Measures of Brain Connectivity through

Permutation Entropy in Epileptic Disorders

Domenico Labate , Giuseppina Inuso, Gianluigi Occhiuto,

Fabio La Foresta, and Francesco C. Morabito

MECMAT - Mediterranea University of Reggio Calabria,

via Graziella Feo di Vito, 89122 Reggio Calabria, Italy
{domenico.labate,giuseppina.inuso,gianluigi.occhiuto,fabio.laforesta,
morabito}@unirc.it

Abstract. Most of the scientist assume that epileptic seizures are trig-
gered by an abnormal electrical activity of groups of neural populations
that yields to dynamic changes in the properties of Electroencephalog-
raphy (EEG) signals. To understand the pathogenesis of the epileptic
seizures, it is useful detect them by using a tool able to identify the dy-
namic changes in EEG recordings. In the last years, many measures in
the complex network theory have been developed. The aim of this paper
is the use of Permutation Entropy (PE) with the addition of a thresh-
old method to create links between the different electrodes placed over
the scalp, in order to simulate the network phenomena that occur in the
brain. This technique was tested over two EEG recordings: a healthy
subject and an epileptic subject affected by absence seizures.

Keywords: Permutation Entropy, EEG, Complex Network.

1 Introduction

EEG is a measure of neurophysiological brain electrical activity recorded through

electrodes placed over the scalp. Neurons, in fact, communicate by means of elec-
trical impulses and generate a bio-electromagnetic field that propagates through
the tissues of the brain, skull and scalp [1]. EEG signal is produced by the ex-
tracellular current flow generated by the sum of the activities of a large number
of neurons, including cortical pyramidal neurons, arranged in correspondence of
the cortex underlying the electrode [2][3].
Epilepsy is one of the most common neurological disorders. Epileptic seizures
are triggered by neurons that exhibit an abnormal electrical activity [4], as if a
critical area is trying to recruit other cerebral areas until the brain can’t longer
bear this hyper-synchronization and triggers the seizure in order to repress this
abnormal state. Then, epileptic phenomena seem a network rather than local
phenomena [5].

Corresponding author.

B. Apolloni et al. (Eds.): Neural Nets and Surroundings, SIST 19, pp. 59–67.
DOI: 10.1007/978-3-642-35467-0_7 c Springer-Verlag Berlin Heidelberg 2013
60 D. Labate et al.

In order to understand the development of epileptic seizures, it is important

to understand how this abnormal synchronization influences the randomness
of the EEG. The possible solution to this issue can be entropy. In particular,
Permutation Entropy is chosen [6][7][8] because, being a symbolic procedure,
pre-normalization step can be discarded and because it was applied to epileptic
EEG analysis obtaining encouraging results.
In this paper, a study of links between different cerebral areas for two types
of subjects is presented: an healthy subject (called ”control”) and an epileptic
subject (called ”patient”). The aim of this paper is to use PE in order to under-
line different network phenomena in epileptic subject during EEG recordings.
The paper is organized as follows: section 2 will address Permutation Entropy;
section 3 will explain the applied method used to analyze the EEG recordings;
section 4 will comment the obtained results and, ultimately, section 5 will draw
the conclusions.

2 Permutation Entropy
PE has been introduced [9] as a robust method to extract information from a
time series. The time series is analyzed from a pure ordinal point of view. PE
is based by counting of the ordinal patterns (called ”motifs”), that describe the
up-and-down in a dynamical signal. Since just ordinal patterns are considered,
the amplitude of the signal has no relevance, thus yielding a structural robust-
ness to noise. In the study of biological signals, like EEG, this also implies the
independence of the choice of the reference electrode [10]. As a consequence,
there is no need of any normalization in the pre-processing step.
PE combines the concept of Shannon Entropy to the ordinal pattern analysis
by analyzing the frequencies of ordinal patterns within a time series. If most of
the ordinal patterns are the same, we have a regular time series; on the contrary,
the presence of different patterns with similar frequency is indicative of high
complexity.
PE is dependent on two-parameters, an embedding dimension, m, and a time-
lag, l ; m is the number of samples belonging to the segment, while l represents
the distance between the samples points spanned by each section of the motif.
The variation of the time-lag basically implies working on different scales of the
time series. In this work, the embedding parameter and time-lag are fixed to 3
and 1 respectively.
The calculation of PE is simple and an example is depicted in Figure 1. First
of all, the EEG signal is fragmented into a sequence of motifs; then, each motif is
identified as belonging to one of the possible types according to their shape; it is
counted the number of motifs to each of the six categories in order to obtain the
probability of occurrence of each motif in the signal, and, finally, is calculated
the PE of the resulting normalized probability distribution of the motifs using
the Shannon uncertainty formula:

pi ∗ ln(pi )
PE = − (1)
ln(number of motifs)
 Measures of Brain Connectivity through Permutation Entropy 61

1000

(a) 0

−1000
1300 1350 1400 1450 1500 1550
Time (sec)

0.9
(b)
0.8

0.7

130 135 140 145 150 155 160

Windows number

0.4 0.4
(c)
0.2 0.2

0 0
#1 #2 #3 #4 #5 #6 #1 #2 #3 #4 #5 #6

(d)

#1 #2 #3 #4 #5 #6

Fig. 1. (a) EEG recording. (b) PE time evolution. (c) Motifs distribution (m=3) of the
two highlighted segments. (d) The six possible motifs.

The normalized entropy is maximal (PE=1) if there is an equal distribution

of motifs between each of the six patterns; on the contrary, when the signal is
dominated by slow waves, there will be a greater occurrence of a few motifs, and
the entropy decreases. The PE of a single motif is zero [6].
PE shows a different dependence with frequency with respect to the selected
time-lag. To understand how, a sequence of pure sine waves time series is simu-
lated and then the PE for each of them is performed. The results are reported
in Figure 2. It is clear that, with l =1, the PE is monotonically decreasing with
decreasing frequency. The frequency dependence of PE with l =2 is markedly dif-
ferent and more difficult to interpret. In Figure 2, it is possible to observe that
the PE function has a minimum for a frequency that is a fraction of the sampling
frequency, Fs . Some authors decided to define a suitably averaged composite PE
index that incorporates both the cases, this way exploiting the effect [6].
The effects of noise on PE have been originally discussed in [9]. If a time series
is perturbed by noise, the PE value strongly increases. Thus, if the added white
noise is of sufficient power, the value of PE tends to its maximum at very low
frequencies. Figure 3 shows the effect of adding white noise to the simulated
sine-wave time series. It is clear that, the added noise leaves undisturbed the PE
curve after a frequency threshold. This threshold level depends on the power of
the noise [11].
62 D. Labate et al.

Frequency dependence
1

0.9

0.8
PE value

0.7

0.6

0.5
l=1
l=2
0.4
0 5 10 15 20 25 30 35 40
Hz

Fig. 2. Frequency dependence of PE. The two distributions for l=1 and l=2, m=3, and
Fs =128 Hz are reported.

Effect of Noise
1

0.9

0.8
PE value

0.7

0.6

0.5 without noise

noise (+ 10%)
noise (+ 5%)
0.4
0 5 10 15 20 25 30 35 40
Hz

Fig. 3. The effect of adding white noise on PE calculation (m=3, l=1 and Fs =128 Hz)

3 Method
The EEG recording is processed by a sliding overlapping windows of 10 seconds
with 1 second of overlap, i.e. each window includes 2560 time samples. The PE
is performed, channel by channel, for each temporal window using m=3 and
l =1. Once PE values are obtained, they are used in order to built a picture in
which a scalp is depicted with all its 19 electrodes. If two electrodes have both
a PE value greater than a fixed threshold, then the two electrodes are linked.
This procedure is repeated for all the time windows obtained in the previous
step. At the end of the analysis a movie with all the obtained pictures is built.
 Measures of Brain Connectivity through Permutation Entropy 63

The purpose of the movie is to better clarify the link time evolution between the
electrodes during the different stages of the recording. To better compare the
obtained results, the two EEG recordings are cut in order to have the same size.
This method is implemented in Matlab code.

4 Results

4.1 Data Description

The analyzed dataset is composed of two EEGs of nineteen channels recorded

simultaneously, the sampling rate is 256 Hz. During the acquisition, Notch filter
was set to ON and the EEGs was band-pass filtered (0.5 - 70 Hz). The analyzed
EEGs are relative to a healthy subject (control) and a patient affected by ab-
sence epilepsy (patient). During EEG recording, the patient has undergone four
seizures.

4.2 Discussion

The EEG recordings was analyzed as described in section 3. PE was performed

for each channel and window by window.
In Figure 4 are reported PE values for the two EEG recordings. In order to
consider all the 19 channels, the average of PE values is reported. It is clear
as for the patient there are four peaks down due to the four absence seizures.
Looking at a typical PE temporal trend, it can probably hypothesized that the
sudden drop during the ictal stage is due to a mechanism of the brain to reset the
high randomness that develops during the inter-ictal stage [5][8]. For the control,
instead, the PE values are mostly the same for the entire duration of the EEG

Permutation Entropy − Patient

0.9
PE value

0.8

0.7

0.6
10 20 30 40 50 60 70 80

Permutation Entropy − Control

0.9
PE value

0.8

0.7

0.6
10 20 30 40 50 60 70 80
Window Number

Fig. 4. (on the top) PE value for the patient; (on the bottom) PE value for the control
64 D. Labate et al.

Fig. 5. The comparison of electrodes linked between the patient and the control for
the window number 27 (pre-ictal stage)

Fig. 6. The comparison of electrodes linked between the patient and the control for
the window number 30 (ictal stage)
 Measures of Brain Connectivity through Permutation Entropy 65

recording. Then PE can be used as a useful tool to identify the dynamic changes
in the EEG.
Figure 5 and Figure 6 show how PE values change during the EEG recording.
In these pictures are reported the scalps for patient and control both. Figure
5 is relative to a pre-ictal stage for the patient; it is clear that some electrodes
are linked. Figure 6, instead, is relative to a ictal stage for the patient. In this
case no electrodes are linked due to fact that all the PE values are lower than
the threshold value; this change is due to absence seizure occurring in Figure
6. Observing the scalp relative to the control, there are no changes between the
two figure because in both the PE values are always greater than the threshold
value. Moreover, analyzing Figure 4, Figure 5 and Figure 6, it is clear that the
PE values relative to the control are greater than the PE values relative to the
patient; this is due to a most important cerebral activity in the healthy subject.
Table 1 and Table 2 report PE values for each electrode and for several EEG
areas for three different thresholds: 0.85, 0.82 and 0.79. This is due to the fact
that there isn’t a fixed threshold value able to distinguish a patient by a control
or a pre-ictal stage by an ictal stage. However the behaviour is similar for all the
thresholds reported.
Observing Table 1 and Table 2, it is clear that the electrodes belonging to
the frontal/temporal area have PE values greater than the other electrodes
throughout the entire EEG recording and not only during the ictal stage. On the

Table 1. In the first column there are the electrodes; in the second, third and fourth
column, there are the number of times in which the Patient’s PE value exceeds the
threshold value set to 0.85, 0.82 and 0.79 respectively

Electrode th=0.85 th=0.82 th=0.79

Fp1 76/85 77/85 79/85
Fp2 76/85 78/85 80/85
F7 68/85 79/85 80/85
F3 24/85 73/85 78/85
Fz 24/85 71/85 76/85
F4 50/85 74/85 77/85
F8 73/85 80/85 80/85
T3 80/85 81/85 84/85
C3 15/85 38/85 66/85
Cz 5/85 27/85 53/85
C4 5/85 53/85 73/85
T4 79/85 80/85 82/85
T5 12/85 39/85 64/85
P3 7/85 20/85 48/85
Pz 2/85 15/85 38/85
P4 12/85 29/85 53/85
T6 35/85 53/85 71/85
O1 2/85 13/85 37/85
O2 3/85 23/85 46/85
66 D. Labate et al.

Table 2. In the first column there are the EEG areas; in the second, third and fourth
column, there are the number of times in which the Patient’s PE value exceeds the
threshold value set to 0.85, 0.82 and 0.79 respectively

Area th=0.85 th=0.82 th=0.79

Fp1-Fp2 76/85 77/85 79/85
F7-F3 24/85 73/85 78/85
F4-F8 50/85 74/85 77/85
T3-C3 15/85 38/85 66/85
T4-C4 22/85 53/85 73/85
P3-T5 7/85 20/85 48/85
P4-T6 12/85 29/85 53/85
O1-O2 2/85 13/85 36/85

contrary, there are some electrodes belonging to the parietal/occipital area, that
have PE values smaller than the average of PE values. However, the analysis of
a unique EEG recording affected by absence seizures is not enough in order to
suggest a clinical reason for this behaviour.
PE values reported in Table 2 are meaningful because they allow detection of
network phenomena relative to different EEG areas by means of a local measure
like PE.

5 Conclusions

In this paper a EEG analysis based on PE was performed in order to study the
evolution of links between the different electrodes placed over the scalp. PE was
proposed to measure the randomness/synchrony of the brain. It has been shown
the different features between an healthy subject and an epileptic subject, and
how there are no links during an epileptic seizure. Observing PE values, it is clear
how the frontal/temporal area has PE values greater than PE values relative to
other cerebral areas. PE has been a good tool to identify dynamical changes in
epileptic EEG.

References

1. Mammone, N., La Foresta, F., Morabito, F.C.: Automatic Artifact Rejection from
Multichannel Scalp EEG by Wavelet ICA. IEEE Sensors Journal 12(3), 533–542
(2012)
2. Labate, D., La Foresta, F., Inuso, G., Morabito, F.C.: Remarks about Wavelet
Analysis in the EEG Artifacts Detection. In: Frontiers in Artificial Intelligence
and Application: Neural Nets WIRN10, vol. 226, pp. 99–106 (2011)
3. Labate, D., La Foresta, F., Inuso, G., Morabito, F.C.: Multiscale Entropy Analysis
of Artifactual EEG Recordings. In: Frontiers in Artificial Intelligence and Appli-
cation: Neural Nets WIRN11, vol. 234, pp. 170–177 (2011)
 Measures of Brain Connectivity through Permutation Entropy 67

4. Mammone, N., Inuso, G., La Foresta, F., Versaci, M., Morabito, F.C.: Clustering
of entropy topography in epileptic electroencephalography. Neural Computing and
Applications 20, 1–9 (2011)
5. Mammone, N., La Foresta, F., Morabito, F.C.: Discovering Network Phenomena in
the Epileptic Electroencephalography through Permutation Entropy Mapping. In:
Frontiers in Artificial Intelligence and Application: Neural Nets WIRN10, vol. 226,
pp. 260–269 (2011)
6. Olofsen, E., Sleigh, J.W., Dahan, A.: Permutation Entropy of the Electroencephalo-
gram: a Measure of Anaesthetic Drug Effect. British Journal of Anaesthesia 101,
810–821 (2008)
7. Bruzzo, A.A., Gesierich, B., Santi, M., Tassinari, C.A., Birbaumer, N., Rubboli,
G.: Permutation entropy to detect vigilance changes and preictal states from scalp
EEG in epileptic patients. A preliminary study. Neurological Science 29, 3–9 (2008)
8. Li, X., Ouyangb, G., Richards, D.A.: Predictability analysis of absence seizures
with permutation entropy. Epilepsy Research 77, 70–74 (2007)
9. Bandt, C., Pompe, B.: Permutation entropy - a natural complexity measure for
time series. Physical Review Lett. 88, 174102 (2002)
10. Mammone, N., Lay-Ekuakille, A., Morabito, F.C., Massaro, A., Casciaro, S., Tra-
bacca, A.: Analysis of absence seizure EEG via Permutation Entropy spatio-
temporal clustering. In: Proceedings of IEEE International Symposium on Medical
Measurements and Applications (MEMEA 2011), Bari, Italy, May 30-31, pp. 532–
535 (2011)
11. Morabito, F.C., Labate, D., La Foresta, F., Bramanti, A., Morabito, G., Pala-
mara, I.: Multivariate Multi-Scale Permutation Entropy for Complexity Analysis
of Alzheimer’s Disease EEG. Entropy 14(7), 1186–1202 (2012)