A Modified XG Boost Classifier Model for Detection of Seizures
and Non-Seizures
RAVEENDRA KUMAR T. H.1, NARAYANAPPA C. K.2, RAGHAVENDRA S.3, POORNIMA G. R.4
1Ramaiah Institute of Technology, VTU Research Centre, Bengaluru, Karnataka, INDIA
2Department of Medical Electronics, Ramaiah Institute of Technology, Bengaluru, Karnataka, INDIA
3Department of CSE, School of Engineering & Technology, CHRIST Deemed to be University, Bengaluru, INDIA,
4 Department of ECE, SVCE, Bengaluru, Karnataka, INDIA
Abstract: Diagnosis of Epilepsy is immensely important but challenging process, especially while using
traditional manual seizure detection methods with the help of neurologists or brain experts’ guidance which are
time consuming. Thus, an automated classification method is require to quickly detect seizures and non-seizures.
Therefore, a machine learning algorithm based on a modified XGboost classifier model is employed to detect
seizures quickly and improve classification accuracy. A focal loss function is employed with traditional XGboost
classifier model to minimize mismatch of training and testing samples and enhance efficiency of the classification
model. Here, CHB-MIT SCALP Electroencephalography (EEG) dataset is utilized to test the proposed
classification model. Here, data gathered for all 24 patients from CHB-MIT Database is used to analyze the
performance of proposed classification model. Here, 2-class-seizure experimental results of proposed
classification model are compared against several state-of-art-seizure classification models. Here, cross validation
experiments determine nature of 2-class-seizure as the prediction is seizure or non-seizure. The metrics results
for average sensitivity and average specificity are nearly 100%. The proposed model achieves improvement in
terms of average sensitivity against the best traditional method as 0.05% and for average specificity as 1%. The
proposed modified XGBoost classifier model outperforms all the state-of-art-seizure detection techniques in
terms of average sensitivity, average specificity.
Keywords: Epilepsy, Seizure detection, XGboost Classifier, CHB-MIT dataset, EEG data.
Received: March 27, 2021. Revised: November 16, 2021. Accepted: December 12, 2021. Published: January 4, 2022.
1 Introduction:
Currently, one of most general and deadliest chronic
disorder of brain is Epilepsy which causes due to the
unexpected and unusual transient disorders of brain
neurons. Epilepsy affects at least 1% of total
population of world [1]. Epilepsy is a temporary
neuronal disease of brain which can last up to several
months or years. Epilepsy word is taken from
‘epilepsia’ which is a Latin and Greek word. The real
meaning of the word ‘epilepsia’ is ‘seizure’ or ‘to
seize upon’. Furthermore, An Epileptic seizure also
known as seizure which is caused due to sudden
uncontrolled electrical activity between brain cells
(also called neurons or nerve cells) that causes
abnormalities in muscle tone or movements
(stiffness, twitching or limpness), behaviors,
sensations or states of awareness which lasts for only
a limited period of time. The term epilepsy can be
dated back to the Babylonian text on medicine (3000
years ago).Epilepsy effects not only humans but also
other species of mammals as well ex. Dogs,
Elephants etc., it is one of the most common
neurological disorder that affects more than 50
million people worldwide [2].
Furthermore, Seizures can be of two types, provoked
and unprovoked i.e., some seizures can be provoked
due to a temporary event such as low blood sodium,
low blood sugar etc., and unprovoked seizures are
those which starts without a known cause.
Unprovoked seizures are likely to be triggered by
stress, diseases of the brain or lack of sleep. When
there has been at least one seizure and a long term risk
of further seizures is known as epilepsy. Epilepsy is a
chronic non-communicable disease. Epilepsy
accounts for 0.5% of the global burden of disease.
Provoked seizures occur in about 3.5 per 10,000
people a year while unprovoked seizures occur in
about 4.2 per 10,000 people a year. After one seizure,
the chance of experiencing a second is about 50%.
Epilepsy affects about 1% of the population at any
given time with about 4% of the population affected
at some point in time. Nearly 80% of those with
epilepsy live in developing countries.
One of the common way to determine the onset of a
seizure before it manifests completely is by using the
analysis of the scalp electroencephalogram (EEG) a
noninvasive(not involving the introduction of
instruments into the body), multi-channel recording
of the brain’s electrical activity. Although invasive
electrodes are sometimes used, as in
electrocorticography, sometimes called intracranial
EEG. EEG is most often used to diagnose epilepsy,
which causes abnormalities in EEG readings. Clinical
EEG recording is usually for about 20–30 minutes
(plus preparation time). Furthermore, EEG is utilized
for the identification of electrical activities of the
brain which can be done by attaching electrodes
(metal discs) to the scalp. Usually, EEG is employed
to diagnose brain disorders by detecting disturbances
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
14
Volume 19, 2022
or changes in brain activities, especially, in case of
epilepsy or another especially epilepsy or another
seizure disorder. EEG characteristics vary among
patients. EEG of a patient with seizure may show
same patterns in the EEG of another patient. Some
EEG monitoring can last up to few hours or even days
and because of this when someone interprets the data
i.e., human intervention it is prone to errors and a lot
of time is wasted.
However, identification of seizures manually is very
challenging and critical due to it requires large period
of time for precise analysis of EEG signals through
visual inspection. Usually, an approximate of 1.20
GB of data is produced by an 18-channel, 36-h digital
recorder and is almost equal to 20 thousand pages of
traditional paper EEG data and it becomes difficult to
review the huge amount of data and its get even more
complicated when the number of channel increases.
Furthermore, EEG contains certain artifacts
(electrical activities arising from sites other than the
brain) and these cause errors by visual inspection of
EEG by experts. Hence automatic methods are being
developed to detect and predict the seizure and is in
high demand for clinical application.
Therefore, in this article, a machine learning
algorithm is employed to detect seizure quickly and
with high accuracy when compared to the previous
methods of seizure detections. The main goal of this
paper is to discover the seizure and epilepsy status
using the prediction algorithm on the test results
received from patient medical reports. Furthermore,
the timely detection of seizures can automatically
play an important part in epilepsy diagnosis. The
identification of seizures and non-seizures in patients
and seizure status knowledge can provide great helps
towards future neurological applications. Therefore,
a novel seizure detection algorithm is presented. This
novel algorithm utilizes modified XGboost classifier
which is modified by using focal loss function to give
better accuracy and results when compared to the
other state-of-art-classification techniques. Here,
seizures are detected for some specific patients from
the available dataset for few seconds. The number of
non-seizure patients are more in contrast to their
counterpart seizure patients. The focal loss function
is utilize to reduce discrepancy between seizures and
non-seizures in classification process. Here, focal
loss function can easily handle the differences of
binary classification operations. Here, machine
learning techniques make implementation of
proposed modified XGboost classifier faster and
efficient. Moreover, the performance results are
evaluated for several patients and compared with
various state-of-art-techniques in terms of sensitivity,
specificity and classification accuracy.
This paper is presented in the following manner.
Section 2, describes about the related work presented
regarding automated detection of seizures and
detection issues and how those issues can be handle
with the help of the proposed epilepsy model based
on machine learning techniques. Section 3, discusses
about the methodology proposed for the effective
implementation of proposed epilepsy model for the
classification analysis of epilepsy. Section 4
discusses about the simulation results and their
comparison with state-of-art classification algorithms
and section 5 concludes the paper.
2 Related Work:
There are various application of machine learning in
different fields of engineering and a significant
development can be seen on health sector and can be
applied on biological data sets for better
outcomes[1][2].Machine learning is also used to find
insights and patterns from different datasets from
different domains[3][4].Applications of machine
learning can also be seen on brain datasets for seizure
detection, epilepsy lateralization, differentiating
seizure sates, and localization [5][6][7][8].
In paper [9-11], feature extraction was not used and
the data was further processed for deep learning
models which was trained with raw EEG signals.
Feature extraction is an important step which can ease
the way to give input to the classifier, but in the
mentioned papers they skipped the process of
classification due to its complexity and fed the raw
data samples to the classifiers. One of the main
difficulties of seizure detection or prediction is that to
find the correlation that is to which EEG timestamp
to be input to the classifier but this process had it
major downside was even with feature extraction
ambiguity it didn’t recognize the patterns of the
temporary signal.
Many machine learning methods for determining
epilepsy collect the emotional condition from the
brain by using an algorithm called as bayes classifier
which contains 1902 statistical and 23 EEG signals
people of age between 10-15 were collected.
Moreover[12] by using wavelet DB Four shanon’s
entropy researches extracted unique features from the
subjects and the method consisted of 4 levels and
when the signal was obtained the features from it was
extracted and a new software was developed called so
which was pre trained to record the changes in the
brain action and by this process the accuracy was
around 75% but this process had it drawbacks such as
that the features were not universal for classification
because there was difference in individual signals.
Seizure prediction is mainly dependent on 2
components one is extracting the requires features
and classifying them and features plays an important
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
15
Volume 19, 2022
role because we need to differentiate the various EEG
signal patterns and by this well get better
classification results on EEG signals based on this
[13] proposed a model which was divided into 2
domains spectral and temporal domain features. This
method is useful even though it differs from various
involving technology.
An algorithm which depends on univariate features
and uses it for machine learning known as
ASPPR(Advanced seizure Prediction via Pre-lctal
Relabeling) [14] and in this process 34 features were
used considering its non-liner dynamics and energy
in which 14 of the features were used to compare
algorithms which in the past used these features and
the rest 20 were constructed on EEG statistical
descriptors and spectral band power which is
calculated over the standard EEG bands and spectral
frequency, “time in advance predictive model” was
introduced and this model used to learn during
training and used to predict the seizure only drawback
of this model was the prediction time was not
accurate and overall accuracy was not satisfying .
In this paper [15], spectral features of intracranial
EEG is patient specific and further trained using
machine learning algorithms a total of 18 patients
data were taken. The noise from the data was
removed between 50Hz and 100Hz using BPS (band
pass filter) and minimized the dominance of low
power frequency power band and power was
normalized across the spectrum. Discriminant
analysis called kernel fisher was used to get best
feature for testing. But the problem with this method
is that it didn’t specify the seizure time and it used pre
optimized parameters.
In [16], Fourier transforms is employed which has a
huge application in detecting EEG since it is a signal
processing method using it can be used to extract the
features. As the amplitude increases it show the
greater the abnormality in the brain hence here is
where the Fourier transforms can be used hence
author used Fourier transforms to extract some
features and complex features by signal processing.
EEG-based epileptic signal classification which
relies on stacked generalization model. In this paper
[17], 5 types of epileptic classification is conducted
with a 20 min scale and various levels of EEG signals
are studied and here the stack generalization model is
developed over a multiple CNNs with various
activation functions are used weighted algorithm and
feature fusion was used. But the drawbacks this
method faced was every methods suffered from
reduction in classification accuracy when applied to
states classifications.
A Unified multi-view deep learning framework was
developed for automatic EEG seizure detection [18],
using clinical scalp multi-channel EEG epilepsy
dataset. Here end to end framework is created which
can learn multi-view hidden representations by
combing inter and intra correlations of EEG channels
and a 2D spectrogram is obtained and further the
features are extracted using deep learning. As this
method is useful in other medical task which has
almost same data structures, but here channel
awareness is still an unsolved problem.
3 Modelling for proposed Modified
XGboost classifier Model:
This section discusses about the mathematical
modelling of proposed Modified XGboost classifier
Model for the identification of seizure onsets quickly
and with high accuracy. In this section, traditional
XGboost classifier is modified with the help of focal
loss function. Generally XGBoost is composed
additive learning method of second order
approximation. Furthermore, here, the 1st order
derivative is called as “gradient” and 2nd order
derivative is called as “hessian” and the loss function
is required to fit the model. Further, following section
demonstrates the mathematical representation of
proposed Modified XGboost classifier Model.
Further, XGboost is a gradient tree boosting approach
which is utilized for handling machine learning
problems. The key idea behind gradient tree boosting
approach is the summation of several tree classifiers.
3.1 Modelling for proposed XGboost
Classifier:
Consider for given number of training samples,
number of generated features are and represented
by the following equation,
(1)
Where, is expressed by , is expressed
by and. Furthermore, traditional
XGboost tree model utilizes additive functions to
estimate the desired result. Then,
(2)
Here, R is expressed as where
represents regression tree space.
Then, pattern of every regression tree is denoted by
which can be used for mapping training samples to
the respective leaf index. The total number of leaves
present in the tree are expressed by. Each
belongs to an individual regression tree pattern and
weights of leaf. Then, every regression tree
provides a constant score on every leaf, unlike the
nature of decision trees. Here, score is represented for
leaf using weights of leaf. For a given
training sample, classification process for leaves is
achieved by following decision procedures and
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
16
Volume 19, 2022
summation of scores which is obtained from weights,
gives the final estimated output for the respective
leaves. Then, the group of functions utilized in this
tree model are given by regularized function using
following equation,
(3)
Where, complexity function for the regression tree
model is defined by,
(4)
Where, is utilized for the evaluation of change
between the estimation and the original and
expressed as convex loss function which can be
differentiated. The smoothness of regularized
function on final estimated weights is achieved with
the help of complexity function to discard over-
fitting. Here, the regularized function selects a
regression tree model which has simple estimated
functions. Here, regression tree is modelled in such a
way that the model can easily parallelize which
improves the efficiency of model unlike other tree
models.
Here, the functions of regression tree model which
shown in equation (3) are difficult to optimize with
the help of traditional optimization approach.
Therefore, regression model is trained in adaptive
mode. Then, assume that considering
iteration, estimated output is
for case,
parameter is required to optimize regularized
function,
(5)
Here, the parameter is used to enhance the
performance efficiency of regression tree model.
Further, second order approximation is performed for
the faster optimization of regularized function which
is demonstrated in below equation,
(6)
Where, gradient statistics of first order and second
order approximation considering loss function are
denoted by and. Here, is expressed as
and is expressed
as
. After simplifying
equation (7) by eliminating constant terms, we get,
(7)
Then, for leaf case set, determine as,
(8)
Then, by simplifying equation (7), we get,
(9)
(10)
Then, the final optimized weights
for leaf can
be evaluated considering a fixed pattern as,
(11)
Then, determine their respective final optimized
value by following equation,
(12)
Where, the quality of tree patterns can be determined
using scoring function which is demonstrated in
above equation (12). This score is used for
classification of tree models and evaluated for
extensive range of regularized functions. Here, the
proposed tree model classify first leaf of the
regression tree and then adds other tree leaves.
Consider that is case set for left side node and
is the case set for right side node after the split.
Assume that then loss minimization
term after the split is given by following equation,
(13)
Here, equation (13) can be utilized for evaluating split
candidates in tree model. The proposed tree model is
utilized for multi-class classification process as well
by combining classification of binary trees.
Equation (14) shows the property of the sigmoid
function and it is used for further derivation of loss
function,
(14)
3.2 Modification derived for proposed
XGboost Classifier Model:
A Modified XGBoost is proposed which uses Focal
losses for classification for binary dataset to reduce
mismatching of training and testing samples in
classification process which generally affects the
prediction accuracy. Since XGBoost is modified
version of tree-boosting, its efficiency enhances to a
high extent. It is used in various fields of study such
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
17
Volume 19, 2022
as medical record analysis or for cancer diagnosis or
for epilepsy while detection of seizures. Thus, Binary
Focal Loss is given by following equation,
(15)
In equation (15), when is set to 0, then above
equation is turned into ordinary cross entropy loss. To
obtain just the cross entropy loss, the sigmoid
activation function can be utilized which is shown in
above equation (14) and using its property, first
derivative of the focal loss can be obtained by using
equation (16) as,
(16)
Then set to 0 in equation (16) to determine cross
entropy loss. On further simplification of equation
(16), we get,
(17)
Substituting the short hand notations of equation (17)
and in equation (16), we get a simplified equation as,
(18)
Then, further derivation w.r.t and combining
equation (14) and (18), 2nd order derivative is
obtained and given as:
(19)
Now when then the obtained second order
derivative is which is similar to 2nd
order derivative of ordinary cross – entropy.
Therefore, this focus loss function can be utilized in
binary classification process to improve accuracy and
performance and can be applied for applications like
medical record analysis and epilepsy seizure
detection.
4 Result and Discussion:
This section discusses about the performance result
of proposed Modified XGboost Classifier model for
the faster detection of seizures through classification
process. The proposed Modified XGboost Classifier
model utilizes an additional focal loss function in
classification process in order to minimize training
and testing inaccuracies which can degrade
prediction results for epilepsy. Furthermore, focal
loss function enhances classification accuracy
performance of proposed classification model.
Furthermore, performance of proposed classification
model is measured using sample data of several
patients from the dataset CHB-MIT SCALP
Electroencephalography (EEG). The desired results
obtained by using an efficient classification process
which can easily differentiate between seizures and
non-seizures. The obtained performance results are
compared with several state-of-art techniques in
terms of average sensitivity and average specificity.
Performance results for several patients are
demonstrated in terms of classification accuracy,
sensitivity and specificity.
4.1 Dataset Details:
In this article, epilepsy samples used was from CHB-
MIT SCALP Electroencephalography (EEG)
database and is a public dataset which is taken from
Physionet. Here, total time duration for the EEG
recording is 983 hrs. EEG epoch contains offset time
intervals, seizure onset ictal activities done manually
by the clinical experts. The CHB-MIT EEG database
is collected by investigators from the Children’s
Hospital Boston (CHB) and Massachusetts Institute
of Technology (MIT) this database includes 23
pediatric patients with intractable seizures in order to
estimate their possibility for surgical intervention.
From those 23 patients, 5 patients were male and 17
patients were females and data of 1 patient was
unknown. All the males are aged between 3 to 22
years and all the females are aged between 1.5 to 19
years. Most of the patients contain 23 types of EEG
signal. However, some of the patients hold 24 or 26
EEG signals. All EEG signals are sampled at the rate
of 256 sample/sec and resolution of 16 bit from
electrodes. Electrodes are used according to
International 10-20 system. In overall 24 cases,
signals are partitioned in 1 hour long epochs. It can
be seen that several epochs are up to 2-4 hours in
duration. Furthermore, all 24 cases are exploring the
frequent changes during EEG recordings. Moreover,
CHB-MIT dataset is huge dataset which provide
several variations of cross-validation methods and
patient-specific as well as used by many researchers
in several works [27-30].
4.2 Performance Evaluation:
This section discusses about the performance
comparison against several state-of-art-seizure
detection techniques in terms of average sensitivity
and specificity for several patients. There are some
essential steps which are necessary for the
implementation of proposed classification model
using proposed Modified Xgboost Classifier to detect
seizures such as addition of channels from one to
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
18
Volume 19, 2022
another epoch and channel selection. Here, only those
channel are selected for classification process which
are available even after completion of training and
testing through cross validation process. However, in
cross validation approach, chosen channels can swap
with each other. The ultimate aim for swapping of
selected channels are to examine the quality of data
heterogeneity. Among the available 24 channels, 18
channels shows the stability which are T7-F7, FP1-
F3, C3-P3, FP2-F4, F4-C4, P3-01, C4-P4, FP2-F8,
T8-P8, FZ-CZ, T7-P7 , CZ-PZ, FP1-F7 ,F3-C3, F8-
T8, P8-02, P7-01,P4-02. Further, those stable 18
bipolar raw EEG channels from the dataset are
selected to obtain classification output of the
proposed classification model.
4.2.1 Performance Metrics:
Furthermore, for classification process, the system
performance is evaluated in terms of following
parameters sensitivity, specificity and accuracy:
(20)
(21)
(22)
Where, TP, TN, FP and FN represent true positive,
true negative, false positive and false negative,
respectively.
4.2.2 Performance Comparison:
Here, the performance results of proposed
classification model through modified XGBoost
classifier are compared against several state-of-art-
seizure detection techniques such as Zabihi et al [21],
Yuan et al [22], Tsiouris et al [23], Selvakumari et al
[24], Difei Zeng et. al. [25], Dinghan Hu et al [26]
and Bhattacharyya et al [2] in terms of Average
sensitivity (%) and Average specificity (%). The
proposed modified XGBoost classifier model obtain
average sensitivity as 99.98%, average specificity as
99.97% and obtained EEG data recordings take 983
hours which is immense compare to other techniques
and demonstrated in Table 1. It is evident from the
performance results that the proposed modified
XGBoost classifier model outperforms all the state-
of-art-seizure detection techniques in terms of
average sensitivity, average specificity and EEG data.
Here, while classification, prediction of seizure or
non-seizure comes under 2-class-seizure for subject-
specific experiments. Moreover, 2-class-seizure
experimental results of proposed classification model
are compared against several state-of-art-seizure
classification models. The metrics results in this task
are nearly 100%. The proposed model achieves
improvement in terms of average sensitivity against
the best traditional method (Difei Zeng et. al.) as
0.05% and for average specificity as 1% as shown in
Table 1. Here, cross validation experiments
determine nature of 2-class-seizure as the prediction
is seizure or non-seizure.
Table 1 comparison of the Performance for different
methods on CHB_MIT Dataset
Method
EEG
Data(h)
Average
sensitivity
(%)
Average
specificity
(%)
Zabihi et al
172
91.5
95.16
Yuan et al
958.2
95.65
95.75
Tsiouris et al
980
-
95.00
Selvakumari
et al
-
97.50
94.50
Difei Zeng et.
al.
-
99.93
98.5
Dinghan Hu et
al
-
98.48
98.97
Bhattacharyya
et al
178
97.91
99.57
Our work
983
99.98
99.97
Here, Table 2 demonstrates performance results of
proposed modified XGboost Classifier model
considering performance metrics like Sensitivity
(%), Specificity (%) and Classification Accuracy
(%).Along with their mean and standard deviation
results are also evaluated. Here, mean results of all 24
patients for sensitivity, specificity and accuracy are
100%, 100% and 99.995% respectively. Moreover,
standard deviation is quite low which concludes the
superiority of proposed modified XGboost Classifier
model. Here, performance result of 24 patients
(i.e. Chb01 to Chb24) considering CHB-MIT
Database are presented. Furthermore, data gathered
for all 24 patients from CHB-MIT Database is used
to analyze the performance of proposed classification
model. Here, among 24 patients, 20 patients achieves
accuracy as 100%. The lowest result considering
classification accuracy is achieved for the patient
Chb14 as 99.96%. Besides, it is evident from Table 2
results that all the metric results are invariably 100%
and their average is higher than 99.99% with
minimum standard deviation. This implies that the
proposed classification model is appropriate for every
patient with high accuracy and resilient stability.
Table 2 Performance Results considering the CHB-
MIT Database using proposed modified XGboost
Classifier
Patient
Sensitivity
(%)
Specificity
(%)
Accuracy
(%)
Chb01
100
100
100
Chb02
100
100
100
Chb03
100
100
100
Chb04
100
100
100
Chb05
100
100
100
Chb06
100
100
100
Chb07
100
100
100
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
19
Volume 19, 2022
Chb08
100
100
100
Chb09
100
100
100
Chb10
100
100
100
Chb11
100
100
100
Chb12
100
100
100
Chb13
100
100
100
Chb14
100
100
99.96
Chb15
100
100
100
Chb16
100
100
100
Chb17
100
100
100
Chb18
100
100
100
Chb19
100
100
99.97
Chb20
100
100
100
Chb21
100
100
100
Chb22
100
100
99.98
Chb23
100
100
100
Chb24
100
100
99.99
Mean
100
100
99.995
STD
0
0
0.22
5. Conclusion:
The significance of accurate and quick seizure
detection is immense. However, efficient
classification of epilepsy is challenging and critical
process. Therefore, a modified XGboost classifier
model is presented for accurate identification of
seizures or non-seizures based on machine learning
algorithms. Moreover, a detailed mathematical
modelling for modified XGboost classifier model is
presented to provide highly efficient results for the
applications like seizure detection or cancer
diagnosis. The proposed XGBoost model is modified
version of gradient tree-boosting classifier.
Moreover, a focal loss function is introduced to
minimize mismatching of training and testing
samples in classification process for binary dataset.
Here, CHB-MIT dataset is utilized for the testing of
proposed classification model. Performance results
for all 24 patients are demonstrated above in terms of
sensitivity, specificity and classification accuracy and
compared against several state-of-art-seizure
detection techniques. The proposed modified
XGBoost classifier model obtain average sensitivity
as 99.98%, average specificity as 99.97% and
obtained EEG data recordings take 983 hours which
is immense compare to other techniques. Among 24
patients, 20 patients achieves accuracy as 100%. All
the metric results are invariably 100% and their
average is higher than 99.99% with minimum
standard deviation. The proposed classification
model is appropriate for every patient with high
accuracy and resilient stability.
References
[1] U. Braun, S. F. Muldoon, and D. S. Bassett,
``On human brain networks in health and disease,’’
in eLS. New York, NY, USA: American Cancer
Society, 2015, pp. 1_9. [Online]. Available:
[2] https://onlinelibrary.wiley.com/doi/abs/10.1
002/9780470015902.a0025783, doi:
10.1002/9780470015902.a0025783.
A. Bhattacharyya and R. B. Pachori, “A
multivariate approach for patient-specific eeg seizure
detection using empirical wavelet transform,” IEEE
Trans. Biomed. Eng., vol. 64, no. 9, pp. 2003–2015,
Sep. 2017.
[3] Mahmud M, Kaiser MS, Hussain A,
Vassanelli S (2018) Applications of deep learning
and reinforcement learning to biological data. IEEE
Trans Neural Networks Learn Syst 29(6):2063–2079
[4] Mahmud M, Kaiser MS, Hussain A (2020)
Deep learning in mining biological data. arXiv
:2003.00108
[5] Fayyad UM, Piatetsky-Shapiro G, Smyth P,
Uthurusamy R (eds.) (1996) Advances in Knowledge
Discovery and Data Mining. American Association
for Artificial Intelligence, Menlo Park, CA, USA
[6] Yin Y, Kaku I, Tang J, Zhu J (2011) Data
mining: concepts, methods and applications in
management and engineering design. Springer, New
York
[7] Fu T-c (2011) A review on time series data
mining. Eng Appl Artif Intell 24(1):164–181
[8] Tzallas AT, Tsipouras MG, Tsalikakis DG,
Karvounis EC, Astrakas L, Konitsiotis S, Tzaphlidou
M (2012) Automated epileptic seizure detection
methods: a review study. In: Epilepsy-histological,
electroencephalographic and psychological aspects.
InTech
[9] Abbasi B, Goldenholz DM (2019) Machine
learning applications in epilepsy. Epilepsia
[10] Siddiqui MK. Brain data mining for epileptic
seizure-detection. Doctoral Dissertation, Charles
Sturt University, Australia
[11] Shoeb A. H., Guttag J. V. (2010).
Application of machine learning to epileptic seizure
detection. In International Conference on Machine
Learning (Haifa: ), 975–982
[12] Patnaik L. M., Manyam O. K. (2008).
Epileptic EEG detection using neural networks and
post-classification. Comput. Methods Programs
Biomed. 91, 100–109. 10.1016/j.cmpb.2008.02.005.
[13] Acharya U. R., Oh S. L., Hagiwara Y., Tan J.
H., Adeli H. (2017). Deep convolutional neural
network for the automated detection and diagnosis of
seizure using EEG signals. Comput Biol. Med. 100,
270–278. 10.1016/j.compbiomed.2017.09.017.
[14] French JA, Kanner AM, Bautista J, et al.
Efficacy and tolerability of the new anti-epileptic
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
20
Volume 19, 2022
drugs II: Treatment of refractory epilepsy.
Neurology. 2004;62(8):1261–1273.
[15] M. Ahmad, M. Saeed, S. Saleem, and A. M.
Kamboh, “Seizure detection using eeg: A survey of
different techniques,” in Emerging Technologies
(ICET), 2016 International Conference on. IEEE,
2016, pp. 1–6.
[16] Moghim N, Corne DW. Predicting epileptic
seizures in advance. PloS ONE 2014;9:e99334.
[17] Park Y, Luo L, Parhi KK, Netoff T. Seizure
prediction with spectral power of EEG using cost-
sensitive support vector machines. Epilepsia
2011;52:1761–70.
[18] Endling F., Bartolomei F., Senhadji L.
(2009). Spatial Analysis of intracerebral
electroencephalographic signals in the time and
frequency domain: identification of epileptogenic
networks in partial epilepsy. Philos. Trans.
Mathemat. Phys. Eng. Sci. 367, 297–316.
10.1098/rsta.2008.0220
[19] J. Cao, J. Zhu, W. Hu and A. Kummert,
“Epileptic Signal Classification With Deep EEG
Features by Stacked CNNs,” in IEEE Transactions on
Cognitive and Developmental Systems, vol. 12, no.
4, pp. 709-722, Dec. 2020, doi:
10.1109/TCDS.2019.2936441.
[20] Y. Yuan, G. Xun, K. Jia and A. Zhang, “A
Multi-View Deep Learning Framework for EEG
Seizure Detection,” in IEEE Journal of Biomedical
and Health Informatics, vol. 23, no. 1, pp. 83-94, Jan.
2019, doi: 10.1109/JBHI.2018.2871678.
[21] M. Zabihi, S. Kiranyaz, V. Jantti, T. Lipping,
and M. Gabbouj, “Patient-Specific Seizure Detection
Using Nonlinear Dynamics and Nullclines,” Ieee
Journal of Biomedical and Health Informatics, vol.
24, no. 2, pp. 543-555, Feb, 2020.
[22] S. S. Yuan, J. X. Liu, J. L. Shang, X. Z. Kong,
Q. Yuan, and Z. Ma, “The earth mover’s distance and
Bayesian linear discriminant analysis for epileptic
seizure detection in scalp EEG,” Biomedical
Engineering Letters, vol. 8, no. 4, pp. 373-382, Nov,
2018.
[23] K. M. Tsiouris, S. Markoula, S. Konitsiotis,
D. D. Koutsouris, and D. I. Fotiadis, “A robust
unsupervised epileptic seizure detection
methodology to accelerate large EEG database
evaluation,” Biomedical Signal Processing and
Control, vol. 40, pp. 275-285, Feb, 2018.
[24] R. S. Selvakumari, M. Mahalakshmi, and P.
Prashalee, “Patient-Specific Seizure Detection
Method using Hybrid Classifier with Optimized
Electrodes,” Journal of Medical Systems, vol. 43, no.
5, May, 2019.
[25] D. Zeng, K. Huang, C. Xu, H. Shen and Z.
Chen, "Hierarchy Graph Convolution Network and
Tree Classification for Epileptic Detection on
Electroencephalography Signals," in IEEE
Transactions on Cognitive and Developmental
Systems, doi: 10.1109/TCDS.2020.3012278.
[26] D. Hu, J. Cao, X. Lai, Y. Wang, S. Wang and
Y. Ding, "Epileptic State Classification by Fusing
Hand-crafted and Deep Learning EEG Features," in
IEEE Transactions on Circuits and Systems II:
Express Briefs, doi: 10.1109/TCSII.2020.3031399.
[27] W. Hu, J. Cao, X. Lai, and J. Liu, “Mean
amplitude spectrum based epileptic state
classification for seizure prediction using
convolutional neural networks,” Journal of Ambient
Intelligence and Humanized Computing, pp. 1–11,
2019.
[28] P. Thodoroff, J. Pineau, and A. Lim,
“Learning robust features using deep learning for
automatic seizure detection,” in Machine learning for
healthcare conference, 2016, pp. 178–190.
[29] K. M. Tsiouris, V. C. Pezoulas, M. Zervakis,
S. Konitsiotis, D. D. Koutsouris, and D. I. Fotiadis,
“A long short-term memory deep learning network
for the prediction of epileptic seizures using eeg
signals,” Computers in biology and medicine, vol. 99,
pp. 24–37, 2018.
[30] Y. Yuan, G. Xun, F. Ma, Q. Suo, H. Xue, K.
Jia, and A. Zhang, “A novel channel-aware attention
framework for multi-channel eeg seizure detection
via multi-view deep learning,” in 2018 IEEE EMBS
International Conference on Biomedical & Health
Informatics (BHI). IEEE, 2018, pp. 206–209
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the Creative
Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en_US
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.3
Raveendra Kumar T. H.,
Narayanappa C. K., Raghavendra S., Poornima G. R.
E-ISSN: 2224-2902
21
Volume 19, 2022
