
0.01). The results of the Kruskal–Wallis analysis of the variance (ANOVA) show that there
are statistically significant differences in SampEn among six age groups during QS and AS,
respectively (P < 0.01). Then, the multi-compare test is performed and the results between
neighboring age groups are shown in tables 2 and 3. It is indicated that SampEn increases
significantly from preterm to term period, during both AS and QS. But it drops a lot in the
fifth group (i.e. infant group) during AS.
3.3. SampEn fluctuation during maturation: interquartile range analysis
The scatter diagrams of interquartile ranges of SampEn statistics in every 5 min EEG recording
are shown in figure 4. Obviously, the SampEn fluctuation that is revealed by the interquartile
range decreases with increasing PMA during both QS and AS. Since EEG and accordingly its
498 D Zhang et al

Figure 4. Relationship between PMA and interquartile range of SampEn statistics in QS and AS.
Each interquartile range value of SampEn is calculated from the 5 min EEG recording selected
from each subject. iqr: interquartile range.

Table 3. Paired ANOVA of the medians of SampEn statistics during AS between age groups. The
same group labels as with table 2.

Pair Difference Confidence interval P

1 and 2 −58.9615 −101.0510 to −16.8720 P < 0.05
2 and 3 −33.9930 −61.9728 to −6.0132 P < 0.05
3 and 4 9.0260 −20.9746 to 39.0266 P > 0.05
4 and 5 43.3545 10.4437 to 76.2653 P < 0.05
5 and 6 16.2963 −27.2406 to 59.8332 P > 0.05

SampEn are nonstationary time series, the interquartile range of SampEn could not decrease
to zero; instead, it tends to remain in a low level (about 0.05) after approximately 50 weeks in
PMA, as shown in figure 4.

4. Discussion

4.1. Neurodevelopment and EEG complexity in newborns

The present paper investigates the neural ontogeny of newborns in view of EEG complexity.
To obtain a universal view, a large population of newborns (168) with a large PMA range
(between 25 and 60 weeks) is included. It is found that EEG complexity increases with PMA
till term age. Yet it decreases gradually to some extent during AS from then on. (At least it
decreases till 60 weeks in PMA.) Similar studies have been done by Janjarasjitt et al (2008a)
and Pereda et al (2006). Our result is consistent with theirs within the neonatal PMA ranges in
their papers, i.e. 28 weeks  PMA  43 weeks in Janjarasjitt’s work and 32 weeks  PMA 
46 weeks in Pereda’s work. Unfortunately, the subjects in their studies (Janjarasjitt et al
2008a, Pereda et al 2006) consisted of only preterm and term neonates, so the decrease of
EEG complexity after term was not discovered.
Obviously, the changes in EEG complexity are influenced by neural structure development
in newborns. Primary neurulation and prosencephalic development occurs before 3 months
of gestation, followed by neuron proliferation (peak time: 7–13 weeks of gestation) and
migration (peak time: 13–22 weeks of gestation). Then a transitory structure in neural
ontogeny, namely subplate zone, appears. It can provide a site for synaptic contact for axons
ascending from thalamus and other cortical sites, since their neuron targets have not yet arrived
Neurodevelopment in newborns 499

or differentiated. After the subplate neuron layer reaches its peak in approximately 22–34
weeks of gestation, programmed cell death begins and this structure breaks down gradually. At
the same time, neuron lamination occurs as neuron migration ceases. Then neurite outgrowth
and axonal development become the dominant developmental activities in cerebral cortex
over the last trimester of gestation (about 30–39 weeks in PMA), followed by postnatal
rapid synaptogenesis. The last crucial procedure in neurodevelopment is myelination, which
progresses most rapidly within the central nervous system (CNS) after birth (Volpe 2001).
Myelination makes axons and dendrites electrically insulated so that neurons can fire with less
electrical generalization.
Not only do the series of neuronic structure changes happen during neonatal and infant
period, but also biochemical features of neurotransmission vary markedly in the immature
brain. As a primary excitatory neurotransmitter in the cerebral cortex, glutamate is highly
active in early life. The expression of glutamate receptors and second-messenger systems is
enhanced especially in perinatal and early postnatal brain. In contrast, the major inhibitory
transmitter (i.e. γ -aminobutyric acid (GABA)) and its expression system exhibit a peak in about
3 weeks after birth (Johnston 1995, Johnston and Silverstein 2004). Therefore, neuroelectric
transmission is more active (due to excitatory neurotransmitters) in preterm and term period
while relatively less active (due to inhibitory neurotransmitters) after term.
The aforementioned neurodevelopment could help in understanding our results
about SampEn tendency shown in figure 3. During the period from early preterm
to term, a neuron network establishes and develops dramatically because of neuron
proliferation, migration, lamination as well as neurite outgrowth and axonal development.
Therefore, SampEn of EEG increases in the first three age groups in figure 3, which
is also illuminated by Janjarasjitt et al (2008a) and Pereda et al (2006). Then, we
suppose that the decrease in EEG complexity after term is due to at least two reasons.
First is the myelination within CNS, which progresses most rapidly after birth. With
myelinated axons and dendrites, electrical activities can transfer much faster, so neuron
networks are able to fire synchronously, which may decrease the SampEn measure.
Neurotransmitter development is another potential reason. Distinct neurotransmitter-specific
neuron groups mature at different rates (Johnston and Silverstein 2004). The transmission
of neuron activities is predominated by excitatory neurotransmitters during preterm and term
period. Then, inhibitory transmitters such as GABA begin to function extensively, which
drives neuron activities into the control of a balance between excitatory transmission and
inhibitory transmission. The latter situation may lead to a lower EEG complexity.
Compared with other investigations of EEG complexity in a whole life span, our result
after term is a little surprising. Meyer-Lindenberg (1996) and Anokhin et al (1996) investigated
the EEG from newborns/children to adults and their results showed a monotone increase in
complexity measure along with brain development. The authors attributed this complexity
rise in EEG to the increasing number of relatively independent oscillating neural networks
contributing to the EEG signal (Lutzenberger et al 1995). However, it is too early to conclude
that our results are incompatible with their because (1) different states were concerned: they
analyzed the EEG during the wake period while we investigated sleep EEG; (2) different time
spans were concerned: they explored the whole life of human while we only focused on the
detailed neural ontogeny of newborn. Since both the structure and the function of CNS change
dramatically during the newborn period, it is highly possible that SampEn of EEG exhibits
a more complicated evolvement in immature brain, compared with the SampEn from mature
brain in child and adult.
In addition to SampEn trend during maturation, this study first analyzes the fluctuations of
EEG complexity in the process of neurodevelopment. Considering that the ‘burst-suppression’
500 D Zhang et al

EEG in very early preterm neonates (refer to figure 2, PMA = 25 weeks) is replaced by
continuous EEG during maturation, it is very logical that SampEn gradually evolves into a
smoother curve. However, EEG signals are typically nonstationary time series, so SampEn
statistics of EEG could not become a straight line with ‘zero fluctuation’. Therefore, the
interquartile range of SampEn should remain in a low level, which is denoted by dots with
PMA more than about 50 weeks in figure 4.

4.2. AS and QS in newborns

Different sleep states reflect different modes of brain-controlled physiologic activities (Peirano
et al 2003). AS is regarded as a sleep state regulated by lower brainstem mechanisms, including
reticular neurons in the pons and spinal cord, while the development of QS is associated with
the formation of thalamocortical and intracortical patterns of innervations (Pace-Schott and
Hobson 2002) and is strongly dependent on the integrity of neocortex. Neurogenesis begins
from the caudal, and then progresses to the rostral part of the brain (Curzi-Dascalova and
Challamel 2000). As a result, the regions correlated with AS mature earlier than those
with QS.
We prove that SampEn of newborn EEG during AS is higher than that during QS, which
is consistent with the results of other researches (Carrozzi et al 2004, Janjarasjitt et al 2008b).
Probably, there are two reasons for the result. Firstly, all physiologic and behavioral parameters
that characterize AS present well before term in human sleep ontogeny (Curzi-Dascalova and
Challamel 2000). In contrast, the maturation of QS is far more behind. Therefore, when the
EEG in AS evolves into a continuous waveform, the background EEG in QS is still rather
flat, which leads to a low SampEn value. This can explain why the SampEn during AS
is much higher than that during QS especially in the preterm group (30 weeks  PMA <
37 weeks). The second reason maybe that the immature nervous system of newborns is
unable to inhibit or control the AS-on center (located in the pontine tegmentum) (Curzi-
Dascalova and Challamel 2000), so redundant desynchronized neuron discharges are produced
during AS.

4.3. Other issues

‘A defining but elusive feature of physiologic systems is their daunting complexity’
(Goldberger et al 2002). In present study, we choose SampEn as a complexity measure
of EEG signals. However, the measures such as SampEn, ApEn and correlation dimension are
designed to quantify the complexity of time series; they are indirect measures of physiological
complexity (Goldberger et al (2002) have explained this matter in detail). This caveat should
be kept in mind when the results computed by SampEn are interpreted.
We chose the SampEn method to investigate the EEG in newborns instead of correlation
dimension used by Janjarasjitt et al (2008a) and Pereda et al (2006). In the view of nonlinear
dynamical analysis (Stam 2005), these two measures have similar outcomes when we describe
a nonlinear dissipative system such as neural networks of the brain. Therefore, the result
obtained from SampEn in this study is similar to that obtained from correlation dimension
(in preterm and term neonates). However, correlation dimension is developed for the purely
deterministic dynamical system setting so that it may produce confusing results for biological
time series, which likely comprise both stochastic and deterministic components (Pincus
1995). In addition, correlation dimension is a time-consuming method (Stam 2005) with
acute sensitivity to noise (Pincus 1995). In contrast, SampEn can distinguish a wide variety
of systems, including mixed (stochastic and deterministic) systems. SampEn is intended for
Neurodevelopment in newborns 501

statistical applications where data sets are noisy and size limited. Therefore, we think SampEn
is more suitable in the present study although correlation dimension also can provide valuable
information.
A question of debate is whether neonatal EEG possesses nonlinearity. In other words,
is it reasonable to use a nonlinear measure such as SampEn to describe EEG time series in
newborns? A statistical control method, namely surrogate analysis, was designed to provide
a framework within which claims of nonlinearity could be evaluated (Theiler et al 1992a,
1992b). Many researchers used this method to evaluate the validity of their results based
on nonlinear EEG analysis (Meyer-Lindenberg 1996, Ferri et al 2003, Shen et al 2003). A
primary conclusion till now is that nonlinear features are quantifiable during QS while the
neonatal EEG during AS appears to be associated with linear processes (Ferri et al 2003,
Janjarasjitt et al 2008b). However, even though it has been unable to find clear indications
of nonlinearity in all the sleep EEG of newborns, nonlinear measures such as correlation
dimension and SampEn are useful to make relative comparisons among groups of data and
could serve as relative indexes of the complexity of brain dynamics (Pritchard and Duke 1995,
Janjarasjitt et al 2008b).
One potential disputation of our data selection is that only a 5 min EEG sample is
considered in each newborn during AS and QS. It is for two reasons: to get noise-free
recordings and to insure a definite sleep state. Similar methods were also adopted by other
researchers (Shen et al 2003, Paul et al 2003, 2006).
It should be noted that ‘AS’ and ‘QS’ are terms to describe sleep states only in neonates,
though they are used throughout this paper for convenient depiction. Beyond the first month of
life, sleep patterns of newborns begin to change. QS, which is called non-rapid eye movement
(NREM) sleep in adult, gradually differentiates into four sub-states (stages 1–4) between 1.5
and 3 months (Sterman et al 1977). Accordingly, AS transforms into REM sleep. It should
be pointed out that the complexity of EEG varies within NREM sleep (decreases from stage
1 to stage 4) in adult (Röschke and Aldenhoff 1991) so more detailed sleep EEG analysis (i.e.
stage 1 to stage 4 in NREM sleep and REM sleep) is suggested when infants grow older.
According to usual rules in EEG ontogeny of newborns (Scher 2008), the present study
investigates the relationship between SampEn statistics and PMA without considering the
influence of extrauterine life. Although we agree with Volpe (2001) and Scher (1996) that
there is little difference in EEG between neonates born at term and those who grow to term after
premature delivery, some studies have proved that there are statistically significant differences
in the dimensional complexity of EEG between these two newborn cohorts (Nunes et al
1997, Scher et al 2005, Janjarasjitt et al 2008a, 2008b). Our subsequent work could further
investigate this issue.

5. Conclusion

The present study investigates the neural ontogeny in newborns using relatively new statistics,
namely SampEn in EEG. Based on EEG recordings from a large population of newborns,
SampEn tendency is explored with PMA ranging from 25 to 60 weeks. SampEn increases
during both AS and QS before about 42 weeks in PMA, which is consistent with other studies.
Yet SampEn of EEG is found to cease its increase during QS and even decreases during AS
after newborns reaching term age. Although some potential neurophysiologic reasons are
given in section 4, this new discovery might require more study to investigate, especially
the question of whether or not the decrease of SampEn after term is a local valley in the
whole span of life. In addition, SampEn fluctuation during neurodevelopment is analyzed for
the first time in this study. A distinct decrease in the interquartile range of SampEn is found
502 D Zhang et al

with PMA from 25 to about 50 weeks, followed by maintenance of less fluctuation in SampEn.
SampEn changes during brain maturation can offer functional clues about neurodevelopment,
based on which further neural ontogeny explorations could be done. And it can be hoped that
SampEn might be a useful parameter against which delays and aberrations in brain maturation
could be tested.

Acknowledgments

This work was supported by the National Natural Science Foundation of China, no 60578004,
and Dr Shun Tak Wu’s Medical Sciences Fund, Tsinghua University. Dr Liu is supported by
the National Science Foundation for Post-doctoral Scientists of China, no 20080430380.

References

Abásolo D, Hornero R, Espino P, Alvarez D and Poza J 2006 Entropy analysis of the EEG background activity in
Alzheimer’s disease patients Physiol. Meas. 27 241–53
Al-Nashash H A, Paul J S, Ziai W C, Daniel F, Hanley D F and Thakor N V 2003 Wavelet entropy for subband
segmentation of EEG during injury and recovery Ann. Biomed. Eng. 31 653–8
Anokhin A P, Birbaumer N, Lutzenberger W, Nikolaev A and Vogel F 1996 Age increases brain complexity
Electroencephalogr. Clin. Neurophysiol. 99 63–8
Anokhin A P, Lutzenberger W, Nokolaev A and Birbaumer N 2000 Complexity of electrical dynamics in children:
developmental aspects Dev. Psychobiol. 36 9–22
Bezerianos A, Tong S and Thakor N 2003 Time-dependent entropy estimation of EEG rhythm changes following
brain ischemia Ann. Biomed. Eng. 31 221–32
Biagioni E et al 2007 Maturation of cerebral electrical activity and development of cortical folding in young very
preterm infants Clin. Neurophysiol. 118 53–9
Carrozzi M, Accardo A and Bouquet F 2004 Analysis of sleep-stage characteristics in full-term newborns by means
of spectral and fractal parameters Sleep 27 1384–93
Casaseca-de-la-Higuera P, Martı́n-Fernández M and Alberola-López C 2006 Weaning from mechanical ventilation:
a retrospective analysis leading to a multimodal perspective IEEE Trans. Biomed. Eng. 53 1330–45
Choi K G, Joo E Y, Yu H J, Oh J Y, Kim E S and Kim J S 2000 A study of fractal dimension and Lyapunov exponent
index in various age groups Neurol. Psychiatry Brain. Res. 8 27–32
Chouvarda I, Papadelis C, Kourtidou-Papadeli C, Bamidis P D, Koufogiannis D, Bekiaris E and Maglaveras N 2007
Non-linear analysis for the sleepy drivers problem Stud. Health Technol. Inform. 129 1294–8
Curzi-Dascalova L and Challamel MJ 2000 Neurophysiological basis of sleep development Sleep and Breathing in
Children: A Developmental Approach ed G M Loughlin, J L Carroll and C L Marcus (New York: Dekker)
pp 3–37
Fele-Žorž G, Kavšek G, Novak-Antolič Z and Jager F 2008 A comparison of various linear and non-linear signal
processing techniques to separate uterine EMG records of term and pre-term delivery groups Med. Biol. Eng.
Comput. 46 911–22
Ferri R, Chiaramonti R, Elia M, Musumeci S A, Ragazzoni A and Stam C J 2003 Nonlinear EEG analysis during
sleep in premature and full-term infants Clin. Neurophysiol. 114 1176–80
Goldberger A L, Peng C K and Lipsitz L A 2002 What is physiologic complexity and how does it change with aging
and disease? Neurobiol. Aging 23 23–6
Hecox K, Nayak S, Gin K, McGee A and van Drongelen W 2003 Linear and nonlinear measures of the human
neonatal EEG Neurocomputing 52–54 779–86
Janjarasjitt S, Scher M S and Loparo K A 2008a Nonlinear dynamical analysis of the neonatal EEG time series: the
relationship between neurodevelopment and complexity Clin. Neurophysiol. 119 822–36
Janjarasjitt S, Scher M S and Loparo K A 2008b Nonlinear dynamical analysis of the neonatal EEG time series: the
relationship between sleep state and complexity Clin. Neurophysiol. 119 1812–23
Jenni O G, Borbély A A and Achermann P 2004 Development of the nocturnal sleep electroencephalogram in human
infants Am. J. Physiol. Regul. Integr. Comp. Physiol. 286 528–38
Johnston M V 1995 Neurotransmitters and vulnerability of the developing brain Brain Dev. 17 301–6
Johnston M V and Silverstein F S 2004 Development of Neurotransmitters Fetal and Neonatal Physiology vol 2
3rd edn ed R A Polin, W W Fox and S H Abman (Philadelphia, PA: Elsevier) pp 1706–10
Neurodevelopment in newborns 503

Jordan K G 1995 Neurophysiologic monitoring in the neuroscience intensive care unit Neurol. Clin. 13 579–626
Lake D E, Richman J S, Griffin M P and Moorman J R 2002 Sample entropy analysis of neonatal heart rate variability
Am. J. Physiol. Regul. Integr. Comp. Physiol. 283 789–97
Lutzenberger W, Preissl H and Pulvermuller F 1995 Fractal dimension of EEG time series and underlying brain
process Biol. Cybern. 73 477–82
Meyer-Lindenberg A 1996 The evolution of complexity in human brain development: an EEG study
Electroencephalogr. Clin. Neurophysiol. 99 405–11
Mizrahi E M, Hrachovy R A and Kellaway P 2004 Elements of the normal neonatal electroencephalogram Atlas of
Neonatal Electroencephalography (Philadelphia, PA: Lippincot Willians & Wilkins) p 56
Nunes M L, Costa J C D and Moura-Ribeiro M V L 1997 Polysomngraphic quantification of bioelectrical maturation
in preterm and fullterm newborns at matched conceptional ages Electroencephalogr. Clin. Neurophysiol.
102 186–91
Pace-Schott E F and Hobson J A 2002 The neurobiology of sleep: genetics, cellular physiology and subcortical
networks Nature Rev. Neurosci. 3 591–605
Papadelis C, Kourtidou-Papadeli C, Bamidis P D, Maglaveras N and Pappas K 2007 The effect of hypobaric hypoxia
on multichannel EEG signal complexity Clin. Neurophysiol. 118 31–52
Paul K, Krajča V, Roth Z, Melichar J and Petránek S 2003 Comparison of quantitative EEG characteristics of quiet
and active sleep in newborns Sleep Med. 4 543–52
Paul K, Krajča V, Roth Z, Melichar J and Petránek S 2006 Quantitative topographic differentiation of the neonatal
EEG Clin. Neurophysiol. 117 2050–8
Peirano P, Algarı́n C and Uauy R 2003 Sleep-wake states and their regulatory mechanisms throughout early human
development J. Pediatr. 143 S70–9
Pereda E, de La Cruz D M, Mañas S, Garrido J M, López S and González J 2006 Topography of EEG complexity in
human neonates: effect of the postmenstrual age and the sleep state Neurosci. Lett. 394 152–7
Pincus S M 1991 Approximate entropy as a measure of system complexity Proc. Natl Acad. Sci. USA 88 3017–26
Pincus S M 1995 Approximate entropy (ApEn) as a complexity measure Chaos 5 110–7
Pincus S M and Goldberger A L 1994 Physiological time-series analysis: what does regularity quantify? Am. J.
Physiol. Heart Circ. Physiol. 266 H1643–56
Prior P F and Maynard D 1986 Technical and practical aspects of monitoring Monitoring Cerebral Function (New
York: Elsevier) p 127
Pritchard W S and Duke D W 1995 Measuring chaos in the brain: a tutorial review of EEG dimension estimation
Brain Cogn. 27 353–97
Ramanand P, Nampoori V P and Sreenivasan R 2004 Complexity quantification of dense array EEG using sample
entropy analysis J. Integr. Neurosci. 3 343–58
Rapp P E, Cellucci C J, Korslund K E, Watanabe T A and Jiménez-Montaño M A 2001 Effective normalization of
complexity measurements for epoch length and sampling frequency Phys. Rev. E 64 016209
Richman J S, Lake D E and Moorman J R 2004 Sample entropy Methods Enzymol. 384 172–84
Richman J S and Moorman J R 2000 Physiological time-series analysis using approximate entropy and sample entropy
Am. J. Physiol. Heart. Circ. Physiol. 278 H2039–49
Röschke J and Aldenhoff J 1991 The dimensionality of human’s electroencephalogram during sleep Biol.
Cybern. 64 307–13
Scher M S 1996 Normal electrographic-polysomnographic patterns in preterm and full-term infants Semin. Pediatr.
Neurol. 3 2–12
Scher M S 1997a Neurophysiological assessment of brain function and maturation: I. A measure of brain adaptation
in high risk infants Pediatr. Neurol. 16 191–8
Scher M S 1997b Neurophysiological assessment of brain function and maturation: II. A measure of brain dysmaturity
in healthy preterm neonates Pediatr. Neurol. 16 287–95
Scher M S 2008 Ontogeny of EEG-sleep from neonatal through infancy periods Sleep Med. 9 615–36
Scher M S, Jones B L, Steppe D A, Cork D L, Seltman H J and Banks D L 2003 Functional brain maturation in
neonates as measured by EEG-sleep analyses Clin. Neurophysiol. 114 875–82
Scher M S, Waisanen H, Loparo K and Johnson M W 2005 Prediction of neonatal state and maturational change
using dimensional analysis J. Clin. Neurophysiol. 22 159–65
Shen Y, Olbrich E, Achermann P and Meier P F 2003 Dimensional complexity and spectral properties of the human
sleep EEG Clin. Neurophysiol. 114 199–209
Stam C J 2005 Nonlinear dynamical analysis of EEG and MEG: review of an emerging field Clin. Neurophysiol.
116 2266–301
Sterman M B, Harper R M, Havens B, Hoppenbrouwers T, McGinty D J and Hodgman J E 1977 Quantitative analysis
of infant EEG development during quiet sleep Electroencephalogr. Clin. Neurophysiol. 43 371–85
504 D Zhang et al

Swartjes J M, van Geijn H P, Mantel R, van Woerden E E and Schoemaker H C 1990 Coincidence of behavioral state
parameters in the human fetus at three gestational ages Early Hum. Dev. 23 75–83
Theiler J, Eubank S, Longtin A, Galdrikian B and Farmer J D 1992a Testing for nonlinearity in time series: the
method of surrogate data Physica D 58 77–94
Theiler J, Galdrikian B, Longtin A, Eubank S and Farmer J D 1992b Using surrogate data to detect nonlinearity in
time series Nonlinear Modeling and Forecasting, SFI Studies in the Sciences of Complexity, proceedings vol 12
ed E S Casdagli (Redwood City, CA: Addison-Wesley) pp 163–88
Vecchierini M F, d’Allest A M and Verpillat P 2003 EEG patterns in 10 extreme premature neonates with normal
neurological outcome: qualitative and quantitative data Brain Dev. 25 330–7
Volpe J J 2001 Neuronicproliferation, migration, organization, and myelination Neurology of the Newborn 4th ed
(Philadelphia, PA: Saunders) pp 45–144
Yum M K, Jung K Y, Kang H C, Kim H D, Shon Y M, Kang J K, Lee I K, Park K J and Kwon O Y 2008 Effect of a
ketogenic diet on EEG: analysis of sample entropy Seizure 17 561–6