Comparative Analysis of Classifiers for Developing an Adaptive Computer-Assisted EEG Analysis System for Diagnosing Epilepsy

Abstract

Computer assisted analysis of electroencephalogram (EEG) has a tremendous potential to assist clinicians during the diagnosis of epilepsy. These systems are trained to classify the EEG based up on the ground truth provided by the neurologists. So, there should be a mechanism in these systems using which a system's incorrect markings can be mentioned and the system should improve its classification by learning from them. We have developed a simple mechanism for neurologists to improve classification rate while encountering any false classification. This system is based on taking discrete wavelet transform (DWT) of the signals epochs which are then reduced using Principal Component Analysis, and then they are fed into a classifier. After discussing our approach, we have shown the classification performance of three types of classifiers; Support Vector Machine (SVM), Quadratic Discriminant Analysis and Artificial Neural Network. We found SVM to be the best working classifier. Our work exhibits the importance and viability of a self-improving and user adapting computer assisted EEG analysis system for diagnosing Epilepsy which process each channel exclusive to each other, along with the performance January 30, 2015 DRAFT 2 comparison of different machine learning techniques in the suggested system.

Full text

Research Article
Comparative Analysis of Classifiers for Developing an Adaptive
Computer-Assisted EEG Analysis System for Diagnosing Epilepsy
Malik Anas Ahmad,1Yasar Ayaz,1Mohsin Jamil,1Syed Omer Gillani,1
Muhammad Babar Rasheed,2Muhammad Imran,3Nadeem Ahmed Khan,4
Waqas Majeed,4and Nadeem Javaid2,5
1SMME, National University of Sciences & Technology, Islamabad 44000, Pakistan
2Department of Electrical Engineering, COMSATS Institute of IT, Islamabad 44000, Pakistan
3King Saud University, P.O. Box 92144, Riyadh 11543, Saudi Arabia
4Lahore University of Management Sciences, Lahore 54000, Pakistan
5Department of Computer Science, COMSATS Institute of IT, Islamabad 44000, Pakistan
Correspondence should be addressed to Nadeem Javaid; nadeemjavaid@comsats.edu.pk
Received  September ; Revised  December ; Accepted  January 
Academic Editor: Tobias Loddenkemper
Copyright ©  Malik Anas Ahmad et al. is is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
Computer-assisted analysis of electroencephalogram (EEG) has a tremendous potential to assist clinicians during the diagnosis of
epilepsy. ese systems are trained to classify the EEG based on the ground truth provided by the neurologists. So, there should
be a mechanism in these systems, using which a system’s incorrect markings can be mentioned and the system should improve its
classication by learning from them. We have developed a simple mechanism for neurologists to improve classication rate while
encountering any false classication. is system is based on taking discrete wavelet transform (DWT) of the signals epochs which
are then reduced using principal component analysis, and then they are fed into a classier. Aer discussing our approach, we have
shown the classication performance of three types of classiers: support vector machine (SVM), quadratic discriminant analysis,
andarticialneuralnetwork.WefoundSVMtobethebestworking classier. Our work exhibits the importance and viability of
a self-improving and user adapting computer-assisted EEG analysis system for diagnosing epilepsy which processes each channel
exclusive to each other, along with the performance comparison of dierent machine learning techniques in the suggested system.
1. Introduction
Epilepsy is a chronic neurological disease. e hallmark
of this disease is recurring seizures. It has been cited that
one out of hundred people suers from this disorder [].
Electroencephalography is the most widely used technique
for diagnosis of epilepsy. EEG signal is the representation
of voltage uctuations which are caused by the ow of
neurons ionic current. Billions of neurons maintain brains
electric charge. Membrane transport proteins pump ions
across their membranes. Neurons are electrically charged by
these membranes. Due to volume conduction, wave of ions
reaches the electrodes on the scalp that pushes and pulls the
electron on the electrode metal. e voltage dierence due to
pullandpushoftheelectronsismeasuredbyvoltmeterwhose
readings are displayed as the EEG potential. Neuron generates
toosmallofachargetobemeasuredbyanEEG,anditisthe
summation of synchronous activity of thousands of neurons
that have similar spatial orientation which is measured by an
EEG. Unique patterns are generated in the EEG during an
epileptic seizure. ese unique patterns help the clinicians
during diagnosis and treatment of this neurological disorder.
at is why EEG is widely used to detect and locate the
epileptic seizure and zone. Localization of the abnormal
epileptic brain activity is very signicant for diagnosis of
epileptic disorder.
Usually the duration of a typical EEG varies from few
minutes to few hours but in case of prolonged EEG it can even
Hindawi Publishing Corporation
BioMed Research International
Volume 2015, Article ID 638036, 14 pages
http://dx.doi.org/10.1155/2015/638036

BioMed Research International
last as long as  hours. is generates an immense amount
of data to be inspected by the clinician which could prove to
be a daunting task.
Advancement in signal processing and machine learning
techniques is making it possible to automatically analyse EEG
data to detect epochs with epileptic patterns. A system based
on these techniques can aid a neurologist by highlighting the
epileptic patterns in the EEG up to a signicant level. Of
course, the task of diagnosis should be le to the neurologist.
However, the task of the neurologist becomes ecient as it
reduces the data to be analysed and lessens up the fatigue.
Along with classication these analysis soware programs
canalsoprovidesimultaneousvisualizationofmultiplechan-
nels which helps the clinician in dierentiating between
generalized epilepsy and focal epilepsy.
Itiswellknownthatanepilepticseizurebringschanges
in certain frequency bands. at is why usually the spectral
content of the EEG is used for diagnosis []. ese are
identied as 𝛿(.– Hz), 𝜃(– Hz), 𝛼(– Hz), and 𝛽
(– Hz). Noachtar and R´
emi mention almost ten types of
epileptic patterns. However, most of the existing work only
focuses on one of the epileptic patterns, that is,  Hz spike and
wave which is a trademark for absence seizure. Other types of
the patterns are rarely addressed [].
Computer-assisted EEG classication involves several
stages including feature extraction, feature reduction, and
feature classication. Wavelet transform has become the most
popular feature extraction technique for EEG analysis due
to its capability to capture transient features, as well as
information about time-frequency dynamics of the signal
[]. Other previously used feature extraction approaches for
epilepsy diagnosis include empirical mode decomposition
(EMD), multilevel Fourier transform (FT), and orthogonal
matching pursuit [–]. Feature extraction is followed by
feature reduction to reduce computational complexity and
avoid curse of dimensionality. Most commonly, the reduced
feature vector consists of statistical summary measures (such
as mean, energy, standard deviation, kurtosis, and entropy) of
dierent sets of original (unreduced) features, although other
methods such as principal component analysis, discriminant
analysis, and independent component analysis have also been
used for feature reduction [,,,]. Feature extrac-
tion/reduction is followed by classication using a machine
learning algorithm, such as articial neural networks (ANN),
support vector machines (SVM), hidden Markov models, and
quadratic discriminant analysis [,–].
A very important and novel phase of our system is user
adaptation mechanism or retraining mechanism. ere are
multiple reasons according to which introduction of this
phase has lots of advantages. During this phase, system will
try to adapt its classication as per users desire. It has been
cited that sometimes even the expert neurologists have some
disagreement over a certain observation of an EEG data.
ere is also a threat of overtting by the classier. In order
to keep the classier improving its performance with the
encounter of more and more examples, we have introduced
this user adaptive mechanism in our system. We consider the
existing systems as dead because they cannot improve their
classication rate aer initial training. ey do not have any
mechanism of learning or improvement from neurologists
corrective marking [–]. e agreement between dierent
EEG readers is low to moderate; our adaption mechanism
helps the user in catering this issue as our system tries to
adapt the detection according to the users corrective marking.
e new corrective markings generate new examples with
improved labels. Hence, it populates the training examples
with newly labelled ones. So aer retraining machine learning
algorithms in the system, users adapt to set of choices.
In the next section we will explain our proposed method
which will be followed by the results. In the results section, we
will explain how SVM performs better than QDA and ANN
in our proposed method. We will also show that exclusive
processing of each channel results in a signicant improve-
ment in the classication rate. Here “epileptic pattern” and
“epileptic spikes” will be used as an alternative to each other.
2. Proposed Method
Computer-aided EEG analysis systems use the neurologists
marking and labelling of the EEG data as a benchmark to
train themselves during initial training phase. But aer initial
training phase, these systems have no simple mechanism
for these neurologists to improve systems classication aer
encountering any false classication. So we have proposed a
method by which systems classication can be improved by
theuserinarelativelysimplerway.isanalysissystemonly
tries to detect the epileptic spikes as mentioned by Noachtar
and R´
emi. Later it adapts its detection of epileptic spikes
exclusively for every user (Figure ).
In this proposed system, we are processing each channel
for each epileptic pattern exclusive to each other. is
exclusiveprocessingofeachchannelnotonlyhelpstheuserin
diagnosing localized epilepsy but also eases up the classiers
job. We have considered that dierent epileptic patterns are
independent to each other and their separate handling will
help us in avoiding error propagation from one epileptic
pattern type detection to the other. Our systems working
has two major phases (A) initial training phase and (B)
adaptation phase. ese two major phases have further three
parts which are () feature extraction, () feature reduction,
and () classication. Next we will briey explain all of these
steps.
2.1. Initial Training
2.1.1. Feature Extraction. To decide which parts of the signal
are epileptic and which are not we rst divided whole of the
signal in small chunks known as epochs. en DWT was
applied on those epochs so that visibility of epileptic activity
canbeenhancedwhichisdistinguishedbysomespectral
characteristics. ese features are then processed to make
them more suitable for the classication technique.
(a) Epoch Size. e rst important part of the feature
extraction is epoch selection. Epoch is a small chunk of the
signal which is processed at a time. e size of the epoch
is very important. e larger it is the less accurate it will
be. e smaller it is the higher the processing time will be.

BioMed Research International 
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
50
100
0
50
100
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
0
10
20
30
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
−100
−50
−100
−50
−40
−30
−20
−10
Epoch number 2
Epoch number 1
F : Epoch size is  sec.
Aer testing dierent epoch sizes, we found epoch size of
nonoverlapping  sec window to be best yielding in terms of
accuracy. It also reestablished the work of Seng et al. []
(Figure ).
(b) DWT. As discussed in Introduction, spectral analysis
is very informative while examining the epilepsy suspected
patients EEG. ere are profound advantages of wavelet
decomposition which is a multiresolution analysis technique.
A multiresolution analysis technique allows us to analyse
a signal for multiple frequency resolutions while maintain-
ing time resolution unlike a normal frequency transform.
Wavelet decomposition allows us to increase frequency res-
olution in the spectral band of our interest while maintaining
the time resolution; in short we can decimate these values
simultaneously in time and frequency domain.
During wavelet transform, the original epoch is split
into dierent subbands: the lower frequency information
is called approximate coecients and the higher frequency
information is called detailed coecients. e frequency
subdivision in these subbands helps us in analysing dierent
frequency ranges of an EEG epoch while maintaining its time
resolution [,,]. e choice of coecients level is very
important as the epileptic activity only resides in the range
of – Hz. Coecients levels of the DWT are determined
with respect to sampling frequency. So, the detailed levels of
interest are adjusted on the run according to the sampling
frequency such that we may get at least one exact value of
the closest separate 𝛿(.– Hz), 𝜃(– Hz), 𝛼(– Hz),
and 𝛽(– Hz) components of the signal. We discarded all
the detailed coecient levels which were beyond the – Hz
range.
enDWTwasappliedoneachepochwithDaubechies-
(db) as mother wavelet. e detailed coecient levels of the
DWT were determined with respect to sampling frequency.
(c) Statistical Features. Aer the selection of detailed coef-
cients which represent the frequency band of our interest,
we calculated the statistical features by calculating the mean,
standard deviation, and power of these selected wavelet
coecients. ese statistical features are inspired from Subasi
and Gursoy work [].
(d) Standardization. ese statistical features were then stan-
dardized. During training stage 𝑧-score standardization was
applied on these features []. is standardization is just
like usual 𝑧-score normalization, but as we do not know
the exact mean and standard deviation of the data (to be
classied) during classication/test stage, we used the mean
and standard deviation of the training examples during
training stage for standardizing (normalizing) the features
during classication stage. We normalized the features by
subtracting and dividing them by training examples mean
and standard deviation, respectively.
2.2. Feature Reduction. In order to avoid overinterpretation
by redundant data and misinterpretation by noisy data we
applied feature reduction method. Inclusion of this part
increases the processing time, thus exacerbating the latency.
Dimensionality reduction using principal component
analysis (PCA) is based on a very important trait that is
variance of the data. PCA develops the nonlinear mapping in
such a way that it maximizes the variance of the data, which
helpsusindiscardingthatpartofthedatawhichismarked
by lesser variances. is reshaping and omission not only
removes the redundant data but also lessens up the noise.
During training stage PCA was applied on these features
in order to reduce the redundant and/or noisy data. We kept
the components which projected the approximate % of the
total variance. We were able to reduce the  features into .
en we fed these reduced features to classiers trainer. Here
as per our observation we again assumed that the EEG data
is stationary for a small length. So, during the testing stage,
we took the PCA coecients matrix from training stage and

BioMed Research International
multiplied it with the standardized statistical features of the
blind test data and then fed the top  features to classier.
2.3. Classication. Classication is a machine learning tech-
nique in which new observations belonging to a category
are identied. is identication is based on the training set
which contains the observations with known labelling of their
category. ese observations are also termed as features. We
tried three types of classication methods: () SVM, () QDA,
and () ANN (Figure ).
e reduced features were fed to these classiers. Here
the reduced features mean that those statistical features of
the selected wavelet coecients are reduced using PCA as
described in previous section. All of the three processing
parts were exclusive for each channel and each epileptic
pattern. So like previous parts the classiers were also trained
and tested exclusively for each channel.
Our system requires individual labelling of channels.
ere is a separate classier for each channel and for each
epileptic pattern type. So, the total number of classiers is
equaltotheproductoftotalnumberofchannelsbytenwhere
ten represents the number of epileptic pattern described by
Noachtar and R´
emi [].
(e) Support Vector Machine. Support vector machine (SVM)
is a supervised learning models machine learning technique.
SVM tries to represent the examples as points in space which
are mapped in a way that points of dierent categories can be
divided by a clear gap that is as wide as possible. Aerwards,
that division is used to categorise the new test examples based
on which side they fall on.
(f) Quadratic Discriminant Analysis. Quadratic discriminant
analysis (QDA) is a widely used machine learning method
among statistics, pattern recognition, and signal process-
ing to nd a quadratic combination of features which are
responsible for characterizing an example into two or more
categories. QDAs combination of discriminating quadratic
multiplication factors is used for both classication and
dimensionality reduction.
(g) Articial Neural Network. Articial neural network
(ANN) is a computational model which is inspired from ani-
mals central nervous system. at is why ANN is represented
by a system of interconnected neurons which are capable of
computing values as per their inputs. In ANN training, the
weights associated with the neurons are iteratively adjusted
according to the inputs and the dierence between the
outputs with expected outputs. e iteration gets stopped
when either the combination of neurons starts generating the
expected results within an error of a tolerable error range or
the iteration limit nishes up.
2.4. Adaption Phase (Retraining/User Adaptation Mecha-
nism). In order to keep the classier improving its perfor-
mance with the encounter of more and more examples, we
have introduced a user adaptive mechanism in our system.
Our system allows the user to interactively select epochs
of his choice by simply clicking on the correction button.
Whileusingoursystem,whenauserthinksthatacertain
epoch is falsely labelled/categorised, our system allows him to
interactively mark mark that label as a mistake. ese details
willbesavedinaloginthebackgroundandtheywillbeused
to retrain the classier to improve its classication rate and
adapt itself according to the user with the passage of time.
When the user is going to select the retraining option in our
system, then classiers will retrain themselves on the previous
and the newly logged training examples. As every user has to
log in with his personal ID, every corrective marking detail
will only be saved in that user’s folder and only classier will
update itself for that user. Hence, the systems classier tries to
adapt itself according to that user without damaging anyone
else classication.
e concept behind the inclusion of the retraining is that
if there is more than one example with same attributes but
dierent labels, the classier is going to get trained to the
one with most population. e user’s corrective marking will
increase the examples of his choice, thus making that classier
adapt itself to the user’s choice in a trivial way. Every user
will have exclusive classiers trained for him and his marking
will not aect other users’ classier. As we know, the users
sometimes do not agree on the choice of the epileptic pattern
or its type. e exclusive processing for each user will help the
same soware keep the system trained for every user and it
will also let dierent users compare their markings with each
other.
Wedonothaveanystandardrightnowtomeasurewhich
neurologist is the most righteous among a disagreeing group
of neurologist users. So we kept the corrective markings of
eachusertohisaccountsothatitmaynotinterferewith
the one who may not agree on his choice. So, the developed
system is used to facilitate the neurologist’s selection to the
user according to his own choice and aer initial training on
every retraining it tries to adopt more users. is system does
not want to dictate to the neurologist but rather learn from
him to adapt him to save his time.
Wewanttheclassiertothinkliketheuserandsupple-
ment him by highlighting the epochs of his choice, so the gold
standard aer few retraining mechanisms will be the user
himself. Already tested examples with new labels inclusion in
the training examples for the retraining will bias the classiers
choice in favour of user.
3. Experimentation
In this section, we will discuss the results in detail. At rst,
we will describe the datasets which we used to train, test, and
validate our method. en we will discuss their versatility
(Figure ).
3.1. Dataset. Two labelled datasets of epilepsy suspected
surface EEG data were available to us. Both of these datasets
have lots of versatility in between them in terms of ethnicity,
age, gender, and equipment. e datasets available to us were
about generalised absence seizure which is characterized by
theHzspikeandwaveepilepticpatterninalmosteach
channel. at is why we have classication results available
only for one type of epilepsy which is absence seizure.

BioMed Research International 
T : is table describes the aliation of detailed coecients
with epileptic frequency band of interest for Hz sampled CHB-
MIT dataset.
Epileptic frequency range Detailed coecients’ level
Beta (𝛽)CD(HztoHz)
Alpha (𝛼)CD(HztoHz)
eta (𝜃)CD(HztoHz)
Delta (𝛿)CD(HztoHz)
Delta (𝛿)CD(HztoHz)
T : is table describes the aliation of detailed coecients
with epileptic frequency band of interest for Hz sampled PIMH
dataset.
Epileptic frequency range Detailed coecients’ level
Beta (𝛽) CD (. Hz to . Hz)
Alpha (𝛼) CD (. Hz to . Hz)
eta (𝜃) CD (. Hz to . Hz)
Delta (𝛿)CD(.HztoHz)
Delta (𝛿)CD(HztoHz)
3.1.1. CHBMIT. is database is the online available surface
EEG dataset [] which is provided by Children Hospital
Boston and Massachusetts Institute of Technology and it is
availableatphysioNetwebsite[]. It contains  hours of
 channels scalp EEG recording from  epilepsy suspected
patients. is ECG recording is sampled at  Hz with -bit
resolution. e rd channel is same as th channel (Tab l e  ).
3.1.2. PIMH. e second database of EEG datasets is pro-
vided by our collaborator at Punjab Institute of Mental Health
(PIMH), Lahore. Its sampling frequency is  Hz and it was
recorded on  channels (among which  channels are for
EEG). is dataset consists of  patients EEG recording.
3.2. Features
3.2.1. Feature Extraction. Data which interests us lies in
between the frequency range of . Hz to  Hz. So aer
applying DWT with db mother wavelet, we have to select
detailed coecients with this frequency range. So in case of
 Hz sampled CHBMIT dataset, we have to go to at least
 levels of decomposition and discard the earlier two as it is
demonstrated in Figure . In order to get the discriminating
information between dierent types of epileptic patterns and
identifying them correctly without mistaking them with each
other, decomposition of this detailed coecient further in
Beta, Alpha, eta, and Delta is hugely helpful. So we further
decomposed them until the th level. Hence, we used the
DWTsdetailedcoecientsoflevels,,,,andforHz
sampled CHB-MIT dataset (Tab l e  ).
Aer the selection of the wavelet coecients, we calcu-
latedthestatisticalfeatureoutofthem.estatisticalfeatures
were the mean, power, and standard deviation of all of the
selected coecients.
In case of Hz sampled PIMH dataset, we used the
DWTsdetailedcoecientsoflevels,,,,and.
Aer the selection of detailed coecients, we calculated
the statistical feature out of them. e statistical features
were the mean, power, and standard deviation of all of the
shortlisted detailed coecients.
3.2.2. Standardization. During training stage, we rst used
simple 𝑧-score normalization to standardize the features []
before applying feature reduction. But the real issue arose
when we tried to normalize them during testing stage. One
way of doing this is that we keep all of the examples and
apply 𝑧-score on them along with the new test data. Instead
of this time taking process, we made an assumption on
our observation that mean and standard deviation does not
deviate a lot. It is analysed in this study that the EEG time
seriesareassumedtobestationaryoverasmalllengthofthe
segments. So we used the mean and standard deviation of the
training examples from the training stage to normalize the
test examples. Figures and illustrate our observation, in
which you can see that there is not much deviation in train
and train + test examples mean and standard deviation.
3.3. Classier. Classication is used in machine learning
to refer to the problem of identifying a discrete category
to which a new observation belongs. Observations with
known labels are used to train a classication algorithm or
classier using features associated with the observation. For
CHBMITdatabase,wehadtotrainclassiersininitial
training stage. e calculation behind  is the  channels
multiplied by  types of epileptic pattern. e rd channel
was same as th channel. For PIMH dataset  classiers
were trained where  channels of EEG were utilized. We
tried three dierent classiers and found SVM to be the most
accurate.
We have used blind validation mechanism for the ten
dierent feature data distributions to estimate the classi-
cation performance. ese  dierent and separate blind
data distributions were taken from a huge set of EEG
dataset. ese  data distributions we randomly divided
into two groups. We trained our classier on one half of the
distribution and tested it on the other half. We repeated that
on all ten distributions. en we calculated the average of the
classication rate for the all ten distributions.
 out of  epochs were randomly taken for ten
times from CHBMIT dataset. Each time half of them were
used to train and half of them were used to test the initial
classication. e average of the sensitivity, specicity, and
accuracy for these ten distributions is considered as the initial
training phase performance.
Same approach was applied on PIMH datasets where 
out of  epochs were randomly taken from PIMH dataset
for the six times instead of ten times.
Due to unavailability of the non- Hz spike and wave
epileptic EEG data, currently we have only classication rates
for generalized absence seizure.
3.3.1. Exclusive Processing. In this study, we have analysed
that even in the case of absence seizure epileptic patterns do
not appear in the exact same way in each channel. Handling
of each channel exclusive to each other was also another very

BioMed Research International
Discarded
Selected
D1 A1
D2 A2
D3
D4 A4
D5
A3
A5
D7 A7
D6 A6
D8 A8
D9 A9
0 200 400
0
0
200
0 100 200
0
50
0 100 200
0 20 40
0
200
0 20 40
0
200
0 20 40
0
500
0 10 20
0
200
0 10 20
0
500
0 5 10
0
100
0 5 10
0
1000
0510
0
500
0 5 10
0
500
0 5 10
0
500
0 5 10
0
1000
0 5 10
0
500
0 5 10
0
0510
0
200
0 5 10
0
−200
−200
−200
−200
−500
−500 −500
−500
−500
0 20 40
0
100
−100
−100 −1000
−1000
−1000
−500
−1000
−500
−500
−200
−50
200
−200
1s wide epoch
F : Selection of DWT detailed coecients for a Hz sampled  sec wide epochs EEG signal.
important decision. We tested the classication in both ways,
that is, one classier for all of the channels at once versus one
separate classier for each channel (Figure ).
isprocessingofeachchannelexclusivetoeachother
improved over average accuracy from approximately % to
approximately % in case of SVM. So for SVM, there is
a signicant improvement of % by this change. In case of
QDAaccuracyrosefrom%to%withanimprovement
of%andincaseofANNitrosefrom.%to.%with
an improvement of .%.
Results show that SVM suites our method in the most e-
cient way. ANN has a lesser classication time and LDA has

BioMed Research International 
No
PCA coefficient
User I/P
Ye s No
Retrain
Tra i n
Select EEG
Select channel
DWT
Mean, power, and standard deviation
PCA
Tra i nin g
Classier
Plot results
Exit
Standardization
Multiplyingcoecients of PCA
User I/P
Ye s No
DWT
Multiplying coecients of PCA
Standardization
Label
Label
Retraining
examples
Initial training
examples
Ye s
Tra i nin g
Mean, power, and standard deviation
Training
examples
classied epochs by the user
Corrective marking
Retraining phase
Tra i nin g
Ye s
User I/P ≥identication of falsely
Extracting the epochs of 1sExtracting the epochs of 1 s
Z-score
F : Workow of a single channel.
F : iNSS interface.
a lesser training time as compared to SVM, but considering
the sensitivity and classication improvement through cor-
rective marking, we think that SVM is the better choice than
LDA and ANN. In upcoming sections, we have shown the
results for all three types of classier.
0 5 10 15 20 25
0
2
4
6
8
10
12
14
16
×104
Mean of the training examples
Mean of the training +test examples
F : Relationship between channel’s number and mean value
(test + training examples).
(a) Adaptation Mechanism. To test the adaptation mechanism
correctiveepochsweremarkedbytheuserforaCHBMIT

BioMed Research International
0 5 10 15 20 25
0
0.5
1
1.5
2
2.5
3
3.5
4
×105
Std. of the training examples
Std. of the training +test examples
F : Relationship between channel’s number and standard deviation value (test + training examples).
EEG
Channel 1 DWT
D1
A1 D2
A2
Mean
Standard deviation
Power
Mean
Standard deviation
Power
Mean
Standard deviation
Power
SVM for pattern 1
SVM for pattern 2
SVM for pattern 3
SVM for pattern 10
O/P
for
channel
1
Channel 2 DWT D1
A1 D2
A2
Mean
Standard deviation
Power
Mean
Standard deviation
Power
Mean
Standard deviation
Power
SVM for pattern 1
SVM for pattern 2
SVM for pattern 3
SVM for pattern 10
O/P
for
channel
2
DWT
D1
A1 D2
A2
Mean
Standard deviation
Power
Mean
Standard deviation
Power
Mean
Standard deviation
Power
SVM for pattern 1
SVM for pattern 2
SVM for pattern 3
SVM for pattern 10
Features
Features
PCA
Reduction
Standardization
PCA
Standardization
Reduction
FeaturesPCA
Reduction
Standardization
Surgeon’s marking
Surgeon’s marking
Surgeon’s marking
O/P
for
channel
n
Al
Dl
Al
Al
Dl
Dl
Z-score
Z-score
Z-score
Channel n
.
.
.
.
.
.
.
.
.
F : Flowchart.
dataset le, and he marked the same amount of epochs for
each channel. ese corrective markings were saved in his
log as training examples. ese corrective markings as the
new examples along with the  epochs of initial training
stagewereusedtoretraintheclassier.enumberhas
come from the  randomly selected epochs from whole of
the CHBMIT dataset for the ten separate times during initial
trainingphase.enlatertheperformanceoftheclassier
aer retraining was judged again on another random 
epochs (Figure ).
In case of PIMH dataset  corrective epochs were
selected for PIMH dataset. is time  epochs of the
PIMH dataset were used along with the  corrective mark-
ings,asforthePIMHwerandomlyselectedthenumbers

BioMed Research International 
89
90
91
92
93
94
95
96
SVM QDA ANN
Processing all channels simultaneously with single classier
Processing all channels separately with separate classier for
each channel
F : Accuracy relationship of dierent classiers and their
classication rate.
84
86
88
90
92
94
96
98
100
Accuracy aer initial training (%)
Accuracy aer retraining (%)
“FP1F7”
“F7T7”
“T7P7”
“P7O1”
“FP1F3”
“F3C3”
“C3P3”
“P3O1”
“FP2F4”
“F4C4”
“C4P4”
“P4O2”
“FP2F8”
“F8T8”
“T8P8”
“P8O2”
“FZCZ”
“CZPZ”
“P7T7”
“T7FT9”
“FT9FT10”
“FT10T8”
F : Relation between average classication rate and accuracy
of the channel aer initial training and retaining.
of epochs for the six times. e retrained classier was tested
on the  remaining epochs.
(b) Support Vector Machine.Weusedthesupportvector
machine classier package available in MATLAB Bioinfor-
matics Toolbox. We found linear kernel to be the most
accurate SVM kernel with  as the box constraint.
(c) CHBMIT. For CHBMIT dataset, initial training of the
classier resulted in .% average accuracy, .% average
0
20
40
60
80
100
120
Accuracy aer initial training (%)
Accuracy aer retraining (%)
“Fp1”
“Fp2”
“F3”
“F4”
“C3”
“C4”
“P3”
“P4”
“O1”
“O2”
“F7”
“F8”
“T3”
“T4”
“T5”
“T6”
“Fz”
“Cz”
“Pz”
“E”
“PG2”
“A 1 ”
“A 2”
“T1”
“T2”
“X1”
“X2”
“X3”
“X4”
“X5”
“X6”
“X7”
“PG1”
F : Relation between average classication rate and accuracy
of the channel aer initial training and retaining.
82
84
86
88
90
92
94
96
98
Accuracy aer initial training (%)
Accuracy aer retraining (%)
“FP1F7”
“F7T7”
“T7P7”
“P7O1”
“FP1F3”
“F3C3”
“C3P3”
“P3O1”
“FP2F4”
“F4C4”
“C4P4”
“P4O2”
“FP2F8”
“F8T8”
“T8P8”
“P8O2”
“FZCZ”
“CZPZ”
“P7T7”
“T7FT9”
“FT9FT10”
“FT10T8”
F : Relation between average classication rate and accuracy
of the channel aer initial training and retaining.
specicity, and .% average sensitivity for  Hz spike and
wave which is a characteristic of absence seizure. Aer
initial training our specicity is better than that of Shoeb
[]andNasehiandPourghassem[]whousedthesame
dataset to validate their technique with dierent features
and application technique. is shows that our technique
is providing better results even at the initial training phase
(Figure ).
In Table  we have shown the average initial classication
and retrained classication results of our system for each
channel.Inthissystem,wehaveshownthataercorrection
of few epochs there is a visible improvement in the systems
classication. e average accuracy of the system rose from
.% to .%.
(d) PIMH. For PIMH dataset, initial training of the classier
resulted in % average accuracy, % average specicity, and
% average sensitivity for  Hz spike and wave which is a
characteristic of absence seizure (Figure ).

 BioMed Research International
0
10
20
30
40
50
60
70
80
90
100
Accuracy aer initial training (%)
Accuracy aer retraining (%)
“F3”
“F4”
“C3”
“C4”
“P3”
“P4”
“O1”
“O2”
“F7”
“F8”
“T3”
“T4”
“T5”
“T6”
“Fz”
“Cz”
“Pz”
“E”
“PG2”
“T1”
“T2”
“X1”
“X2”
“X3”
“X4”
“X5”
“X6”
“X7”
“PG1”
“A 1 ”
“A 2”
“Fp1”
“Fp2”
F : Relation between average classication rate and accuracy
of the channel aer initial training and retaining.
80
82
84
86
88
90
92
94
96
98
Accuracy aer initial training (%)
Accuracy aer retraining (%)
“FP1F7”
“F7T7”
“T7P7”
“P7O1”
“FP1F3”
“F3C3”
“C3P3”
“P3O1”
“FP2F4”
“F4C4”
“C4P4”
“P4O2”
“FP2F8”
“F8T8”
“T8P8”
“P8O2”
“FZCZ”
“CZPZ”
“P7T7”
“T7FT9”
“FT9FT10”
“FT10T8”
F : Relation between average classication rate and accuracy
of the channel aer initial training and retaining.
In Table , we have shown the average initial classication
and retrained classication results of our system for each
channel. Table  showsthatourtechniqueisrobustandit
works also on a dierent dataset. e average accuracy of the
system rose from approximately % to %.
3.3.2. Discriminate Analysis. We used the discriminant anal-
ysis package available in MATLAB Statistics Toolbox. We
found pseudoquadratic to be the best performing discrimi-
nate type with uniform probability.
(a) CHBMIT. For CHBMIT dataset, initial training of the
classier resulted in % average accuracy, % average
specicity, and % average sensitivity for  Hz spike and
wave which is a characteristic of absence seizure. Aer initial
training our specicity is better than that of Shoeb []and
Nasehi and Pourghassem [](Figure ).
In Table ,wehaveshowntheaverageinitialclassication
and retrained classication results of our system for each
channel.Inthissystem,wehaveshownthataercorrection
0
10
20
30
40
50
60
70
80
90
100
Accuracy aer initial training (%)
Accuracy aer retraining (%)
“F3”
“F4”
“C3”
“C4”
“P3”
“P4”
“O1”
“O2”
“F7”
“F8”
“T3”
“T4”
“T5”
“T6”
“Fz”
“Cz”
“Pz”
“E”
“PG2”
“T1”
“T2”
“X1”
“X2”
“X3”
“X4”
“X5”
“X6”
“X7”
“PG1”
“A 1 ”
“A 2”
“Fp1”
“Fp2”
F : Relation between average classication rate and accuracy
of the channel aer initial training and retaining.
T : First column shows the channel label, the second column
shows the initial training accuracy, and third one shows the marked
correction by a neurologist and the last one shows the nal accuracy.
Channel Accuracy aer
initial training (%)
Number of epochs
marked by the user
Accuracy aer
retraining (%)
FPF .  .
FT .  .
TP .  .
PO .  .
FPF .  .
FC .  .
CP .  .
PO .  .
FPF .  .
FC .  .
CP .  .
PO .  .
FPF .  .
FT .  .
TP .  .
PO .  .
FZCZ .  .
CZPZ .  .
PT .  .
TFT .  .
FTFT .  .
FTT .  .
of few epochs there is visible improvement in the systems
classication. e average accuracy of the system rose from
% to %.
(b) PIMH. For PIMH dataset, initial training of the classier
resulted in % average accuracy, % average specicity, and
% average sensitivity for  Hz spike and wave which is a
characteristic of absence seizure.
In Table , we have shown the average initial classication
and retrained classication results of our system for each

BioMed Research International 
T : First column shows the channel label, the second column
shows the initial training accuracy, and third one shows the marked
correction by a neurologist and the last one shows the nal accuracy.
Channel Accuracy aer
initial training (%)
Number of epochs
marked by the user
Accuracy aer
retraining (%)
Fp .  .
Fp .  .
F .  .
F .  .
C .  .
C .  .
P .  .
P .  .
O .  .
O .  .
F .  .
F .  .
T .  .
T .  .
T .  .
T .  .
FZ .  .
CZ .  .
PZ .  .
E.  .
PG .  .
PG .  .
A .  .
A .  .
T .  .
T .  .
X .  .
X .  .
X .  .
X .  .
X .  .
X .  .
X .  .
channel. Table  showsthatourtechniqueisrobustandit
works also on a dierent dataset. e average accuracy of the
system rose from approximately % to %.
3.3.3. Articial Neural Network. We used feedforward back-
propagation package available in MATLAB Neural Network
Toolbox and found Levenberg-Marquardt to be the best
method, with . learning rate.
(a) CHBMIT. For CHBMIT dataset, initial training of the
classier resulted in .% average accuracy, .% average
specicity, and .% average sensitivity for  Hz spike and
wave which is a characteristic of absence seizure (Figure ).
T : First column shows the channel label, the second column
shows the initial training accuracy, and third one shows the marked
correction by a neurologist and the last one shows the nal accuracy.
Channel Accuracy aer
initial training (%)
Number of epochs
marked by the user
Accuracy aer
retraining (%)
FPF .  .
FT .  .
TP .  .
PO .  .
FPF .  .
FC .  .
CP .  .
PO .  .
FPF .  .
FC .  .
CP .  .
PO .  .
FPF .  .
FT .  .
TP .  .
PO .  .
FZCZ .  .
CZPZ .  .
PT .  .
TFT .  .
FTFT .  .
FTT .  .
In Table , we have shown the average initial classication
and retrained classication results of our system for each
channel.Inthissystem,wehaveshownthataercorrection
of few epochs there is visible improvement in the systems
classication. e average accuracy of the system rose from
.% to .%.
(b) PIMH. For PIMH dataset, initial training of the classier
resulted in % average accuracy, .% average specicity,
and .% average sensitivity for  Hz spike and wave which
is a characteristic of absence seizure (Figure ).
In Table , we have shown the average initial classication
and retrained classication results of our system for each
channel. Table  showsthatourtechniqueisrobustandit
works also on a dierent dataset. e average accuracy of the
system rose from approximately % to .%.
4. Discussion and Future Work
Computer-assisted analysis of EEG has tremendous potential
for assisting the clinicians in diagnosis. A very important
andnovelphaseofoursystemisuseradaptationmechanism
or retraining mechanism. Introduction of this phase has
importance in many aspects. In this phase, system tries to
adapt its classication according to users desire. Moreover,
this technique personalizes the classiers classication. It has

 BioMed Research International
T : First column shows the channel label, the second column
shows the initial training accuracy, and third one shows the marked
correction by a neurologist and the last one shows the nal accuracy.
Channel Accuracy aer
initial training (%)
Number of epochs
marked by the user
Accuracy aer
retraining (%)
Fp .  .
Fp .  .
F .  .
F .  .
C .  .
C .  .
P .  .
P .  .
O .  .
O .  .
F .  .
F .  .
T .  .
T .  .
T .  .
T .  .
Fz .  .
Cz .  .
Pz .  .
E.  .
PG .  .
PG .  .
A .  .
A .  .
T .  .
T .  .
X .  .
X .  .
X .  .
X .  .
X .  .
X .  .
X .  .
been cited that sometimes even the expert neurologists have
some disagreement over a certain observation of an EEG data.
issystemwillbeusefulfordisagreeingusersanditwillalso
help them in comparing their results with each other.
ere is also a threat of overtting by the classier.
In order to keep the classier improving its performance
with the encounter of more and more examples, we have
introduced this user adaptive mechanism in our system. We
consider the existing systems as dead because these cannot
improve their classication rate aer initial training (during
soware development). e self-improving mechanism aer
deployment makes our tool alive. is system can be made
part of the whole epileptic diagnosis process. It will highlight
T : First column shows the channel label, the second column
shows the initial training accuracy, and third one shows the marked
correction by a neurologist and the last one shows the nal accuracy.
Channel Accuracy aer
initial training (%)
Number of epochs
marked by the user
Accuracy aer
retraining (%)
FPF .  .
FT .  .
TP .  .
PO .  .
FPF .  .
FC .  .
CP .  .
PO .  .
FPF .  .
FC .  .
CP .  .
PO .  .
FPF .  .
FT .  .
TP .  .
PO .  .
FZCZ .  .
CZPZ .  .
PT .  .
TFT .  .
FTFT .  .
FTT .  .
the epileptic spikes among the whole EEG, thus leading to
reduced fatigue and time consumption of a user. We obtained
high classication accuracy on datasets obtained from two
dierent sites, which indicates reproducibility of our results
and robustness of our approach.
Inthefuture,weareplanningtomakethisawebbased
application; neurologists can log in and consult each other’s
reviews about a particular subject. is will make our system
experience a whole versatility of examples and learn from all
of them. Integration of the video and its automatic analysis
(video EEG) can help a neurologist in diagnosing epilepsy in
a better way, whereas this can also help him in distinguishing
between psychogenic and epileptic seizures. We would also be
investigating how much overtting is an issue in the reported
performances which are now even touching % based on
some claims. ere is a need for method/criteria which could
limit these algorithms improving their detection on a limited
number of available examples.
is system is made keeping in mind that we have to facil-
itate the neurologist by supplementing him in the analysis of
the EEG. We do not want to enforce the classication of the
EEG data on a user.
Inthefuture,wewillalsoincludeasliderinthesystem
which will allow the user to adjust the sensitivity and speci-
city before retraining. is assisting system is more like a
detection tool which is continuously learning with encounter

BioMed Research International 
T : Classication rate improvement caused by retraining. First
column shows the channel label, the second column shows the initial
training accuracy, and third one shows the marked correction by a
neurologist and the last one shows the nal accuracy.
Channel Accuracy aer
initial training (%)
Number of epochs
marked by the user
Accuracy aer
retraining (%)
Fp .  .
Fp .  .
F .  .
F .  .
C .  .
C .  .
P .  .
P .  .
O .  .
O .  .
F .  .
F .  .
T .  .
T .  .
T .  .
T .  .
FZ .  .
CZ .  .
PZ .  .
E.  .
PG .  .
PG .  .
A .  .
A .  .
T .  .
T .  .
X .  .
X .  .
X .  .
X .  .
X .  .
X .  .
X .  .
of better examples. More and better examples will certainly
improve its performance. e agreement between dierent
neurologists over the EEG readings is low to moderate. If
wecouldndtheagreementonatleastfewoftheepileptic
patterns correspondence with epileptic disease then we can
take this tool further ahead and use it for diagnosis instead of
just assistance.
One of the biggest limitations to this study is the unavail-
abilityofnon-Hzspikeandwavedata.Eventhoughwehave
included the data features of the entire epileptic frequency
ranges exclusive to each other, proof testing on the data will
certainly prove worthy for the progress of these assisting tools
toward a diagnostic tool.
Conflict of Interests
e authors declare that there is no conict of interests
regarding the publication of this paper.
Acknowledgment
e authors would like to extend their sincere appreciation to
the Deanship of Scientic Research at King Saud University
for funding this research through Research Group Project
(RG no. -).
References
[] H. Adelia, Z. Zhoub, and N. Dadmehrc, “Analysis of EEG
records in an epileptic patient using wavelet transform,” Journal
of Neuroscience Methods,vol.,no.,pp.–,.
[] M. A. Ahmad, W. Majeed, and N. A. Khan, “Advancements in
computer aided methods for EEG-based epileptic detection,” in
Proceedings of the 7th International Conference on Bio-Inspired
Systems and Signal Processing (BIOSIGNALS '14),Angers,
France, .
[] S. Noachtar and J. R´
emi, “e role of EEG in epilepsy: a critical
review,” Epilepsy and Behavior,vol.,no.,pp.–,.
[] E. B. Petersen, J. Duun-Henriksen, A. Mazzaretto, T. W. Kjar, C.
E. omsen, and H. B. D. Sorensen, “Generic single-channel
detection of absence seizures,” in Proceedings of the Annual
International Conference of the IEEE Engineering in Medicine
and Biology Society (EMBS ’11),pp.–,Boston,Mass,
USA, September .
[] M. Kaleem, A. Guergachi, and S. Krishnan, “EEG seizure
detection and epilepsy diagnosis using a novel variation of
Empirical Mode Decomposition,” in Proceedings of the 35th
Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC ’13),pp.–,Osaka,
Japan, July .
[] S. M. S. Alam and M. I. H. Bhuiyan, “Detection of seizure and
epilepsy using higher order statistics in the EMD domain,” IEEE
Journal of Biomedical and Health Informatics,vol.,no.,pp.
–, .
[] H.T.Ocbagabir,K.A.I.Aboalayon,andM.Faezipour,“Ecient
EEG analysis for seizure monitoring in epileptic patients,” in
ProceedingsoftheLongIslandSystems,ApplicationsandTech-
nology Conference (LISAT ’13), pp. –, IEEE, Farmingdale, NY,
USA, May .
[] A. A. Abdullah, S. A. Rahim, and A. Ibrahim, “Development of
EEG-based epileptic detection using articial neural network,”
in Proceedings of the International Conference on Biomedical
Engineering (ICoBE ’12),pp.–,Penang,Malaysia,.
[] P.Guo,J.Wang,X.Z.Gao,andJ.M.A.Tanskanen,“Epileptic
EEG signal classication with marching pursuit based on har-
mony search method,” in Proceedings of the IEEE International
Conference on Systems, Man, and Cybernetics (SMC ’12),pp.
–, Seoul, Republic of Korea, October .
[] A. Shoeb, “CHB-MIT scalp EEG database,” , http://
physionet.org/pn/chbmit/.
[] A. S. M. Murugavel, S. Ramakrishnan, U. Maheswari, and B.
S. Sabetha, “Combined Seizure Index with Adaptive Multi-
Class SVM for epileptic EEG classication,” in Proceedings

 BioMed Research International
of the International Conference on Emerging Trends in VLSI,
Embedded System, Nano Electronics and Telecommunication
System (ICEVENT ’13), pp. –, iruvannammalai, India,
Januar y .
[] M. H. Abdullah, J. M. Abdullah, and M. Z. Abdullah, “Seizure
detection by means of hidden Markov model and station-
ary wavelet transform of electroencephalograph signals,” in
Proceedings of the IEEE-EMBS International Conference on
Biomedical and Health Informatics (BHI ’12),pp.–,Hong
Kong, January .
[] A. Subasi and M. I. Gursoy, “EEG signal classication using
PCA, ICA, LDA and support vector machines,” Expert Systems
with Applications,vol.,no.,pp.–,.
[] E.Sezer,H.Is¸ik, and E. Saraco˘
glu, “Employment and compari-
son of dierent articial neural networks for epilepsy diagnosis
from EEG signals,” Journal of Medical Systems,vol.,no.,pp.
–, .
[] Brain Products GmbH/Products and Applications/Analyzer ,
http://www.brainproducts.com/productdetails.php?id=.
[] NeuroExplorer Home, http://www.neuroexplorer.com/.
[] Neuralynx, Spike S ort D Soware, , http://neuralynx.com/
research soware/spike sort d.
[]C.H.Seng,R.Demirli,L.Khuon,andD.Bolger,“Seizure
detection in EEG signals using support vector machines,” in
Proceedings of the 38th Annual Northeast Bioengineering Con-
ference (NEBEC ’12), pp. –, Philadelphia, Pa, USA, March
.
[] Z.A.Khan,S.B.Mansoor,M.A.Ahmad,andM.M.Malik,
“Input devices for virtual surgical simulations: a comparative
study,” in Proceedings of the 16th International Multi Topic Con-
ference (INMIC ’13), pp. –, Lahore, Pakistan, December
.
[] A. L. Goldberger, L. A. Amaral, L. Glass et al., “PhysioBank,
PhysioToolkit, and PhysioNet: components of a new research
resource for complex physiologic signals.,” Circulation,vol.,
no.,pp.E–E,.
[] S. Nasehi and H. Pourghassem, “Epileptic seizure onset detec-
tion algorithm using dynamic cascade feed-forward neural net-
works,” in Proceedings of the International Conference on Intelli-
gent Computation and Bio-Medical Instrumentation (ICBMI ’11),
pp. –, December .