ROBUST DETECTION OF EPILEPTIC SEIZURES USING DEEP NEURAL NETWORKS

Ramy Hussein ?§ , Hamid Palangi † , Z. Jane Wang § , and Rabab Ward §

§
Electrical and Computer Engineering, University of British Columbia, Vancouver, BC, Canada
†
Microsoft Research AI, Redmond, WA 98052, United States

ABSTRACT have been investigated in the literature. The majority of these

use EEG features that are manually extracted from the time
Robust detection of epileptic seizures in the presence of in-
domain [4], frequency domain [5], time-frequency domain
evitable artifacts in Electroencephalogram (EEG) signals is
[6] and sometimes from a combination of two domains [7].
addressed. The EEG dataset considered contains 300 signals
In practice, these hand-crafted features experience three main
recorded from 15 volunteers. Current seizure detection sys-
challenges. First, they are very sensitive to variations in
tems achieve good performance when the EEG data is entirely
seizure activities since EEG is a non-stationary signal and
free of noise. However, their performance drastically decays
the seizure behavior varies across different patients and over
with authentic EEG data polluted by real artifacts. We intro-
time for the same patient. Secondly, the EEG measurements
duce a robust seizure detection method that can address clean
are highly susceptible to different sources of noise such as
and noisy data. The proposed method uses Long Short-Term
muscle activities, eye-blinking/movement, and environmental
Memory (LSTM) neural networks to extract the representa-
white noise. These artifacts interfere with the EEG data and
tive EEG features pertinent to seizures. Experimental results
seriously impact the detection accuracy of epileptic seizures
show that the proposed method beats existing methods by
[8]. Finally, most of the existing seizure detection methods
achieving 100% classification accuracy. Our method is also
have been built based on limited EEG data taken from few
shown to be robust against the common EEG artifacts (e.g.,
subjects, making them less practical for clinical applications.
muscle activities and eye-blinking) and white noise.
To solve these problems, we introduce a robust deep learn-
Index Terms— Epilepsy, seizure detection, EEG, LSTM ing approach for robust detection of epileptic seizures. The
time-series EEG signals are first divided into small EEG seg-
1. INTRODUCTION ments, which are then presented as an input to a Long Short-
Term Memory (LSTM) network to learn the discriminative
Epilepsy is a neurological brain disorder that affects around characteristics of seizure and non-seizure activities. The soft-
70 million people worldwide [1]. The defining characteristic max model is then used for EEG classification. The overall
of epilepsy is recurrent seizures that strike without warning. model is examined, under ideal and real-life conditions, us-
Symptoms may range from brief suspension of awareness ing the popular clinical dataset provided by Bonn University
to violent convulsions and sometimes loss of consciousness [9]. The results demonstrate that, under ideal conditions, the
[2]. Routine Electroencephalogram (EEG) has been widely proposed approach outperforms the baseline methods in terms
used in clinical settings for the diagnosis of epileptic seizures. of the sensitivity, specificity, and classification accuracy. It is
However, the visual inspection of EEG is laborious and time- also shown that the proposed approach maintains its superior
consuming. Even worse, around 75% of people with epilepsy performance in noisy environments.
live in low and middle-income countries and may not afford
to consult clinicians and neurologists [3]. Automatic seizure
2. MATERIALS AND PRIOR WORK
detection systems, however, can minutely identify the EEG
seizure patterns and help patients watch their own risks and
2.1. EEG Dataset
improve their quality of life.
Several methods have been developed for automatic In this work, the proposed seizure detection method is ex-
seizure detection using EEG signals. Extracting the best amined on the well known EEG dataset provided by Bonn
features that describe the characteristics of EEG seizure activ- University [9]. In this study, we address the classification
ities has a significant influence on the performance of seizure problem between the following three different EEG sets:
detection systems. Numerous feature extraction approaches Normal EEG taken from five healthy volunteers, Inter-ictal
This work was supported by Vanier Canada Graduate Scholarship from
EEG recorded from five epileptic patients during seizure-free
the Natural Sciences and Engineering Research Council of Canada (NSERC). intervals, and Ictal EEG taken from five epileptic patients
? Corresponding author: [email protected] while having active seizures. Each set contains 100 single-

978-1-5386-4658-8/18/$31.00 ©2018 IEEE 2546 ICASSP 2018

 200 2000

Amplitude ( V)
(a) (a)
Amplitude ( V)

100
0

-2000
-100 0 5 10 Time (Sec) 15 20

2000

Amplitude ( V)
-200 (b)
0 5 10 15 20
Time (Sec)
200 0
(b)
Amplitude ( V)

100
-2000
0 5 10 Time (Sec) 15 20
0
2000

Amplitude ( V)
(c)
-100
0
-200
0 5 10 15 20
Time (Sec)
-2000
(c) 0 5 10 Time (Sec) 15 20
1000
Amplitude ( V)

2000

Amplitude ( V)
(d)
0
0

-1000
-2000
0 5 10 15 20
0 5 10 15 20 Time (Sec)
Time (Sec)

Fig. 1. Time-series EEG plots: (a), (b), and (c) examples of Fig. 2. Time-series EEG plots: (a) clean ictal signal; (b),
normal, inter-ictal, and ictal EEG activities, respectively. (c), and (d) ictal signals corrupted with muscle artifacts, eye-
blinking, and white noise, respectively.
channel EEG signals, each of 23.6 seconds duration. All
the EEG signals had already been denoised, amplified, sam- 2.2. Prior Work
pled at 173.6Hz and digitized using a 12-bit analog-to-digital
converter. Figures 1(a), (b), and (c) show examples of these The automatic detection of epileptic seizures using EEG sig-
time-series EEG signals corresponding to normal, inter-ictal, nals has been broadly investigated over the past three decades.
and ictal activities, respectively. In this study, we will cover the prior work that has been re-
In practice, several sources of noise affect the EEG sig- cently developed to discriminate between three classes: Nor-
nals contaminating the seizure manifestations and negatively mal, Inter-ictal, and Ictal EEG patterns [11]-[22]. In [11], the
affecting the detection accuracy of epileptic seizures. The potential of wavelet transform to obtain and analyze the main
authors of [10] reviewed the most common EEG artifacts spectral rhythms of the EEG signals is investigated. Then, the
and developed adequate models to mimic their behavior. In statistical features that characterize the behavior of the EEGs
this study, we focus on the most three critical and inevitable were extracted and tested using a neural network-based clas-
sources of artifacts, which are: sification model called mixture of experts (ME). The results
showed a classification accuracy of 93.17%. This was im-
• Muscle Artifacts: As shown in [10], muscle activities
proved one year later to 94.83% by using the same features
can be modeled by random noise filtered with a band-
to examine the performance of a multilayer perceptron neural
pass filter (BPF) of 20Hz and 60Hz cut-off frequencies.
network (MLPNN) classifier [12]. In [13], a feature extrac-
• Eyes Movement/Blinking: The eye blinks mixed with tion method based on the sample entropy was used together
EEG data can be modeled as a random noise signal fil- with the extreme learning machine (ELM) classifier, and clas-
tered with a BPF of 1Hz and 3Hz cut-off frequencies. sification accuracy of 95.67% was reached. Besides, in [14],
a set of temporal and spectral EEG features was fed into an
• White Noise: The electrical and environmental noises
MLPNN for EEG classification. This resulted in a classifica-
are characterized as white Gaussian noise [10].
tion accuracy of 97.50%.
Fig. 2(a) shows an arbitrary noise-free EEG signal corre- In an effort to alleviate the computational complexity bur-
sponding to seizure (ictal) activities, while figures 2(b), (c), den in seizure detection systems, Acharya et al. relaxed the
and (d) show the noisy versions of the same signal corrupted need for any pre-processing techniques and worked directly
with muscle artifacts, eye-blinking, and white noise, respec- on the raw EEG data [15, 16, 17]. For example, in [15],
tively. The amplitude of each of these artifacts can be adjusted they extracted a set of 10 robust statistical features from the
to produce noisy EEG signals with different signal-to-noise- time-series EEG recordings in the absence of any filtering or
ratios (SNRs). The SNR of the noisy signals shown in Fig. 2 denoising approaches. The effectiveness of the selected fea-
is set to 0dB. This is where the power values of the EEG sig- tures was examined along with seven different classifiers. The
nal and the noise signal are equal. support vector machine (SVM) achieved better performance

2547
than the other classifiers with an average classification accu- length can lessen the computational complexity of the LSTM
racy, 94.40%. In [16], Acharya et al. significantly improved models but at the cost of the detection accuracy [28]. Hence,
the performance of their seizure detection model by introduc- each EEG segment is designed to have only 2 data-points out
ing more representative EEG features. These features were of 4096, producing 2048 segments for each EEG signal.
approximate entropy, sample entropy, and phase entropies, In the second stage, the deep learning model of LSTM
and were computed from the recorded EEG signals and then was deployed to extract the discriminative EEG features that
fed into fuzzy Sugeno classifier (FSC) for EEG classification. best describe seizure characteristics. The segmented EEG
This approach notably boosted the classification accuracy to data samples were initially shuffled randomly to remove any
98.10%. Additionally, Acharya et al. proposed the use of possible drifts. Then, we design our deep neural network to
wavelet packet transform (WPT) to analyze the EEG signals include three layers, with a MaxPooling layer as the top layer.
into eight approximation and detail wavelet bands [17]. The The shuffled EEG data samples were first passed through a
wavelet coefficients of these bands were then used to infer the fully connected (Dense) layer, which performs a linear op-
eigenvalues which were used as the input to the Gaussian mix- eration from the input to the output. The LSTM layer was
ture model (GMM) classifier achieving an outstanding classi- adopted afterward to learn the most robust EEG features perti-
fication accuracy of 99.00%. Comparable classification accu- nent to seizures. Subsequently, the output of the LSTM layer
racy of 98.67% was achieved in [18] by using a feature ex- was fed into the MaxPooling layer to help reduce the over-
traction method based on recurrence quantification analysis fitting by providing an abstracted form of the EEG representa-
integrated with a two-stage classifier named error-correction tions. MaxPooling layer also reduces the computational cost
output code (ECOC). by reducing the number of parameters to learn and provide
Chiang et al. developed an energy-efficient EEG moni- basic translation invariance to the internal representation.
toring system for epileptic seizure detection [19]. They intro- In the third stage, the output of the MaxPooling layer
duced a novel EEG feature extraction method. The features is presented as an input to a probabilistic classification
were extracted on the sensor side, and the seizure detection model of softmax to create label predictions [29]. The
was performed on the server side. This resulted in a signifi- class labels are assumed to be: y (i) ∈ 1, · · ·, K, where
cant saving in the power consumption, and a detection accu- K is the total number of classes. Given a training set
racy of 95.61% was obtained. In [20], a piecewise quadratic {(x(1) , y (1) ), (x(2) , y (2) ), · · · , (x(N ) , y (N ) )} of N labeled
(PQ) classifier was built for detecting epileptic EEG episodes, samples, where x(i) ∈ <(N ) . For each test sample x, the
and it was integrated with a combination of temporal, spec- softmax hypothesis evaluates the probability that P(y =
tral, and non-linear features to reach 98.70% classification ac- k|x(t), x(t − 1), x(t − 2), · · · ) for each class label k =
curacy. In [21], a feature extraction method based on the dis- 1, · · · , K; where t shows the time step of each EEG seg-
crete short-time Fourier transform was adopted together with ment. The summations of these K-probability values should
an MLPNN classifier to achieve a high detection accuracy of equal to 1 and the highest probability belongs to the predicted
99.10%. In [22], a hybrid scheme based on some statistical class. The cost function of the softmax classifier is the cross
features and a least-square SVM (LSSVM) classifier showed entropy, denoted by J(θ), is defined as follows:
an average classification accuracy of 97.19%. In [23], the
energy features of EEG-based harmonic WPT and fractal di-  
mensions were computed and tested using a relevance vector N X K
machine (RVM) to obtain the highest classification accuracy
X exp(θkT x(i) ) 
J(θ) = − 

1{y(i) = k} log K 
 (1)
of 99.80%.
exp(θjT x(i) )
 i=1 P 
k=1
j=1
3. METHODOLOGY
where 1{.} is the “indicator function”, which equals to 1 if
Deep learning has proven to achieve promising results in dif- the statement is true and 0 if the statement is false; θ =
ferent research problems such as information retrieval [24], {θ1 , θ2 , · · · , θK } are the softmax model parameters.
speech recognition [25], and image classification [26]. In this Then, an iterative optimization method such as the stochas-
study, we propose the use of deep recurrent neural networks, tic gradient descent [30] is used to minimize the cost function
particularly the long short-term memory (LSTM) [27], for and maximize the probability of the correct class label.
epileptic seizure diagnosis. Our model follows three main
stages. In the first stage, the time-series EEG signals are 4. RESULTS AND DISCUSSION
divided into non-overlapping segments of a specific length.
Given the sampling rate of 173.6Hz and signal duration of To evaluate the effectiveness of the proposed deep learning-
23.6 seconds, the total number of data-points in each EEG based seizure detection method, we compare its performance
signal is 4096. In our experiments, we tested a wide range of to those of the baseline methods that use the same benchmark
segments’ lengths. We concluded that increasing the segment dataset. The detection performance was assessed using the

2548
Table 1. Seizure detection results of the proposed and state- Table 2. Seizure classification accuracies of the proposed
of-the-art methods. method under noisy conditions.
Methods Year Classifier Sens (%) Spec (%) Acc (%) Noise SNR (dB)
beyli et al. [11] 2008 ME 92.75 94.00 93.17 Source 20 15 10 5 0 −5 −10 −15
beyli et al. [12] 2009 MLPNN 96.00 94.00 94.83 Muscle 100.0 100.0 100.0 100.0 99.50 98.00 97.33 91.00
Song et al. [13] 2010 ELM 97.26 98.77 95.67 Eye-blinking 100.0 100.0 100.0 100.0 100.0 100.0 99.67 98.83
White noise 100.0 100.0 100.0 99.33 97.65 95.33 83.33 80.50
Naghsh et al. [14] 2010 MLPNN 97.46 98.74 97.50
Acharya et al. [15] 2011 Fuzzy 97.70 94.70 94.40
Acharya et al. [16] 2012 FSC 99.40 100.0 98.10
Acharya et al. [17] 2012 GMM 99.00 99.00 99.00 other non-Gaussian artifacts as well as the Gaussian white
Niknazar et al. [18] 2013 ECOC 98.55 99.33 98.67
noise. In this work, we propose a practical seizure detection
Chiang et al. [19] 2014 SVM 91.82 99.40 95.61
Gajic et al. [20] 2015 PQ 98.60 99.33 98.70 system that can address noisy EEG data corrupted with real
Samiee et al. [21] 2015 MLPNN 99.20 98.90 99.10 physical noise (e.g., muscle artifacts, eye-blinking, and white
Behara et al. [22] 2016 LSSVM 96.96 99.66 97.19 electrical noise).
Vidyaratne et al. [23] 2017 RVM 99.00 100.00 99.80
Proposed Method 2017 Softmax 100.0 100.0 100.0
Table 2 investigates the performance of the proposed
method against two common EEG artifacts and white noise
at different SNR levels. It can be observed that the proposed
standard metrics of sensitivity (Sens), specificity (Spec), and method maintains its superior performance when applied to
classification accuracy (Acc). noise-corrupted data of SNRs above 0dB. The main reason
is that the LSTM network can effectively learn the most dis-
criminative and robust EEG features associated with seizures,
4.1. Classification of Clean EEG Data even under the abnormal conditions. The performance of our
The proposed method is first examined for the ideal condi- model starts to decline when applied to noisy EEG data of
tions, i.e., when the EEG signals are completely free of noise. SNRs below 0dB, particularly when the data is polluted with
Our proposed method uses an LSTM network that finds the white noise. Better performance can be achieved for the case
correlation between the EEG signals taken from different sub- of muscle artifacts since the muscle activities interfere with
jects as well as the dependencies between EEG segments of the EEG signals within a limited frequency band of 20-60Hz.
the same subject. Table 1 demonstrates the effectiveness of The superior performance is achieved for the case of eye-
the LSTM network to learn the representative EEG features blinking artifacts that interfere with the EEG signals in the
that best describe the behavior of normal, inter-ictal and ictal low frequency band of 1-3Hz. Table 2 verifies the robustness
EEG activities. of the proposed approach against eye-blinking artifacts, even
We compare the performance of the proposed deep learn- at extremely low SNRs. It can accurately identify the seizure
ing method to those of the state-of-the-art methods that have activities immersed in noise with acceptable classification
been developed in the last 10 years [11]-[22]. The perfor- accuracies.
mance metrics are reported in Table 1. It is worth highlight-
ing that the proposed method outperforms all others in terms 5. CONCLUSION
of the sensitivity, specificity, and classification accuracy. It
yields a seizure sensitivity of 100%, which is superior to In this paper, we introduce a deep learning approach for the
any of the baseline methods. Further, the proposed method automatic detection of epileptic seizures using EEG signals.
produces an outstanding seizure specificity of 100%, which This approach can learn the high-level EEG representations
is similar to the recent results obtained by Vidyaratne et al. and effectively discriminate between the seizure and non-
[23], and is better than those of the reference methods. Also, seizure activities. Another advantage of this approach lies
amongst all existing seizure detection methods, the proposed in its robustness against common EEG artifacts (muscle ac-
scheme achieves the highest classification accuracy of 100%. tivities and eye-blinking) and white noise. The proposed
approach was examined using the Bonn EEG dataset and
4.2. Classification of Noisy EEG Data compared to several baseline methods. The experimental
results demonstrate the effectiveness and superiority of the
We further examine the robustness of the proposed deep proposed method in detecting epileptic seizures. It is also
learning method against two common EEG artifacts as well shown that our method achieves a robust performance with
as white noise. In our previous work, we developed a re- noisy EEG data.
liable EEG feature learning method capable of performing
on noisy signals and achieving reasonable seizure detection 6. REFERENCES
accuracies [31]. This method, however, assumed that the
noise introduced during EEG data acquisition had a Gaus- [1] R. Appleton and A. G. Marson, “Epilepsy,” Oxford University
sian distribution. However, in real life situations, there are Press, third edition, 2009.

2549
 [2] U. R. Acharya, S. V. Sree, G. Swapna, R. J. Martis and J. [17] U. R. Acharya, S. V. Sree, A. P. Alvin and J. S. Suri, “Use of
S. Suri, “Automated EEG analysis of epilepsy: A review,” principal component analysis for automatic classification of
Knowledge-Based Systems, vol. 45, pp. 147-165, 2013. epileptic EEG activities in wavelet framework,” Expert Sys-
[3] F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehnertz, tems with Applications, vol. 39, no. 10, pp. 9072-9078, 2012.
“Seizure prediction: the long and winding road,” Brain, vol. [18] M. Niknazar, S. R. Mousavi, B. V. Vahdat and M. Sayyah, “A
130, no. 2, pp. 314-333, 2017. new framework based on recurrence quantification analysis
[4] G. R. Minasyan, J. B. Chatten, M. J. Chatten and R.N. Harner, for epileptic seizure detection,” IEEE Journal of Biomedical
“Patient-Specific Early Seizure Detection from Scalp EEG,” and Health Informatics, vol. 17, no. 3, pp. 572-578, 2013.
Journal of Clinical Neurophysiology, vol. 27, no. 3, pp. 163- [19] J. Chiang and R. K. Ward, “Energy-efficient data reduction
178, 2010. techniques for wireless seizure detection systems,” Sensors
[5] K. Polat and S. Gne, “Classification of epileptiform EEG us- (Basel, Switzerland), PMC. Web., vol. 14, no. 2, pp. 2036-
ing a hybrid system based on decision tree classifier and fast 2051, 2014.
Fourier transform,” Applied Mathematics and Computation, [20] D. Gajic, Z. Djurovic, J. Gligorijevic, S. D. Gennaro and I.
vol. 187, no. 2, pp. 1017-1026, 2007. Savic-Gajic, “Detection of epileptiform activity in EEG sig-
[6] A. T. Tzallas, M. G. Tsipouras and D. I. Fotiadis. “Automatic nals based on time-frequency and non-linear analysis,” Fron-
Seizure Detection Based on Time-Frequency Analysis and tiers in Computational Neuroscience, pp. 1-6, 2015.
Artificial Neural Networks,” Computational Intelligence and [21] K. Samiee, P. Kovcs and M. Gabbouj, “Epileptic Seizure
Neuroscience, vol. 80. no. 519, 2007. Classification of EEG Time-Series Using Rational Dis-
[7] J. Mitra, et al., “A multistage system for the automated de- crete Short-Time Fourier Transform,” IEEE Transactions on
tection of epileptic seizures in neonatal electroencephalogra- Biomedical Engineering, vol. 62, no. 2, pp. 541-552, 2015.
phy,” Journal of Clinical Neurophysiology, vol. 26, no. 4, pp. [22] D. S. Behara, A. Kumar, P. Swami, B. K. Panigrahi and T.
218-226, 2009. K. Gandhi, “Detection of epileptic seizure patterns in EEG
[8] K. Abualsaud, M. Mahmuddin, M. Saleh and A. Mohamed, through fragmented feature extraction,” International Con-
“Ensemble classifier for epileptic seizure detection for imper- ference on Computing for Sustainable Global Development,
fect EEG data,” The Scientific World Journal, vol. 2015, no. New Delhi, pp. 2539-2542, 2016.
945689, pp. 1-15, 2015. [23] L. S. Vidyaratne, and K. M. Iftekharuddin, “Real-Time
[9] R. G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. Epileptic Seizure Detection Using EEG,” IEEE Transactions
David, and C. E. Elger, “Indications of nonlinear determin- on Neural Systems and Rehabilitation Engineering, vol. 1, no.
istic and finite-dimensional structures in time series of brain 99, pp. 1-1, 2017.
electrical activity: dependence on recording region and brain [24] H. Palangi, L. Deng, Y. Shen, J. Gao, X. He, J. Chen, X.
state,” Physical Review, Statistical, Nonlinear and Soft Matter Song, and R. Ward, “Deep Sentence Embedding Using the
Physics, Dec. 2001. Long Short-Term Memory Networks: Analysis and Appli-
[10] A. Delorme, T. Sejnowski, and S. Makeig, “Enhanced detec- cation to Information Retrieval,” IEEE/ACM Transactions on
tion of artifacts in EEG data using higher-order statistics and Audio, Speech, and Language Processing, vol. 24, no. 4, pp.
independent component analysis,” NeuroImage, vol. 34, no. 694-707, 2016.
4, pp. 1443-1449, 2007. [25] A. Graves, A. R. Mohamed, and G. Hinton, “Speech recog-
[11] E. D. beyli, “Wavelet/mixture of experts network structure nition with deep recurrent neural networks,” IEEE Interna-
for EEG signals classification,” Expert Systems with Appli- tional Conference on Acoustics, Speech and Signal Process-
cations, vol. 34, no. 3, 2008, pp. 1954-1962, 2008. ing (ICASSP), pp. 6645-6649, 2013.
[12] E. D. beyli, “Combined neural network model employing [26] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ”Imagenet
wavelet coefficients for EEG signals classification,” Digital classification with deep convolutional neural networks,” Ad-
Signal Processing, vol. 19, no. 2, pp. 297-308, 2009. vances in neural information processing systems, pp. 1097-
[13] Y. Song and P. Lio, “A new approach for epileptic seizure de- 1105, 2012.
tection: sample entropy based feature extraction and extreme [27] S. Hochreiter and J. Schmidhuber, “Long Short-Term Mem-
learning machine,” Journal of Biomedical Science and Engi- ory,” Neural Computation, vol. 9, no. 8, pp. 1735-1780, Nov.
neering, vol. 3, pp. 556-567, 2010. 1997.
[14] A. R. Naghsh-Nilchi and M. Aghashahi, “Epilepsy seizure [28] R. Hussein, R. Ward, “https://github.com/ramyh/
detection using eigen-system spectral estimation and multiple Epileptic-Seizure-Detection.git”, 2017.
layer perceptron neural network,” Biomedical Signal Process-
[29] J. Friedman, T. Hastie, and R. Tibshirani, “Regularization
ing and Control, vol. 5, no. 2, pp. 147-157, 2010.
paths for generalized linear models via coordinate descent,”
[15] U. R. Acharya, S. V. Sree, S. Chattopadhyay, W. YU and P. Journal of Statistical Software, vol. 33, no. 1, pp. 1-22, 2010.
C. ANG, “Application of Recurrence Quantification Analysis
[30] L. Bottou, “Large-Scale Machine Learning with Stochastic
for the Automated Identification of Epileptic EEG Signals,”
Gradient Descent,” International Conference on Computa-
International Journal of Neural Systems, vol. 21, no. 3, pp.
tional Statistics, pp. 177-186, France, Sept. 2010.
199-211, 2011.
[16] U. R. Acharya, F. Molinari, S. V. Sree, S. Chattopadhyay, K. [31] R. Hussein, Z. J. Wang, and R. Ward, “Ll-regularization based
Ng and J. S. Suri, “Automated diagnosis of epileptic EEG EEG feature learning for detecting epileptic seizure,” IEEE
using entropies,” Biomedical Signal Processing and Control, Global Conference on Signal and Information Processing
vol. 7, no. 4, pp. 401-408, 2012. (GlobalSIP), Washington, DC, pp. 1171-1175, 2016.

2550