2017 International Seminar on Sensor, Instrumentation, Measurement and Metrology (ISSIMM)

Surabaya, Indonesia, August 25th - 26th, 2017

Electroencephalogram Analysis with Extreme

Learning Machine as a Supporting Tool for
Classifying Acute Ischemic Stroke Severity
Osmalina N. Rahma1 Sastra K. Wijaya2a
1
Biomedical Engineering, Department of Physic, 2
Departement of Physics
Universitas Airlangga, Universitas Indonesia
Surabaya, Indonesia Depok, Indonesia
Email: 1osmalina.n.rahma@fst.unair.ac.id Email: 2skwijaya@sci.ui.ac.id

Prawito3 Cholid Badri4

3
Departement of Physics 4
Departement of Biomedical Engineering, Graduate Program,
Universitas Indonesia Universitas Indonesia
Depok, Indonesia Jakarta, Indonesia
Email: 3prawito@sci.ui.ac.id Email: 4cholid.badri11@ui.ac.id

Abstract—Stroke is one of the highest causes of history. The prevalence of stroke in the group increases with
death in adults and disability in Indonesia, even in the age, the highest is at age ≥75 years (43.1) and (67.0‰)[1].
world. Therefore, it is necessary to diagnose stroke in the Based on two main types of stroke that exist, at about
early stage and give accurate prognosis assessment to 85% of all strokes are ischemic which is caused due to
improve stroke management. This study tried to automati- clogged blood vessel in the brain and almost 66,67% of post-
cally classify AIS severity based on EEG signals by using stroke patients leave the hospital with a disability [2], [3].
digital signal processing such as Wavelet transform and Stroke attack could give a psycho-social burden in the com-
feedforward type of neural network with ELM algorithm. munity [4]–[7]. Therefore, it is necessary to diagnose stroke
In this study, Delta Alpha Ratio (DAR), especially in the early stage condition [8], [9].
(Delta+Theta)/(Alpha+Beta) Ratio (DTABR) and Brain EEG is a potential tool to monitor brain activity in real-
Symmetry Index (BSI)'s value were used as the ELM input time, especially in intensive care unit. Continuous EEG
feature score, which were obtained by using Wavelet (CEEG) is neurodiagnostic method which has high sensitivity
transformation (Daubechies 4) and Welch's method to for detecting AIS [10]. In patients with cerebral ischemic
classify the acute ischemic stroke severity which refers to injury or ischemic stroke, EEG correlated with cerebral blood
the National Institutes of Health Stroke Scale (NIHSS). It flow (CBF) level. CBF In normal people are 50 to 70
had shown that the performance of system test accuracy, and EEG also showed normal pattern
the sensitivity and specificity were above 72%. These re- (alfa and beta wave), whereas CBF in patients with ischemic
sults were useful for classifying AIS based on EEG signals. stroke declines to 25 to 30 and EEG
recordings also showed slow waves (delta and theta wave) as
Keywords—Electroenchepalogram (EEG), Acute Ischemic well as the asymmetry wave of right and left hemisphere.
Stroke (AIS), Extreme Learning Machine (ELM) These indicate that EEG recording had a correlation to brain
ischemia, cerebral blood flow (CBF) and brain metabolism
[10]–[12].
I. INTRODUCTION
There were some studies of electroencephalograph that
related to stroke. Kenneth G. Jordan [10] suggested that elec-
Stroke is one of the highest cause of death that need
troencephalography can help to confirm or detect acute is-
special attention. The result of Indonesian Health Research in
chemic stroke which is shown by the presence of the slow
2013 was performed on people aged over 15 years, showed
waves (theta-delta activity) on the electroencephalogram
that the prevalence of stroke was 7% based on the health pro-
spectral and reduced cerebral blood flow. Michael J.A.M. van
fessionals statement and 12.1% based on patients' symptoms'
Putten and Tavy [13] used the BSI in monitoring the signal

978-1-5386-0745-9/17/$31.00 ©2017 IEEE 180

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:26:19 UTC from IEEE Xplore. Restrictions apply.
 181

from hemispheres in stroke patients using Quantitative elec-

EEG signal (.Edf)
troencephalograph. Furthermore, QEEG has improved to de-
tect and localize brain ischemia [14]–[16]. Rosmehah Wan
Omar Wan, et al [17] used k-Nearest Neighbors (k-NN) classi-
Pre-processing (Signal decompotition)
fication technique for classifying brain waves in acute ischem-
ic stroke. Huang, Shen and Duong used Support Vector Ma-
chine (SVM) for predicting acute ischemic tissue [18]. While Feature Extraction
Rajini and Bhavani used segmentation and texture features for
ischemic stroke detection [19].
This study was conducted based on previous studies Classification (Extreme Learning Machine)
that had successfully recognized the EEG signal of AIS pa-
tients and normal EEG with accuracy and sensitivity above
85% [20]. The aim of this study was to classify the severity of
stroke using digital signal processing such as wavelet trans- research, EEG recordings were done for about 30 minutes with
form and ELM algorithm, based on training data sets from 32 channel electrode during patients’ conscious condition. The
normal patients and patients with acute ischemic stroke which electrodes were placed in accordance with international 10-20
are complemented with information about each patient's NIHS system configuration
scale The research flow diagram is shown in Fig. 1. The EEG
DAR, DTABR and BSI values were used for the ELM signals in European Data Format (.edf) were converted into
input feature, which were obtained based on the Wavelet ASCII using Matlab script (edfread) from Brett Shoelson
transformation (Daubechies 4) and Welch's method to classify [22], then filtered into 4-sub wave based on frequency ranges
level of stroke severity into four classes such as no stroke (alpha, beta, theta and delta) using Power spectral density.
(normal), mild stroke, moderate stroke and heavy stroke. The Spectral power for each electrode were performed using
DAR and DTABR features are required to know the ratio of Discrete Wavelet Transform (DWT) to decompose EEG signal
the appearance of the slow waves, such as delta and theta and into 4-sub wave (alpha, beta delta and theta) for calculating the
fast waves, such as alpha and beta, while the BSI feature is DAR and DTABR. Delta, theta, alpha and beta’s spectral
used to determine the symmetry of brainwave between the left power were over the range of frequency (1-3 Hz), (4-7 Hz), (8-
and right hemisphere. 13 Hz) and (14-20 Hz), respectively [5].
(1)
II. PATIENT & METHODS
(2)
This research was conducted and approved by the Eth-
ics Committees of the National Brain Center Hospital (RS A signal has to be passed in two DWT filtrations, the
PON), Jakarta. The EEG signals were obtained using the EEG low pass filter ( ) and high pass filter ( ) so that the frequen-
machine from Biologic Netlink System and Xltek EEG 32U, cies of the signal can be analyzed. The analysis of the frequen-
Natus. The recordings were sampled using frequency sampling cy component was done by using the result of resolution ( )
of 256 Hz and 512 Hz and saved into European Data Format after the signal passes through those filtrations, which usually
(.Edf). There were 57 subjects both of 30 normal patients and called as multi-resolution analysis [23]. The DWT of data
27 patients with AIS, which were confirmed by CT-scan.
sequence with multi resolution analysis is defined as:
Normal patients were the subjects who not experiencing a
[24]
stroke during EEG recording, while patients with acute
ischemic stroke are patients who suffered a stroke due to
blockage in the blood vessels of the brain within <72 hours (3)
after stroke attack. The inclusion criteria for this research
were: (1) the age of 40-60 years old, (2) not having seizures or where and are approximation coefficient and
dementia before and during stroke, (3) willing to provide in- detail coefficient at resolution of k with is time sequence.
formed consent, (4) willing to do both CT scan and EEG ex- While the parameters of and are defined as following
aminations. The examinations included a medical history and Equation:
other routine clinical examinations.
Electroencephalogram (EEG) is a bioelectric signal
coming from the electrical activity in the cortex or scalp sur- (4)
face, caused by the brain physiological activity [21]. In this

181

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:26:19 UTC from IEEE Xplore. Restrictions apply.
 182

The selection of families type is the substantial thing in

the Wavelet transformation. Daubechies 4 (db4) has the
orthogonal shape that approximates the EEG signal and has
the smallest error value in the EEG signal processing
compared to other types of mother wavelet [25]. MATLAB
provides several multi resolution wavelet analysis functions:
[C,L] = wavedec(x, N, ‘wname’)returns the wave-
let decomposition of the signal x at level N, using type of
wavelet family ‘wname’. This study decomposed EEG signal
at level 6 (for sampling frequency 256 Hz) and at level 7 (for
sampling frequency 512 Hz) using Daubechies 4 as a type of
Wavelet families (Fig. 2).
Meanwhile, BSI is used to measure the symmetry value
of left and right brain activities. The symmetry level value is
measured based on the average absolute of the power spectral Fig. 2. Signal decomposition using DWT level 6&7
density from both brain hemisphere in the frequency range of
1 to 25 Hz. The BSI that is used to look for abnormalities is and N samples which can be modeled mathematically by the
defined as the following Equation: [26] following Equation:

(5) (6)
where is the weight vector which connects the ith hidden
Where M : The number of Fourier coefficients, neuron and the output node, with is the weight vector input
: The total of channel pairs, and is the impact factor of the ith hidden neuron or
: EEG signal on the right hemisphere, threshold. [28]
: EEG signal on the left hemisphere
is equivalent to
The EEG signal which recorded from a 68-year-old
woman with acute stroke on the left hemisphere in a sleep where :
stage spectra showed a striking asymmetry of the waves and
the suppression of all frequency from all over the left hemi-
sphere [27]. Waveform asymmetry between the left and right (7)
hemisphere also shown during clamping of the right carotid by
the appearance of slow-wave on the right hemisphere [10]. and
Brain Symmetry Index (BSI) is one way to determine the
presence of ischemic brain to assess the symmetry of right and and (8)
left brain waves. Zero is the lowest value of BSI which is
infallible symmetry in all channels. While the highest value of
BSI is equaled 1, which implies maximal asymmetry [26]. The hidden layer output matrix of the neural network is
BSI feature’s values were obtained using Matlab called parameter H, which each column contains of ith hidden
toolbox " pxx = pwelch (x)" based on Welch method node output with reference to inputs x1, x2, ... , xN.
then calculated with Equation 5 to obtain estimates of the ELM was used to minimize the training error with
signal power at various frequencies. The Pxx (results Welch Moore Penrose Pseudo Inverse theory, so in order to obtain the
method) of the input signal (x) was found by dividing the output vector based on equation (3) and (4) is by
signal into overlap segment. In this study, one segment multiplying the inverse matrix ( . [28]
contains the data for 20 seconds. BSI value of each segment
(9)
Welch method results then averaged to be used as input
feature on artificial neural networks ELM along with the value ELM parameters such as input weight ( ) and hidden
of DAR and DTABR features to classify the level of stroke node biases ( ) were selected at random so that the ELM has
severity. a fast learning and able to produce high accuracy.
ELM is a type of feedforward neural network with a
single hidden layer with L hidden neuron, active function g(x)

182

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:26:19 UTC from IEEE Xplore. Restrictions apply.
 183

III. RESULT AND DISCUSSION matrix size (Fig.3), where n is the amount of data during the
30 minutes recording.
An example of the EEG recordings is shown in Table 1. The Relative Power Ratios (RPR) as a result of signal
Both the EEG machines have 32 channels but only used 9 decomposition were calculated using DWT. These ratios
pairs of channels are used, which are frontoparietal (FP1-FP2), compared the different frequencies spectrum of EEG, such as
central (C3-C4), frontal (F3-F4 and F7-F8), temporal (T1-T6), RPR alpha, RPR beta, RPR delta and RPR theta between
occipital (O1-O2), as well as parietal (P3-P4). EEG signals in normal patients and patients with ischemic stroke patients as
the form of .Edf then were converted to ASCII to n-by-18 shown in Fig. 4.

TABLE 1. CLINICAL CONDITIONS OF PATIENTS

183

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:26:19 UTC from IEEE Xplore. Restrictions apply.
 184

Fig. 3. EEG signal in European data Format (.edf) converted to matrix

Based on Fig. 4, it can be seen that there were

differences in RPR values between patients with AIS and
normal people. The delta RPR in patients with AIS tend to be
higher than in normal patients, while the RPR alpha and RPR
beta tend to be lower.
Fig. 5 shows the result of DAR’s and DTABR’s calcu-
lation in AIS patients, compared to normal. DAR and DTABR
in Fig. 5 indicate the ratio of slow wave appearance (delta and
theta) in patients with AIS tend to be higher than in normal
patients. The average values of BSI in patients with different
level of stroke severity are shown in Fig. 6. The highest value
of BSI was in severe stroke, while the lowest value of BSI was
in mild stroke. On the whole, BSI’s value increased along with
the increment of stroke severity, even though its values fluctu-
ated in each channel.
The following process after obtaining the feature value
was normalized all three feature values in the range of 1 and -
Fig. 4. Comparison of patients’ RPR 1. Normalization of data was done using MATLAB toolbox
with syntax: [y,PS]= mapminmax(x,-1,1), where x
was the input data and y was the result of data normalization.
The result of data normalization was used to trains the data to
obtain the input and output weight that used for classifying the
severity of stroke.
The results of testing data performed by a ELM training
algorithm using sigmoid activation function based on input
data in the form of featuers’ value DAR, DTABR and BSI as
shown in Table 2. Thus, based on 3-fold cross validation
analysis method using 36 data as a training data and 22 data as
a testing data, the minimum hidden neuron for achieving
100% training accuracy was 50 and the overall average value
of software/system accuracy was 72,73%. Detailed perfor-
mance measurement of accuracy, sensitivity, and specificity at
each level of severity is shown in Table 3 and Table 4.

Fig. 5. Comparison of DAR and DTABR of normal patients and patient

with acute ischemic stroke

184

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:26:19 UTC from IEEE Xplore. Restrictions apply.
 185

These accuracies were still below expectations because ACKNOWLEDGMENT

of the disproportionate amount of data among stroke's severity
levels. In this study, the number of patients with moderate AIS This study was supported by the Department of Education
(16 patients) were far more than the number of mild stroke (8 and Culture of the Republic of Indonesia
patients) and severe stroke (5 patients). Therefore, the system (0402/UN2.R12/HKP.05.00/2015) and University of Indonesia
performance and stability were not done by varying the ratio
of training and testing data, but by repeating the set data based
on k-fold cross-validation method. REFERENCES
[1] Badan Penelitian dan Pengembangan Kesehatan, “Riset Kesehatan Dasar
(RISKESDAS) 2013,” Lap. Nas. 2013, pp. 1–384, 2013.
IV. CONCLUSION [2] Stroke Association, “State of the Nation Stroke Stratistics,” no. January,
London: © Stroke Association, 2017 Stroke, 2017.
Discrete Wavelet Transform using Daubechies (DB4) [3] Intercollegiate Stroke Working Party, “National clinical guideline for
stroke, 5th edition,” London: Royal College of Physicians 2016, 2016.
as a mother wavelet and Brain Symmetry Index using Welch [4] R. J. Adams, G. Albers, M. J. Alberts, O. Benavente, K. Furie, L. B.
method followed by a feedforward neural network with Goldstein, P. Gorelick, J. Halperin, R. Harbaugh, S. C. Johnston, I.
Extreme Learning Machine could be used to classify the level Katzan, M. Kelly-Hayes, E. J. Kenton, M. Marks, R. L. Sacco, and L. H.
of stroke severity,
Fig. 6. BSI'seven though
value from thelevel
different highest accuracy
of stroke severity only Schwamm, “Update to the AHA/ASA recommendations for the
prevention of stroke in patients with stroke and transient ischemic attack,”
reaches 72% but with further development and more data
TABLE 2. PERFORMANCE MEASUREMENT FOR ACCURACY.
provided, hopefully this would be beneficial for the country,
Stroke, vol. 39, no. 5, pp. 1647–1652, 2008.
[5] R. V. A. Sheorajpanday, G. Nagels, A. J. T. M. Weeren, and P. P. De
especially patients in rural areas, where CT-scan facilities are Deyn, “Quantitative EEG in ischemic stroke: Correlation with infarct
limited. volume and functional status in posterior circulation and lacunar
syndromes,” Clin. Neurophysiol., vol. 122, no. 5, pp. 884–890, 2011.
[6] J. Claassen, F. S. Taccone, P. Horn, M. Holtkamp, N. Stocchetti, and M.
Oddo, “Recommendations on the use of EEG monitoring in critically ill
patients: Consensus statement from the neurointensive care section of the
ESICM,” Intensive Care Med., vol. 39, no. 8, pp. 1337–1351, 2013.
[7] R. L. Sacco, R. Adams, G. Albers, M. J. Alberts, O. Benavente, K. Furie,
L. B. Goldstein, P. Gorelick, J. Halperin, R. Harbaugh, S. C. Johnston, I.
Katzan, M. Kelly-Hayes, E. J. Kenton, M. Marks, L. H. Schwamm, and T.
Tomsick, Guidelines for prevention of stroke in patients with ischemic
TABLE 3 PERFORMANCE MEASUREMENT WITH HIDDEN stroke or transient ischemic attack: a statement for healthcare
NEURON 50. professionals from the American Heart Association/American Stroke
Association Council on Stroke: co-sponsored by the Council on C, vol.
113. 2006.
[8] Anindito, “Expert-System Based Medical Stroke Prevention,” J. Comput.
Sci., vol. 9, no. 9, pp. 1099–1105, 2013.
[9] M. K. Kiymik, I. Güler, A. Dizibüyük, and M. Akin, “Comparison of
STFT and wavelet transform methods in determining epileptic seizure
activity in EEG signals for real-time application,” Comput. Biol. Med.,
vol. 35, no. 7, pp. 603–616, 2005.
[10] K. G. Jordan, “Emergency EEG and continuous EEG monitoring in acute
ischemic stroke.,” J. Clin. Neurophysiol., vol. 21, no. 5, pp. 341–52, 2004.
[11] S. K. Wijaya, C. Badri, J. Misbach, T. P. Soemardi, and V. Sutarmo,
“Electroencephalography (EEG) for Detecting Acute Ischemic Stroke,” in
2015 4th International Conference on Instrumentation, Communications,
Information Technology, and Biomedical Engineering (ICICI-BME),
2015, vol. 4, pp. 42–48.
[12] S. Finnigan and M. J. A. M. van Putten, “EEG in ischaemic stroke:
Quantitative EEG can uniquely inform (sub-)acute prognoses and clinical
management,” Clin. Neurophysiol., vol. 124, no. 1, pp. 10–19, 2013.
[13] M. J. A. M. Van Putten and D. L. J. Tavy, “Continuous quantitative EEG
monitoring in hemispheric stroke patients using the brain symmetry
index,” Stroke, vol. 35, no. 11, pp. 2489–2492, 2004.
[14] S. Bhattarai, Z. Xiao-ning, and T. Tuerxun, “Neurology &
Neurophysiology EEG and SPECT Changes in Acute Ischemic Stroke,” J
TABLE 4. AVERAGE PERFORMANCE MEASUREMENT WITH Neurol Neurophysiol, vol. 5, no. 2, 2014.
HIDDEN NEURON 50 [15] L. Murri, S. Gori, R. Massetani, E. Bonanni, F. Marcella, S. Milani,
“Original article Evaluation of acute ischemic stroke using quantitative
EEG : acomparison with conventional EEG and CT scan,” pp. 249–257,
1998.
[16] N. Thakor, Quantitative EEG analysis methods and clinical applications.

185

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:26:19 UTC from IEEE Xplore. Restrictions apply.
 186

2009. [23] A. Subasi, “EEG signal classification using wavelet feature extraction and
[17] W. R. W. Omar, N. Fuad, M. N. Taib, R. Jailani, R. M. Isa, Z. Mohamad, a mixture of expert model,” Expert Syst. Appl., vol. 32, no. 4, pp. 1084–
and Z. Sharif, “Brainwave Classification for Acute Ischemic Stroke 1093, 2007.
Group Level Using k-NN Technique,” 2014 5th Int. Conf. Intell. Syst. [24] A. K. Verma and A. K. Mangaraj, “Analysis and Classification of
Model. Simul., pp. 117–120, 2014. Electroencephalography Signal,” National Institute of Technology,
[18] S. Huang, Q. Shen, and T. Q. Duong, “Quantitative prediction of acute Rourkela, 2010.
ischemic tissue fate using support vector machine,” Brain Res., vol. 1405, [25] S. Z. M. Tumari, R. Sudirman, and A. H. Ahmad, “Selection of a Suitable
pp. 77–84, 2011. Wavelet for Cognitive Memory Using Electroencephalograph Signal,”
[19] N. Hema Rajini and R. Bhavani, “Computer aided detection of ischemic Engineering, vol. 5, no. 5, pp. 15–19, 2013.
stroke using segmentation and texture features,” Meas. J. Int. Meas. [26] C. C. De Vos, S. M. Van Maarseveen, P. J. a M. Brouwers, and M. J. a M.
Confed., vol. 46, pp. 1865–1874, 2013. Van Putten, “Continuous EEG Monitoring During Thrombolysis in
[20] O. N. Rahma, S. K. Wijaya, and C. Badri, “Electroencephalogram Acute,” J. Clin. Neurophysiol., vol. 25, no. 2, pp. 77–82, 2008.
Analysis With Wavelet Transform and Neural Network As a Tool for [27] L. J. Hirsch and R. P. Brenner, Atlas of EEG in Critical Care, vol. 53.
Acute,” IASTEM -The 9th Int. Conf. Medical, Biol. Pharm. Sci., no. 3, pp. John Wiley & Sons, Inc., 2010.
108–113, 2016. [28] G.-B. Huang, Q. Zhu, C. Siew, G. H. Ã, Q. Zhu, C. Siew, G.-B. Huang,
[21] E. C. Djamal and H. A. Tjokronegoro, “Identifikasi dan Klasifikasi Sinyal Q. Zhu, and C. Siew, “Extreme learning machine: Theory and
EEG terhadap Rangsangan Suara dengan Ekstraksi Wavelet dan Spektral applications,” Neurocomputing, vol. 70, no. 1–3, pp. 489–501, 2006.
Daya,” vol. 37, no. 1, pp. 69–92, 2005.
[22] B. Shoelson, “edfRead,” MatlabCentral, 2011. .

186

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:26:19 UTC from IEEE Xplore. Restrictions apply.