31st Annual International Conference of the IEEE EMBS

Minneapolis, Minnesota, USA, September 2-6, 2009

Detecting Complexity Abnormalities in Dyslexia Measuring

Approximate Entropy of Electroencephalographic Signals
Ioannis I. Andreadis, Giorgos A. Giannakakis, Charalabos Papageorgiou, and Konstantina S. Nikita,
Senior Member, IEEE

Abstract— Dyslexia constitutes a specific reading disability, a generators which begin to act together in a coherent way
condition characterized by severe difficulty in the mastery of producing the event-related potentials (ERP). This can be
reading despite normal intelligence or adequate education. thought as the transition of a system from a general disorder
Electroencephalogram (EEG) signal may be able to play an to a state of increased order.
important role in the diagnosis of dyslexia. The Approximate
In addition, neuronal systems have been shown to exhibit
Entropy (ApEn) is a recently formulated statistical parameter
used to quantify the regularity of a time series data of some kind of nonlinear or chaotic behavior, which makes it
physiological signals. In this paper, we initially estimated the reasonable to apply methods from the theory of non-linear
ApEn values in signals recorded from controls subjects and dynamics, such as the entropy [5], to the EEG signal.
dyslectic children. These values were firstly used for the Entropy has been applied in many fields of science like
statistical analysis of the two groups and secondly as feature information theory, signal analysis and computer science.
input in a classification scheme. We also used the cross-ApEn
High entropy values indicate high level of disorder of a
methodology to get a measure of the asynchrony of the signals
recorded from different electrodes. This preliminary study system, whereas low values describe a more ordered system
provides promising results towards correct identification of capable to produce some work.
dyslexic cases, analyzing the corresponding EEG signals. Approximate entropy is a statistical parameter proposed
by Pincus to quantify the regularity of a time series data of
I. INTRODUCTION physiological signals [6]. It has already been used in many

T HE human brain function is determined by activation

and interaction mechanisms of the millions of neurons
from which it is constituted. Their oscillatory activity is
applications such as analysis of heart rate variability [7]-
[10], detection of epilepsy [11]-[12] and analysis of the
endocrine hormone release pulsatility [13]. Higher ApEn
increasingly thought to get synchronized during means wide disorder while lower ApEn means regularity.
physiological or pathological brain states, at stimulation or ApEn method can be used only for the evaluation of the
during the performance of certain tasks (e.g. sleep-wake regularity of a single time series. However, in many
states, increased attention tasks, optical stimulation, epileptic applications, a measure of similarity between different time
seizures, etc.) [1]. Dyslexia constitutes a specific reading series is needed. To this end, Cross Approximate Entropy
disability, a condition characterized by severe difficulty in (cross-ApEn) can be used, providing a measure of the
the mastery of reading despite normal intelligence or asynchrony of a pair of signals [14]. The method has already
adequate education [2]. Electrophysiological studies have been applied in the field of endocrinology [15]-[16]. As in
shown that there are physiological deficits in dyslectic the case of ApEn, higher cross-ApEn values imply greater
subjects [3]-[4], which may affect cognitive functions of the degree of asynchrony between the two time-series and
brain such as selective attention, working memory, audio or reversely.
language process. In the present work, ApEn has been used in order to decide
It is questionable whether the electroencephalogram whether an EEG/ERP signal corresponds or not to a
signal can be used for the purpose of the dyslectic cases dyslectic case. Furthermore, we applied the cross-ApEn
detection as dyslexia is probably a condition related to the method to quantify the degree of asynchrony between EEG
brain’s electrical activity. In order to examine the signals of the same subject, recorded at different electrodes,
electrophysiological background of dyslectic condition, we in order to perform statistical analysis of the two groups
investigate the regularity and the degree of order of the EEG (controls-dyslectics).
signal. If a stimulus takes place, brain’s response activates
II. MATERIALS AND METHODS
Manuscript received June 18, 2009.
I. I. Andreadis, G. A. Giannakakis and K. S. Nikita are with the A. Selection of patients and controls
Biomedical Simulations and Imaging Laboratory, National Technical We studied 38 patients (twenty six boys and twelve girls
University of Athens, Greece (corresponding author: I. Andreadis, phone:
30-210-7722968; e-mails: [email protected], [email protected], with mean age and standard deviation 11.47±2.12 years)
[email protected]). fulfilling the criteria of dyslexia as described in the 10th
C. Papageorgiou is with the University of Athens, Department of edition of the International Classification of Diseases (ICD-
Psychiatry, Eginition Hospital, Athens, Greece and University Mental
Health Research Institute (UMHRI), Athens, Greece (e-mail: 10). The patients were recruited from the Department of
[email protected] ).

978-1-4244-3296-7/09/$25.00 ©2009 IEEE 6292

Psychiatry of the Eginition Hospital in Athens, where the d [ X (i ), X ( j )] = max1≤ k ≤ m { x (i + k − 1) − x ( j + k − 1) } (1)
EEG was recorded.
The control group consisted of 19 children, including 7
3. For each X(i) we estimate the number Νim(r) of vectors
boys and 12 girls with mean age and standard deviation
such that
12.21±2.25 years. The statistical t-test for the mean ages of
d [ X (i ), X ( j )] ≤ r (2)
the two groups did not show any significant differences.
The local ethics committee approved the study. All control with r representing the noise filter level.
subjects and all caregivers of the demented patients gave Then, we estimate the parameters Cim as,
m
their informed consent for participation in the current study. m Ni ( r )
An EEG was recorded from all patients and controls. All of Ci ( r ) = (3)
N − m +1
them had undergone assessment of educational attainment
including reading, comprehension, spelling and arithmetic
4. We define φm(r) as the mean value of the parameters
ability. Participants with hearing difficulties, history of head m
Ci :
injury, attention deficit disorders or neurological disease N = m +1
were not included in the study.
∑ ln C i
m
(r )
(4)
φ (r ) =
m
B. Data recording and acquisition i =1

The EEGs and ERPs were recorded at the 15 scalp loci of N − m +1

the international 10–20 system (channels Fp1, F3, C5, C3,
Fp2, F4, C6, C4, O1, O2, P4, P3, Pz, Cz, Fz), with all 5. ApEn(m,r,N) is calculated using φm(r) and φm+1(r) as
electrodes referenced to the chin. An electrode placed on the follows:
ApEn ( m, r , N ) = φ ( r ) − φ
m m +1
subject's forehead served as ground. With the subjects in a (r ) (5)
relaxed state, awake and with closed eyes so as to minimize
eye movements, data were recorded from each subject. Eye Briefly, ApEn measures the logarithmic likelihood that
movements were recorded through electro-oculogram (EOG) runs of patterns that are close (within r) for m contiguous
and recordings with EOG higher than 75 μV were rejected. observations remain close (within the same tolerance width)
EEG data were first sampled at frequency of 1 kHz so that on subsequent incremental comparisons. Comparisons
for signals in the frequency range 0–35 Hz the Shannon between time series segments can only be made with the
theorem is over satisfied. For each trial of the experiment, same values of m and r.
rest EEG signal was recorded for 500 msec before the
stimulus and ERP was recorded for 1000 msec after the
stimulus onset. A single sound tone of either high (3000 Hz) D. Cross Approximate Entropy
or low frequency (500 Hz) was presented to the subjects Cross Approximate Entropy (Cross-ApEn) is a two
through earphones, followed by numbers which had to be parameter family of statistics introduced as a measure of
memorized. Noise was considered to be a random process asynchrony between two paired time series [14]. Cross-
with zero man value. For this reason, the signal's SNR (ERP ApEn can be employed to compare sequences from two
in relation to rest EEG) was improved by averaging across distinct yet intertwined variables in a network, herein
the 52 trials of the experiment. applied to the EEG signals as recorded from different
electrodes.
C. Approximate Entropy
Given two N-length paired time series u and v, cross-
ApEn was introduced as a quantification of regularity in ApEn(m,r,N) measures, within tolerance r, the (conditional)
sequences and time series data, initially motivated by regularity or frequency of v-patterns similar to a given u-
applications to relatively short, noisy data sets [6]. pattern of window length m. Larger cross-ApEn values
Approximate entropy returns a non-negative single value indicate greater signal asynchrony. The precise mathematical
reflecting the predictability of future values in a time series definition is thematically similar to that for ApEn:
on the basis of previous values, with larger values 1. We form the vector sequences x(i)=(u(i), u(i+1),...,
corresponding to more complexity or irregularity in the data. u(i+m-1)) and y(j)=(v(j), v(j+1),..., v(j+m-1)) which
Given N data points from a time series {x(n)}={x(1), x(2),., represent m consecutive values, commencing with the ith and
x(N)}, the ApEn value is calculated through the following jth point, respectively.
steps: 2. We define the distance between x(i) and y(j), defined as
1. We form the vector sequences X(1), X(2),…, X(N-
m+1), defined by X(i)=(x(i), x(i+1)…, x(i+m-1)), which
d [ x (i ), y ( j ) ] = max1≤ k ≤ m { x (i + k − 1) − y ( j + k − 1) } (6)
represent m consecutive values, commencing with the ith
point.
2. We estimate the distance between X(i) and X(j), 3. For each x(i) we compute the number Νim(r) of vectors
defined by such that

6293
 d [ x (i ), y ( j )] ≤ r (7) Finally, accuracy is a related measure that quantifies the
number of subjects (dyslectics and control subjects)
accurately classified. The results are presented in Table 2.
We compute then the parameters Cim defined as We observe that a satisfactory value of sensitivity is
m
N i (r ) achieved, which is encouraging as it indicates that dyslexia
Ci ( r ) ( v || u ) = (8)
m

N − m +1 may be identified through the analysis of the EEG signal

using the ApEn parameter. The results achieved at the end of
this process are comparable with those reported on previous
4. We define φ m
(r) as the mean value of the parameters C: studies [19]-[20].
N = m +1

∑ ln C i
m
(r) ( v || u )
(9)
TABLE I
(r) ( v || u ) =
THE AVERAGE APEN VALUES (MEAN ± S.D.) OF THE EEG/ERP
φ m i =1
SIGNAL FOR THE CONTROLS AND DYSLECTICS
N − m +1

5. Cross-ApEn(m,r,N) is calculated as: Electrode Controls Dyslectics p-value

m m+1
crossApEn(m, r , N )(v || u ) = φ ( r )(v || u ) − φ ( r )(v || u ) (10)
Fp1 0,095 ± 0.056 0,104 ± 0.097 0,6953
F3 0,112 ± 0.046 0,105 ± 0.034 0,4806
Typically, Cross-ApEn is applied to normalized {u*(i),
C5 0,141 ± 0.054 0,114 ± 0.027 0.014*
v*(i)} time series, where u*(i) = [u(i)−mean(u)]/s.d.(u) and C3 0,128 ± 0.043 0,122 ± 0.039 0,6026
v*(i) = [v(i)−mean(v)]/s.d.(v) (s.d. is the standard deviation). Fp2 0,096 ± 0.052 0,096 ± 0.045 0,9463
We should also note here that there is a direction F4 0,123 ± 0.056 0,113 ± 0.065 0,5713
dependence, and as a result φ m
(r)(v || u ) will not be C6
C4
0,127 ± 0.048
0,123 ± 0.054
0,114 ± 0.036
0,114 ± 0.038
0,2473
0,4808
generally equal to φ m
(r)(u || v ) . O1 0,140 ± 0.082 0,112 ± 0.066 0,1688
O2 0,164 ± 0.079 0,123 ± 0.059 0.0299*
P4 0,148 ± 0.056 0,121 ± 0.042 0.0453*
P3 0,141 ± 0.060 0,126 ± 0.039 0,2557
III. RESULTS Pz 0,143 ± 0.056 0,128 ± 0.043 0,2586
ApEn and Cross-ApEn were estimated for 15 channels Cz 0,114 ± 0.043 0,106 ± 0.025 0,359
Fz 0,111 ± 0.052 0,094 ± 0.032 0,1377
(Fp1, F3, C5, C3, Fp2, F4, C6, C4, O1, O2, P4, P3, Pz, Cz,
Fz and T6) with m=2 and r=0.2 of each waveform data.
Each signal was preprocessed by subtracting its mean value
and dividing by its standard deviation (s.d.) providing TABLE II
normalized waveforms. The results were averaged between SYSTEMS’ PERFORMANCE MEASURES
groups (controls, dyslectics) within the EEG/ERP Performance Measures
waveforms (1500 samples). The average ApEn values (mean Sensitivity Specificity Accuracy
± s.d.) for the controls and dyslexics are shown in Table 1. It SVM classification
89.47% 57.89% 78.95%
scheme
can be observed that there are three electrodes (C5, O2, P4)
that appear to discriminate significantly the two groups.
Then, we evaluated the ability of ApEn to discriminate
At the next step, Cross-ApEn values were calculated for
dyslectics from control subjects at the statistically significant
each pair of electrodes. As the Cross-ApEnij is different than
electrodes. We constructed a classifier based on the Support
Cross-ApEnji of i,j electrodes, a total number of 15*(15-
Vector Machines (SVM) methodology [17]. We trained the
1)=210 Cross-ApEn values were calculated for each subject.
classifier with a dataset constituted from the ApEn values of
It was tested that data follow normal distribution. To this
the significant electrodes for all 57 subjects, control and
end, our goal was to investigate whether there are
dyslectics. To this end, the library LIBSVM [18] was used in
statistically significant pairs of electrodes. These results may
our experiments. Due to the limited size of the available
give us important information on possible connections
dataset, we used the Leave One Out technique which is
between different compartments in the brain. The cross-
mainly used in such situations because of its good
ApEn values can also be used as input in a new classification
generalization ability.
scheme for discrimination between dyslectic and controls.
To evaluate the classifier, we measured both the
Due to the great number of statistically significant pairs,
sensitivity and specificity achieved. Sensitivity is the
further analysis is needed in order to combine the values of
percentage of dyslectics correctly recognized by the
the two methodologies in a unified classification scheme.
classifier, whereas specificity represents the percentage of
Fig. 1 indicates the p-values of equality test between
control subjects classified correctly by the classifier, using
controls and dyslectics groups (two tailed t-test).
the ApEn values from the statistically significant electrodes.

6294
 CrossApEn statistics which aims at the improvement of the classification results
1
Fp1 and at a complete framework for the diagnosis and
F3 0.9 identification of dyslexia, based on the EEG signal’s
C5 analysis.
0.8
C3
Fp2 0.7 REFERENCES
F4 [1] O. A. Rosso, S. Blanco, J. Yordanova, V. Kolev, A. Figliola, M.
0.6
C6 Schurmann, and E. Basar, “Wavelet entropy: a new tool for analysis
of short duration brain electrical signals,” J. Neurosci. Methods, vol.
C4 0.5
105 (1),pp. 65-75, Jan. 2001.
O1 [2] E. Grikorenko, “Developmental Dyslexia: An Update of Genes,
0.4
O2 Brains and Enviroment”, J. Child. Psychol. Psychiat., vol. 42, no. 1,
pp. 91 – 125, 2001.
P4 0.3
[3] M. L. Bonte, L. Blomert, “Developmental dyslexia: ERP correlates of
P3 anomalous phonological processing during spoken word recognition,”
0.2
Pz Brain Res Cogn Brain Res, vol. 21, issue 3, pp. 360-376, 2004.
[4] V. Csepe, D. Szucs, F. Honbolygo, “Number-word reading as
Cz 0.1
challenging task in dyslexia? An ERP study,” International Journal of
Fz Psychophysiology, vol. 51, issue 1, pp. 69-83, December 2003.
0
Fp1 F3 C5 C3 Fp2 F4 C6 C4 O1 O2 P4 P3 Pz Cz Fz [5] B. Berthold. Entropy. Best Practice & Research Clinical
Fig.1. Differentiation results between controls and dyslectics for all pairs Anaesthesiology, Vol. 20, No. 1, pp. 101–109, 2006.
of electrodes (colorbar indicates p-values) [6] S. Pincus “Approximate entropy as a measure of system complexity”.
Proc. Nati. Acad. Sci. USA, Vol. 88, pp. 2297-2301, March 1991.
[7] S. M. Pincus, T. R. Cummins, and G. G. Haddad,“Heart rate control in
IV. DISCUSSION normal and aborted SIDS infants,” Amer. J. Physiol., vol. 264, no. 3,
pp. R638–R646, 1993.
In this study, the application of nonlinear analysis [8] S. M. Pincus and A. L. Goldberger, “Physiological time series
methods ApEn and Cross-ApEn on EEG and ERP signals analysis: what does regularity quantify?,” Amer. J. Physiol., vol. 266,
was investigated. Our study involved 19 controls and 38 pp. H1643– Hl656, 1994.
[9] J. S. Richman and J. R. Moorman,“Physiological time-series analysis
dyslectic children whose EEG and ERP signals were using approximate entropy and sample entropy,” Amer. J. Physiol.,
recorded at 15 electrodes according to 10-20 system. It was vol. 278, pp. H2039–H2049, 2000.
found that there are some specific electrodes (C5, O2, P4) [10] W. S. Kim, Y. Z. Yoon, J. H. Bae, and K. S. Soh, “Nonlinear
characteristics of heart rate time series: Influence of three recumbent
that could discriminate statistically the average ApEn values positions in patients with mild or severe coronary artery disease,”
of the two groups with dyslectics having lower mean values Physiol. Meas., vol. 26, pp. 517–529, 2005.
as compared to controls. [11] V. Srinivasan, C. Eswaran, N. Sriraam. Approximate Entropy-Based
Epileptic EEG Detection Using Artificial Neural Networks.
An SVM classification scheme has been implemented in Biomedicine, Vol. 11, No. 3, May 2007.
order to classify the cases, using the ApEn values as [12] Ocak, H., Automatic detection of epileptic seizures in EEG using
estimated at the statistically significant electrodes. The discrete wavelet transform and approximate entropy, Expert Systems
results are encouraging, indicating that the analysis of the with Applications, 2008, doi:10.1016/j.eswa.2007.12.065.
[13] S.M. Pincus,“Older males secrete luteinizing hormone and
EEG signal may reveal valuable information for the testosterone more irregularly and joint more asynchronously, than
identification of dyslexia. The sensitivity achieved is younger males,” Proc. Natl. Acad. Sci. U.S.A., vol. 93, pp. 14100–
89.47%, whereas specificity is equal to 57.89%. 14105, 1996.
[14] S.M. Pincus, B.H. Singer , “Randomness and degrees of irregularity”.
Using Cross-ApEn which is a measure of irregularity Proc Natl Acad Sciences USA. 93:2083–2088, 1996.
between two time series, it was found that electrodes C5, Cz, [15] F. Roelfsema , S. Pincus , J. Veldhuis , “Patients with Cushing’s
C4, P4, O2 show differences between controls and Disease Secrete Adrenocorticotropin and Cortisol Jointly More
Asynchronously than Healthy Subjects.” Journal of Clinical
dyslectics. Although Cross-ApEn appears to provide higher Endocrinology and Metabolism Printed in U.S.A., Vol 83, No2, 1998.
discrimination ability, there is still no clear pattern and [16] S.M. Pincus, T. Mulligan , A. Iranmanesh ,S. Gheorghiu , M.
further analysis is needed to interpret the results. Godschalk, J.D.Veldhuis, “Older males secrete luteinizing hormone
(LH) and testosterone more irregularly, and jointly more
asynchronously, than younger males”. Proc Natl Acad Sci USA.
V. CONCLUSIONS 93:14100 –14105.1996.
[17] C. J. C. Burges, “A Tutorial on Support Vector Machines for Pattern
The application of methods able to measure the different Recognition,” Knowledge Discovery Data Mining, 2:1-43, 1998.
degree of complexity of the signals recorded appears to [18] Chih-Chung Chang and Chih-Jen Lin, LIBSVM: a library for support
provide useful tools for the discrimination of dyslexic and vector machines, 2001. Software available at
http://www.csie.ntu.edu.tw/~cjlin/libsvm.
control subjects.
[19] D.L. Molfese, “Predicting Dyslexia at 8 Years of Age Using Neonatal
Although ApEn and Cross ApEn cannot yet be applied as Brain Responses”. Brain and Language 72,pp. 238–245, 2000.
diagnostic tools, they can provide useful information about [20] F.H. Duffy, M.B. Denckla, P.H. Bartels, G. Sandini, L.S. Kiessling,
signals’ properties that remains hidden when classical and “Dyslexia: Automated diagnosis by computerized classification of
brain electrical activity”. Annals of Neurology 7, pp. 421-428, 1980.
conventional statistical methods are used. These preliminary
results could be proven quite helpful in the understanding of
deeper mechanisms of neurological/ neuropsychological
disorders. Our future goals include methodology refinement

6295