AN EVALUATION OF STROKE SEVERITY

UTILISING A FIRST-ORDER MULTILAYER NEURAL

NETWORK USING EEG DATA.

Prasath Alias Surendhar S 1 , Sandhiya R 2 , Bhuvana R 3

Associate Professor , 1 , 2Assistant Professor 3
Department of Biomedical Engineering, 1,2,3
Aarupadai Veedu Institute of Technology, 1, 2
Vinayaka Missions Research Foundation ( DU), Salem, Tamilnadu, India. 1, 2
G.K.M College Of Engineering And Technology , Perungalathur , Chennai ,
Tamilnadu 3
[Link]@[Link] 1, [Link]@[Link] 2, bhuvana2089@[Link] 3

Abstract. Acute Ischaemic Stroke (AIS) is to get this result.

one sort of stroke that happens the most. 1. Introduction
Stroke is the leading cause of mortality. It Stroke is a disease that cannot be
can restrict blood flow and deprive the transmitted from one person to another,
brain of oxygen. Patients can receive and it has become one of the major causes
medical care more quickly if they are of death worldwide. According to a report
diagnosed early, which can help prevent published by The Institute for Health
harm to the brain. An Metrics and Evaluation (IHME) in 2017, it
electroencephalogram (EEG) is a different is the leading cause of mortality,
method for detecting AIS than the premature death, and death and disability
traditional methods, such as MRI or CT combined in Indonesia. Strokes can be
scan. In this study, we attempt to classified into two categories: ischaemic
categorise the severity of strokes using a and hemorrhagic. Acute ischaemic stroke
one-dimensional convolutional neural (AIS) is present in over 80% of cases.
network (CNN). The suggested technique When a clot or other blockage closes the
analyses the power spectral density (PSD) channel within an artery that travels to the
of EEG recordings from normal and stroke brain, it lowers blood flow and deprives
individuals, as the model’s inputs and the brain of oxygen. This condition is
extracts feature automatically using CNN. referred to as AIS.
The final feature maps were trained in a In order to reduce the negative effects on
fully connected layer to categorise four the patients, it is important to make an
classes: normal, mild, moderate, and early prediction. Computed tomography
severe stroke. The study is carried out in (CT scan) and magnetic resonance
order to get the best feasible computation imaging (MRI) are the two most frequent
time with an accuracy of 97.3%. This is diagnostic technologies. They are both
done by using 64-second segmentation and capable of producing fairly reliable
50 convolutional filters with a kernel size findings. There are still some small
of 1x120. The EEG information from four possible dangers associated with exposing
channels—C3, C4, O1, and O2—was used the patient's body to X-rays, even when the

1
safest protocols are utilised to conduct the discovered that there was a positive
surgery. In Indonesia, big hospitals still correlation between the Brain Symmetry
have limited access to the equipment since Index (BSI) and the National Institute of
it is expensive to use and maintain. In Health Stroke Scale (NIHSS) scores [4].
addition, these technologies may take Finnigan et al. used quantitative
many hours to provide findings, while electroencephalography (QEEG) to
critical first help may need a rapid reply measure the Delta/Alpha Ratio (DAR) and
[1]. relative alpha power. They discovered that
In order to meet their needs, community every one of these parameters was
hospitals require several types of associated with the NIHSS score [5].
diagnostic technology, such as CT and Osmalina combined DAR, DTABR, and
MRI machines. The proposed solution is to BSI [6]. In order to enhance accuracy
utilise an electroencephalogram (EEG), while using fewer channels and the same
which is both safer and less expensive. information, Fitriah adopted Principal
During an EEG, electrodes are placed on Component Analysis [7]. Most of the
the patient's head to collect brain activity study was carried out using handmade
without the need for any invasive features, which need specialised expertise
procedures. The electrodes are able to in the field to be attained in the first place.
detect these electrical impulses, which are In this study, the classification job was
generated by the functions of the human carried out without the handmade features,
brain. and the learning process was accomplished
We want to use brain wave signals to using one of the deep neural network
identify people who have had an ischaemic approaches, namely the Convolutional
stroke since the cerebral blood flow (CBF) Neural Network (1D-CNN).
of stroke patients is different from that of Deep learning is a cutting-edge machine
healthy persons. For individuals who have learning (ML) method that automatically
experienced an ischaemic stroke, the CBF encodes a hierarchy of characteristics and
value typically ranges from 50 to 70 mL • is customised to the input [8]. This
100 g-1 • min-1. That said, the CBF value capability enables the model to learn only
can fall below 25 to 30 mL • 100 g-1 • from "raw" data. Convolutional neural
min-1 [2]. These changes are closely networks (CNNs) [1] are one of the most
connected to the EEG signals because the often utilised types of neural networks in
rapid wave activities, such as beta and deep learning. In recent years, CNN has
alpha frequencies, are decreasing while the achieved a great deal of success in the
slow wave activities, such as delta and field of computer vision. This is due to its
theta frequencies, are rising [1]. ability to extract local features from
nearby objects in the input data, where the
For example, some studies are done to same patterns may emerge in various
observe the pattern of these signals in locations. The number of learnable
order to find out if a person has had an parameters may be decreased to the size of
ischaemic stroke or not. Omar et al. used the kernel by utilising a filter kernel, and
the Relative Power Ratio (RPR) delta, this is not influenced by the input data. On
theta, alpha, and beta to classify three the other hand, as the amount of input
different stages of stroke [3]. Putten increases, the number of parameters in a

2
normal shallow neural network also frequency analysis of the EEG is a simpler
increases. method for identifying this phenomenon
In addition to being used with picture data, than the time domain analysis. It is
CNN models may also be generated with recommended to utilise the power spectral
one-dimensional data, such as EEG density (PSD) for spectrum analysis since
signals. There are just a few research that EEG signals represent a stochastic process.
are related to 1D-CNN for EEG data. For This is due to the fact that the PSD
example, there is research on averages the signal instead of utilising a
automatically grading sleep phases [9], straightforward FFT.
extracting features for categorising
epilepsy [10], detecting and diagnosing There are several different methods that
seizures automatically [8], and identifying may be utilised to estimate PSD. In this
ischaemic strokes [11]. analysis, we used the Welch technique,
2. Method which has the main advantage of being
This study included EEG data from strong enough to guarantee that the
previous studies [6], [7], [11]. The data estimated power spectral density (PSD)
was acquired from 31 healthy people and does not include any incorrect frequency
27 patients who had an acute ischaemic peaks. On the other hand, the windowing
stroke. The Pusat Otak Nasional (PON), or method may cause the resulting PSD to
National Brain Centre Hospital, located in become distorted, which is a drawback.
Jakarta, was responsible for gathering the The Welch approach is made up of the
data. Two different tools were utilised: following steps: dividing signals into a
Xltek was utilised for a portion of the data, number of overlapping sections, acquiring
and Biologic was used for the rest. With periodograms for each section, and then
the exception of a few subjects, who were averaging the periodograms to get spectral
sampled at a rate of 256 Hz, the sample estimates. The length of the window that is
rate for the acquisitions was 512 Hz. Using used in segmentation has an effect on the
the International 10-20 approach for frequency resolution. As a result, a bigger
electrode insertion, it took around 30 window will produce a smaller frequency
minutes to capture EEG data for all peak. This research is centred on four
patients. The data was then stored in the frequency bands that are lower than 20 Hz.
European Data Format (.edf file). We were As a result, we reduce the 512 Hz to 32 Hz
limited to choosing only four channels in order to decrease the amount of
from the many that were available: C3, C4, computing needed. Welch's PSD is
O1, and O2. This was based on the calculated with a window length of 64 and
channel reduction research that Fitriah [7] an overlap of 50%. The CNN model is
did using PCA. then given the estimated PSDs, which are
referred to as "raw inputs."
In people with acute ischaemic stroke Before calculating the PSD, the EEG of
(AIS), the CBF is generally lower than that one subject is divided into several portions
of non-AIS participants. As a result, that are shorter in duration and equal in
slower wave frequencies, such as delta and length. This is done in order to have a big
theta, appear, whereas higher frequencies, number of inputs. The length of each
such as beta and alpha, disappear [1]. The segment is then modified such that the

3
optimal duration that delivers the best Figure 1. A block diagram showing the
degree of accuracy may be determined. steps required in preprocessing
After segmentation, the new inputs are
split into two sets: a training set that makes
up 80% of the total and a testing set that
makes up 20% of the total. The schematic
of the preprocessing operation is shown in
Figure 1.
This method is distinct from previous
studies since it automatically extracts
characteristics using a learning algorithm
instead of deriving them from manual
computations that need some subject
experience. The CNN architecture is
responsible for both the feature extraction
and classification stages. A convolutional Figure 2. The 1D-CNN architecture is made up of one
neural network (CNN) is made up convolutional
of layer with 50 1x40 kernel filters, followed by a
layers that are predominantly
Rectified Linear Unit (ReLU) activation, one pooling layer that
convolutional layers (CL) with 50 filters outputs the maximum value of two adjacent elements, and one
and a kernel size of 1x120, pooling layers fully connected layer that converts 2250 flattened data into
(PL) with a scaling factor of 1/2, and fully
four stroke detection classes.
connected (FC) layers, in that order. The
training would be completed in around 300 3. Result and Discussion
epochs. For the sake of visualisation, Figure 3 shows a part of the PSD that was
Figure 2 displays the architecture of the acquired from the EEG for each lesson. As
CNN, which has one CL and one PL. The the severity of strokes has grown, the
created model is implemented using the frequency of the second peak for each
Keras Python library. class has reduced. This supports the notion
that Delta and Theta waves would be the
most prevalent during a stroke, whereas
Beta and Alpha waves would become less
prominent. In most courses, the second
peak often occurs at a frequency that is
greater than 12 Hz. After that, it gradually
changes to below 12 Hz for every kind of
stroke. The second peak frequency will
move further to the left when the NIHSS
score increases or the severity of the stroke
worsens.

The overall accuracy for each segment

variation is shown in Table 1. The
accuracy of the data is influenced by the

4
length of the segment. As more data because of the downsampling process of
becomes available for training, we may smaller segments, which would distort the
expect to see improvements in accuracy. actual signal. The best accuracy was
However, the accuracy started to decrease attained in 64 s segment.
when the duration was 64 seconds or less

Figure 3. Illustrations of PSD acquired for normal individuals, mild, moderate, and severe
strokes.

Table 1. Number of Data and Testing Accuracy from Segmentation Variation

Segment Length (s) #Training #Testing Accuracy (%)
1024 48 12 58.3
512 136 34 79.4
256 316 79 89.9
128 644 161 93.8
64 1316 329 97.3
32 2305 577 96.8
16 5300 1325 94.6
The performance of the classification model can be seen from the confusion matrix shown in
Table 2-5 for some segmentation interval. Each class has different accuracy shown in Figure 4.
Above 256 s segmentation, the available data for testing are very limited and the per-class-
accuracy drop significantly, except for the mild class which is predicted correctly but the number
of testing set is not enough to represent the performance.

5
 Table 2. Confusion Matrix of 128 s Segment
Predicted Class
Normal Mild Moderate Severe
Normal 87 0 2 3
Actual Mild 2 22 0 0
Class
Moderate 1 2 28 0
Severe 0 0 0 14

Table 3. Confusion Matrix of 64 s Segment

Predicted Class
Normal Mild Moderate Severe
Normal 41 1 3 0
Actual Mild 0 7 2 0
Class
Moderate 1 0 15 0
Severe 0 1 0 8

Table 4. Confusion Matrix of 32 s Segment

Predicted Class
Normal Mild Moderate Severe
Normal 17 0 0 0
Actual Mild 0 2 3 0
Class
Moderate 4 0 3 0
Severe 0 0 0 5

Figure 4. Accuracy of Each Class: Normal, Mild, Moderate and Severe Stroke

The accuracy are calculated from architecture in Fig 2 which consists of 1 CL and 1 PL. For
additional CL-PL in 128 s segment data, number of parameters and accuracy are changed. Second CL
uses 20 filters with 1x20 kernel size, and the third uses 15 filters with 1x12 kernel size. The accuracies
for using 1, 2 and 3 consecutively are 93.8%, 93.2% and 95.8%. Here we can see using only 1 CL still
gives good accuracy where the trainable parameters are nearly half of using more CL. To reduce the
trainable parameters, we need to decrease number of filters or increase the kernel size. Here we vary

number of filters into 5, 10, 20, 40, 50, 80 and the consecutive accuracies are 84.8%, 89.8%, 92.5%,
6
91.9%, 93.8%, 91.4%. It turns out that the maximum accuracy is reached when 50 filters are used.
With the same filter number, we vary kernel size to 20, 40, 80 and 120 data, and the accuracies
achieved are 90.1%, 92.6%, 93.6%, 93.4%, and 93.8%.
The execution time of training for each segment variation are 115.8 s, 59.6 s, 31.0 s, 14.3 s, 7.8 s,
4.4 s, and 2.8 s. consecutively from the shortest to the longest segment. This time is nearly
proportional to the number of input data.

1. Conclusion
From varying the length of segmentation in preparing the PSD calculation, number of CL, number of
filter and kernel size the optimal accuracy achieved is when using 64 second segmentation, 1 CL, 50
filters and 1x120 kernel size. 97.3% accuracy is high enough considering the channels used in this
study are only C3, C4, O1, and O2.

References
[1] M. A. Nielsen, Neural Networks and Deep Learning. San Fransisco: Determination Press, 2015.
[2] K. G. Jordan, “Emergency EEG and Continuous EEG Monitoring in Acute Ischemic Stroke,” vol.
21, no. 5, pp. 341–352, 2004.
[3] W. R. W. Omar et al., “Acute Ischemic Stroke Brainwave Classification Using Relative Power
Ratio Cluster Analysis,” Procedia - Soc. Behav. Sci., vol. 97, pp. 546–552, 2013.
[4] 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.
[5] S. P. Finnigan, M. Walsh, S. E. Rose, and J. B. Chalk, “Quantitative EEG indices of sub-acute
ischaemic stroke correlate with clinical outcomes,” Clin. Neurophysiol., vol. 118, no. 11, pp.
2525–2532, 2007.
[6] O. N. Rahma, S. K. Wijaya, Prawito, and C. Badri, “Electroencephalogram analysis with extreme
learning machine as a supporting tool for classifying acute ischemic stroke severity,” Proc. -
2017 Int. Semin. Sensor, Instrumentation, Meas. Metrol. Innov. Adv. Compet. Nation,
ISSIMM 2017, vol. 2017-Janua, pp. 180–186, 2017.
[7] N. Fitriah, S. K. Wijaya, M. I. Fanany, C. Badri, and M. Rezal, “EEG channels reduction using
PCA to increase XGBoost’s accuracy for stroke detection,” AIP Conf. Proc., vol. 1862, no.
July, 2017.
[8] I. Ullah, M. Hussain, E. Qazi, and H. Aboalsamh, “An Automated System for Epilepsy Detection
using EEG Brain Signals based on Deep Learning Approach Insight Centre for Data Analytics
, National University of Ireland , Galway , Ireland Visual Computing Lab , Department of
Computer Science , College of Com,” Expert Syst. Appl., no. 107, pp. 61–71, 2018.
[9] O. Tsinalis, P. M. Matthews, Y. Guo, and S. Zafeiriou, “Automatic Sleep Stage Scoring with
Single-Channel EEG Using Convolutional Neural Networks,” 2016.
[10] T. Wen and Z. Zhang, “Deep Convolution Neural Network and Autoencoders-Based Unsupervised
Feature Learning of EEG Signals,” IEEE Access, vol. 6, pp. 25399–25410, 2018.
[11] E. P. Giri, M. I. Fanany, A. M. Arymurthy, and S. K. Wijaya, “Ischemic stroke identification
based on EEG and EOG using ID convolutional neural network and batch normalization,”
2016 Int. Conf. Adv. Comput. Sci. Inf. Syst. ICACSIS 2016, pp. 484–491, 2017.