Content uploaded by Walid El-Shafai
Author content
All content in this area was uploaded by Walid El-Shafai on May 14, 2022
Content may be subject to copyright.
BASIC RESEARCH
Deep-learning-based seizure detection and prediction from
electroencephalography signals
Fatma E. Ibrahim
1
| Heba M. Emara
1
| Walid El-Shafai
1,2
|
Mohamed Elwekeil
1,3
| Mohamed Rihan
1,3
| Ibrahim M. Eldokany
1
|
Taha E. Taha
1
| Adel S. El-Fishawy
1
| El-Sayed M. El-Rabaie
1
|
Essam Abdellatef
4
| Fathi E. Abd El-Samie
1,5
1
Department of Electronics and Electrical
Communications Engineering, Faculty of
Electronic Engineering, Menoufia
University, Menouf, Egypt
2
Security Engineering Lab, Computer
Science Department, Prince Sultan
University, Riyadh, Saudi Arabia
3
Department of Electrical and
Information Engineering (DIEI),
University of Cassino and Southern Lazio,
Cassino, 03043, Italy
4
Delta Higher Institute for Engineering
and Technology (DHIET), Mansoura,
Egypt
5
Department of Information Technology,
College of Computer and Information
sciences, Princess Nourah Bint
Abdulrahman University, Riyadh, Saudi
Arabia
Correspondence
Fatma E. Ibrahim, Department of
Electronics and Electrical
Communications Engineering, Faculty of
Electronic Engineering, Menoufia
University, Menouf 32952, Egypt.
Email: eng.fatmaelsayed@gmail.com
Abstract
Electroencephalography (EEG) is among the main tools used for analyzing
and diagnosing epilepsy. The manual analysis of EEG must be conducted by
highly trained clinicians or neuro-physiologists; a process that is considered to
have a comparatively low inter-rater agreement. Furthermore, the new data
interpretation consumes an excessive amount of time and resources. Hence, an
automatic seizure detection and prediction system can improve the quality of
patient care in terms of shortening the diagnosis period, reducing manual
errors, and automatically detecting debilitating events. Moreover, for patient
treatment, it is important to alert the patients of epilepsy seizures prior to sei-
zure occurrence. Various distinguished studies presented good solutions for
two-class seizure detection problems with binary classification scenarios. To
deal with these challenges, this paper puts forward effective approaches for
EEG signal classification for normal, pre-ictal, and ictal activities. Three
models are presented for the classification task. Two of them are patient-spe-
cific, while the third one is patient non-specific, which makes it better for the
general classification tasks. The two-class classification is implemented
between normal and pre-ictal activities for seizure prediction and between nor-
mal and ictal activities for seizure detection. A more generalized three-class
classification framework is considered to identify all EEG signal activities. The
first model depends on a Convolutional Neural Network (CNN) with residual
blocks. It contains thirteen layers with four residual learning blocks. It works
on spectrograms of EEG signal segments. The second model depends on a
CNN with three layers. It also works on spectrograms. On the other hand, the
third model depends on Phase Space Reconstruction (PSR) to eliminate the
limitations of the spectrograms used in the first models. A five-layer CNN is
used with this strategy. The advantage of the PSR is the direct projection from
the time domain, which keeps the main trend of different signal activities. The
third model deals with all signal activities, and it was tested for all patients of
the CHB-MIT dataset. It has a superior performance compared to the first
models and the state-of-the-art models.
Received: 5 September 2021 Revised: 19 January 2022 Accepted: 19 January 2022
DOI: 10.1002/cnm.3573
Int J Numer Meth Biomed Engng. 2022;e3573. wileyonlinelibrary.com/journal/cnm © 2022 John Wiley & Sons Ltd. 1of30
https://doi.org/10.1002/cnm.3573
KEYWORDS
Convolutional Neural Network (CNN), electroencephalography, epilepsy, Phase Space
Reconstruction (PSR), seizure prediction, spectrogram
1|INTRODUCTION
About 50 million people around the world are affected by epilepsy disorders of different kinds. Any person, of any age,
gender, race, or class, may be affected by epilepsy. In addition, epilepsy seizures can also vary in frequency of occur-
rence. Such seizures sometimes cause cognitive disorders, which may lead to physical injury of the patients.
1
Epilepsy
is recognized by the World Health Organization (WHO) as a public health concern because of its physical and psycho-
logical consequences. Moreover, epilepsy may lead to premature death, loss of work productivity, and increased
healthcare needs and expenditure.
2
For diagnosing epileptic seizures, distinct screening techniques have been devel-
oped; including Electroencephalography (EEG), positron emission tomography, magneto encephalography, and mag-
netic resonance imaging. EEG signals are characterized by being easily acquired with portable devices.
3
EEG can be defined as an electrophysiological exploration method by which electrical activities of the brain are
measured using electrodes fixed on the scalp.
4
These electrodes may be bulky for patients. Utilization of EEG signals
for diagnosing epilepsy is time- and effort-consuming; as epileptologists have to screen EEG signals minute by minute.
Furthermore, human error is inevitable. Hence, a computer-based diagnosis, by which epileptic seizures can be early
detected, is expected to help the patients.
5–9
Artificial intelligence covers several areas and includes several branches
such as Machine Learning (ML) and Deep Learning (DL).
Conventional ML algorithms, including feature extraction and classification, were formerly used before the appear-
ance of DL. Hand-crafted features limit the performance of the classification algorithms, but deep features are preferred
due to their better representation of signals and images. Such techniques have achieved great progress, when used in
many aspects of medicine, especially in the diagnosis of epileptic seizures. In many fields, such as anomaly detection
from medical signals and images, feature learning, target monitoring, and recognition; DL has achieved great
advances.
10–12
In this paper, we propose an efficient strategy for both seizure detection and prediction from medical EEG signals.
Three models are presented for the classification task. Two of them are patient-specific, while the third one is patient
non-specific. EEG signals for epilepsy patients can be divided into three states: normal (inter-ictal), ictal (seizure), and
pre-ictal which represents the period of 30–60 min before the ictal state.
13
We assumed in this paper that the pre-ictal
state occurs 30 min before the ictal state. The two-class classification is implemented between normal and pre-ictal
activities for seizure prediction and between normal and ictal activities for seizure detection. A more generalized three-
class classification framework is considered to identify all EEG signal activities. For the first two proposed models, the
spectrogram estimation process is performed on EEG signal segments to convert them to 2D images. Deep features are
extracted from the spectrogram images using a Convolutional Neural Network (CNN). The extracted features are used
for the classification process. Besides, we propose two CNN models that differ in the number of deep layers. The first
model consists of thirteen deep layers and four residual learning blocks, while the second model includes three deep
layers, only. The third model depends on Phase Space Reconstruction (PSR) to eliminate the limitations of the spectro-
grams used in the first two models. A five-layer CNN is used with this strategy. The advantage of the PSR is the direct
projection from the time domain, which keeps the main trend of different signal activities. The third model deals with
all signal activities, and it has been tested for all patients of the CHB-MIT dataset. The main objectives of this study are
as follows:
•Proposal of efficient seizure detection and prediction models by generating spectrogram images from EEG signals.
•Proposal of CNNs that differ in the number of deep layers, in order to study the effect of varying the number of layers
on the classification performance.
•Getting the best possible results from the proposed CNNs by studying their performance with different Optimization
Algorithms (OAs) and Learning Rates (LRs).
•Proposal of an efficient patient non-specific approach that depends on 2D PSR images, in order to eliminate the limi-
tations of the spectrograms.
2of30 IBRAHIM ET AL.
•Comparison of the proposed CNNs with various state-of-the-art CNNs, in order to ensure the superiority of the pro-
posed models.
In this paper, Section 2covers the main ideas of the related work. Section 3describes the main architecture of the
proposed patient-specific approach. Section 4presents the main architecture of the proposed patient non-specific
approach. Section 5gives the conclusion of this work.
2|RELATED WORK
M. Talha et al.
14
have developed an end-to-end CNN for seizure detection. Their main target was decreasing the num-
ber EEG signal channels to only two channels and comparing the results with those obtained from all channels. EEG
data of 29 pediatric patients, who were admitted to KK Women and Children's Hospital, has been used. The SeizNet
model was used. It yields a sensitivity of 95.8% and 0.17 false alarms per hour on full channel data, while on two-chan-
nel data, SeizNet yields a sensitivity of 93.3% and a 0.58 false alarms per hour. In,
15
before the Nested Long Short-Term
Memory (NLSTM) model, a CNN with three convolution layers is firstly used. Then, the high-level features obtained
from the NLSTM model are fed into the softmax layer to obtain the predicted labels. The accuracy range of this method
is 98.44% to 100% in ten different experiments on Bonn University database. Moreover, the authors built on a larger
EEG database. The average sensitivity was 97.47%, specificity was 96.17%, and the false detection rate was 0.487 per
hour. For the CHB-MIT scalp EEG dataset, this method achieved a segment-level sensitivity of 94.07% with a false
detection rate of 0.66 per hour.
SeizureNet, as a framework of DL, was introduced by using convolutional layers with dense connections on the
TUH EEG seizure corpus dataset. An overall weighted Fscore of up to 0.90 was achieved by this framework.
16
Covert
et al.
17
proposed the Temporal Graph Convolutional Network (TGCN) for automatic seizure detection. For the seizure
detection task on the epilepsy dataset collected at the Boston Children's Hospital, a 2D CNN model was used to obtain
frequency-domain and time-domain features of EEG signals to characterize seizures. This method, for cross-patient
EEG data, gave an overall sensitivity of 90.00%, a specificity of 91.65%, and an overall accuracy of 98.05%, for the whole
dataset of 23 patients.
12
AlexNet has been applied through the transformation of 1D signals to 2D images.
18
B. Bouaziz et al.
19
managed to segment the EEG signals of CHB-MIT dataset into 2-second frames, and then
transformed them with a spatial representation by producing a set of intensity images. These images are fed to a
CNN, which has a total of eight layers, comprising one initial input layer, five hidden layers, one fully-connected
layer, and an output layer. This approach achieved an accuracy of 99.48%. To reduce the dimensionality, and
then allow the Genetic Algorithm (GA) classification, Rajaguru et al.
20
adopted Multi-Layer Auto-Encoders
(MLAEs) and Expectation Maximization merged with Principal Component Analysis (EM-PCA). The perfor-
mance index represented in classification accuracy was 93.78%. Normal and abnormal brain activities were
detected with a great focus by Roy et al.
21
as they used four different deep learning schemes. They developed the
ChronoNet model, which achieved 90.60% and 86.57% for training and testing accuracies, respectively.
A multi-scale 3D-CNN with a bi-directional gated recurrent unit model was introduced by Choi et al.
22
for cross-
patient seizure detection. Short-Time Fourier Transform (STFT) was used to get spectral and temporal features from
EEG signals. The authors evaluated their proposed model on two datasets: the CHB-MIT dataset and the Seoul National
University Hospital (SNUH) scalp EEG dataset. Their approach achieved sensitivities of 89.4% and 97% on the CHB-
MIT and the SNUH datasets, respectively. Truong et al.
23
used the Freiburg Hospital intracranial EEG (iEEG) dataset,
the CHB-MIT dataset, and the American Epilepsy Society (AES) seizure prediction challenge dataset. To construct spec-
trograms, the STFT algorithm has been applied on raw data. This patient-oriented model gave sensitivities of 81.4%,
81.2%, and 75% and FPR values of 0.06/h, 0.16/h, and 0.21/h, for the prementioned datasets, respectively.
In addition, Z. Wang et al.
24
presented an approach in which data is encoded in the form of different types of
images, namely Gramian Angular Fiields (GAFs) and Markov transition fields. S. Barra et al.
25
presented a framework
in which GAF images are used to train an ensemble of CNNs. They also provided a multi-resolution imaging approach
to feed each CNN. A. Emamia et al.
26
applied CNNs on long-term EEG recordings that include epileptic seizures. After
filtering, CNNs are used to classify EEG images as seizure or non-seizure cases. R. Yuvaraj et al.
27
proposed a CNN-
based unsupervised feature learning framework for automated seizure onset detection. M. Zhou et al.
28
designed a
framework to distinguish ictal, pre-ictal, and inter-ictal segments for epileptic seizure detection from raw EEG signals
instead of using manual feature extraction.
IBRAHIM ET AL.3of30
Subasi et al.
29
introduced a hybrid model for epileptic seizure detection using GAs and Particle Swarm Optimization
(PSO) for EEG data classification. The SVM parameters are optimized using GA- and PSO-based approaches. The PSO-
based approach outperforms the GA-based approach in terms of classification accuracy. This hybrid SVM classifier
achieved a classification accuracy of 99.38%. Alickovic et al.
8
proposed an automated seizure detection and prediction
approach for EEG signals based on de-noising of these signals. Multi-Scale Principal Component Analysis (MSPCA) is
applied. After that, the EEG signal is decomposed using Empirical Mode Decomposition (EMD), Discrete Wavelet
Transform (DWT), or wavelet packet decomposition. Then, to extract relevant features, statistical metrics are applied.
Finally, classification is performed using an ML algorithm. For ictal versus inter-ictal classes, 100% accuracy was
achieved. This approach distinguished between inter-ictal, pre-ictal, and ictal EEG states, with an overall accuracy
of 99.77%.
Qaisar et al.
30
employed intelligent event-driven EEG signal acquisition to achieve real-time compression as well as
effective signal processing and data transmission. The experimental results showed that the developed approach
achieves an overall 3.3-fold compression and transmission bandwidth reduction. It guarantees a significant decrease in
post-analysis and classification processing activities. The system performance has been investigated on a standard
three-class EEG epileptic seizure dataset. For the three classes, the best average classification accuracy of 96.4% has
been achieved on the Andrzejak dataset. For EEG signal classification, a feature extraction network based on Local
Graph Structure (LGS) has been introduced in ref. [31]. Logically extended LGS, symmetric LGS, vertical LGS, vertical
symmetric LGS, zigzag horizontal LGS, zigzag horizontal middle LGS, zigzag vertical LGS, and zigzag vertical middle
LGS were exploited. An ensemble feature extraction network has been formed by combining these LGS methods with
the DWT. LGS was used for feature extraction, and 2D-DWT was used for pooling. Two well-known feature reduction
techniques, relief and Neighborhood Component Analysis (NCA) were used in tandem during the feature reduction
phase. Two publicly available EEG datasets were used in the experiments to test the ensemble LGS feature extraction
method.
Tunable-Q factor Wavelet Transform (TQWT) and bootstrap aggregation using EEG signals have been used to
develop an automated epilepsy detection approach in ref. [32]. The TQWT should be used first to decompose the EEG
signal segments into subbands. The TQWT sub-bands are then analyzed for various spectral properties. Statistical
methods and graphical analysis are used to determine whether spectral properties are suitable for the TQWT domain.
The epileptic seizure classification is then done using bagging. The effectiveness of bagging in the detection strategy has
been investigaed. An accuracy of 98.4% has been achieved for the three-class classification.
Hassan et al.
33,34
introduced a Complete Ensemble Empirical Mode Decomposition with Adaptive Noise
(CEEMDAN) calculation to allow useful feature extraction from physiological signals. The CEEMDAN solves the prob-
lem of mode mixing and improves the mode spectrum separation. The CEEMDAN has been used to develop an auto-
matic epileptic seizure detection method to illustrate the feasibility of the CEEMDAN-based features. Various statistical
features are retrieved from the decomposed EEG signal segments by CEEMDAN, and seizure classification is performed
using an artificial neural network. The authors also presented an automated technique for detecting epileptic seizures
based on CEEMDAN. CEEMDAN has been used to break down segments of EEG signals into Intrinsic Mode Functions
(IMFs). Normal Inverse Gaussian (NIG) probability density function parameters were used to model the IMFs. Intui-
tive, graphical, and statistical assessments illustrate the usefulness of the NIG parameters in the CEEMDAN domain.
To perform classification, adaptive boosting, an eminent ensemble learning-based classification model, has been used.
This method has been evaluated on Bonn University dataset. An average accuracy of 97.6% has been obtained for three-
class classification.
35
Sharma et al.
36
provided features based on 2D PSRs of IMFs of EEG signals for automatic epileptic seizure classifica-
tion. For the creation of PSRs of IMFs, fixed values of time lag and embedding dimension have been suggested. For dif-
ferent window sizes, the 2D PSR images of IMFs have been developed as the input feature set for the LS-SVM classifier
for the identification of ictal and normal classes of EEG signals. The performance of various kernels, such as the radial
basis function, Mexican hat wavelet kernel, and Morlet wavelet kernel, was evaluated in a combination with the LS-
SVM classifier for the classification of ictal and normal EEG signals. A maximum accuracy of 98.67% has been obtained
using the Morlet wavelet kernel. Shankar et al.
37
presented the PSR technique to generate 2D input images from EEG
signals for specific brain rhythms. A CNN model with five convolutional layers and three fully-connected layers has
been used. Both Bonn University and CHB-MIT datasets were considered for experimental validation. Accuracies of
93% and 85% have been obtained on Bonn University and CHB-MIT datasets, respectively.
The proposed work in this paper is mainly concerned with EEG signal classification with simple and highly efficient
models based on signal transformations such as spectrogram and PSR transformations.
4of30 IBRAHIM ET AL.
3|THE PROPOSED PATIENT-SPECIFIC APPROACH
Figure 1depicts the main architecture of the proposed approach for seizure detection and prediction. Firstly, the pro-
cess of spectrogram estimation is applied on the EEG signal segments to transform them into image-like format. It is
well known that spectrogram gives a good time-frequency representation of the signals. The spectrogram estimation
produces images from which deep features can be extracted using a CNN. The CNN is responsible for feature extrac-
tion, and hence classification. Different layers including convolutional, pooling, and depth concatenation layers are
used for the process of feature extraction. Finally, the extracted deep features are used for classification to obtain the
detection and prediction results.
3.1 |EEG signals
The proposed models in this approach are evaluated on the publicly available CHB-MIT scalp EEG epilepsy dataset,
which was collected at Children's Hospital Boston (CHB) in collaboration with the Massachusetts Institute of Technol-
ogy (MIT).
38
The data included in this dataset was recorded for 23 epilepsy patients (7 males, and 17 females). Patients
were prevented from taking the anti-seizure medications, and then EEG signals were recorded for a couple of days after
prevention. This dataset was used for analyzing seizures of patients to determine their need for surgery. To record this
dataset, the 10-20 international systems of an EEG electrode montage scheme were used. The sampling frequency of
the signals for all patients is 256 Hz with a 16-bit resolution.
39
In this work, the proposed patient-specific approach is
applied on eight patients that were chosen randomly. The selected patients are chb01, 02, 03, 04, 06, 08, 11, and 14. The
objective is to introduce a proof of concept. Table 1illustrates the gender, age, number of hours, and number of seizures
for each patient.
FIGURE 1 Block diagram for the patient-specific approach
IBRAHIM ET AL.5of30
3.2 |Spectrogram estimation
A spectrogram is a time-frequency representation of time-domain EEG signals. The signal is windowed, and the spec-
trum is estimated for each window. The windowed spectra are used to build the spectrogram as an image. The color in
this image represents the power distributed over the time-frequency axes of the signal.
40
Spectrogram images can be fed
to the CNN as input. Figure 2illustrates the spectrograms of different EEG signal activities. There is a difference in
color between normal, pre-ictal, and ictal activities.
3.3 |CNN-based feature extraction
The CNN depends on relevant filters to extract the temporal and spatial dependencies in an image. The CNNs contain
several layers such as convolutional, pooling, residual learning, and softmax layers. The main essential roles of different
CNN layers are described briefly as follows.
3.3.1 | Convolutional layer
This layer has several learnable filters to compute dot products between the corresponding entries of both the filter and
the input image. The output feature map fC,l
x,y,kfor a particular layer land the input fOp,l1
x,ycan be computed as
39,41
:
fC,l
x,y,k¼wlT
kfOp,l1
x,yþbl
kð1Þ
TABLE 1 Dataset description
Patient No. No. of seizures No. of hours Gender Age
1 7 40.55 F 11
2 3 35.16 M 11
3 7 36 F 14
4 4 150.7 M 22
55 39F7
6 10 68.24 F 1.5
7 3 67.05 F 15.5
8 5 20 M 3.5
9 4 65.02 F 10
10 7 50.02 M 3
11 3 34.62 F 12
12 40 23.67 F 2
13 12 32 F 3
14 8 25.85 F 9
15 40 39.42 M 16
16 10 19 F 7
17 3 22 F 12
18 6 35.63 F 18
19 3 29.93 F 19
20 8 27.59 F 6
21 4 31.81 F 13
22 3 32 F 9
23 7 25.73 F 6
24 16 12 M 12.5
6of30 IBRAHIM ET AL.
where wl
krepresents the shared weights, bl
krepresents the bias and Cdenotes convolution. O
p
represents the input
image, for l=1, while it represents convolution, pooling, or activation, for l>1.
3.3.2 | Maximum pooling layer
Maximum pooling is performed by computing the maximum value in a local spatial neighborhood, and then reducing
the spatial resolution
39,41
:
fP,l
x,y,k¼max
m,nðÞNx,y
fOp,l1
m,n,kð2Þ
where the pooling operation is denoted by Pand the local spatial neighborhood of (x,y) coordinates is denoted by N
x,y
.
3.3.3 | Residual learning
This strategy is used to optimize the loss of CNNs, efficiently. The output of a residual learning block Rcan be expressed
as
39,41
:
fR,l
x,y,k¼fOp,lq
x,yþFf
Op,lq
x,y,wk
ð3Þ
where fOp,lq
x,yis the input feature map, F(.) is the residual mapping to be learned, and qis the total number of stacked
layers.
FIGURE 2 Various spectrogram images: (A) Normal case, (B) Pre-ictal case, and (C) Ictal case
IBRAHIM ET AL.7of30
3.3.4 | Fully-connected layer
This layer is mainly used for classification purposes. Let layer (l-1) be a fully-connected layer. Layer lexpects ml1ðÞ
1fea-
ture maps having a size of ml1ðÞ
2ml1ðÞ
3as input. The following equation can be used to compute the i
th
unit in layer
l
39,41
:
YlðÞ
i¼fZ
lðÞ
i
,ZlðÞ
i¼X
ml1ðÞ
1
j¼1X
ml1ðÞ
2
r¼1X
ml1ðÞ
3
s¼1
Wl
i,j,r,sYl1
j
r,sð4Þ
where Wl
i,j,r,srepresents the weight used to connect the unit at position (r,s) in the i
th
unit in layer land the j
th
feature
map of layer (l1).
3.3.5 | Softmax layer
This layer is used for computing the loss. The form of softmax loss is given as
39,41
:
Lsoftmax ¼X
N
i¼1
log ewT
yifiþbyi
PK
j¼1ewT
jfiþbjð5Þ
where f
i
denotes features and y
i
is the true class label of the i
th
image. w
j
and b
j
are the weights and bias of the j
th
class,
respectively. Nis the number of training samples and Kis the number of classes.
In this work, we propose two CNNs that are different in the number of deep layers. The first model, namely pro-
posed CNN_1, adopts residual learning and depth concatenation strategies. It includes 13 deep layers and 4 residual
learning blocks. Figure 3presents the main architecture of the proposed CNN_1, which consists of two main residual
learning blocks; ResBL_1 and ResBL_3, with ResBL_2 and ResBL_4 included in ResBL_1 and ResBL_3, respectively.
Moreover, Figure 4shows the structure of ResBL_1, ResBL_3, ResBL_2, and ResBL_4. In the prementioned figures,
conv_1 refers to the first convolutional layer; conv_2 refers to the second convolutional layer, and so on. Each con-
volutional layer is followed by a Rectified Linear Unit (ReLU) activation function. Residual learning blocks allow the
flow of data from initial layers to last layers. Furthermore, these blocks are used to optimize the loss of CNNs through
FIGURE 3 The main architecture of the proposed CNN_1
8of30 IBRAHIM ET AL.
the concatenation of the extracted features from various convolutional layers. Depth concatenation is an essential com-
ponent in the residual learning block. Depth concatenation is used to increase the depth of the feature map. Addition-
ally, Table 2presents the filter specifications of the proposed CNN_1. On the other hand, the second model, namely
proposed CNN_2, is a very simple model which consists of 3 layers. Figure 5shows the main architecture of the pro-
posed CNN_2 that includes 3 convolutional blocks: convolutional block_1, convolutional block_2, and convolutional
block_3. Each convolutional block consists of a convolutional layer, a batch normalization layer, and a ReLU function.
Table 3introduces the filter specifications of the proposed CNN_2 model.
3.4 |Experimental results
In this paper, three different classes were examined for classifying the segments as ictal, normal, or pre-ictal states. The
performance of the proposed spectrogram-based DL approach was tested using standard metrics such as classification
Recall (REC), Specificity (SPE), Accuracy (ACC), Precision (PRE), and F
score
for all binary classes, as shown below.
42
Recall is given by:
REC ¼Tp
TpþFn
100 ð6Þ
Specificity is given by:
SPE ¼Tn
TnþFp
100 ð7Þ
Accuracy is given by:
(A) (B)
(C) (D)
FIGURE 4 Main architecture of the proposed (A) ResBL_1 (B) ResBL_3, (C) ResBL_2 and (D) ResBL_4
IBRAHIM ET AL.9of30
ACC ¼TpþTn
TpþTnþFpþFn
100 ð8Þ
Precision is given as:
PRE ¼Tp
TpþFp
100 ð9Þ
Fscore is given by:
Fscore ¼Tp
Tpþ1
2FpþFn
100 ð10Þ
TABLE 2 Filters Specifications of the Proposed CNN_1
Layer name No. of filters Filter sze
Input layer of size [227,227 3]
Conv_1 192 3 3
ReLU
Conv_2 64 1 1
ReLU
Conv_3 128 3 3
ReLU
Conv_4 64 3 3
ReLU
Conv_5 128 3 3
ReLU
Concatenation of the output of convolutional layers “3 and 5”
Conv_6 192 3 3
ReLU
Concatenation of the output of convolutional layers “1 and 6”
Conv_7 192 3 3
ReLU
Conv_8 64 1 1
ReLU
Conv_9 128 3 3
ReLU
Conv_10 64 3 3
ReLU
Conv_11 128 3 3
ReLU
Concatenation of the output of convolutional layers “9 and 11”
Conv_12 192 3 3
ReLU
Concatenation of the output of convolutional layers “7 and 12”
Conv_13 192 3 3
ReLU
Max. Pooling 1 7 7
Fully connected layer
Softmax layer
Output layer
10 of 30 IBRAHIM ET AL.
FIGURE 5 The main architecture of the proposed CNN_2
TABLE 3 Filters Specifications of the Proposed CNN_2
No. Name Type Activations
1 Image input
224 224 3 images
Image Input 224 224 3
2 Convolutional_1
16 3 33 convolutions with stride [1 1] and padding [1 1 1 1]
Convolution 224 224 16
3 Batchnorm_1
Batch normalization with 16 channels
Batch 224 224 16
4 Relu_1 ReLU 224 224 16
5 Maxpool_1
22 max pooling with stride [2 2] and padding [0 0 0 0]
Max pooling 112 112 16
6 Convolutional_2
32 3 316 convolutions with a stride of [1 1] and a padding of [1 1 1 1]
Convolution 112 112 32
7 Batchnorm_1
Batch normalization with 32 channels
Batch 112 112 32
8 Relu_2 ReLU 112 112 32
9 Maxpool_2
22 max pooling with a stride of [2 2] and a padding of [0 0 0 0]
Max pooling 56 56 32
10 Convolutional_3
64 3 332 convolutions with a stride of [1 1] and a padding of [1 1 1 1]
Convolution 56 56 64
11 Batchnorm_1
Batch normalization with 64 channels
Batch 56 56 64
12 Relu_3 ReLU 56 56 64
13 Fc Fully connected —
14 softmax Softmax —
15 classoutput Output —
IBRAHIM ET AL.11 of 30
T
p
denotes the number of images that are truly epileptic and forecasted as epileptic, whereas F
p
denotes the
number of images that are nonepileptic and forecasted as being epileptic. T
n
stands for really nonepileptic clas-
ses predicted to be nonepileptic, whereas F
n
stands for actually epileptic classes predicted to be nonepileptic.
The proposed strategy mainly depends on a CNN to extract deep features from spectrogram images. In this sec-
tion, the experimental results of the proposed approach using various CNNs are provided. The proposed CNNs are
compared with state-of-the-art CNNs such as; VGG-19,
43
ResNet-101,
44
Inception-v3,
45
GoogLeNet,
46
DenseNet-
201,
47
and DarkNet-53.
48
In addition, the performance of the proposed approach is compared with that of another
EEG seizure detection and prediction method for Zhou.
28
He designed a CNN with no more than three layers. The
comparison is held in terms of accuracy, specificity, precision, sensitivity, and F score. The CHB-MIT dataset sig-
nals are divided into three classes: normal, pre-ictal, and ictal. Each class contains 1000 spectrogram images,
which are divided into 700 images for training and 300 images for testing. Experiments are carried out on 8
patients. Moreover, we perform two classification scenarios for each patient: normal versus pre-ictal and normal
versus ictal.
Our target is to reach the optimal performance of the proposed CNNs. To achieve that objective, we have to select
the OAs to be used and adjust the values of various hyperparameters such as weight decay, momentum value, mini-
batch size, maximum number of epochs, and Learning Rates (LRs). We apply L
2
regularization with the weight decay
set to 5 10
4
. In addition, the momentum value is set to 0.9, the mini-batch size is set to 32, and the maximum num-
ber of epochs is adjusted to 5. Tables 4and 5show the performance of the proposed CNN_1 and the proposed CNN_2,
respectively. In Tables 4and 5, three different OAs are tested: Adaptive Moment (Adam), Root Mean Square propaga-
tion (RMS prop), and Stochastic Gradient Descent with Momentum (SGDM). Furthermore, for each algorithm, three
LRs are tested: 0.1, 0.01, and 0.001. We consider the data of patient 1 and the classification scenario of normal versus
pre-ictal.
From Tables 4and 5, it is clear that the optimal performance of the proposed CNN_1 could be achieved by using the
SGDM algorithm for optimization and adjusting the LR to 0.001. Using Rmsprop algorithm for optimization and adjusting
the LR to 0.001 give promising results for CNN_2. It is also clear that the proposed CNN_1 is superior to the proposed
CNN_2, and this is expected due to the utilization of a larger number of deep layers and the application of residual learning
and depth concatenation strategies. On the other hand, the proposed CNN_2 is simpler and less time-consuming than the
proposed CNN_1. For more clarification, Figure 6depicts a graphical comparison between the proposed CNNs.
As mentioned before, experiments have been carried out on 8 patient signals, and we performed two classification
scenarios for each patient: normal versus pre-ictal and normal versus ictal. Tables 6–13 provide the classification results
of various state-of-the-art CNNs for patients 1, 2, 3, 4, 6, 8, 11, and 14, respectively.
From Tables 6–13, we can notice that the proposed CNN_1 is superior to the other state-of-the-art CNNs. The pro-
posed model achieves the highest results for all patients except patient 4. The rationale behind these findings is the
hyper-parameter optimization process involved in the proposed model. Finally, to further emphasize the superiority of
the proposed approach, a comparison between the proposed CNN_1 and Zhou's CNN
28
is presented in Table 14. The
comparison between CNNs is performed in terms of Accuracy, Specificity, and Recall. This comparison confirms the
superiority of the proposed model.
From all the previous results, we can say that the proposed approach succeeds in giving higher classification rates than
those of the state-of-the-art methods. These results also ensure that the spectrogram representation of EEG segments reflects
TABLE 4 Performance of the proposed CNN_1 for different optimization algorithms and learning rates
OA LR Accuracy Specificity Recall Precision F
score
SGDM 0.1 0.9573 0.9586 0.9432 0.9469 0.945
0.01 0.9697 0.9713 0.9529 0.9618 0.9573
0.001 0.9788 0.9798 0.9626 0.9721 0.9673
RMSprop 0.1 0.9548 0.956 0.9457 0.9524 0.949
0.01 0.9558 0.9567 0.9437 0.9517 0.9476
0.001 0.9536 0.9542 0.9429 0.9476 0.9452
Adam 0.1 0.9655 0.9667 0.9445 0.9624 0.9533
0.01 0.9682 0.9695 0.9515 0.9637 0.9575
0.001 0.9673 0.9685 0.9523 0.9632 0.9577
12 of 30 IBRAHIM ET AL.
all activities in these segment in a time, frequency and power representation that can be used as a visual pattern for signal clas-
sification. Furthermore, these high percentage results ensure the ability to develop EEG signal classification approaches based
on spectrogram estimation and CNNs.
4|THE PROPOSED PATIENT NON-SPECIFIC APPROACH
Figure 7shows an overview of the proposed patient non-specific approach, in which raw EEG signals are normalized
using a Z-score. The PSR technique is then used to produce 2D images for non-overlapping segments. The PSR images
are assessed using entropy and global statistics. Then, the images are fed into the CNN_3 model to classify epileptic sei-
zures by extracting features, automatically. Signal preprocessing, input image synthesis, CNN classification, and input
image quality evaluation are the primary phases, which are detailed as follows:
4.1 |Signal preprocessing
First, the EEG signals are preprocessed with a low-pass Finite Impulse Response (FIR) filter with 0.3 and 40 Hz cut-off frequen-
cies to remove noise and artifacts. The Z-score is then used to remove the DC components and normalize the EEG data. Feature
scaling is essential in ML, since it ensures that all feature sets are assessed on the same scale. Statistics have brought to light the
notion of feature scaling. Statistical analysis is a method for putting different variablesonthesamescale.Itisfrequentlyused,
when the dataset has various scales. The features of a dataset may have extensive and substantial variances across their ranges
at different times. So, in this scenario, standardization is imposed on the dataset to get all on the same scale. Z-score can be cal-
culated as
49
:
TABLE 5 Performance of the proposed CNN_2 for different optimization algorithms and learning rates
OA LR Accuracy Specificity Recall Precision F
score
SGDM 0.1 0.88 0.91 0.85 0.9043 0.8763
0.01 0.85 0.95 0.75 0.9375 0.8333
0.001 0.8833 0.90 0.8667 0.8966 0.8814
RMSprop 0.1 0.7583 0.9333 0.5833 0.8974 0.7071
0.01 0.7383 0.8333 0.50 0.8498 0.6565
0.001 0.8883 0.9767 0.9433 0.9554 0.8942
Adam 0.1 0.865 0.91 0.82 0.9011 0.8586
0.01 0.8717 0.9067 0.8367 0.8996 0.867
0.001 0.8683 0.96 0.7767 0.951 0.855
FIGURE 6 Graphical comparison between the proposed CNN models
IBRAHIM ET AL.13 of 30
Z¼Xμ
σ
ð11Þ
where X,μ, and σrepresent the input data, the mean, and the standard deviation, respectively. Figure 8shows EEG sig-
nals and their Z-score for normal, ictal, and pre-ictal activities.
TABLE 6 Experimental results for patient 1
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.9035 0.9044 0.8848 0.8961 0.8904
ResNet-101 0.9383 0.9399 0.9196 0.9286 0.924
Inception-v3 0.9407 0.9421 0.9216 0.9322 0.9268
GoogLeNet 0.9183 0.9198 0.8911 0.9016 0.8963
DenseNet-201 0.9507 0.9518 0.9281 0.9077 0.9177
DarkNet-53 0.9234 0.9246 0.9009 0.8852 0.8929
Proposed CNN_1 0.9788 0.9798 0.9626 0.9721 0.9673
Proposed CNN_2 0.8883 0.9767 0.9433 0.9554 0.8942
Normal
versus
Ictal
VGG-19 0.8152 0.8165 0.796 0.806 0.8009
ResNet-101 0.8673 0.8683 0.8476 0.8586 0.853
Inception-v3 0.8847 0.8855 0.8657 0.8764 0.871
GoogLeNet 0.8209 0.8222 0.79 0.8067 0.7983
DenseNet-201 0.8632 0.8643 0.8367 0.84 0.8383
DarkNet-53 0.8266 0.8271 0.7992 0.82 0.8095
Proposed CNN_1 0.9049 0.9061 0.8842 0.8918 0.8879
Proposed CNN_2 0.8415 0.8586 0.8321 0.8542 0.843
TABLE 7 Experimental results for patient 2
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.964 0.9655 0.9482 0.9603 0.9542
ResNet-101 0.9751 0.9768 0.959 0.9698 0.9643
Inception-v3 0.9777 0.9785 0.9612 0.9733 0.9672
GoogLeNet 0.9507 0.9514 0.9243 0.9478 0.9359
DenseNet-201 0.9728 0.9744 0.9524 0.9636 0.9579
DarkNet-53 0.9465 0.9482 0.9061 0.9273 0.9165
Proposed CNN_1 0.9954 0.9967 0.9758 0.9877 0.9817
Proposed CNN_2 0.9657 0.9457 0.9867 0.9433 0.965
Normal
versus
Ictal
VGG-19 0.9157 0.9171 0.8995 0.9093 0.9043
ResNet-101 0.9506 0.9521 0.9346 0.9464 0.9404
Inception-v3 0.9509 0.9517 0.9351 0.9473 0.9411
GoogLeNet 0.9393 0.9417 0.9118 0.9272 0.9194
DenseNet-201 0.9138 0.9148 0.8847 0.903 0.8938
DarkNet-53 0.952 0.9534 0.9253 0.9395 0.9323
Proposed CNN_1 0.9756 0.977 0.9535 0.9626 0.958
Proposed CNN_2 0.9085 0.907 0.91 0.9067 0.9083
14 of 30 IBRAHIM ET AL.
4.2 |Phase Space Reconstruction (PSR)
The PSR of a signal is a visual representation of the signal dynamic behavior as it changes over time. Phase space may
be used to rebuild a univariate time series in a chaotic system. In a dynamic system, the univariate time series contains
all information about variables. Each point in the phase space represents a state of the system, while the phase space
TABLE 8 Experimental results for patient 3
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.6757 0.6769 0.6507 0.6629 0.6567
ResNet-101 0.7008 0.7019 0.6759 0.6891 0.6824
Inception-v3 0.7253 0.7265 0.7011 0.7131 0.707
GoogLeNet 0.7122 0.7142 0.7031 0.7114 0.7072
DenseNet-201 0.7341 0.7354 0.7293 0.739 0.7341
DarkNet-53 0.7162 0.7177 0.7007 0.7047 0.7027
Proposed CNN_1 0.7794 0.7805 0.7635 0.7774 0.7703
Proposed CNN_2 0.6456 0.6445 0.6467 0.6433 0.645
Normal
versus
Ictal
VGG-19 0.7624 0.7641 0.7382 0.7507 0.7443
ResNet-101 0.7893 0.7908 0.7647 0.7769 0.7707
Inception-v3 0.8246 0.8264 0.8002 0.8124 0.8062
GoogLeNet 0.8091 0.8153 0.8111 0.8072 0.8091
DenseNet-201 0.8128 0.8144 0.8026 0.8034 0.8028
DarkNet-53 0.8003 0.8062 0.8081 0.8028 0.8054
Proposed CNN_1 0.8416 0.8427 0.8248 0.835 0.8298
Proposed CNN_2 0.7981 0.7575 0.8433 0.73 0.7867
TABLE 9 Experimental results for patient 4
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.6949 0.6962 0.6685 0.6824 0.6753
ResNet-101 0.7331 0.7344 0.7054 0.7181 0.7116
Inception-v3 0.7564 0.7576 0.7289 0.7436 0.7361
GoogLeNet 0.7683 0.943 0.9675 0.9693 0.9683
DenseNet-201 0.781 0.7892 0.7503 0.7719 0.7609
DarkNet-53 0.7372 0.7425 0.7329 0.7417 0.7373
Proposed CNN_1 0.7931 0.7942 0.7791 0.783 0.781
Proposed CNN_2 0.7143 0.7119 0.7167 0.71 0.7133
Normal
versus
Ictal
VGG-19 0.7116 0.7128 0.6847 0.6978 0.6911
ResNet-101 0.7451 0.7465 0.7174 0.7301 0.7236
Inception-v3 0.7637 0.7649 0.7366 0.7493 0.7428
GoogLeNet 0.7223 0.7365 0.7235 0.7212 0.7223
DenseNet-201 0.804 0.8236 0.8067 0.8014 0.804
DarkNet-53 0.7761 0.7857 0.7412 0.7311 0.7361
Proposed CNN_1 0.7988 0.7997 0.774 0.7812 0.7775
Proposed CNN_2 0.7319 0.7406 0.7233 0.7467 0.735
IBRAHIM ET AL.15 of 30
trajectory reflects the system temporal progress under various initial conditions.
50
A phase space can be generated from
a one-dimensional time series using the taken time-delay embedding theorem.
50
This theorem may be used to analyze
chaotic time series. If a scalar time series T
t
=N
1
,N
2
,N
3
,…….N
n
from a chaotic system is given, then reconstruction is
possible in terms of the phase space vectors X(t) expressed as X(t)=[x(t),x(t+τ),……x(t+[m1]τ)], t=1,2,….M;
M=N(m1)τ. Here, τis the time delay, mis the embedding dimension of the PSR, and Mis the number of phase
TABLE 10 Experimental results for patient 6
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.5414 0.5429 0.5205 0.5307 0.5255
ResNet-101 0.5574 0.5584 0.5374 0.5478 0.5425
Inception-v3 0.5769 0.5776 0.5583 0.5681 0.5631
GoogLeNet 0.6285 0.6404 0.6229 0.6343 0.6285
DenseNet-201 0.6133 0.6252 0.6072 0.6195 0.6133
DarkNet-53 0.6204 0.6323 0.6148 0.6262 0.6204
Proposed CNN_1 0.6377 0.6389 0.6125 0.6224 0.6174
Proposed CNN_2 0.5364 0.5437 0.4567 0.6167 0.4967
Normal
versus
Ictal
VGG-19 0.648 0.6499 0.6287 0.6385 0.6335
ResNet-101 0.6896 0.6908 0.6688 0.6788 0.6737
Inception-v3 0.7114 0.7125 0.6909 0.6997 0.6952
GoogLeNet 0.7395 0.7436 0.7338 0.7454 0.7395
DenseNet-201 0.7283 0.7349 0.7221 0.7346 0.7283
DarkNet-53 0.733 0.7455 0.7279 0.7383 0.733
Proposed CNN_1 0.7636 0.7649 0.7467 0.7553 0.7509
Proposed CNN_2 0.7459 0.6795 0.8267 0.61 0.7183
TABLE 11 Experimental results for patient 8
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.7691 0.7704 0.7505 0.7632 0.7567
ResNet-101 0.8108 0.8117 0.792 0.8025 0.7972
Inception-v3 0.8446 0.8459 0.8254 0.8369 0.8311
GoogLeNet 0.8416 0.8542 0.8356 0.8478 0.8416
DenseNet-201 0.8469 0.8554 0.8414 0.8525 0.8469
DarkNet-53 0.8491 0.8565 0.8437 0.8547 0.8491
Proposed CNN_1 0.9033 0.9046 0.8776 0.8867 0.8821
Proposed CNN_2 0.7483 0.769 0.71 0.7867 0.7383
Normal
versus
Ictal
VGG-19 0.7107 0.7123 0.6914 0.7027 0.697
ResNet-101 0.7526 0.7541 0.7342 0.7458 0.7399
Inception-v3 0.7937 0.7952 0.7748 0.7859 0.7803
GoogLeNet 0.8428 0.8523 0.8374 0.8484 0.8428
DenseNet-201 0.8396 0.8448 0.8338 0.8456 0.8396
DarkNet-53 0.8446 0.8557 0.8386 0.8508 0.8446
Proposed CNN_1 0.8773 0.8782 0.8515 0.8628 0.8571
Proposed CNN_2 0.8126 0.7701 0.86 0.7433 0.8017
16 of 30 IBRAHIM ET AL.
points of the reconstructed phase space. Computation of τand mvalues is very essential in PSR. To estimate the value
of τfor a nonlinear time series, the Average Mutual Information (AMI) method can be used.
51
The equation used to cal-
culate the AMI is given as follows:
IτðÞ¼
X
Nτ
t¼1
PX
t,Xtþτ
ðÞ:log PX
t,Xtþτ
ðÞ
PX
t
ðÞ:PX
tþτ
ðÞ
ð12Þ
TABLE 13 Experimental results for patient 14
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.6542 0.6559 0.6399 0.6508 0.6453
ResNet-101 0.704 0.7057 0.6913 0.6978 0.6945
Inception-v3 0.7417 0.7432 0.7288 0.7322 0.7304
GoogLeNet 0.6427 0.6538 0.6367 0.6488 0.6427
DenseNet-201 0.6389 0.6497 0.6329 0.645 0.6389
DarkNet-53 0.6445 0.6545 0.6385 0.6507 0.6445
Proposed CNN_1 0.7683 0.7698 0.7362 0.7472 0.7416
Proposed CNN_2 0.6938 0.7057 0.6633 0.7233 0.6833
Normal
versus
Ictal
VGG-19 0.6438 0.6451 0.6291 0.6373 0.6331
ResNet-101 0.6804 0.6815 0.6657 0.6734 0.6695
Inception-v3 0.7163 0.7176 0.7022 0.7101 0.7061
GoogLeNet 0.6363 0.6351 0.6339 0.6389 0.6363
DenseNet-201 0.6282 0.6343 0.6254 0.6312 0.6282
DarkNet-53 0.6431 0.6374 0.6452 0.6412 0.6431
Proposed CNN_1 0.7462 0.7471 0.7134 0.7282 0.7207
Proposed CNN_2 0.7321 0.7081 0.8733 0.64 0.7167
TABLE 12 Experimental results for patient 11
Scenario CNN Accuracy Specificity Precision Recall F
score
Normal
versus
Pre-ictal
VGG-19 0.5136 0.5148 0.4903 0.5007 0.4954
ResNet-101 0.5452 0.5463 0.519 0.5285 0.5237
Inception-v3 0.5723 0.5739 0.5495 0.5594 0.5544
GoogLeNet 0.538 0.5494 0.5324 0.5438 0.538
DenseNet-201 0.5279 0.5381 0.5218 0.5341 0.5279
DarkNet-53 0.5325 0.5437 0.5267 0.5384 0.5325
Proposed CNN_1 0.6013 0.6024 0.5726 0.5845 0.5784
Proposed CNN_2 0.56 0.572 0.4767 0.6433 0.52
Normal
versus
Ictal
VGG-19 0.6469 0.6481 0.6235 0.6354 0.6293
ResNet-101 0.6755 0.677 0.6515 0.6614 0.6564
Inception-v3 0.7067 0.7084 0.6847 0.6946 0.6896
GoogLeNet 0.6418 0.6525 0.6361 0.6477 0.6418
DenseNet-201 0.6489 0.659 0.6429 0.6551 0.6489
DarkNet-53 0.6499 0.6599 0.6438 0.6562 0.6499
Proposed CNN_1 0.7332 0.7344 0.7111 0.7238 0.7173
Proposed CNN_2 0.7159 0.6319 0.91 0.47 0.69
IBRAHIM ET AL.17 of 30
where P(x
t
) is the probability density of X
t
.P(X
t
,X
t+τ
) is the joint probability density of X
t
and X
t+τ
.I(τ) is a measure
of the statistical dependence of the reconstruction variables. For a nonmonotonous decrease of I(τ), the location of the
first local minimum is considered as the suitable value of τ.
52
Kennel et al.
52
introduced a False Nearest Neighbor
(FNN) method to compute the optimum m. The FNN method calculates the distance D(m) to the nearest neighbor and
TABLE 14 Comparison between the proposed CNN_1 and Zhou's CNN
28
Patient ID
Accuracy Specificity Recall
Proposed CNN_1 Zhou's CNN
28
Proposed CNN_1 Zhou's CNN
28
Proposed CNN_1 Zhou's CNN
28
1 0.9788 0.905 0.9798 0.9 0.9721 0.91
2 0.9954 0.623 0.9967 0.607 0.9877 0.64
3 0.7794 0.623 0.7805 0.677 0.7774 0.57
4 0. 7931 0.515 0. 7942 0.573 0. 783 0.457
6 0.6377 0.53 0.6389 0.547 0.6224 0.513
8 0.9033 0.508 0.9046 0.547 0.8867 0.47
11 0.6013 0.548 0.6024 0.607 0.5845 0.49
14 0.7683 0.503 0.7698 0.513 0.7472 0.493
FIGURE 7 Block diagram of the proposed patient non-specific approach
18 of 30 IBRAHIM ET AL.
the distance D(m+1) between two points in m+1 dimensions for each point in the m-dimensional image. The neigh-
bors are said to be false, if the following condition is met
52
:
Dmþ1ðÞ
RA
>Rtol,Atol ð13Þ
whereR
tol
and A
tol
are constant thresholds, while R
A
is the time-series standard deviation. The procedure is continued
for the dimensions, and it is terminated when the proportion of FNN becomes zero or negligible.
4.3 |Evaluation of input quality
In this paper, the PSR images for all EEG segments are created. EEG signals are regarded as time series data with
nonlinear and nonstationary properties. The time series incorporates recurrent properties such as nonlinearity and cha-
otic behavior of the time series. The PSR is a good candidate for creating 2D input images. It also has the capability of
graphically analyzing hidden patterns, pattern correlations, predictive characteristics, stability, fundamental similarity,
and structural changes in time series data across time. The entropy and global statistical values of produced PSR images
are used to assess quality. Entropy (E
t
) is a statistical measure of randomness that can be exploited to depict the texture
of an image and to compare image information. E
t
is calculated for images as:
FIGURE 8 EEG signals and their Z-scores for (A) Normal EEG signals, (B) Ictal EEG signals, and (C) Pre-ictal EEG signals
IBRAHIM ET AL.19 of 30
FIGURE 9 The main architecture of the proposed CNN_3
FIGURE 10 Samples of (A) ictal, (B) pre-ictal and (C) normal PSR images for τ=1 and m=2
20 of 30 IBRAHIM ET AL.
Et¼X
N
i¼1
hilog hi
ðÞ ð14Þ
where h
i
is the count of the normalized histogram and Nis the number of levels in the image. Furthermore, image
global statistics such as mean, variance, and standard deviation are calculated and used.
4.4 |CNN-based feature extraction
A simple CNN model, namely proposed CNN_3 has been used to classify EEG signal segments. The 2D PSR images
serve as input to the CNN_3 model. Figure 9shows the proposed CNN_3 architecture for ictal, normal, and pre-ictal
activity classification based on PSR images. The CNN_3 model has five convolution blocks (each with a convolutional
TABLE 15 Experimental results of the two-class classification with the CNN_3 model from spectrogram images
Scenario CNN Accuracy Recall Specificity Precision Fscore Fpr Mcc Kappa
Normal
versus
Pre-ictal
VGG-19 0.8335 0.8761 0.8244 0.844 0.8604 0.12 0.844 0.8744
ResNet-101 0.893 0.8886 0.8765 0.8675 0.8863 0.057 0.89 0.8678
Inception-v3 0.8409 0.8342 0.8451 0.8276 0.8298 0.089 0.825 0.862
GoogLeNet 0.813 0.805 0.8104 0.7931 0.7923 0.175 0.798 0.7998
DenseNet-201 0.8506 0.8071 0.851 0.821 0.857 0.0856 0.834 0.8653
Inception-ResNet-v2 0.829 0.7872 0.8696 0.8001 0.793 0.134 0.729 0.8234
Proposed CNN _3 0.9128 0.911 0.9058 0.892 0.9256 0.0589 0.820 0.9034
Normal
versus
Ictal
VGG-19 0.87 0.856 0.878 0.8234 0.8409 0.154 0.842 0.876
ResNet-101 0.8973 0.8886 0.8983 0.8876 0.873 0.0899 0.797 0.9053
Inception-v3 0.8847 0.8764 0.8855 0.8657 0.871 0.098 0.797 0.8971
GoogLeNet 0.8909 0.8967 0.8822 0.89 0.8983 0.075 0.877 0.884
DenseNet-201 0.9032 0.94 0.943 0.9367 0.9383 0.063 0.8756 0.9150
Inception-ResNet-v2 0.9266 0.92 0.9271 0.8992 0.9095 0.0143 0.813 0.935
Proposed CNN_3 0.9249 0.9318 0.9061 0.9042 0.9079 0.0112 0.8126 0.0937
FIGURE 11 Confusion matrix and ROC curve for two-class classification with the proposed CNN_3 model from spectrogram images, 1
refers to pre-ictal class and 2 refers to normal class
IBRAHIM ET AL.21 of 30
layer, batch normalization, and a max-pooling layer) and three Fully-Connected (FC) layers. A 2 2 kernel is used to
perform convolution with the 128 128 input images. In the convolution and FC layers, the ReLU is exploited as an
activation function, but the final output layer employs a softmax function.
The ReLU activation function is used in the CNN_3 model, because it reduces the computing complexity. The
CNN_3 extracts features automatically by using multiple kernel functions to execute convolution at multiple levels.
53
The outputs of several convolution blocks produce smooth and detailed feature vectors from lower to higher levels of
input images, which helps in the classification and detection operations. Furthermore, batch normalization reduces the
FIGURE 12 Training progress for CNN_3 model for two-class classification from spectrogram images
TABLE 16 Experimental results for two-class classification with CNN_3 model from PSR images
Scenario CNN Accuracy Recall Specificity Precision Fscore Fpr Mcc Kappa
Normal
versus
Pre-ictal
VGG-19 0.964 0.9603 0.9655 0.9482 0.9542 0.013 0.82 0.956
ResNet-101 0.9751 0.9698 0.9768 0.959 0.9643 0.011 0.816 0.9612
Inception-v3 0.9777 0.9733 0.9785 0.9612 0.9672 0.008 0.807 0.9681
GoogLeNet 0.9507 0.9478 0.9514 0.9243 0.9359 0.0091 0.858 0.936
DenseNet-201 0.9728 0.9636 0.9744 0.9524 0.9579 0.006 0.817 0.959
Inception-ResNet-v2 0.9865 0.979 0.9873 0.9921 0.9865 0.0053 0.854 0.983
Proposed CNN _3 1 1 1 1 1 0 1 1
Normal
versus
Ictal
VGG-19 0.9757 0.9693 0.9871 0.9795 0.9743 0.007 0.867 0.9748
ResNet-101 0.9806 0.9864 0.9921 0.9846 0.9804 0.0034 0.856 0.9861
Inception-v3 0.9809 0.9873 0.9917 0.9751 0.9611 0.0067 0.952 0.9773
GoogLeNet 0.9793 0.9772 0.9717 0.9618 0.9694 0.0082 0.987 0.9718
DenseNet-201 0.9738 0.973 0.9848 0.9647 0.9538 0.00865 0.978 0.9898
Inception-ResNet-v2 0.992 0.9895 0.9934 0.9753 0.9723 0.002 0.912 0.9921
Proposed CNN_3 1 1 1 1 1 0 1 1
22 of 30 IBRAHIM ET AL.
internal covariance shift in the convolution layer output.
54
Additionally, to minimize complexity, spatial size, and fea-
ture variances, the max-pooling of size 2 2 is considered.
55
Overfitting has been addressed by setting a predetermined
dropout rate before each FC layer. Finally, in the output layer, the classification is performed by using the softmax acti-
vation function.
4.5 |Experimental results
In this approach, the segments are classified as ictal, normal, or pre-ictal states. The performance of the proposed
PSR-based DL approach has been investigated on the signals of all patients of the CHB-MIT dataset.
38
The proposed
PSR-based approach has been tested using the same metrics adopted with the patient-specific approach in addition to
Matthews Correlation Coefficient (MCC), False Positive Rate (F
pr
), and Kappa for all binary and ternary classes, as
shown below.
42
The MCC is defined as:
MCC ¼TpTnFpFn
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
TpþFp
TpþFn
TnþFp
TnþFn
ðÞ
q100 ð15Þ
False-Positive Rate is given by:
Fpr ¼Fp
TnþFp
ð16Þ
Kappa coefficient is defined as:
Kappa ¼2TpTnFnFp
TpþFp
FpþTn
TpþFn
FnþTn
ðÞ
100 ð17Þ
To create the PSR images for different segments, three essential parameters are determined: ϵ,τand m. First, the
value of τis calculated using the AMI function. The optimum value is chosen based on the point of rising of AMI at ini-
tial local minima. AMI values often decrease with an increase of τ. The optimum values for normal, ictal, and pre-ictal
FIGURE 13 Confusion matrix and ROC curve for the two-class classification with CNN_3 model on PSR images, 1 refers to pre-ictal class and 2
refers to normal class
IBRAHIM ET AL.23 of 30
states are 1–12, 1–25, and 1–15, respectively. The EEG signal points are mapped to a 2D space by setting mto 2. This
value has been considered for generating 2D representations of EEG signal segments for epileptic seizure detec-
tion.
36,56,57
The PSR images of EEG signals for the ictal state are more regular than those for the normal state, as seen
in Figure 10. Visual evaluation of the PSR images of the EEG signals reveals that the dispersion of points in the PSR
images of normal intervals is significantly greater than that of PSR images of ictal intervals. In other words, as com-
pared to normal states, the majority of data points in ictal states are around the coordinate center. Furthermore, the
deviation of data points from the center of the coordinate plane is greater in normal states than in ictal states. The diag-
onal plots represent the histograms of x(t) and x(t+1).
FIGURE 14 Training progress for CNN_3 model for two-class classification from PSR images
TABLE 17 Experimental results of three-class classification with with CNN_3 model from spectrogram and PSR images
Preprocessing CNN Accuracy Recall Specificity Precision Fscore Fpr Mcc Kappa
Spectrogram VGG-19 0.8904 0.8961 0.8848 0.9044 0.9035 0.173 0.821 0.856
ResNet-101 0.924 0.9286 0.9196 0.9399 0.9383 0.121 0.909 0.9112
Inception-v3 0.9068 0.9322 0.9216 0.9421 0.9407 0.089 0.9012 0.8981
GoogLeNet 0.8963 0.9016 0.8911 0.9198 0.9183 0.191 0.812 0.826
DenseNet-201 0.9077 0.9077 0.9281 0.9518 0.9507 0.016 0.872 0.909
Inception-ResNet-v2 0.9065 0.939 0.9373 0.9521 0.9565 0.0083 0.907 0.893
Proposed CNN_3 0.9121 0.9446 0.9524 0.9447 0.9446 0.0076 0.9370 0.9256
Phase Space Reconstruction VGG-19 0.9257 0.913 0.871 0.9195 0.8943 0.125 0.8562 0.9048
ResNet-101 0.9406 0.9164 0.9021 0.9146 0.9104 0.0134 0.883 0.9061
Inception-v3 0.9309 0.943 0.9317 0.9451 0.9311 0.0067 0.9352 0.9273
GoogLeNet 0.9193 0.9172 0.9017 0.9218 0.9394 0.0082 0.907 0.9118
DenseNet-201 0.9438 0.953 0.9348 0.9547 0.9238 0.00865 0.9056 0.9298
Inception-ResNet-v2 0.952 0.9695 0.9534 0.9453 0.9423 0.002 0.9134 0.921
Proposed CNN_3 0.9989 0.9989 0.9995 0.9989 0.9989 0.00005 0.9984 0.9975
24 of 30 IBRAHIM ET AL.
4.5.1 | Results of two-class classification
Six pre-trained models, namely VGG-19, ResNet-101, Inception-v3, GoogLeNet, DenseNet-201 and Inception-ResNet-v2
are used in this task. Moreover, the proposed CNN_3 trained from scratch is also used. Table 15 presents the detection
results obtained for different pre-trained models and the proposed CNN_3 based on spectrograms. It is clear that the
proposed CNN_3 outperforms the pre-trained models for normal-pre-ictal and normal-ictal classification with
FIGURE 15 Confusion matrix and ROC curve for three-class classification with the CNN_3 model from spectrogram images, 1 refers to
ictal class, 2 refers to pre-ictal class and 3 refers to normal class
FIGURE 16 Training progress of the CNN_3 model for three-class classification from spectrogram images
IBRAHIM ET AL.25 of 30
accuracies that reach 90.28% and 92.49%, respectively. The corresponding confusion matrix and ROC curve of the pro-
posed CNN_3 model on spectrogram images for per-ictal-normal classification are shown in Figure 11. Figure 12 dis-
plays the performance of the proposed CNN_3 model in terms of both accuracy and loss. There is a coincidence in
performance between validation and training accuracies as well as validation and training loss. The Minimal Square
Error (MSE) has been chosen as the loss function. Based on the MSE, a distance minimization approach is used.
FIGURE 17 Confusion matrix and ROC curve for three-class classification with CNN_3 model from PSR images, 1 refers to ictal class,
2 refers to pre-ictal class, and 3 refers to normal class
FIGURE 18 Training progress for the CNN_3 model for three-class classification from PSR images
26 of 30 IBRAHIM ET AL.
Table 16 presents the detection results obtained from different pre-trained and the proposed CNN_3 model from
PSR images. It is clear that the proposed CNN_3 model outperforms the pre-trained models with accuracies that reach
100% for normal-pre-ictal and normal-ictal classification. The corresponding confusion matrix and ROC curve of the
proposed CNN_3 model on PSR images for per-ictal-normal classification are shown in Figure 13. Figure 14 displays
the performance of the CNN_3 model in terms of both accuracy and loss. There is a coincidence in performance
between validation and training accuracies as well as validation and training loss.
4.5.2 | Results for three-class classification
Table 17 presents the detection results obtained from spectrogram and PSR images for three-class classification. It is
clear that the proposed CNN_3 model performs better on PSR images. Moreover, the proposed CNN_3 model outper-
forms the other pre-trained models on PSR images. It achieves an average accuracy of 99.89%. The corresponding
TABLE 18 Comparison of the proposed CNN_3 model with the state-of-the-art models
Authors Technique Dataset Classes Accuracy %
M. Talha et al.
14
CNN KK women and children's hospital Normal-ictal 93.3
Yang et al.
15
NLSTM Boon University Normal-ictal 98.44
CHB-MIT 97.47
Covert et al.
17
TGCN Boston children's hospital Normal-ictal 98.05
B. Bouaziz et al.
19
CNN CHB-MIT Normal-ictal 99.48
Rajaguru et al.
20
MAE +EM-PCA CHB-MIT Normal-ictal 93.78
Roy et al.
21
ChronoNet CHB-MIT Normal-ictal 86.57
Choi et al.
22
3D-CNN +GRU CHB-MIT Normal-ictal 89.4
SNUH 97
Truong et al.
23
Spectrograms +STFT Freiburg hospital intra-cranial EEG Normal-ictal 81.4
CHB-MIT 81.2
Subasi et al.
29
GA +PSO Boon University Normal-ictal 99.38
Alickovic et al.
8
MSPCA +DWT CHB-MIT Normal-ictal 100
Normal-ictal-pre-ictal 99.77
M. Zhou et al.
28
CNN CHB-MIT Ictal-pre-ictal
Normal-ictal
Normal-ictal-pre-ictal
95.6
97.5
93
Freiburg Ictal-pre-ictal
Normal-ictal
Normal-ictal-pre-ictal
96.7
95.4
92.3
Qaisar et al.
30
CNN Andrzejak Normal-ictal-pre-ictal 96.4
Hassan et al.
34
TQWT Boon University Normal-ictal-pre-ictal 98.4
Hassan et al.
35
CEEMDAN +NIG Boon University Normal-ictal-pre-ictal 97.6
Sharma et al.
36
2D PSRs +EMD +LS-SVM CHB-MIT Normal-ictal 98.67
Shankar et al.
37
PSR +CNN Boon University Normal-ictal 93
CHB-MIT Normal-ictal 85
Proposed CNN_3 model Spectrogram images CHB-MIT Normal-ictal
Normal-pre-ictal
Normal-ictal-pre-ictal
91.28
92.49
90.21
PSR images Normal-ictal
Normal-pre-ictal
Normal-ictal-pre-ictal
100
100
99.89
IBRAHIM ET AL.27 of 30
confusion matrices and ROC curves of the proposed CNN_3 model on spectrogram and PSR images are shown in Fig-
ures 15, 16 and 17. The performance of the CNN_3 model on spectrogram and PSR images in terms of both accuracy
and loss is shown in Figures 16 and 18, respectively. It is clear from Table 17 that the utilization of PSR images
improves the accuracy by 5.7%, recall by 5.8%, kappa with 7.77% compared to the results obtained with spectrogram
images. In addition, with PSR images the CNN_3 model achieves a reduction of the F
pr
by about 99% of the value
obtained on spectrogram images.
4.6 |Discussion and comparison with the-state-of-the-art methods
Asymptotically, chaotic signals like EEG signals may evolve towards lower-dimensional attractors. The PSR can be used
to visualize these attractors, indirectly. We can observe from Figure 10 that such lower-dimensional attractors exist for
EEG signals, especially for the ictal class. This was also shown in previous research,
36,37
where the PSR was utilized as
a starting point for manually extracting features, and subsequently classifying them. The choice of time delay and
embedding dimension are essential aspects in PSR and its chaotic features. If the range of τis too small, the attractor
will be condensed into the area of the coordinate system diagonal line. On the other hand, if the range of τis too large,
the data point trajectory will twist and fold, making obvious projection correlations difficult to create. The experimental
results reveal that working on spectrogram images gives good results, but the phase information is ignored which limits
the accuracy levels to 91.28%, 92.49%, and 90.21% for ictal-pre-ictal, normal-ictal, and normal-pre-ictal-ictal classifica-
tion, respectively. Hence, we thought to avoid this limitation, by working with PSR, which gives accuracy levels up to
100%, 100%, and 99.89% for ictal-preictal, normal-ictal, and normal-pre-ictal-ictal classification, respectively, due to the
kept phase information.
The CNN_3 takes 1125 s for the two-class classification and 2015 s for the three-class classification from PSR
images.
The accuracy level with the CNN_3 model reaches 100%, which is higher than the levels of the traditional methods
given in Table 18, for two-class classification. Moreover, an accuracy level of 99.89% has been obtained for the three-
class classification. These findings confirm the efficiency of the CNN_3 model on PSR images for the classification of
EEG signals.
5|CONCLUSIONS
This paper dealt with the problem of EEG signal classification for two purposes: seizure detection and seizure predic-
tion. Two patient-specific CNN models were presented for seizure detection and seizure prediction. They work on spec-
trogram images of EEG signal segments. The third CNN model is patient non-specific. It has the ability to classify two
and three states of EEG signals. Moreover, it is valid for operation on spectrogram as well as PSR images of EEG signal
segments. Experimental results revealed the best classification performance with the patient non-specific CNN model
from PSR images of EEG signal segments. This is attributed to the elegant EEG signal representation with PSR. On the
other hand, the first two models are acceptable for patient-specific applications, but the dependence on spectrogram
images limits their performance to some extent. In future work, the proposed models will be tested on different datasets
and different signal representations.
DATA AVAILABILITY STATEMENT
The data that support the findings of this study are openly available in CHB-MIT dataset at https://archive.physionet.
org/pn6/chbmit/, reference number 56.
ORCID
Walid El-Shafai https://orcid.org/0000-0001-7509-2120
Mohamed Rihan https://orcid.org/0000-0003-4030-2559
REFERENCES
1. WHO, “Programmes and Projects.”http://www.who.int/mediacentre/factsheets/fs999/en/, 2017.
2. Thurman DJ, Logroscino G, Beghi E, et al. The burden of premature mortality of epilepsy in high-income countries: a systematic review
from the mortality task force of the international league against epilepsy. Epilepsia. 2017;58(1):17-26.
28 of 30 IBRAHIM ET AL.
3. Shoeibi A, Khodatars M, Ghassemi N, et al. Epileptic seizures detection using deep learning techniques: a review. Int J Environ Res
Public Health. 2021;18(11):5780.
4. Blinowska KJ, Zygierewicz J. Practical Biomedical Signal Analysis Using MATLAB R. CRC Press; 2011.
5. Acharya UR, Hagiwara Y, Adeli H. Automated seizure prediction. Epilepsy Behav. 2018;88:251-261.
6. Sharma M, Bhurane AA, Acharya UR. Mmsfl-owfb: a novel class of orthogonal wavelet filters for epileptic seizure detection. Knowledge-
Based Sys. 2018;160:265-277.
7. Acharya UR, Hagiwara Y, Deshpande SN, et al. Characterization of focal eeg signals: a review. Future Generation Comput Sys. 2019;91:
290-299.
8. Alickovic E, Kevric J, Subasi A. Performance evaluation of empirical mode decomposition, discrete wavelet transform, and wavelet
packed decomposition for automated epileptic seizure detection and prediction. Biomed Signal Process Control. 2018;39:94-102.
9. Emara HM, Elwekeil M, Taha TE, et al. Hilbert transform and statistical analysis for channel selection and epileptic seizure prediction.
Wireless Personal Commun. 2021;116(4):3371-3395.
10. Emara HM, Shoaib MR, Elwekeil M, et al. Deep convolutional neural networks for covid-19 automatic diagnosis. Microsc Res Tech.
2021;84:2504-2516.
11. Emara HM, Elwekeil M, Taha TE, et al. Efficient frameworks for EEG epileptic seizure detection and prediction. Annals of Data Science.
Springer; 2021:1-36.
12. Hossain MS, Amin SU, Alsulaiman M, Muhammad G. Applying deep learning for epilepsy seizure detection and brain mapping
visualization. ACM Transac Multimedia Comput Commun Appl (TOMM). 2019;15(1s):1-17.
13. Parvez MZ, Paul M, Antolovich M. Detection of pre-stage of epileptic seizure by exploiting temporal correlation of emd decomposed
EEG signals. J Med Bioeng. 2015;4(2):110-116.
14. Avcu MT, Zhang Z, Chan DWS. Seizure detection using least eeg channels by deep convolutional neural network. ICASSP 2019–2019
IEEE IntConf Acoustics Speech Signal Process (ICASSP). 2019;7:1120-1124.
15. Li Y, Yu Z, Chen Y, et al. Automatic seizure detection using fully convolutional nested lstm. Int J Neural Syst. 2020;30(04):2050019.
16. Asif U, Roy S, Tang J, Harrer S. Seizurenet: multi-spectral deep feature learning for seizure type classification. Machine Learning in
Clinical Neuroimaging and Radiogenomics in Neuro-Oncology. Springer; 2020:77-87.
17. Covert IC, Krishnan B, Najm I, et al. Temporal graph convolutional networks for automatic seizure detection. Machine Learn Healthcare
Confe. 2019;5:160-180.
18. Bizopoulos P, Lambrou GI, Koutsouris D. Signal2image modules in deep neural networks for EEG classification. 2019 41st Ann Int Conf
IEEE Eng Med Biol Soc (EMBC). 2019;6:702-705.
19. Bouaziz B, Chaari L, Batatia H, Quintero-Rincon A. Epileptic seizure detection using a convolutional neural network. Digital Health
Approach for Predictive, Preventive, Personalised and Participatory Medicine. Springer; 2019:79-86.
20. Rajaguru H, Prabhakar SK. Multilayer autoencoders and empca with genetic algorithm for epilepsy classification from EEG. 2018
Second Int Conf Electron Commun Aerospace Technol (ICECA). IEEE; 2018:353-358.
21. Roy S, Kiral-Kornek I, Harrer S. Chrononet: a deep recurrent neural network for abnormal EEG identification. Conf Artif Intelligence
Med Eur. 2019;11:47-56.
22. Choi G, Park C, Kim J, et al. A novel multi-scale 3d CNN with deep neural network for epileptic seizure detection. 2019 IEEE Int Conf
Consum Electron (ICCE). 2019;3:1-2.
23. Truong ND, Nguyen AD, Kuhlmann L, et al. Convolutional neural networks for seizure prediction using intracranial and scalp
electroencephalogram. Neural Netw. 2018;105:104-111.
24. Wang Z, Oates T. “Encoding time series as images for visual inspection and classification using tiled convolutional neural networks,”in
Workshops at the Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015.
25. Barra S, Carta SM, Corriga A, Podda AS, Recupero DR. Deep learning and time series-to-image encoding for financial forecasting.
IEEE/CAA J Automatica Sinica. 2020;7(3):683-692.
26. Emami A, Kunii N, Matsuo T, Shinozaki T, Kawai K, Takahashi H. Seizure detection by convolutional neural network-based analysis of
scalp electroencephalography plot images. NeuroImage: Clinical. 2019;22:101684.
27. Yuvaraj R, Thomas J, Kluge T, Dauwels J. A deep learning scheme for automatic seizure detection from long-term scalp EEG. 2018 52nd
Asilomar ConfSig Sys Comput. 2018;9:368-372.
28. Zhou M, Tian C, Cao R, et al. Epileptic seizure detection based on EEG signals and CNN. Front Neuroinform. 2018;12:95.
29. Subasi A, Kevric J, Canbaz MA. Epileptic seizure detection using hybrid machine learning methods. Neural Comput Appl. 2019;31(1):
317-325.
30. Qaisar SM, Subasi A. Effective epileptic seizure detection based on the event-driven processing and machine learning for mobile
healthcare. J Ambient Intelligence Human Comput Secur. 2020;8:1-13.
31. Tuncer T, Dogan S, Ertam F, Subasi A. A novel ensemble local graph structure-based feature extraction network for EEG signal analysis.
Biomed Signal Process Control. 2020;61:102006.
32. Hassan AR, Siuly S, Zhang Y. Epileptic seizure detection in EEG signals using tunable-q factor wavelet transform and bootstrap
aggregating. Comput Methods Programs Biomed. 2016;137:247-259.
33. Hassan AR, Haque MA. “Epilepsy and seizure detection using statistical features in the complete ensemble empirical mode decomposi-
tion domain,”in TENCON 2015–2015 IEEE region 10 conferences, pp. 1–6, 2015.
34. Hassan AR, Subasi A. Automatic identification of epileptic seizures from EEG signals using linear programming boosting. Computer
Methods and Programs in Biomedicine. Vol 136. Elsevier; 2016:65-77.
IBRAHIM ET AL.29 of 30
35. Hassan AR, Subasi A, Zhang Y. Epilepsy seizure detection using complete ensemble empirical mode decomposition with adaptive noise.
Knowledge-Based Sys. 2020;191:105333.
36. Sharma R, Pachori RB. Classification of epileptic seizures in EEG signals based on phase space representation of intrinsic mode
functions. Expert Sys Appl. 2015;42(3):1106-1117.
37. Shankar A, Khaing HK, Dandapat S, Barma S. Analysis of epileptic seizures based on EEG using recurrence plot images and deep
learning. Biomed Signal Process Control. 2021;69:102854.
38. PhysioNet, “CHB-MIT Scalp EEG Database.”https://www.physionet.org/pn6/chbmit/, 2000.
39. Hasnat A, Bohn'e J, Milgram J, Gentric S, Chen L. Deepvisage: making face recognition simple yet with powerful generalization skills.
Proc IEEE Int Conf Comput Vision Workshops. 2017;3:1682-1691.
40. Pham L, Phan H, Nguyen T, Palaniappan R, Mertins A, McLoughlin I. Robust acoustic scene classification using a multispectrogram
encoder-decoder framework. Digital Signal Process. 2021;110:102943.
41. Abdellatef E, Omran EM, Soliman RF, et al. Fusion of deep-learned and hand-crafted features for cancelable recognition systems. Soft
Comput. 2020;24(20):15189-15208.
42. Sokolova M, Lapalme G. A systematic analysis of performance measures for classification tasks. Inf Process Manag. 2009;45(4):427-437.
43. Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv. 2014;12:1409.1556.
44. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Proc IEEE Conf Comput Vision Pattern Recog. 2016;6:
770-778.
45. Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the inception architecture for computer vision. Proc IEEE Conf Comput
Vision Pattern Recog. 2016;5:2818-2826.
46. Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions. Proc IEEE Conf Comput Vision Pattern Recog. 2015;7:1-9.
47. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. Proc IEEE Conf Comput Vision Pattern
Recog. 2017;11:4700-4708.
48. Redmon J, Farhadi A. Yolov3: an incremental improvement. arXiv. 2018;8:1804.02767.
49. Polat K, Nour M. Epileptic seizure detection based on new hybrid models with electroencephalogram signals. IRBM. 2020;41(6):331-353.
50. Salini GA, P'erez P. A study of the dynamic behavior of fine particulate matter in Santiago, Chile. Aerosol Air Qual Res. 2015;15(1):
154-165.
51. Fraser AM, Swinney HL. Independent coordinates for strange attractors from mutual information. Phys Rev A. 1986;33(2):1134-1140.
52. Kennel MB, Brown R, Abarbanel HD. Determining embedding dimension for phase-space reconstruction using a geometrical construc-
tion. Phys Rev A. 1992;45(6):3403-3411.
53. Madhavan S, Tripathy RK, Pachori RB. Time-frequency domain deep convolutional neural network for the classification of focal and
non-focal EEG signals. IEEE Sensors J. 2019;20(6):3078-3086.
54. Zhang Y, Guo Y, Yang P, Chen W, Lo B. Epilepsy seizure prediction on EEG using common spatial pattern and convolutional neural
network. IEEE J Biomed Health Inform. 2019;24(2):465-474.
55. Acharya UR, Oh SL, Hagiwara Y, Tan JH, Adeli H. Deep convolutional neural network for the automated detection and diagnosis of
seizure using EEG signals. Comput Biol Med. 2018;100:270-278.
56. Akbari H, Saraf Esmaili S, Farzollah Zadeh S. Detection of seizure EEG signals based on reconstructed phase space of rhythms in ewt
domain and genetic algorithm. Signal Process Renewable Energy. 2020;4(2):23-36.
57. Yakopcic C, Alom MZ, Taha TM. Extremely parallel memristor crossbar architecture for convolutional neural network implementation.
2017 Int Joint Conf Neural Networks (IJCNN). 2017;8:1696-1703.
How to cite this article: Ibrahim FE, Emara HM, El-Shafai W, et al. Deep-learning-based seizure detection and
prediction from electroencephalography signals. Int J Numer Meth Biomed Engng. 2022;e3573.
doi:10.1002/cnm.3573
30 of 30 IBRAHIM ET AL.
