You are on page 1of 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331245903

Feature Extraction and Classification of EEG Signals for Seizure Detection

Conference Paper · February 2019


DOI: 10.1109/ICREST.2019.8644337

CITATIONS READS

2 794

6 authors, including:

Abdullah Nahid Md.abdul Awal


Macquarie University Khulna University
50 PUBLICATIONS   244 CITATIONS    31 PUBLICATIONS   152 CITATIONS   

SEE PROFILE SEE PROFILE

Mohammad Ashik Alahe


Khulna University
1 PUBLICATION   2 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Image Analysis View project

Fault Classification Using Machine Learning View project

All content following this page was uploaded by Abdullah Nahid on 02 March 2019.

The user has requested enhancement of the downloaded file.


Feature Extraction and Classification of EEG
Signals for Seizure Detection
Apu Nandy Mohammad Ashik Alahe S.M.Nasim Uddin
Electronics and Communication Electronics and Communication Electronics and Communication
Engineering (ECE) Discipline Engineering (ECE) Discipline Engineering (ECE) Discipline
Khulna University Khulna University Khulna University
Khulna-9208,Bangladesh Khulna-9208,Bangladesh Khulna-9208,Bangladesh
Apu.ku.19@gmail.com ashik.ece140903@gmail.com unasim4o@gmail.com

Shafiqul Alam Adullah-Al-Nahid Md. Abdul Awal*


Electronics and Communication Electronics and Communication Electronics and Communication
Engineering (ECE) Discipline Engineering (ECE) Discipline Engineering (ECE) Discipline
Khulna University Khulna University Khulna University
Khulna-9208,Bangladesh Khulna-9208,Bangladesh Khulna-9208,Bangladesh
tushar140933@gmail.com nahidku@yahoo.com awalece04@yahoo.com;
m.awal@ece.ku.ac.bd
*Corresponding Author.

Abstract— Epileptic seizure is a neurological disorder imaging (MRI), computed tomography (CT) scan, positron
characterized by abnormal synchronous discharge of the emission tomography (PET), ultrasound,
neuronal activities in the brain structures. These abnormal Electroencephalogram (EEG). MRI, CT and Ultrasound are
electrical activities can be recorded via multi-channel costly and cannot be used for long term evaluation. On the
electroencephalography (EEG) signals placed on the scalp of the
brain. Usually, these signals, recorded from this EEG device, are
other hand, EEG is a low cost tool and non-invasive tool
interpreted by the neurologist which require their availability which can be used for long-term evaluation. Therefore, EEG
and it is very time consuming especially for long duration is the most useful tool for the diagnosis of epilepsy. In this
signals. This study presents a fully automatic system for the test, electrodes are attached on the scalp of the brain via
detection of seizure from non-seizure signals. Firstly, it pre- adhesive gel. If a person has epilepsy, his/her EEG recording,
processes the signal to remove noise and artefacts from the raw- and EEG pattern will be different from normal pattern.
EEG signals and then extracts features. Features are extracted Neurologist reads that EEG recoding and take decision based
from time-domain, spectral domain, wavelet domain. In on the pattern. Therefore, epileptic seizure is connected with
addition, connectivity and entropy based feature have also been the primary parameter which is provided by the EEG signal
extracted. After that, prominent features have been selected
from this large feature set by a multi-objective evolutionary
[1]. Epileptic activity associated with the actual detector can
algorithm and finally, Support Vector Machine (SVM) classifier be integrated into a cellular feedback program for daily
has been used for classification. A Bayesian optimization utilization [2-4]. After all the consequences, epileptic seizure
algorithm has been used to optimize the hyper-parameters of can be shortened lifespan and include social and
SVM. Linear Discriminant Analysis (LDA) and Quadratic neuropsychological disability, excessive bodily injury.
Linear Discriminant Analysis (QLDA) have also been used for Performance detection of existing algorithms may be affected
comparison. The proposed system is tested on a publicly by such kind of differences. So, EEG is more conventional
available CHB-MIT database and results show the significance than other diagnosing systems [5]. Different features have
of the proposed system. The distinguished accuracy of the
been extracted from discrete wavelet transform (DWT) and
classifier is 76.41%, 80.79% and 97.05% in LDA, QLDA and
SVM, respectively.
artificial neural network (ANN). For example, standard
deviation, mean of absolute value, average power and ratio of
Keywords— Bayesian optimization, Classifier, absolute mean values of DWT were applied in [6], relative
Electroencephalogram (EEG), Feature selection, LDA, QLDA, energy was applied in [7] and line length of DWT was studied
SVM, Seizure. in [8] and energy of DWT was used in [9]. On the other hand,
SVM and DWT features have been used by [10][11]. Sample
I. INTRODUCTION entropy (SampEn) and distribution entropy (DistEn) and
SVM was applied in [12]. As can be seen that, different
Epilepsy is a brain neurological disorder and often studies use different type of features. Therefore, it is
observed in the ictal period and between inter-ictal period. It important to explore and extract a large set of features and
is a common neurological disorder which affects people of all use feature selection algorithm in order to get prominent
ages [1, 2]. It is one of the most common neural disorder that feature subset. Motivated by this, we have extracted different
affects about 1% of the whole world population and among features from different domain. The major contributions and
them about 0.2% affected individuals lose their lives. A key topics covered by this study are as follows:
report reveals that approximately 50 million people of • We have collected publicly available bipolar EEG
worldwide have epilepsy and approximately 1.38 million database and selected channel to reduce
people are suffering from epilepsy in our country. Mostly 1 computational complexity (Section II.A).
in every 200 newborn babies face some form of seizures. • Prominent feature subset has been selected using a
Different diagnostic tools such as magnetic resonance multi-objective feature selection algorithm that

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE


simultaneously select least features subset with it requires left hemispheric and right hemispheric channel
minimum error (Section II.D). information. Therefore, for left hemispheric and right
• A Bayesian optimization algorithm have been used to hemispheric EEG channel are formed by the following
optimise the radial basis function kernel-support equations:
vector machine (RBF-SVM) classifier in section IV.
• The result of the proposed method has been compared 𝐿
1
with other methods in section V. 𝐿𝑒𝑓𝑡ℎ𝑒𝑚𝑖𝑠𝑝ℎ𝑒𝑟𝑖𝑐 [𝑛] = (∑ 𝑥𝑖 [𝑛]) (2)
𝐸𝐸𝐺 𝐿
𝑖=1
II. METHODOLOGY
𝑅
Electroencephalography (EEG) is used in the neonatal 1
𝑅𝑖𝑔ℎ𝑡ℎ𝑒𝑚𝑖𝑠𝑝ℎ𝑒𝑟𝑖𝑐 [𝑛] = (∑ 𝑥𝑖 [𝑛]) (3)
intensive care environment to monitor brain signal of ill new- 𝐸𝐸𝐺 𝑅
𝑖=1
born babies. This technology provides continuous assessment
of cortical function. Specialists are required to identifying
where, 𝐿 = 2 is the number of left hemispheric channels
seizures if present [13], assessing brain maturation and
grading brain injury [14]. The process of epilepsy diagnosis and 𝐿 ∈ (F3 − C3), (C3 − P3) and 𝑅 = 2 is the number of
checks whole parts that is shown in Fig 1: right hemispheric channels and 𝑅 ∈ (F4 − C4), (C4 − P4) .

Fig. 1: Simplified block diagram of this study.

At the beginning of the categorized process collecting


appropriate data is the main concern. For this purpose, we
have used CHB-MIT database related epileptic seizure which
is publicly available at [15]. After collecting the data all the
physical-biological data is converted into MATLAB data Fig. 2: 10-20 system of electrode placement.
using MATLAB EEG LAB tool kit. In pre-processing stage,
the artefacts are removed from the main signal to find the B. Feature Extraction
artefact free signal. This process is done by filtering the data
using band-pass filter (0.5 − 30 𝐻𝑧). Here, the major After channel selection the features of the test signals are
artefacts are removed by the method presented in [16]. After extracted from it using various equation and algorithm. The
removing artefacts all pure channels data are low passed filter extracted features are amplitude features, range EEG or EEG
using finite impulse response (FIR) filter. features, connectivity features, spectral features, wavelet and
entropy features where the number of the features found from
A. Channel Selection each section is 24, 32, 20, 21, 44 and 4 respectively. The total
Channel selection can reduce the system complexity as a number of features extracted is 145.
smaller number of channels is used in the analysis. Greene et.
al. showed that Channel C3-C4 provided the high accuracy of C. Features Normalization and Data Balancing
90.77% in seizure classification [17]. Reference [18]
reviewed that P3 and P4 position of the brain is more prone After finding all the features these are need to be
to seizure and it is over the cerebrovascular watershed region normalized. To normalize these features rescaling and mean
which is the high risk for brain injury. Considering these, we normalize equations are used that is given bellow:
have selected four channels i.e., (𝐹3 − 𝐶3) , (𝐶3 − 𝑃3) ,
(𝐹4 − 𝐶4), (𝐶4 − 𝑃4) in this study; see Fig. 2. To reduce the 𝑥−𝑎𝑣𝑒𝑟𝑎𝑔𝑒(𝑥)
computational complexity these 4 channels are averaged by 𝑥 ′ = max(𝑥)−min(𝑥) (4)
the following equation:
Where, 𝑥 represents an original value and 𝑥′ is the
1
𝐴𝑣𝑒𝑟𝑔𝑎𝑟𝑒𝑑𝐸𝐸𝐺 [𝑛] = (∑𝐶𝑖=1 𝑥𝑖 [𝑛]) (1) normalized value. The number non-seizure segments are
𝐶
higher than seizure segments i.e. dataset is unbalanced. To
balance this data, we have used Adaptive Synthetic
This single 𝐴𝑣𝑒𝑟𝑔𝑎𝑟𝑒𝑑_𝐸𝐸𝐺[𝑛] has been used for feature (ADASYN) algorithm [19].
extraction. For the extraction of connectivity-based features,
D. Feature Selection Here, training vectors 𝑥𝑖 are mapped into a higher (maybe
infinite) dimensional space by the function ∅. 𝐶 > 0 is the
All the features have been tested on the EEG data set. penalty parameter of the error terms. Furthermore,
Result of which are seen to vary considerably from feature to 𝐾(𝑥𝑖 , 𝑥𝑗 ) ≡ ∅(𝑥𝑖 )𝑇 ∅(𝑥𝑗 ) is called the kernel function. It can
feature. Feature selection provides less data and hence the be seen from TABLE I that the accuracy of the LDA is lower
classification system will be less complex and boost the due to the problem that the data and feature vectors are not
computation of machine learning algorithms. In this study, linearly separable. Kernel trick can able to transfer non-
Multi-objective evolutionary algorithm (MOEA) has been separable features into separable feature space and hence,
used for feature selection [20]. This algorithm simultaneously improve the accuracy. Popular kernel is Radial Basis
optimizes two objectives: (i) selects lower number of features Function (RBF) kernel which can be expressed as:
and (ii) with minimum classification error.
2
−‖𝑥𝑖 −𝑥𝑗 ‖
III. CLASSIFICATION 𝐾(𝑥𝑖 , 𝑥𝑗 ) = exp ( ),σ > 0 (8)
2𝜎 2

Classification is used to classify the test data with the 2


trained data of the various classifiers. By which it is possible The ‖𝑥𝑖 − 𝑥𝑗 ‖ is the squared Euclidian distance between
to find whether the data is a seizure or non-seizure. Three two features and can be interpreted as similarity measure and
classifiers, Linear Discriminant Analysis (LDA), Quadratic 𝜎 is the controlling parameter. To summarize, in the RBF-
Linear Discriminant Analysis (QLDA), Support Vector SVM, 𝐶 and 𝜎 parameters need to optimize for better
Machine (SVM), have been used in this study. classification performance.
D. Classification Performance Evaluation
A. Linear Discriminant Analysis (LDA)
Different performance evaluation matrices such as
LDA is achieved by comparing the variation of the test sensitivity, specificity, accuracy, F-measured, kappa index,
data with the predetermined dataset of LDA. At the Mathew’s correlation coefficient (MCC), Precision, and Area
conclusion from the procedure it's wished that every course under the curve (AUC) have been used. They are defined by
may have an ordinary submission associated with the following equations (Eq. 9 – Eq. 14).
discriminant ratings. This discriminant rating is tended to be
determined with a discriminant perform that has the shape: Sensitivity (SNV) or true positive rate
𝑆𝑁𝑉 = 𝑇𝑃⁄(𝑇𝑃 + 𝐹𝑁) (9)
D = w1Z1 + w2 Z2 + w3 Z3 … … … … . wp Zp (5)
Specificity (SPC) or true negative rate (TNR)
B. Quadratic Linear Discriminant Analysis (QLDA) 𝑆𝑃𝐶 = 𝑇𝑁⁄(𝑇𝑁 + 𝐹𝑃) (10)

QLDA is not so different from LDA except the Precision or positive predictive value (𝑃𝑃𝑉)
covariance matrix. The covariance matrix ∑k is separated for 𝑃𝑃𝑉 = 𝑇𝑃⁄(𝑇𝑃 + 𝐹𝑃) (11)
each class, k = 1,2, … … K. Quadratic discriminant function:
Accuracy (𝐴𝐶𝐶)
1 (𝑇𝑃 + 𝑇𝑁)
𝐴𝐶𝐶 = ⁄(𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁 + 𝑇𝑁) (12)
𝛿𝑘 = − 𝑙𝑜𝑔 |∑ |
2
𝑘
−1
1 Matthews’s correlation coefficient (𝑀𝐶𝐶)
− (𝑥 − 𝜇𝑘 )𝑇 ∑(𝑥 − 𝜇𝑘 ) + 𝑙𝑜𝑔𝜋𝑘 (6) 𝑇𝑃 × 𝑇𝑁 − 𝐹𝑃 × 𝐹𝑁
2 𝑀𝐶𝐶 =
𝑘 √(𝑇𝑃 + 𝐹𝑃)(𝑇𝑃 + 𝐹𝑁)(𝑇𝑁 + 𝐹𝑃)(𝑇𝑁 + 𝐹𝑁)
(13)
C. Support Vector Machines (SVM) 𝐹1 Score is the harmonic mean of precision and sensitivity
𝐹1 = 2𝑇𝑃⁄(2𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁) (14)
SVM tries to build an optimal hyper plane from a given
training set This optimal model has been used for the Here, 𝑇𝑃, 𝐹𝑃, 𝑇𝑁, 𝐹𝑁 represent true positve, false
classification of new examples. Given a training set of positive, true negative and false negative, respectively.
instance-label pairs(𝑥𝑖 , 𝑦𝑖 ),𝑖 = 1,2, … … . . , 𝑙 where 𝑥𝑖 ∈ 𝑅 𝑛
and 𝑦 ∈ (1, −1)𝑙 the support vector machines (SVM) [21] IV. RESULTS
require the solution of the following optimization problem:
𝑙
𝑚𝑖𝑛
1 𝑇 In this study, EEG based seizure prediction result will be
𝑤,𝑏,𝜖 𝑊 𝑊 + 𝐶 ∑ 𝜀𝑖 (7)
2 presented. To perform this, we have collected data from
𝑖=1
CHB-MIT base database. The raw EEG data was pre-
processed to remove major artefact.
Subject to 𝑦𝑖 (𝑊 𝑇 ∅(𝑥𝑖 ) + 𝑏) ≥ 1 − 𝜀𝑖 , 𝜀𝑖 ≥ 0
TABLE I. COMPARISION AMONG LDA, QLDA AND SVM

Classification performance indexes (in %)


Classifier Name

Correlation
Coefficient
Specificity
Sensitivity
Accuracy

F1_Score
Precision

Mathews
Positive

Kappa
Error

False

AUC
LDA 76.41 23.59 76.09 76.73 76.61 23.27 76.35 52.82 83.05 52.82

QLDA 80.79 19.21 73.40 88.20 86.17 11.80 79.27 62.28 89.98 61.59

SVM 97.05 2.95 98.99 95.12 95.29 4.88 97.11 94.17 99.54 94.10

After that, it was band-passed filtered to get only (0.5-30) whose objective function is the minimum one. In this figure,
Hz EEG. We know that, the most of the activities live the lowest point which represents the lowest classifier error
between this frequency bands. Different time domain value. The lowest of error is the highest of the accuracy of
features, frequency feature, time-frequency features (wave- this objective function.
length) as well as entropy base features has been extracted. A
total of 145 features have been extracted. The dataset was
imbalanced i.e., the number of non-seizure data is greater
than seizure data. The imbalanced dataset was balanced by
using ADASYN algorithm. A feature selection algorithm
called Multi-objective evolutionary feature selection based
on ENORA (Evolutionary Non-dominated Radial slots based
Algorithm) has been used [20]. The calling parameters are:
Generation=100, Population size=200, Report frequency=10,
Seed=1. The selected features using this algorithm are given
TABLE II:

TABLE II: SELECTED FEATURES

Serial number Features#


1 amplitude_skewness4
2 spectral_entropy1
3 spectral_entropy4 Fig. 3: SVM parameter optimization using Bayesian optimization
algorithm.
4 Spectral-edge-frequency
5 rEEG_median4
6 connectivity_BSI1
7 connectivity_corr2
8 connectivity_coh_freqmax1
9 connectivity_coh_freqmax2
10 connectivity_coh_freqmax3
11 connectivity_coh_freqmax4
12 energy4
Here, # In the feature’s column, (∗)𝑛 , 𝑛 = 1,2,3,4 represents Delta,
Theta, Alpha, Beta band signal. For example, 𝑠𝑝𝑒𝑐𝑡𝑟𝑎𝑙_𝑒𝑛𝑡𝑟𝑜𝑝𝑦 1 presents
in Delta band EEG signal.

In TABLE I, the performance of different classifiers has


been represented. The accuracy of the classifiers is 76.41%,
80.79% and 97.05% in LDA, QLDA and SVM, respectively.
Besides accuracy, the sensitivity is 76.09%, 73.40% and Fig. 4: Confusion matrix of LDA.
98.99% and the AUC is 83.05%, 89.98% and 99.54% of
LDA, QLDA and SVM, respectively. The optimized SVM From Fig. 4, it can be seen that LDA can classify 75.6%
outperformed LDA and QLDA. of the non-seizure data accurately whereas it misclassifies to
24.4%. Again, it can classify 75.3% of the seizure data
Fig. 3 shows that how the objective function i.e. classifier accurately and it misclassifies 24.7% of the seizure data.
error is minimized by optimizing 𝐶 and 𝜎. Here, different
values give different minimization and finally we got the In Fig. 5, it is seen that the overall accuracy is 80.80%.
minimum value using the Bayesian optimization algorithm Here, QLDA can classify 86.20% of the non-seizure data
accurately whereas it misclassifies to 13.80%. Again, it can Fig. 7, represents the Receiver Operating Characteristics
classify 76.8% of the seizure data accurately and it (ROC) curves for SVM, QLDA and LDA. It can be seen that
misclassifies 23.2% of the seizure data. SVM is out performing over QLDA and LDA.
V. COMPARISION WITH OTHER METHODS

To compare our proposed optimized RBF-SVM method,


we have compared our results with time domain, frequency
domain, connectivity based and entropy-based features.
Proposed method obtained 97.05% accuracy whereas time
domain, frequency domain, connectivity based, entropy-
based feature obtained 67.14%, 64.87%, 79.76% and 57.76%
accuracy respectively (TABLE III). This demonstrate the
superior performance of our proposed methods.

In addition, to demonstrate the effect of features selection


method, we compared our proposed result with features
selection to the result without features selection. In both case
Fig. 5: Confusion matrix of QLDA.
optimized SVM has been used and the accuracy improved to
5.90% in the result with features selection; see TABLE III.
Similarly, from Fig. 6, it is seen that the overall accuracy
is 97.1%. The confusion matrix shows that optimized SVM VI. DISSCUSSION AND CONCLUSION
can classify 95.3% of the non-seizure data accurately whereas
it misclassifies to 4.7%. Again, it can classify 98.9% of the Epileptic seizure is one of the major cause of morbidity
seizure data accurately and it misclassifies 1.1% of the
and mortality around the globe. The normal procedure
seizure data. Analysis of confusion matrix reveals that SVM
includes the acquisition of EEG which is then observed by an
can classify accurately non-seizure and seizure data than
expert neurologist who terms it as normal or patient. This
LDA and QLDA. Among these classifiers we can estimate
procedure is both time consuming and requires expert
that the SVM is the most optimized classifier.
neurologists as well which is not available in many countries,
especially in the developing countries like Bangladesh. To
help the patients, an automated system is required which can
replace this entire procedure. The proposed automated
system presented in this study focuses on automatic seizure
detection using four distinct processes. They are: (i) pre-
processing, (ii) feature extraction, (iii) feature selection and
(iv) classification. The pre-processing generates noise-free
EEG signals whereas feature extraction part extracts several
(145) features. The best features are selected from the
available set of features and passed over to classification
block. The classification of the signal is governed by a RBF-
SVM classifier, which optimized for the best results. The
proposed method has achieved an overall value of AUC,
accuracy, sensitivity and specificity, as 99.54%, 97.05%,
98.99% and 95.12%, respectively.

The results have shown that it can detect a larger number


of seizure samples correctly. The features used in this
Fig. 6: Confusion matrix of SVM.
research work are from time, frequency, connectivity and
entropy based features. We believed that this work is one of
the suitable solutions for classifying the EEG signals for
seizure detection.

In the present work, only three classifiers have been used.


Some other classifiers such as random forest, Xgboost can be
used. In addition, limited number of datasets have been used.
In future, we will address these limitations, compare the
performance of the proposed classifier with other classifiers,
add more analysis and in-depth discussion.

Fig. 7: ROC of different classifiers.


TABLE III. COMPARISION WITH O THER METHODS

Classification performance indexes (in %)


SVM operation
with different

Specificity
Sensitivity
Accuracy

F1_Score
Precision

Kappa
Positive

MCC
Error

False

AUC
features
domain

Time 67.14 32.86 73.74 60.54 65.18 39.46 69.19 34.58 72.23 34.28
Frequency 64.87 35.13 71.89 57.84 63.07 42.16 67.19 30.03 70.45 29.73
Connectivity 79.76 20.24 83.64 75.89 77.62 24.11 80.52 59.71 88.22 59.53
Entropy 57.76 42.24 75.38 40.13 55.74 59.83 64.09 16.58 58.70 15.51
Without feature selection 91.15 8.85 82.29 100 100 0 90.29 83.61 99.62 82.29
SVM with feature selection (Proposed) 97.05 2.95 98.99 95.12 95.29 4.88 97.11 94.17 99.54 94.10

REFERENCES

level decomposition," in Machine Learning for Signal


[1] U. R. Acharya, S. V. Sree, G. Swapna, R. J. Martis, and J. S. Processing (MLSP), 2015 IEEE 25th International Workshop
Suri, "Automated EEG analysis of epilepsy: a review," on, 2015, pp. 1-6.
Knowledge-Based Systems, vol. 45 pp. 147-165, 2013.
[12] M. Li, W. Chen, and T. Zhang, "Automatic epilepsy detection
[2] S. Ramgopal, S. Thome-Souza, M. Jackson, N. E. Kadish, I. S. using wavelet-based nonlinear analysis and optimized SVM,"
Fernández, J. Klehm, et al., "Seizure detection, seizure Biocybernetics and Biomedical Engineering, vol. 36 (4), pp.
prediction, and closed-loop warning systems in epilepsy," 708-718, 2016.
Epilepsy & behavior, vol. 37 pp. 291-307, 2014.
[13] G. B. Boylan, N. J. Stevenson, and S. Vanhatalo, "Monitoring
[3] N. Kovacevic, P. Ritter, W. Tays, S. Moreno, and A. R. neonatal seizures," Seminars in Fetal and Neonatal Medicine,
McIntosh, "‘My Virtual Dream’: Collective Neurofeedback in vol. 18 (4), pp. 202-208, 2013.
an Immersive Art Environment," PloS one, vol. 10 (7), p.
e0130129, 2015. [14] M. André, M.-D. Lamblin, A.-M. d’Allest, L. Curzi-
Dascalova, F. Moussalli-Salefranque, M.-F. Vecchierini-
[4] M. E. Menshawy, A. Benharref, and M. Serhani, "An Blineau, et al., "Electroencephalography in premature and full-
automatic mobile-health based approach for EEG epileptic term infants. Developmental features and glossary," Clinical
seizures detection," Expert Systems with Applications, vol. 42 Neurophysiology, vol. 40 (2), pp. 59-124, 2010.
(20), pp. 7157-7174, 2015.
[15] (May 22, 2018). CHB-MIT Scalp EEG. Available:
[5] I. Lushin, "Detection of epilepsy with a commercial EEG https://www.physionet.org/pn6/chbmit/
headband," 2016. [16] J. M. Toole and G. B. Boylan, "NEURAL: quantitative
[6] A. Subasi, "EEG signal classification using wavelet feature features for newborn EEG using Matlab," arXiv preprint
extraction and a mixture of expert model," Expert Systems with arXiv:1704.05694, 2017.
Applications, vol. 32 (4), pp. 1084-1093, 2007. [17] B. Greene, G. Boylan, W. Marnane, G. Lightbody, and S.
[7] S.-H. Lee and J. S. Lim, "Minimum feature selection for Connolly, "Automated single channel seizure detection in the
epileptic seizure classification using wavelet-based feature neonate," in 30th Annual International Conference of the IEEE
extraction and a fuzzy neural network," Applied Mathematics Engineering in Medicine and Biology Society,, 2008, pp. 915-
& Information Sciences, vol. 8 (3), p. 1295, 2014. 918.

[8] L. Guo, D. Rivero, J. Dorado, J. R. Rabunal, and A. Pazos, [18] M. A. Awal, M. M. Lai, G. Azemi, B. Boashash, and P. B.
"Automatic epileptic seizure detection in EEGs based on line Colditz, "EEG background features that predict outcome in
length feature and artificial neural networks," Journal of term neonates with hypoxic ischaemic encephalopathy: A
neuroscience methods, vol. 191 (1), pp. 101-109, 2010. structured review," Clinical Neurophysiology, vol. 127 (1), pp.
285-296, 2016.
[9] I. Omerhodzic, S. Avdakovic, A. Nuhanovic, and K.
Dizdarevic, "Energy distribution of EEG signals: EEG signal [19] H. He, Y. Bai, E. A. Garcia, and S. Li, "ADASYN: Adaptive
wavelet-neural network classifier," arXiv preprint synthetic sampling approach for imbalanced learning," in
arXiv:1307.7897, 2013. IEEE International Joint Conference on Neural Networks,
(IJCNN), 2008, pp. 1322-1328.
[10] C. A. Lima, A. L. Coelho, and M. Eisencraft, "Tackling EEG
signal classification with least squares support vector [20] F. Jiménez, G. Sánchez, J. M. García, G. Sciavicco, and L.
machines: A sensitivity analysis study," Computers in Biology Miralles, "Multi-objective evolutionary feature selection for
and Medicine, vol. 40 (8), pp. 705-714, 2010. online sales forecasting," Neurocomputing, vol. 234 pp. 75-92,
2017.
[11] D. Chen, S. Wan, and F. S. Bao, "Epileptic focus localization
using EEG based on discrete wavelet transform through full- [21] C. Cortes and V. Vapnik, "Support-vector networks," Machine
learning, vol. 20 (3), pp. 273-297, 1995.

View publication stats

You might also like