Lo

Ca

b

c

a

ARRAA

KEHTFE

1

etanm[poga

s(

h1

Biomedical Signal Processing and Control 47 (2019) 262–275

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

jo ur nal homep age: www.elsev ier .com/ locate /bspc

esion localization algorithm of high-frequency epileptic signal basedn Teager energy operator

hunsheng Wanga, Hui Yia,∗, Wei Wangb,∗, Palaniappan Valliappanc

School of Information Science and Engineering, Central South University, Changsha 410083, ChinaSchool of Electrical Engineering, Guangxi University, Nanning 530004, ChinaSchool of Chemical Engineering and Analytical Science, The University of Manchester, UK

r t i c l e i n f o

rticle history:eceived 15 August 2017eceived in revised form 27 July 2018ccepted 27 August 2018vailable online 5 September 2018

eywords:pileptic EEGigh frequency oscillationeager energy operatoreature extractionpilepsy positioning

a b s t r a c t

High-frequency oscillations reflect the abnormal brain electrical activity in patients with epilepsy. It issignificant to research the relationship between high-frequency oscillation and epilepsy originating areafor the diagnosis and treatment of epilepsy. In view of the identification of epileptic EEG and the location ofepileptic foci, a localization algorithm based on Teager operator is proposed. Firstly, the original epilepticEEG data are preprocessed, the wavelet weighted threshold and the frequency notch method are usedto denoise the original epileptic high frequency oscillation signal. Taking into account the frequencycharacteristics of the high frequency oscillating signal itself, the FIR (Finite Impulse Response) digitalfilter is used to filter the epileptic high-frequency oscillation signal. Secondly, because the high frequencyoscillation rhythm has the characteristics of high frequency, high energy and low amplitude, the Teagerenergy operator and curve length method are used to extract the characteristics. The PSD (Power SpectralDensity) method is applied to qualitative analysis of epileptic lesion location. Finally, the EMD (EmpiricalMode Decomposition) algorithm is used to decompose the high-frequency oscillation signal. Combined

with Teager energy operator, Teager-huang transform is used to analyze the signal by time-frequencyenergy analysis. Quantitative analysis of epileptic lesion location is made by using EMD energy entropymethod for different lead epileptic signals. The algorithm can effectively locate the location of the focus,independent of individual parameters and high degree of automation. It has good clinical applicationprospects.

© 2018 Elsevier Ltd. All rights reserved.

. Introduction

The annually incidence of epilepsy is 50–70 in 100,000 peoplevery year, and the prevalence rate is 0.5% to 2%. It is estimatedhat 1 to 2 percent of the US population suffers from epilepsy [1],nd in a large-scale survey organized by the World Health Orga-ization shows that the prevalence rate is about 8% in China. Thiseans that there are about more than ten million epilepsy patients

2,3], in which there are about 2.5 million intractable epilepsyatients. Epilepsy is already a deadly disease which is only sec-nd to cerebrovascular disease in neurological diseases [4]. It has

reat significance to study various abnormal EEG for the diagnosisnd treatment of epilepsy [5].

∗ Corresponding authors.E-mail addresses: wangcsu@csu.edu.cn (C. Wang), anything@csu.edu.cn (H. Yi),

pecialwangwei@163.com (W. Wang), palaniappan.valliappan@manchester.ac.ukP. Valliappan).

ttps://doi.org/10.1016/j.bspc.2018.08.033746-8094/© 2018 Elsevier Ltd. All rights reserved.

Currently, the research for epilepsy EEG is mostly concentratedin the frequency band below 100 Hz. The study shows that, beforeand after the seizures of epilepsy, a higher frequency band sig-nal which is acquired by intracranial electrode can be found inEEG with a high sampling rate [6], and the signal is known ashigh-frequency oscillation (HFO) rhythm. It is closely related toseizures [7], reflecting the important features of abnormal EEG sig-nals during seizures [8,9]. Among the EEG, the oscillation rhythmwhose frequency range from 80 to 200 Hz is called ripples, canbe regarded as a normal HFO rhythm [10]. The oscillation signalsmore than 200 Hz frequency are usually considered as pathologicalHFO rhythm, which is closely linked to abnormal EEG and physio-logical changes produced during the seizures [11]. The oscillationsignal including the frequency range of 200–600 Hz called “fast rip-ples” (FRS) and frequency range of 1000–2500 Hz called “very highfrequency oscillations” (VHFO) [12].

The major difference between pathological HFO and normalHFO, is the different parts of its occurrence. The normal HFOsare unlikely to appear in dentate gyrus structure [13,14]. In the

C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275 263

gorith

poHpiaUahtocFbto

fpfdwtpfhtchdo[

hsttltSnfteoocsweh

Fig. 1. Al

resence of seizures, accompanied by the production and spreadf HFOs. In seizures, usually with the generation and spread ofFOs. When the motor cortex of the brain is affected, it shows theatient’s local movement symptoms, which is the so-called clin-

cal manifestations [15]. During the period of seizures, HFO have specific distribution, which is the occurrence frequency of HFO.sually increasing from outside to inside in epilepsy originatingrea, especially FRS, the closer to the epileptic foci core area, theigher frequency FRS is recorded [16]. From the perspective of time,here is a diffusion process which the HFOs spread from inside toutside during the seizures period, while in terms of space, theloser to the epileptic foci originating area, the higher frequencyRS will be recorded [17]. The research for VHFO started late, andecause of its very high frequencies and difficult to acquisition, nowhe study of high-frequency oscillation signal mainly concentratedn the ripples and FRS [18].

Many scholars have stated exploratory research on highrequency oscillation analysis algorithms. Zelmann used high sam-ling rate of 2000 Hz, with a 500 Hz low pass filter to obtain therequency characteristics of ripples and fast ripples present in EEGuring an epilepsy episode [19]. This was then combined withavelet transform entropy and energy feature extraction to obtain

he high frequency oscillation rhythm [19]. Smart et al. used high-recision multi-lead (64 leads) EEG to acquired epileptic EEG datarom a plurality of epilepsy patients, and through analysis withigh-pass filter and data segmentation method were able to detecthe characteristic wave of ripples in HFO rhythm [20]. Chaibi et al.ooperate with Montreal Neurological Institute in Canada, usedigh sampling rate of 2000 Hz to obtained epilepsy EEG data, andesigned a software based on MATLAB to detect the high-frequencyscillation rhythm automatically by matching tracking method21].

Staba et al. at the University of California collected a series ofigh-frequency oscillation signals with intracranial electrodes at aampling rate of 10 KHz. The acquired signal contains 16-leads, andhe frequency of the signal ranges from 0.1 Hz to 5 KHz. Therefore,he signal contains various components of high-frequency oscil-ating rhythm, such as ripples, fast ripples, etc. In order to extracthe high-frequency oscillation signal corresponding rhythm band,taba et al. used a FIR band-pass filter in the 80 to 200 Hz range andonlinear fitting Lorentz distribution method to get ripples rhythm

rom the background signal [22]. Makeyev et al. used hardware fil-ering method to obtain ripples and fast ripples rhythm signals frompilepsy patients and Laplace transform to extract high-frequencyscillation rhythm [23]. Allison et al. analysed the high-frequencyscillation data during seizures. According to the morphologicalharacteristics of the high-frequency oscillation rhythm of epileptic

eizures, the high-frequency oscillatory rhythm feature extractionas performed by three methods: support vector machine, K near-

st neighbour, and logistic regression [24]. Jose et al. also used theigh sampling rate of 20 KHz to acquire the 16-lead high-frequency

m ideas.

oscillation signal. After the FIR band-pass filter was used to obtainthe fast ripples signal of the corresponding frequency band, theHilbert transform method was used to obtain the high-frequencyoscillation signal [25].

At present, the study of high-frequency oscillation signal is notmature [26], the method mainly used are FFT power spectrum anal-ysis, FIR filtering, quantitative analysis, Laplace transform, wavelettransform entropy, support vector machines and MP algorithm, theresearch results are still not used for clinical diagnosis [27–29].

Due to the difficulty in identifying high-frequency oscillationrhythm, the remainder of this paper is organized as follows. Chap-ter 2 is the pre-treatment process for epileptic EEG, at first handlethe original data with wavelet threshold method and 50 Hz octavenotch to eliminate frequency interference signals. Then, accordingto its own rhythm ripples and FRS frequency characteristics, pro-cess the signal which is already noise reduction with FIR digitalband-pass filter. Chapter 3 is the feature extraction and lesion loca-tion of high-frequency oscillatory rhythm in epileptic EEG signals.The Teager energy operator is first used to calculate the energy ofeach channel signal, and then combined with the length of the curveof the Teager energy operator to define the threshold and enhancethe separation effect. In addition, the Empirical Mode Decomposi-tion (EMD) decomposition is introduced to perform Teager-huangtransform on the signal to obtain the time-frequency-energy dis-tribution characteristics of the signal. Finally, a qualitative powerspectral density method and a quantitative EMD energy entropymethod was used to compare and analyse different leads of epilepsydata to locate epilepsy lesions. The algorithm ideas in this paper areshown in Fig. 1.

2. The acquisition and pre-processing of epilepsy EEGsignal for high frequency rhythm

2.1. The acquisition of high-frequency epilepsy EEG oscillationsignal

In general, there are two main methods to extract EEG. Oneis to put the disc electrode on the scalp, which requires that theelectrode is connected to the scalp reliably, and the EEG machineis recorded by the electroencephalograph, which is called elec-troencephalogram (EEG). The other way is by putting the electrodedirectly into the exposed cerebral cortex surface to record the elec-trical activity when an animal test or a craniotomy is performed in aclinical case, which is called cortical electroencephalogram (ECOG).Both of them reflect the self-generating activities of cerebral cortex.Because the high frequency oscillatory rhythm signal in epileptic

EEG is weak and its frequency is high, it is usually placed on thesurface of the cerebral cortex after craniotomy, or the depth elec-trode is directly placed on the cerebral cortex, and then the highsampling rate is used to collect it.

264 C. Wang et al. / Biomedical Signal Processi

Tls

2

hhtpeedrmpfh

2

dfiltwf

mnhtsdltnto

s

Fig. 2. The distribution diagram of the electrodes.

In this paper, the data is collected through in-depth electrodes.he sampling frequency is 2000 Hz, and the number of leads col-ected is 64 leads. The distribution diagram of the electrodes ishown in Fig. 2.

.2. Pre-processing of epilepsy EEG

According to the noise components exists in the originaligh-frequency epilepsy oscillation signal, this paper presents aigh-frequency oscillation epilepsy EEG pre-process method. Ini-ially, the wavelet weighted threshold method was utilized torocess the original epileptic EEG data to get a clear high-frequencypilepsy EEG signal. Then, in order to filter the frequency interfer-nce signal which is mixed during the collecting of epilepsy EEG,esign a frequency notch filter which is based on infinite impulseesponse (IIR) digital filter whose frequency is 50 Hz and frequencyultiplication. Finally, based on the band characteristics of rip-

les and FRS, design band-pass filter-based FIR digital filter whoserequency is 80–200 Hz and 200–600 Hz, ensuring the quality ofigh-frequency epilepsy EEG oscillation after pretreatment.

.2.1. De-noising with weighted wavelet thresholdIn this research, the weighted wavelet threshold was used to

ecompose the signal to get useful information and power coef-cient of noise. Generally, useful information is focused on a

imited number of coefficients of wavelet decomposition, whereashe energy coefficient of noise exists in all the coefficients of theavelet decomposition, and several coefficients that contains use-

ul information are greater than the coefficient of the noise signal.After the original epilepsy EEG data was decomposed by wavelet

ethod, effective EEG information will be stored in a limitedumber of wavelet coefficients, which consists of both low andigh-frequency oscillation signal, and noise signal is decomposedo lower the wavelet coefficients. Therefore, it was necessary toet threshold value for the wavelet coefficients after the signal wasecomposed by wavelet method, which would remove coefficients

ess than threshold and retain wavelet coefficients greater thanhe threshold. Then the effective information will be achieved andoise information will be separated. Finally, use inverse wavelet

ransform to rebuild signal and obtained epilepsy high-frequencyscillation signal after noise removal.

Based on the election law of wavelet threshold after decompo-ition, Wavelet threshold de-noising mainly includes two different

ng and Control 47 (2019) 262–275

method of hard threshold and soft threshold. Assuming eachwavelet coefficient after wavelet decomposition is d, the giventhreshold is �, the new wavelet coefficients after threshold pro-

cessing is︷︸︸︷d j,k, and the hard threshold equation is expressed

as:︷︸︸︷d j,k = {

dj,k |dj,k| ≥ �

0 |dj,k| < �(1)

The soft threshold equation is expressed as follows:

︷︸︸︷d j,k = {

sgn (dj,k)(|dj,k| − �) |dj,k| ≥ �

0 |dj,k| < �(2)

In Eq. (1), it can be seen that the hard threshold method is simplyremoving the coefficients less than the given threshold. Since Eq. (1)is not continuous in the threshold � point, the signal after treatmentappears to be rough. Soft threshold method as shown in Eq. (2) onlyprocesses the part larger than the given threshold, so the signal hasa certain sleek character after processing, the overall effect will bebetter than the hard threshold. Nevertheless, when

∣∣dj,k∣∣ ≥ �, thereis always a constant error between the estimated and actual values.In order to compensate, in this research both hard and soft thresh-old were combined to get a weighted method to process the signal,the threshold weighted equation can be expressed as follows.

︷︸︸︷d j,k = {

(1 − �) • dj,k + � • sgn (dj,k)(|dj,k| − �) |dj,k| ≥ �

0 |dj,k| < �(3)

Eq. (3) shows that, after the introduction of a weighting factor,which combines the properties of hard and soft threshold methodcan overcome the shortcomings of soft and hard threshold to someextent. In practice, the weighting coefficient � is generally taken tobe 0.5. Compared with the traditional hard and soft threshold de-noising method, the weighted threshold de-noising method have abetter effect of filter.

2.2.2. De-noising with notch filterNotch filter was designed to filter out 50 Hz and its harmonics

which was introduced during the acquisition process. The filter isessentially a digital band-stop filter, only that the stopband rangeof the filter is very small. In this research, the stopband range wasselected as 50 ± 2 Hz, and similar for other multiples of 50 Hz. Inorder to preserve the other effective information in EEG during theprocess of notch filtering, the requirements of an ideal notch filterfrequency response are:

H(jω) ={

1 ω /= ω0

0 ω = ω0

(4)

Where, ω0is the center frequency, in ideal condition, the responsefrequency at the center frequency should be zero, the frequency innon-central is 1.

In this paper, an IIR digital filter was designed based on Matlabto filter out interference signal of 50 Hz and its harmonics. The IIRdigital filter can be expressed by the following linear differentialequation with constant coefficients:

y(n) =M∑i=0

aix(n − i) −N∑i=1

biy(n − i) (5)

Where, x(n)and y(n) respectively stand for the sequence of the inputand output signal, ai and bi represent the filter coefficients, M andN is order of the input and output signal.

ocessi

2

8ttfifilfi

y

H

3a

aef

3c

ctrw

O

E

W

sTv

a

Ao

a

tn

C. Wang et al. / Biomedical Signal Pr

.2.3. FIR band-pass filterThe high-frequency oscillation rhythm ripples are in the

0–200 Hz band and FRS in the 200–600 Hz band. However, most ofhe collected epileptic EEG signal were below 100 Hz frequency, andherefore it was necessary to design the corresponding band-passlter to extract an effective high-frequency oscillation rhythm. So,nite impulse response (FIR) digital filter was designed with Mat-

ab for band-pass filtering. The differential equation of FIR digitallter is:

(n) =N−1∑i=0

aix(n − i) (6)

The impulse response of the system is:

hd(n) =N−1∑i=0

aiı(n − i) = a0ı(n) + a1ı(n − 1) + · · ·

+aN−1 [n − (N − 1)] (7)

The frequency response of the system is:

d(ejω) =

N−1∑n=0

hd(n)e−jωn (8)

. Feature extraction of high-frequency oscillation rhythmnd lesion localization

High frequency oscillation rhythm is a high frequency, lowmplitude, high energy characteristics signal, in this paper Teagernergy operator and EMD method were used to extract the usefuleatures from the epilepsy EEG signal.

.1. Feature extraction based on Teager energy operator andurve length method

According to high-frequency, high-energy and low-amplitudeharacteristics of the high-frequency oscillation signal, it is easy forhe Teager energy operator to detect the high-frequency oscillationhythm from the background signal for a set of known time seriesith high frequency oscillation rhythm.

For a limited band signal, the original definition of Teager Energyperator can be described as:

T = A2 sin (˝)2, = 2�f

fs(9)

here, A is the amplitude, is the digital angular frequency.Eq. (9) can be further expressed as a continuous signal in as

hown in Eq. (10) and as a discrete signal as shown in Eq. (11) below.he output value of Teager energy operator is usually called TEOalue [30].

For any continuous signalx(t), Teager energy operator is defineds:

[x (t)] = [x (t)]2 − x (t) x (t) (10)

mong them, x(t)and x(t)are the first and second derivatives of theriginal signal respectively, which is a nonlinear energy operator.

For the discrete signal, the definition of the Teager energy oper-tor is:

2

[x (n)] = [x (n)] − x (n − 1) x (n + 1) (11)

For discrete signals, it is possible to get the signal energy at anyime only to calculate the samples of three consecutive instanta-eous data points, which means good transient.

ng and Control 47 (2019) 262–275 265

Calculate Teager energy value by Eq. (11), the entire signalincludes non-high-frequency oscillation signals, and the value ofthese TEO value is smaller than the high-frequency oscillationrhythm. Therefore, it could set threshold for the signal after Tea-ger energy calculation to remove the Teager energy value of lowfrequency oscillation signal, which will improve the separation per-formance of Teager energy operator.

In this paper, the Teager energy operator values of several con-secutive points are calculated curve length to obtain the thresholdwhich TEO separate high-frequency and non-high-frequency oscil-lation signal, the expression of curve length can be described asfollows:

˚[x(n)] =∑n+N

k=n

∣∣x(k − 1) − x(k)∣∣ (12)

Which x(n)represents the calculated Teager energy values of thesignal and N represents the number of consecutive Teager energyoperators.

From Eq. (12) can know, at any point in the Teager energy sig-nal n, the abscissa difference before and after only 1. Therefore, thelength of the curve between two points in a row can be approxi-mately represented by its ordinate, at the same time, the value Ncould be well controlled and the length of the data could meet therequirement to calculate curve.

3.2. Feature extraction based on Teager-huang transform

Take the advantage of the Teager energy operator sensitive forfrequency and energy of the signal, it could obtain Teager energyvalue of the signal. The EMD transform could decompose the signalinto a series of intrinsic mode functions (IMF). The IMF should meetthe following two conditions: (1) the number of the extreme pointsand zero points should be equal or differ by one at most; (2) thesignal should be locally symmetrically distributed on the time scale.Finally, the time spectrum of the signal is obtained by Teager-huangtransform.

For a continuous signal x(k), it could finally get n componentswhich meet the conditions of IMF and a residual component afterthe EMD method:

x(t) =n∑i=1

ci(t) + rn(t) (13)

Where ci(t) represents the IMF component and rn(t) represents theresidual component. The IMF of n contains information about theoriginal signal in different frequency bands, which means the num-ber of decomposed layers and the information they contain varieswith the signal itself adaptively.

From Eq. (10), the instantaneous frequency f (t) and the instan-taneous amplitude

∣∣a(t)∣∣ can be obtained as follows:

f (t) ≈ 12�

√ (x(t)) (x(t))

(14)

∣∣a(t)∣∣ ≈ (x(t))√ (x(t))

(15)

Combined with Eqs. (14) and (15), calculate the instantaneous fre-quency and amplitude of every intrinsic mode function in formula(13), at last use Teager-huang transform to process all the IMF com-ponents and obtain:

n∑j∫f (t) dt

x(t) = Re

i=1

a(t)e (16)

Where Re means to take a real part. As the meaning of Hilbert spec-trum, the instantaneous amplitude and instantaneous frequency

2 rocessi

ocsd

3r

lchscpifhtfri

3

pdspo

dmsrtt

aieci

I

o

P

odwir

I

66 C. Wang et al. / Biomedical Signal P

f each IMF component in Eq. (16) are related to time. Thus, itould obtain the amplitude and frequency characterization of theignals in time domain, which is the three-dimensional and two-imensional time-frequency-energy distribution of the signal.

.3. Lesion localization based on high-frequency oscillationhythm

The appearance of high-frequency oscillation rhythm and lesionocation of epileptic are closely related to each other. As dis-ussed previously, in patients with epilepsy the emergence ofigh-frequency oscillation rhythm during the seizures which thenpread to different locations of the brain over time. This signalan be collected by multi-lead depth electrode distributed at everyosition of brain. From the view of space, the signal distribution

s specific, in epilepsy originating area from outside to inside, therequency of the high-frequency oscillation rhythms increasinglyigher, especially FRS, the closer to the core area of the epileptic foci,he higher the frequency FRS will be recorded. Therefore, epilepticoci could be located by comparing the differences between the FRSecorded at the various leads, where the highest frequency of FRSs recorded can be considered the epileptic focus position.

.3.1. Lesion localization based on power spectral density methodFor a lead that contains a high frequency oscillating signal, the

ower it carries is greater than the power carried by the lead thatoes not contain the high frequency oscillating signal, and its powerpectral density (PSD) value should be greater. Based on this, thisaper proposes a lesion localization algorithm from high-frequencyscillation rhythm based on PSD of the obtained signals.

Traditionally, the PSD value of the signal is calculated by theirect method. For the direct method of the power spectrum esti-ation, when the data length N is too large, the fluctuation of

pectral curve will increase. If N is too small, the power spectralesolution is low. Because these problems cause the value of PSDo be difficult to obtain, this paper introduces the Bartlett methodo obtain the PSD value.

Bartlett method is used to calculate the diagram and then aver-ge, which the limited long sequence x(n) with N points is dividednto a series of data segment. The data is divided into L sections,ach section have M samples, thusN = LM, the ith sample sequencean be written xi(n) = x(n + (i − 1)M), and the ith the periodograms:

iM(ω) = 1

M

∣∣∣∣∣∣M−1∑n=0

xi(n)e−jωn2∣∣∣∣∣∣ (17)

Spectral estimation is defined as the average of L segment peri-dogram:

ˆxx(ω) = 1L

L∑i=1

IiM(ω) (18)

Welch method is utilized to amend Bartlett law in two aspects,ne is to select the appropriate window function ω(n) and thenirectly added to the calculation before calculate the periodogram,hich makes the estimation of spectrum non-negative. The other

s overlap between segments when the divide segments, which willeduce the variance.

iM(ω) = 1

MU

∣∣∣∣∣N−1∑n=0

xi(n)ω(n)e−jωn

∣∣∣∣∣2

(19)

ng and Control 47 (2019) 262–275

Which U = 1M

N−1∑n=0

W2(n)is normalization factor, then estimate the

power spectrum with Eq. (18).

3.3.2. Lesion localization based on EMD energy entropyEpileptic EEG signals collected from different leads contain dif-

ferent kind of frequencies, when a lead appear high-frequencyoscillation rhythm, the proportion of different signal componentswill be different. Therefore, the energy distribution of lead whichcontains high-frequency oscillation signal components will be dif-ferent. In order to verify the change of energy, here introduce theconcept of EMD energy entropy. Calculate the energy of each IMFfunction to obtain the energy distribution vector of the signal afterthe signal decomposed by EMD, and then calculate the value ofEMD energy entropy on different leads by energy feature vectors. Bycompare the difference between EMD energy entropy and obtainedFRS distribution of different leads, ultimately locate the epilep-tic focus. This paper proposes a new method for epileptic focuslocalization based on empirical mode energy entropy.

Decompose the epilepsy EEG high-frequency oscillation signalx(t) which is preprocessed with EMD method, and n IMF compo-nents of different frequency bands will be acquired. The energydistribution of signal on different leads will be different, since theEMD itself orthogonal, the total energy of all the IMF componentsand the energy of the original signal will be equal on each lead.Because each IMF component contain different frequency compo-nents, and thus has a different energy, the EMD energy entropy canbe defined as:

H = −n∑i=1

pi log pi (20)

Where, pi = Ei/E represent the proportion of the energy of each IMFcomponent in total energy E.

The smaller value of H indicates the more concentrated ofepileptic EEG energy, effective information concentrated in a lim-ited number of IMF components. At the same time, indicate thatcomponent contains effective information dominates in the epilep-tic EEG, that is, ripples or FRS composition is in the majority.Conversely, if the larger the value H, the more divergent the energyof Epileptic EEG, and no specific frequency rhythm occurs meansthe energy of each IMF component uniform distribution, which themain component of the signal at this time is the background noise.

4. Simulations analysis and discussion

The data used in this paper acquired from a hospital which is col-lected by using the depth electrodes mode. The sampling frequencyis 2000 Hz, and the number of leads collected for the 64 leads.The data used were divided into two groups. The first patient’sepileptic high-frequency oscillation data was used as the verifica-tion data of the proposed method. The second patient’s epileptichigh-frequency oscillation data was used as the application data ofthis method.

4.1. The pretreatment results

In this paper, Daubechies wavelets used for wavelet decompo-sition. 5 layers of epileptic signals are decomposed by DB3 wavelet.The data mainly selected for pretreatment is the 60,000 data points

which the sampling time is 30 s acquired from the first patient, theselected lead is clearly known that contains high-frequency oscil-lation rhythm and be able to locate the signal of epileptic foci fromthe third lead.

C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275 267

Fig. 3. Comparison of the results before and after signal weighted wavelet threshold de-noising.

Fp

sadlrd

ttnsa

wmb

by frequency notch filter, the frequency interference can be well

ig. 4. The results of denoise signal, notch data, and bandpass filtered data of thereprocessing process.

The weighted wavelet threshold method is utilized as the firsttep of de-noising, the weighting coefficient � selected for hardnd soft threshold is taken as 0.5. Fig. 3 shows the before and aftere-noising results of the third lead signals, in which the Fig. 3 on

eft side shows the de-noising results of 60,000 points data and theight half of the Fig. 3 shows the comparison of 10,000 samples ofata.

From Fig. 3, it can be seen that after the wavelet-weightedhreshold denoising of epileptic EEG data, partial noise can be effec-ively suppressed. The signal will become smoother and there willot be too much signal lost at the tip. After processing, it can retainome useful information, and the overall noise reduction effect islso better.

Use 50 Hz octave frequency notch filter to process the signal

hich is already pretreatment by the weighted wavelet thresholdethod. Fig. 4 shows the contrast before and after using FIR digital

and-pass filter (200–600 Hz), the spectrum is shown in Fig. 5.

Fig. 5. The spectrum results of denoise signal, notch data, and bandpass filtered dataof the preprocessing process.

As is known from Fig. 4, the magnitude of the effective informa-tion that acquired from epileptic EEG data are concentrated in the−300 uV–300 uV. There is a widely difference before and after pro-cess the data with notch filter, especially in the vicinity of 40,000data points i.e. nearby 20 s, which means the frequency interferencecomponents mainly generated within this time, and this interfer-ence will have some impact on the data behind 40,000 points. Afterprocess by 200–600 Hz band-pass filter, it is obviously that thereis a clear FRS shock wave whose amplitude is between −10 uV and10 uV.

The first picture in Fig. 5 shows that there still contains a largeamount of frequency interference signal after the collected epilepsyEEG data process by weighted wavelet threshold method, espe-cially the signals of 50 Hz and its harmonic frequency. After process

suppressed. It can be seen from the middle picture of Fig. 5, the maincomponents of the signal which is handled by frequency notch filterare mostly distributed at low frequencies below 100 Hz. This paper

268 C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275

Fig. 6. Comparison of third lead bandpass filter data before and after processing byTeager energy operator.

FT

fahth

gocrs

4r

uHstt

Fig. 8. The curve length on the third lead.

ig. 7. Comparison of first lead bandpass filter data before and after processing byeager energy operator.

ocus on the high-frequency (200–600 Hz) information of FRS, after band-pass filtering, the data after preprocessing mainly containsigh-frequency oscillation rhythm. As is known from Figs. 4 and 5,he characteristics of the high-frequency oscillation rhythm areigh-frequency, high-energy and low-amplitude.

In fact, after the band-pass filter analysis of all 64 leads of thisroup, the ripples of 80–200 Hz exists on many leads, while thebvious FRS of 200–600 Hz exists only on the third lead. It indi-ates that FRS can directly respond to epileptic seizures, and theelationship between ripples and epileptic seizures needs furthertudy.

.2. Feature extraction results of high-frequency oscillationhythm

Firstly, Teager energy operator and curve length method aresed for feature extraction from high-frequency oscillation rhythm.

ere, the abnormal third lead and the non-abnormal first lead are

elected for comparative analysis. Teager energy operator on thehird lead is shown in Fig. 6, while the Teager energy operator onhe first lead that does not contain FRS rhythm is shown in Fig. 7.

Fig. 9. Comparison of Threshold teager energy operator, Teager energy operator andbandpass filtered date in the third lead.

From the results of the third lead in Fig. 6, it can be seen thatat the first 40,000 sampling points of the data, each larger TEOvalue corresponds to each FRS rhythm of the original data. The moreobvious the FRS performance is, the greater the TEO value is, andthe amplitude of the latter 20,000 sampling points is not smallerthan that of the FRS. This shows that the amplitude of the signalis similar to the high frequency oscillating rhythm signal, but itsfrequency should be lower than that of FRS, so it does not showa large Teager energy value. It also shows that the characteristicsof high frequency FRS can be displayed well by using the Teagerenergy operator.

From the results of Fig. 7, it can be seen that for the firstlead without FRS rhythm, the amplitude of the signal is within−2 uV–2 uV and the amplitude of the signal is very low after band-pass filtering. The signal itself does not show the characteristicsof FRS rhythm apparently. It is considered that the lead does notcontain high-frequency oscillation rhythm signals but only con-tains some of the background signal or a normal EEG which doesnot have the high-frequency, high-energy and other features. The

calculated Teager energy values is obviously different with the val-ues on the third lead that contains the FRS rhythm, indicating theeffectiveness of Teager energy operator to detect the presence ofFRS rhythm.

C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275 269

Fig. 10. The 1–4 layer EMD results on third lead.

EMD

aiFcl

ms

itdF

Fig. 11. The 5–8 layer

In order to further improve the spatial separation effect of TEOnd eliminate the effect of low TEO on the result, the TEO thresholds determined by the length of the curve, and the TEO value of non-RS is abandoned. In the formula (12), when the value of N is 3, theurve length of three consecutive points TEO is calculated, and theength chart of third lead TEO is shown in Fig. 8.

It can be seen from Fig. 8 that the TEO values on the third leadainly distributed below 4.5. The results plus threshold value are

hown in Fig. 9.Fig. 9 shows that the separation results of Teager energy value

n the feature space could be better enhanced after the addition of

hreshold in the energy Teager results, which makes the results ofistinct, especially have a very good inhibitory effect for the non-RS value after 40,000 data points.

results on third lead.

Finally, Teager-huang method is utilized for feature extractionof high-frequency oscillation rhythm. There are 16 IMF componentsafter EMD decomposition for the signal on the third lead and theresults are shown in Figs. 10–13.

From Figs. 10–13 can be seen, each of the IMF componentsdecomposed by EMD contains different components of the ampli-tude and frequency, and the amplitude and frequency of the IMFcomponents vary from high to low. This shows that after the firstlead is decomposed by EMD, the effective components of IMF com-ponents of signals are mainly concentrated on the first 8 leads. Interms of the characteristics of EMD decomposition, the frequency

resolution of each frequency segment varies with the signal itselfand has the characteristics of adaptive multiresolution analysis.

270 C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275

Fig. 12. The 9–12 layer EMD results on third lead.

er EMD

tcels

tatiqI

I

Fig. 13. The 13–16 lay

Seen from the first to the ninth IMF component, the amplitude ofhe IMF drops from about 10 uV to about 0.2 uV. On the various IMFomponents, each FRS rhythm also performed less and less obvious,specially after the eighth IMF components the amplitude is muchower, and the IMF component signals become gentler, even cannotee FRS rhythm.

As far as the spectrum of each IMF component is concerned,he frequency of the IMF component is getting lower and lower,nd the proportion of the frequency is becoming less and less. Ashe number of EMD decomposition layers increases, the effectivenformation of the signal is less and less, and the main high fre-

uency oscillating components should be concentrated in the front

MF component.Use Teager energy operator to process the AM-FM signal of each

MF component after EMD decompose with Teager-huang trans-

results on third lead.

form, which is intend to separate the instantaneous frequency andamplitude of the signal, then get the spectrum of signal on the timeaxis. The spectrum of three-dimensional time and two-dimensionaltime are shown in Figs. 14 and 15 respectively. The data of the thirdlead after processed by band-pass filter is shown in Fig. 16.

As can be seen from the three-dimensional and two-dimensional time spectrum in Figs. 14 and 15, the timespectrum after Teager-huang transform contains the charac-terizes of high-frequency oscillation rhythm on lead from thetime-frequency-energy of the signal. The three-dimensional time-frequency spectrum in Fig. 14 also show that every time the FRS

appears, the energy value has a significantly enhance, which per-form ups and downs the same as the waves. The crest on behalf ofthe appearance of the FRS rhythm, while the trough means no FRSrhythm.

C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275 271

Fig. 14. The spectrum of three-dimensional time on third lead.

Fig. 15. The spectrum of two-dimensional time on third lead.

Fig. 16. The band-pass filtered data on third lead.

Fig. 17. The spectrum of three-dimensional time on first lead.

From the contrast diagram of the two-dimensional time-frequency spectrum of Fig. 15 and the third lead band pass filterof Fig. 16, it can be seen that with the time of the FRS rhythm in theoriginal signal, the two-dimensional time spectrum is character-ized by high energy and high frequency at the corresponding time.Especially around 20 s, or near the 40,000 data points, the frequencyand energy change is very obvious.

Both the three-dimensional spectrum chart in Fig. 14 and thetwo-dimensional spectrum chart in Fig. 15 shows that, based onthe high-frequency, high-energy, low-amplitude characteristics ofhigh-frequency oscillation rhythm, Teager energy operator and theEMD method are utilized to obtain the distribution of the signal atthe time-frequency- energy, which is clearly token the character-istic of high-frequency oscillation rhythm.

Similarly, for comparison purposes, the same approach that areTeager energy operator and EMD method are taken to analyze theresults of Teager-huang transform on the lead (first leads is chosehere) does not contain the high-frequency oscillation rhythm. Itsthree-dimensional and two-dimensional spectral diagram shownin Figs. 17 and 18.

From the results of the first lead three-dimensional time

spectrum and the two-dimensional time-frequency spectrum ofFigs. 17 and 18, the frequency and energy of the first lead timefrequency diagram without high frequency oscillation rhythm sig-nal is relatively slow and the fluctuation is relatively small. This

Fig. 18. The spectrum of two-dimensional time on first lead.

272 C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275

if

4o

4m

atstltIf

eldr

htrtintblr

deprp

tsw

rhythm. At this time, the energy will be concentrated in the corre-sponding frequency band of the high-frequency oscillation rhythm,so that the energy distribution will be concentrated, so that theEMD energy entropy will decrease. Therefore, the minimum energy

Table 1The partial energy entropy of the first patient.

lead Ordinary energy entropy Teager energy entropy

1 0.4135 0.19732 0.3879 0.21403 0.3328 0.07416 0.4070 0.21929 0.3920 0.225413 0.3863 0.225316 0.3825 0.255320 0.3917 0.226926 0.4044 0.195930 0.4074 0.249933 0.3951 0.224038 0.4135 0.2008

Fig. 19. The PSD spectrum comparison of the first patient.

s because there is no obvious FRS rhythm in the first lead, so therequency and energy values are mainly concentrated at low values.

.3. The lesion localization application of high-frequencyscillation rhythm

.3.1. Results of focal localization by power spectral densityethod

The power spectrum density method was used to qualitativelynalyze the suspected signals containing high-frequency oscilla-ory rhythm, and the largest PSD value was chosen to be the mostignificant lesion location. In fact, after the Teager energy calcula-ion of all 64 leads of the patient, it was found that the only thirdead signals suspected to be the FRS rhythm were very strong, andhe Teager energy values of the other lead showed a very low value.t shows that the data of this patient is relatively good, and the highrequency oscillation signal is not strong during epileptic seizure.

Therefore, after feature extraction and screening by the Teagernergy operator, the PSD power spectral density analysis (the firstead, the third lead and the thirty-eighth lead) is selected from theifferent lead, which is suspected to have high frequency oscillationhythm. The PSD spectrum contrast diagram is shown in Fig. 19.

As can be seen from Fig. 19, the PSD value of the third lead withigh frequency oscillation rhythm is larger than the PSD value ofhe first and thirty-eighth leads without high frequency oscillationhythms at the high frequency section. The closer the FRS rhythm iso the position of the epileptic focus, the greater the probability ofts occurrence. The largest PSD lead is considered a lead with a largeumber of high frequency oscillation rhythms, and the position ofhe brain is the location of the epileptic focus. Therefore, it cane considered that the third lead of PSD value corresponds to the

ocation of epileptic foci in the brain, which is consistent with theesults of the previous doctor.

In order to verify the effectiveness of the method, the epilepsyata of the second patient is taken to process, it is not known thepileptic foci on which leads before processing. Take the same pre-rocessing and feature extraction methods to handle the data, theesults show that the leads suspected epileptic foci of the secondatient are 42-lead, 50-lead, 55-lead, 60-lead and 61-lead.

The epilepsy data of the second patient is unlike the data of

he first patient that is easily to identification. Five leads of theecond patient are detected to contain the FRS in simultaneous,hich illustrate that FRS spread to a region relatively quickly and

Fig. 20. The PSD spectrum comparison of the second patient.

detected by several leads at the same time. The PSD comparisonchart is shown in Fig. 20.

As can be seen from Fig. 20, for the detected FRS, PSD valuesof 55-lead is the largest, indicating the high-frequency componenton the 55-lead is the largest, which means that the epileptic focusoriginating region of the second patient is the 55-lead. The result isthe same with the conclusion of the doctor afterwards.

4.3.2. Results of focal localization by EMD energy entropy methodThe EMD energy entropy method is used to quantitatively ana-

lyze the signals, and the entropy of EMD energy on different leadsis obtained. According to the previous analysis, the leader withthe minimum EMD energy entropy was selected as the locationof epileptic foci. The EMD energy entropy of the signal is calculatedby the ordinary energy algorithm and the Teager energy algorithmrespectively. The EMD energy entropy results of the signal are cal-culated by calculating the energy distribution vector of each IMF,as shown in Table 1.

As can see from the results in Table 1, the average energy entropyand Teager energy entropy are all minimum in the third lead. Thisis due to the relative average and uncertainty of the energy distri-bution of epileptic EEG signals under normal conditions, and thefrequency of high frequency oscillation occurs in the IMF compo-nent when the epileptic seizures appear high frequency oscillation

44 0.4013 0.235550 0.4075 0.223355 0.3990 0.2548

C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275 273

Table 2The partial energy entropy of the second patient.

lead Ordinary energy entropy Teager energy entropy

3 0.4078 0.19267 0.4056 0.183210 0.3859 0.172313 0.3828 0.171617 0.3923 0.183422 0.4085 0.191030 0.3968 0.173834 0.4116 0.175938 0.4079 0.163942 0.3726 0.123950 0.3761 0.136755 0.3137 0.0680

eqete

eeT

epcfT

rbttdobr

4

caphBhpfiSeo

Table 3Comparison of the characteristics of different method.

Feature

Method Degree ofautomation

Degree ofdistinction

accuracy speed

Method of the paper I I I II

This work was supported by the National Natural Science Foun-dation of China under Grant No. 61573381 and Fundamental

56 0.3959 0.180860 0.3824 0.128061 0.3858 0.1392

ntropy of EMD can be considered as the lead of the highest fre-uency oscillatory rhythm components, namely the location ofpileptic foci. Moreover, the energy entropy of Teager is better thanhe general energy entropy, and it also reflects the instantaneousfficiency of Teager energy operator.

Similarly, in order to verify the validity of the EMD energyntropy method, the EMD energy entropy is calculated for secondpileptic patients with undetermined position of epileptic focus.he results are shown in Table 2.

It is seen from the results in Table 2, both the ordinary energyntropy and the Teager energy entropy on 55-lead of the secondatient are minimum. According to the conclusions ahead, it isonsidered that in the 55-lead detective a large number of high-requency oscillation rhythm and the 55-lead is the epileptic focus.his result is the same with conclusion the doctor given afterwards.

Power spectral density (PSD) and EMD energy entropy algo-ithm have their advantages. The implementation of PSD algorithmelongs to a qualitative analysis method. The algorithm is simplerhan the EMD energy entropy algorithm, and the disadvantage ishat the results are not very clear when analyzing too many leadata at the same time. The EMD energy entropy algorithm is a kindf quantitative analysis method. The algorithm is more complex,ut it can get the EMD energy entropy of all lead data, and theesult is straightforward.

.4. Discussion

In this paper, the high-frequency oscillation rhythm of clini-al epilepsy EEG is used for the study. According to the sourcesnd characteristics of the noise may exist during the acquisitionrocess of epileptic EEG, an effective pretreatment method forigh-frequency oscillation rhythm in Epileptic EEG is proposed.ased on the amplitude, frequency and energy characteristics ofigh-frequency oscillation signal after pre-treatment, this paperroposes the Teager energy operator and EMD Transform algorithmor feature extraction. For the difficult issues of epileptic focus local-

zation, a lesion localization algorithm is raised based on Powerpectrum and EMD energy entropy. With the simulation for realpileptic EEG data demonstrate the effectiveness and superiorityf the proposed approach.

Support vector machine I II I IIIK nearest neighbor I II II IILogistic regression II III III I

Although the amount of patient data used in this study is limited,the potential performance of the algorithm has been fully demon-strated. It can be seen from Tables 1 and 2 that the lead of thefocus is distinctly different from the non-focus lead. At the sametime, the PSD of the algorithm is used as a qualitative criterion, andenergy entropy is used as a quantitative criterion, which ensuresthe correctness of location from different angles. In addition, thisalgorithm has a high degree of automation, and it does not needto adjust the individual parameters of the brain electroencephalo-gram of different patients to realize the effective identification ofthe epileptic focus. Existing algorithms based on high-frequencyoscillation signals need to be assisted by manual pre-parameteradjustment and can’t achieve good discrimination. At the sametime, the frequency threshold set is unified and the nonlinearityis poor. Based on the data in this paper, several popular methodssuch as support vector machine, K nearest neighbor, and logisticregression and so on, are implemented in paper [24]. Comparedwith the degree of automation, degree of distinction, accuracy andspeed, the results are divided into three grades, I, II and III fromhigh to low, as shown in Table 3. It can be seen from the Table 3that this method is better than the current mainstream algorithm.

5. Conclusions

Based on epileptic EEG in clinical medicine, a qualitative andquantitative analysis algorithm for high frequency oscillatoryrhythm is proposed. Through the data analysis process and calcu-lation results of two patients, to a certain extent, it shows that themethod is feasible and effective.

Although this algorithm can reflect the relationship betweenepileptic high frequency oscillation rhythm and epileptic diseaseto a certain extent, there are still some shortcomings. In the future,the frequency band and lead number of conventional EEG acqui-sition device will be improved. In addition, as the results of thisalgorithm are only compared with the results of neurosurgeon dis-criminant, a large number of comparative studies will be furthercompared with the clinical outcomes in the future to carry out amore comprehensive performance evaluation.

Acknowledgments

Research Funds for the Central Universities of Central South Uni-versity under Grant No. 2018zzts528.

2

A

74 C. Wang et al. / Biomedical Signal Processing and Control 47 (2019) 262–275

ppendix A.

The first patient’s entire lead EMD energy entropy table

Lead Ordinary energy entropy Teager energy entropy Lead Ordinary energy entropy Teager energy entropy

1 0.4135 0.1973 33 0.3951 0.22402 0.3879 0.2140 34 0.3995 0.21263 0.3328 0.0741 35 0.4097 0.20134 0.3916 0.2039 36 0.4101 0.20025 0.3856 0.1923 37 0.4099 0.20116 0.4070 0.2192 38 0.4135 0.20087 0.4025 0.2012 39 0.4111 0.20108 0.3987 0.1986 40 0.4123 0.21169 0.3920 0.2254 41 0.4105 0.212210 0.3984 0.2158 42 0.4098 0.215911 0.3953 0.2049 43 0.4102 0.224712 0.3956 0.2163 44 0.4013 0.235513 0.3863 0.2253 45 0.4025 0.226714 0.3889 0.2178 46 0.4087 0.221415 0.3847 0.2114 47 0.4142 0.225116 0.3825 0.2553 48 0.4069 0.215617 0.3920 0.2254 49 0.4088 0.222418 0.3941 0.2196 50 0.4075 0.223319 0.3915 0.2173 51 0.4015 0.215620 0.3917 0.2269 52 0.4007 0.221921 0.3926 0.2047 53 0.4026 0.224722 0.3989 0.2061 54 0.3985 0.231323 0.3984 0.2109 55 0.3990 0.254824 0.3952 0.2094 56 0.4008 0.241025 0.4026 0.1987 57 0.4023 0.235926 0.4044 0.1959 58 0.4019 0.224727 0.4078 0.2046 59 0.4037 0.215928 0.4015 0.2197 60 0.3987 0.225429 0.4027 0.2276 61 0.3968 0.217430 0.4074 0.2499 62 0.4078 0.201431 0.4022 0.2315 63 0.4055 0.209932 0.4014 0.2284 64 0.4017 0.2015

The second patient’s entire lead EMD energy entropy table

Lead Ordinary energy entropy Teager energy entropy Lead Ordinary energy entropy Teager energy entropy

1 0.4012 0.1856 33 0.4091 0.17712 0.4023 0.1963 34 0.4116 0.17593 0.4078 0.1926 35 0.4015 0.17124 0.4067 0.2016 36 0.4009 0.17105 0.4010 0.1923 37 0.3995 0.17056 0.4026 0.1957 38 0.4079 0.16397 0.4056 0.1832 39 0.3926 0.15238 0.3979 0.1846 40 0.3855 0.15169 0.3947 0.1816 41 0.3759 0.140210 0.3859 0.1723 42 0.3726 0.123911 0.3854 0.1710 43 0.3715 0.132912 0.3889 0.1723 44 0.3728 0.147913 0.3828 0.1716 45 0.3801 0.142614 0.3896 0.1779 46 0.3798 0.140715 0.3901 0.1796 47 0.3719 0.140116 0.3919 0.1816 48 0.3695 0.139817 0.3923 0.1834 49 0.3708 0.132218 0.3958 0.1898 50 0.3761 0.136719 0.4012 0.1923 51 0.3718 0.135920 0.4029 0.1982 52 0.3846 0.144221 0.4057 0.1928 53 0.3866 0.145722 0.4085 0.1910 54 0.3812 0.150223 0.4012 0.1927 55 0.3137 0.068024 0.4059 0.1936 56 0.3959 0.180825 0.4002 0.1917 57 0.3885 0.177226 0.3994 0.1897 58 0.3892 0.171027 0.3988 0.1846 59 0.3814 0.165328 0.3971 0.1815 60 0.3824 0.1280

29 0.3961 0.1791

30 0.3968 0.1738

31 0.3987 0.1713

32 0.4017 0.1792

61 0.3858 0.139262 0.3796 0.145863 0.3819 0.152064 0.3756 0.1468

ocessi

R

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[29] D.R. Macedo, G.M. Silva, A.P. Fortes, et al., High-Frequency Oscillations (HFOs)in Electroencephalography: Review and Case Study, Biosignals and

C. Wang et al. / Biomedical Signal Pr

eferences

[1] Kirsten M. Fiest, Sauro, et al., Prevalence and incidence of epilepsy: Asystematic review and meta-analysis of international studies, Neurology 88(3) (2017) 296–303.

[2] X.M. Gu, C.Y. Ding, N. Wang, et al., Influence of Occupational Status on theQuality of Life of Chinese Adult Patients with Epilepsy, Chin. Med. J. 129 (11)(2016) 1285–1290.

[3] Zhongmou Han, Xiaoqin Zhang, The study progress of epilepsy epileptic foci, J.Int. Neurol. Neurosurg. 37 (5) (2010) 461–465.

[4] E. Ferlazzo, S. Gasparini, E. Beghi, et al., Epilepsy in cerebrovascular diseases:review of experimental and clinical data with meta-analysis of risk factors,Epilepsia 57 (8) (2016) 1205–1214.

[5] M. Urekar, P. Sovilj, EEG dynamic noise floor measurement with stochasticflash A/D converter, Biomed. Signal Process. Control 38 (2017) 337–345.

[6] C. Stamoulis, B. Chang, Non-invasively recorded transient pathologicalhigh-frequency oscillations in the epileptic brain: a novel signature of seizureevolution, BMC Neurosci. 16 (Suppl 1) (2015) P32.

[7] M. Pail, P. Rehulka, J. Cimbálník, et al., Frequency-independent characteristicsof high-frequency oscillations in epileptic and non-epileptic regions, Clin.Neurophysiol. 128 (1) (2017) 106–114.

[8] M.N. Tibdewal, H.R. Dey, M. Mahadevappa, et al., Multiple entropiesperformance measure for detection and localization of multi-channelepileptic EEG, Biomed. Signal Process. Control 38 (2017) 158–167.

[9] M. Li, W. Chen, T. Zhang, Classification of epilepsy EEG signals usingDWT-based envelope analysis and neural network ensemble, Biomed. SignalProcess. Control 31 (2017) 357–365.

10] U.K. Yusuf, G. Jean, High frequency oscillations in intracranial recordings ofepileptic patients, in: International Conference on Multimedia, SignalProcessing and Communication Technologies, 2011, pp. 126–129.

11] W.C. Stacey, S. Kellis, B. Greger, et al., Potential for unreliable interpretation ofEEG recorded with microelectrodes, Epilepsia 54 (8) (2013) 1391–1401.

12] Xuanji Dong, Yuxiu Peng, Jiongxing Wu, The high-frequency oscillationrhythm’s origin, positioning and significance, Prog. Physiol. Sci. 43 (3) (2012)235–237.

13] A. Bragin, J. Engel Jr, R.J. Staba, High-frequency oscillations in epileptic brain,Curr. Opin. Neurol. 23 (2) (2010) 151–156.

14] G.P. Krishnan, G. Filatov, M. Bazhenov, Dynamics of high-frequencysynchronization during seizures, J. Neurophysiol. 109 (10) (2013) 2423–2437.

15] J. Cimbalnik, M.T. Kucewicz, G. Worrell, Interictal high-frequency oscillationsin focal human epilepsy, Curr. Opin. Neurol. 29 (2) (2016) 175–181.

16] S. Patidar, T. Panigrahi, Detection of epileptic seizure using Kraskov entropyapplied on tunable-Q wavelet transform of EEG signals, Biomed. SignalProcess. Control 34 (2017) 74–80.

[

ng and Control 47 (2019) 262–275 275

17] B. Milan, H. Jose, J. Pavel, et al., Interictal high-frequency oscillations indicateseizure onset zone in patients with focal cortical dysplasia, Epilepsy Res. 90(1-2) (2010) 28–32.

18] H. Kalbkhani, M.G. Shayesteh, Stockwell transform for epileptic seizuredetection from EEG signals, Biomed. Signal Process. Control 38 (2017)108–118.

19] R. Zelmann, F. Mari, J. Jacobs, et al., Automatic detector of High frequencyoscillations for human recordings with microelectrodes, IEEE 31 (4) (2010)2329–2333.

20] O.L. Smart, G.A. Worrell, G.J. Vachtsevanos, et al., Automatic detection of highfrequency epileptiform oscillations from intracranial EEG recordings ofpatients with neocortical epilepsy, Technical, Professional and StudentDevelopment Workshop (2005) 53–58.

21] S. Chaibi, R. Bouet, J. Jung, et al., Developement of Matlab-based GraphicalUser Interface (GUI) for detection of high frequency oscillations (HFOs) inepileptic patients, in: Emerging Signal Processing Applications (ESPA). LasVegas: IEEE, 2012, pp. 56–62.

22] R.J. Staba, C.L. Wilson, A. Bragin, et al., Quantitative analysis of high-frequencyoscillations (80-500 Hz) recorded in human epileptic hippocampus andentorhinal cortex, Neurophysiology 88 (4) (2002) 1743–1752.

23] O. Makeyev, X. Liu, L.L. Wang, et al., Feasibility of recording high frequencyoscillations with tripolar concentric ring electrodes duringpentylenetetrazole-induced seizures in rats, in: Engineering in Medicine andBiology Society (EMBC). California: IEEE, 2012, pp. 4599–4602.

24] P. Allison, W. Drausin, L. Brian, et al., Does the morphology of high-frequency(100-500Hz) brain oscillations change during epileptic seizures? in: Signals,Systems and Computers (ASILOMAR). Asilomar: IEEE, 2011, pp. 461–465.

25] I. Jose, F. Guglielmo, C. Elena, Emergent dynamics of fast ripples in theepileptic hippocampus Neuroscience 30 (48) (2010) 16249–16261.

26] L.M. de la Prida, R.J. Staba, J.A. Dian, Conundrums of high-frequencyoscillations (80-800 Hz) in the epileptic brain, J. Clin. Neurophysiol. 32 (3)(2015) 207–219.

27] A. Bragin, S.K. Benassi, F. Kheiri, et al., Further evidence that pathologichigh-frequency oscillations are bursts of population spikes derived fromrecordings of identified cells in dentate gyros, Epilepsies 52 (1) (2011) 45–52.

28] A. Bragin, C.L. Wilson, Voltage depth profiles of high frequency oscillationsafter clinic acid-induced status epileptics, Epilepsies 48 (5) (2007) 35–40.

Biorobotics Conference 18 (20) (2013) 1–5.30] R.A. Thuraisingham, Estimation of Teager energy using the Hilbert–Huang

transform, IET Signal Process. 9 (1) (2015) 82–87.