A novel LSTM-based data synthesis approach for performance

improvement in detecting epileptic seizures

W.M.N.D.KULASINGHE, MAHESHI B. DISSANAYAKE

Department of Electrical and Electronic Engineering, Faculty of Engineering,

University of Peradeniya,

Peradeniya, 20400

SRI LANKA

Abstract: Bio-electrical time signals play a significant role in assisting non-invasive observational procedures in

healthcare. These bioelectrical signals are weak signals with inherently low voltage and low frequency, hidden

mostly under relatively large high-voltage noise signals. Hence it is extra challenging to analyze them. In modern

clinical data analysis, these signals could be further analyzed using conventional machine learning (ML) methods.

Also, in the recent past, two-dimensional spectrum-based classification, predominantly with Convolutional Neural

Networks (CNN), has been tried with time-series data. One of the objectives of this study is to find which approach

would suit better for biomedical signal analysis when data are scarce and signals are weak. Also, in bio-medical

signal analysis data is scarce. Yet, to effectively train either an ML or a deep learning (DL) model, a sample

clinical dataset of a significant size is required. Hence, the second objective of this research is to present a novel

data synthesis method to address data scarcity. With these objectives, the study compares the performance of the

time-series-based classification with traditional ML approaches, against the 2D spectrum-based classification for

bio-electrical signal classification. For this purpose the study utilizes learning models; Multi-layer Perceptron

(MLP), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), Long Short-Term Memory Networks

(LSTMs), Auto Encoder (AE), and Convolutions Neural Network (CNN). Also, the authors propose a novel data

synthesis method based on LSTMs to improve the sample size of the standard CHB-MIT Scalp EEG dataset.

The results show that with the expanded dataset, the two-dimensional spectrum-based classification architecture

was able to achieve a precision level of 85% at the classification. The conventional ML-based methods showed

on average a precision level of 82%. In conclusion with the proposed virtual sample generation approach, 2D

spectrum-based classification with Convolutional Neural Networks showed promising performances.

Key-Words: Bio-electrical time signal classification, Long Short-Term Memory Networks (LSTMs),

Convolutional Neural Networks (CNN), Time series based classification, Two-dimensional spectrum based

classification, Virtual Sample Generation

Received: May 27, 2022. Revised: August 29, 2023. Accepted: September 23, 2023. Published: October 10, 2023.

1 Introduction

The human body is a complex biological system. Its

neural intercommunication system could be treated as

a low-voltage and low-frequency complex electrical

signaling system. The nervous system is the central

communication network of the human body, which

consists of billions of neurons that are responsible

for carrying small electrical signals. While the cen-

tral nervous system (CNS) of the body carries out the

functionalities in the brain and the spinal cord, the pe-

ripheral nervous system (PNS) transports signals be-

tween the CNS and the rest of the body. As the human

body works as a conductor, the neural signals can be

recorded from the skin via surface electrodes. Such

signals observed, especially as the changes in the elec-

trical potential across selected locations of the human

body, are referred to as “bioelectrical time signals”.

Examples of bioelectrical time signals are the elec-

troencephalogram (EEG), electrocardiogram (ECG),

and nerve conduction studies (NCS) which are the

recordings of the electrical activity of the body ob-

tained on the scalp, over the heart, and nerves respec-

tively.

In medical sciences, these bioelectrical time sig-

nals play a significant role in detecting abnormali-

ties, such as seizure disorders, encephalopathies, and

structural lesions. When making a diagnosis, these

signals’ frequency and amplitude fluctuations are ex-

amined in comparison to the norm. Yet, these bioelec-

trical signals are extremely weak by nature and are

concealed by large noise signals. They are difficult

to isolate and process. As a result, signal prepossess-

ing techniques like filtering are used to qualitatively

improve the analysis and to carry out the feature ex-

tractions efficiently.

Machine learning and deep learning are sub-fields

of artificial intelligence (AI), which has exhibited a

dramatic development cycle in the last decade. These

models possess the capacity to learn patterns and ex-

tract information from given data with minimum hu-

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

132

Volume 20, 2023

man intervention. They thereby open the door for

real artificial intelligence in machines and electronic

devices. AI has influenced and found application

in modern medical science extensively in tasks such

as disease diagnosis, prognosis, and treatment. In

this context, biomedical signal analysis with AI has

emerged as an exciting and rapidly evolving field that

fuses the learning architectures of AI with biomedi-

cal signal processing to revolutionize healthcare. It

integrates sensors and acquisition systems used in di-

agnosis with AI models for effective prepossessing,

precise categorization, and meaningful interpretation,

which would result in better healthcare solutions.

There are currently two fundamental approaches

for AI-based bioelectrical signal analysis: time series-

based analysis and two-dimensional spectrum-based

analysis. The authors of this study compared these

two methods for the use of EEG-based seizure de-

tection on a qualitative and quantitative level. The

CHB-MIT dataset, [1],[2], which is a publicly avail-

able standardized seizure detection dataset, was used

by the authors to train each learning algorithm. Fur-

thermore, this study utilizes learning models Multi-

layer Perceptron (MLP), Recurrent Neural Network

(RNN), Gated Recurrent Unit (GRU), Long Short-

Term Memory Networks (LSTMs), and Auto En-

coder (AE) to perform time series-based analysis,

while Convolutions Neural Network (CNN) is used

to model two-dimensional spectrum based analysis.

The lack of data, as in any AI-based research, is one

of the key limitations of bioelectrical signal analysis.

Deep learning models in particular need a lot of data

to make appropriate choices. The authors propose a

novel data synthesis method based on LSTMs to solve

the bottleneck caused by data scarcity. The proposed

method was able to qualitatively improve the size of

the standard dataset. Finally, through simulations, the

authors demonstrate the suitability of the proposed

data synthesis method for improving the detection ac-

curacy of epileptic seizures using EEG signals.

2 Background on Bio-electrical time

signal analysis methods

More often than not, bioelectrical time signal anal-

ysis is carried out by first viewing the received sig-

nals as groups of time series. These groups of signals

are either directly visualized or analyzed using tech-

niques such as time-frequency distributions (TFD),

fast Fourier transform (FFT), eigenvector methods

(EM), and autoregressive methods (ARM), [3].

In the research presented in this manuscript, the

authors analyze the EEG signals for seizure classifi-

cation. According to the literature, the most popular

machine learning algorithms for seizure classification

are the support vector machine (SVM), extreme learn-

ing machine (ELM), and linear discriminant analy-

sis (LDA), [3]. Along with the wave decomposition

techniques (WDT), SVM, ELM and LDA have shown

63.85%, [4], 42.8%, [5], and 50.14%, [6], of

accuracy in EEG based seizure classification,

respectively.

Many supervised deep-learning techniques are

also tested for seizure classification. For instance,

in [7], 74.3% of seizure classification accuracy was

recorded with stacked auto-encoders. The deep be-

lief networks (DBN), [8], and multi-layer

perceptron neural networks (MLPNN), [9], have

performed with 80.4% and 85.0 % accuracy in

seizure detection, re-spectively. The long short-

term memory network (LSTM) has returned an

accuracy of 87.0% for recog-nition of emotion from

raw EEG signals, [10].

Fig. 1. The two main approaches of EEG classification ar-

chitectures.

The literature highlighted above, analyzes EEG

signals as time series representations. An emerging

alternative approach for classifying EEG signals is to

first map the time series signal to a two-dimensional

(2D) image base representation as shown in Fig. 1,

and then to perform classification using 2D image

analysis models. In this alternative approach, the

mapping of a one-dimensional time series signal into

a two-dimensional spectrum is carried out using pre-

processing methods such as FFT, and power spectral

density (PSD). Next, the resultant spectrum images

are analyzed using either machine learning or deep

learning-based image analysis techniques.

The commonly used image classification method

in literature is the CNN, which has shown 87.3%

accuracy in EEG seizure classification, [11].

Also, more promising results were obtained, when the

CNN models were combined with other pre-

processing and post-processing techniques. For

instance, combining a sparse representation

classification (SRC) model with a fast compression

residual convolutional neu-ral networks (FCRes-

CNN) led to an improvement in average accuracy

from 88.79% to 98.82% in the de-tection of EEG

seizures, [12].

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

133

Volume 20, 2023

In the presented study, the authors, present a qual-

itative and quantitative comparison of the classifi-

cation performances of time series-based and two-

dimensional spectrum-based machine learning and

deep learning architectures for EEG-based seizure de-

tection. This objective is approached by applying

those two types of classification architectures on the

same standard seizure detection dataset (CHB-MIT)

[1],[2], under a similar environment and comparing

the classification performances against each other.

3 Dataset

The famous EEG classification dataset, CHB-MIT

Scalp EEG Database, [1],[2], provided by the

Mas-sachusetts Institute of Technology (MIT, USA)

is uti-lized in this study. This dataset is used to

differentiate epileptic seizures from the normal state

in pediatric and young adolescent patients.

The EEGs were recorded at the Children’s Hos-

pital Boston with 23 children with epileptic seizures

using scalp electrodes. The study included 17 female

patients with ages ranging from 1.5 to 19 years and

5 male patients with ages ranging from 3 to 22 years.

All subjects stopped the related treatments and medi-

cations several days before the data collection.

Most EEG files in the dataset contain EEG signals

with 23 leads per seizure detection instance, while

very few cases consist of 24 or 26 leads. These groups

of EEG signals were collected using the International

10-20 system of EEG electrode positions and nomen-

clature, [1]. The signals were recorded at a 256 Hz

sampling rate with 16-bit resolution. Throughout the

844 hours, i.e. over 900,000 time points, of observa-

tions presented in the dataset, a total of 686 recordings

with 198 seizures were captured. The database con-

sists of onsets and ends of 182 annotated seizures.

4 Methodology

It was observed that this dataset largely contains EEG

signals of a ”normal” state compared to the interested

”epileptic seizure” state, with an average distribution

of 98:2 of the total dataset. Due to the class imbalance

present within the dataset, in general, machine learn-

ing or deep learning classification models would fail

to learn the correct patterns and they would eventually

produce a biased output. 5

To address this shortcoming in the dataset, first,

outlier detection methods such as isolation forest and

local outlier factors were tested by considering the

seizures as the positive outliers of the dataset. Al-

ternatively, corrective sampling techniques such as

classifications with sample weights and random over-

sampling (ROS), synthetic minority oversampling

technique (SMOTE), and adaptive synthetic method

(ADASYN) were also tried in order to adjust the

shortage of samples in the seizure class. Out of these

tested methods, ROS and SMOTE showed promising

results as seen in Fig. 2. It can be observed that the

amplitudes of the oversampled data points lie within

the same range as the actual seizure signals. How-

ever, since these tested methods do not preserve the

temporal properties of the original samples and only

consider each instance as an individual sample, the re-

constructed time signals appear completely random.

This shortcoming has given rise to the necessity of

developing a new data reconstruction method to over-

come class imbalance.

Fig. 2. Original and reconstructed signals of random over-

sampling (ROS) and synthetic minority oversampling tech-

nique (SMOTE).

4.1 Data reconstruction

As presented earlier most EEG files in the dataset con-

tain EEG signals with 23 leads per each seizure de-

tection instance, while very few cases consist of 24,

25, or 26 leads, as extra leads or dummy signals. At

the data reconstruction, the first filter-based prepos-

sessing was utilized to remove these dummy signals

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

134

Volume 20, 2023

Fig. 3. Sample of extracted seizure and normal time signal.

and occasional extra lead signals. Next, frequencies

between 0.2Hz – 25Hz were extracted from the sig-

nals of all EEG leads. Thereafter the segment of the

seizure signal, as well as a random segment from the

normal state section of the EEG, were extracted. The

time duration of the normal state signal was made

equal to the time duration of the seizure signal. Both

the normal and seizure signals were combined to pro-

duce a new EEG signal of twice the length of the

seizure segment. This process was repeated with all

the EEG signals and all EEG leads. A sample output

of the newly generated EEG signal data is shown in

Fig. 3.

Next, an LSTM network was trained with the new

dataset to predict the next observation of the input sig-

nal when the previous observation was known to the

system. A sample of the training data for this LSTM

network is shown in Fig. 4. Once the network is

trained, the test input signals are fed to the network

to synthesize new signals similar to the original sig-

nals. A sample of the synthesized new EEG signal

generated by the LSTM network is shown in Fig. 5.

The proposed new EEG signal synthesizing process

was repeated for all the usable seizures present in the

original dataset and a newly synthesized signal was

created for each seizure signal in the dataset. All the

original and reconstructed signals were separated into

two classes; normal and seizure, in order to create a

balanced dataset with more volume of data, especially

from the seizure class.

The proposed new data synthesizing process was

further improved by randomly grouping seizure and

normal signal segments to create a more generalized

dataset. In this refinement stage, three segments of

either seizure (S) or normal (N) EEG signals from

the same patient were randomly selected and concate-

nated into one continuous signal. This approach of-

fers 8 possible combinations i.e NNN, NNS, NSN,

NSS, SNN, SNS, SSN, and SSS, per EEG data record-

ing of a single patient. This in turn helps to enhance

the dataset in a qualitative manner. Since data in med-

ical research is scarce, this proposed approach pro-

vides a practical solution to the problem identified.

Fig. 4. Sample of training data for the LSTM network.

Fig. 5. Reconstruction of a new signal.

4.2 Time series based classification

In order to analyze the EEG signals in practical en-

vironments such as at healthcare facilities for diag-

nostics, it is necessary to have a trained classifier

that could deliver excellent performance with unseen

data. For this purpose, state-of-the-art time series-

based deep learning classifiers, namely MLP, RNN,

GRU, LSTM, and AE were first trained and tested us-

ing the original dataset. It was observed that feeding

the original data directly to machine learning classi-

fiers did not return the expected performance. This

was mainly due to the class imbalance and the smaller

sample size associated with the original dataset. To

address this bottleneck, the offline synthesized dataset

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

135

Volume 20, 2023

created above, with improved 2000 data points from

each class, was utilized to retrain the MLP, RNN,

GRU, LSTM, and AE classifiers.

The created dataset was split into train-

ing:validation:testing as in 60:20:20 ratio. The

training subset of the created dataset was used to train

the models along with the Random Search and Tree

of Parzen Estimators as the parameter optimization

methods. The trained classifier was tested using

the testing subset of the dataset. The results were

averaged over 10 iterations of training and testing

sessions.

4.3 Two-dimensional (2D) spectrum based

classification

The 2D spectrum-based classification has emerged

in the literature as an alternative analysis method for

time-series data, [13]. In such an approach, the time

series signal is first converted to the 2D representa-

tions using either spectrograms or Frequency Spec-

trum (FS) with FFT or Power Spectral Density Spec-

trum (PSDS). However, as of literature, [13], [14],

[15], we have noticed that the performance of the

CNN-based classifiers with the spectrogram

converted 2D representation was very poor. This

probably may be due to the fact that spectrograms

illustrate only signal strengths.

The FS-based imaging visualizes frequencies of

the signals at each instance along the time axis. It

was observed that FS imaging-based classification

exhibited better performance than spectrogram im-

ages. The PSDS image of each class shows, the power

present in the signal as a per unit frequency at a time.

Although PSDS is almost similar to the correspond-

ing FS, PSDS itself adds slightly more contrast to the

spectrum image by revealing extra features of the time

series signal. This results in PSDS performing better

with CNN-classifier than FS.

In order to address the lack of sample data, data

augmentation approaches are widely utilized in 2D

space-image-based classification techniques. Fol-

lowing this trend, we also propose a novel augmen-

tation technique to create new 2D spectrum data sam-

ples out of the previously created dataset which com-

prised both actual and synthesized EEG signals. This

approach provides a solution to the key bottleneck as-

sociated with deep learning models in the healthcare

sector, i.e. scarcity of data. The augmented dataset

is used to train the CNN models, while the testing of

the trained model is carried out using only the origi-

nally acquired EEG data. The proposed augmentation

mechanism is presented as follows.

1. For every other 25 time points of the input sig-

nal, the immediate next 100 neighboring time

points of the same signal are appended to create

an epoch of 125-time points of length as shown

in Fig. 6. In this approach, the current 25-time

points will be repeated 4 more times (i.e. the

same 25-time points contained in 4 consecutive

epochs (=100/25)), creating redundancy in the

dataset.

Fig. 6. Example of epoch separation for the spectrum cal-

culation.

2. Next, FFT and PSDS were created for each epoch

as shown in Fig. 7.

Fig. 7. Sample frequency spectrum of each class.

3. The target class probabilities for each epoch were

calculated as the percentage of the seizure time

points contained inside the epoch. So, it indi-

cates the fraction of seizure signals present in

each spectrum in comparison to the normal sig-

nals as shown in Fig. 8.

Fig. 8. Original classes and target values for the CNN.

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

136

Volume 20, 2023

4. The new dataset with the spectrum and targeted

class probabilities was utilized to train a CNN

model as the classifier.

5 Results and Discussion

As presented in the introduction, two hypotheses re-

lated to biomedical signal classification are addressed

in this research. The authors first explore the pos-

sibilities of using a novel LSTM-based EEG signal

generation method to address the data shortage prob-

lem in EEG research. The study then assesses the

hypothesis of whether the conventional time-series-

based EEG classification approach is superior to the

two-dimensional spectrum-based approach. In or-

der to achieve this, we first train the chosen machine

learning and deep learning algorithms using the new

dataset that has been synthesized, and then we test

each of them using authentic EEG data from clinical

collections available through the CHB-MIT dataset.

This decision was made because we want to see how

well the classification algorithms that were developed

using the synthesis dataset, perform during deploy-

ment.

Table. 1 and Table 2 shows the classification

results for each classifier, namely MLP, RNN, GRU,

LSTM, AE and CNN. The classifiers MLP, RNN.

GRU, LSTM and AE were trained using the time-

series data han-dling approach, whereas CNN was

trained using the two-dimensional spectrum-based

approach. The pre-sented results are averaged

values over 10 iterations of independent training

and testing sessions. Fur-thermore, the two-

dimensional spectrum-based clas-sification

architectures appeared to be doing well in the

seizure classification, according to the results pre-

sented in Table. 1 and Table 2. The

performance sum-mary of the CNN-based

classification has reached al-most 98% accuracy with

better precision in each class. Also, we did not

notice a significant change in the performance of

either FFT or PSDS spectrum-based analysis.

The enhanced performance displayed by 2D

spectrum-based classification approach has been

caused by two aspects. One is that each 2D spectrum

sends the classifier a single input containing all the

frequencies and amplitude information of each epoch,

and the classifier automatically extracts the features

to learn from these inputs. The classifier that was

trained using input data from a sequential time series,

however, relies on manually extracted input features.

The redundancy added by the overlapping epochs at

the generation of 2D spectrum data is the second fac-

tor that contributed to the performance enhancement.

Due to this factor, the classifiers could further validate

the occurrences of seizures by looking at the subse-

quent spectrum data and assessing the temporal do-

main behavior.

Table 1: Time series-based deep learning classification re-

sults.

Classif

-ier

Precis

-ion

class

Precis

-ion

class

Recall

class

Recall

class

Accur

-acy

MLP 0.99 0.72 0.98 0.61 0.94

RNN 1.00 0.78 0.99 0.73 0.95

GRU 0.99 0.80 0.99 0.74 0.97

LSTM 1.00 0.82 0.99 0.79 0.98

AE 0.98 0.71 0.98 0.31 0.94

Table 2: Two-dimensional spectrum-based deep learning

classification results.

Classif

-ier

Precis

-ion

class

Precis

-ion

class

Recall

class

Recall

class

Accur

-acy

FFT+

CNN

1.00 0.83 0.99 0.79 0.99

PSDS+

CNN

1.00 0.85 0.99 0.80 1.00

As of Table. 1 RNN and LSTM showed better per-

formance under the time series-based signal analysis

approach. According to Table. 2 CNN which utilized

the 2D spectrum generated with PSDS outperformed

the scenario where the 2D Spectrum was generated

using FFT. Fig. 9 presents a detailed classification

report of RNN while Fig. 10 presents a detailed clas-

sification report of CNN with PSDS to give further

insight into the performance achieved by the two clas-

sifiers under the two different training scenarios, with

the same input.

The classification report of an RNN classifier

trained exclusively utilizing the original data, which

had a sizable class imbalance, is shown in Fig. 11.

Due to the intrinsic class imbalance in the original

data, the classification accuracy of this approach is

close to 100 %. This in turn shows that the EEG-

based seizure detection applications necessitate a data

reconstruction technique, and the suggested data syn-

thesis strategy aids in enhancing the overall effective-

ness of the AI models employed for this purpose.

6 Conclusion

This study makes two contributions to the field of

bio-electrical signal processing, particularly for auto-

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

137

Volume 20, 2023

Fig. 9. Classification results of the RNN network with syn-

thesis data.

Fig. 10. Classification results of the CNN network with

synthesis data.

Fig. 11. Classification results of the RNN network with

original data (without synthesis data).

mated EEG seizure detection using artificial intelli-

gence (AI). When actual clinical data are hard to get

by for EEG research, it first offers a time-series data

synthesis method. The proposed remedy focuses in

particular on the high-class imbalance, which is a non-

seizure-to-seizure ratio of 98:2 in the CHB-MIT Scalp

EEG database. Modern deep-learning-based classi-

fication architectures that use 2D spectrum have the

potential to improve classification results by automat-

ically extracting and processing more features from

the inputs at once. In order to categorize bioelectrical

signals, this research suggests with evidence that the

2D spectrum-based deep learning classification ap-

proach is an efficient and effective substitute for con-

ventional time-series-based machine learning.

The LSTM network is adopted in the proposed

novel seizure signal synthesizing architecture. The

proposed architecture is able to synthesize new

seizure signals that replicate and keep epileptic char-

acteristics while randomly injecting some fluctua-

tions to the signal to create the flaws that are intrin-

sically present in a real seizure signal. The study

then examines the effectiveness of the suggested data

synthesis approach in the time-series classification

of bioelectrical signals. To assess the validity of

the research hypothesis, the performance of the 2D

spectrum-based CNN technique and the traditional

time series-based machine learning and deep learn-

ing classification architectures were also evaluated.

For this analysis, 2D spectrums were generated us-

ing FFT and PSDS methods. Also, classifiers MLP,

RNN, GRU, LSTM, and AE were trained using time

series data analysis methods.

According to the results and observations, the 2D

spectrum-based deep learning classification architec-

ture overpowers the time series-based deep learning

classification architectures in terms of recall, preci-

sion, and accuracy. Additionally, both approaches

were able to prevent the negative effects of biaseness

found in the original dataset thanks to the data synthe-

sis architecture. The 2D spectrum-based deep learn-

ing classification architectures, however, were shown

to use up more processing power and memory due to

their intricate computations. The needs for process-

ing and memory may be easily met thanks to the de-

velopment of the modern hardware sector. In future

research, the authors wish to investigate further, novel

signal synthesis approaches for biomedical time se-

ries data analysis to assist the data scarcity present at

the biomedical signal process.

References:

[1] Ali Shoeb. Application of Machine Learning to

Epileptic Seizure Onset Detection and Treatment.

PhD Thesis, Massachusetts Institute of Technol-

ogy, September 2009.

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

138

Volume 20, 2023

[2] Goldberger, A., Amaral, L., Glass, L., Haus-

dorff, J., Ivanov, P. C., Mark, R. and Stanley,

H. E. Components of a new research resource

for complex physiologic signals. PhysioBank,

PhysioToolkit, and PhysioNet. (2000). Circula-

tion [Online]. 101 (23), pp. e215–e220.

[3] Prasanna J, Subathra M S P, Mohammed M A,

Damaševičius R, Sairamya N J, George S T.

Automated Epileptic Seizure Detection in Pe-

diatric Subjects of CHB-MIT EEG Database—

A Survey. J. Pers. Med. 2021, 11(10), 1028;

https://doi.org/10.3390/jpm11101028

[4] Park Y, Luo L, Parhi KK, Netoff T. Seizure

prediction with spectral power of EEG using

cost-sensitive support vector machines. Epilep-

sia. (2011) 52:1761–70. doi: 10.1111/j.1528-

1167.2011.03138.x

[5] Yuan Q, Zhou W, Zhang J, Li S, Cai D, Zeng

Y. EEG classification approach based on the ex-

treme learning machine and wavelet transform.

Clin EEG Neurosci. (2012) 43:127–32.

[6] Gadhoumi K, Lina JM, Gotman J. Seizure pre-

diction in patients with mesial temporal lobe

epilepsy using EEG measures of state similar-

ity. Clin Neurophysiol. (2013) 124:1745–54. doi:

10.1016/j.clinph.2013.04.006

[7] Yin Z and Zhang J, Recognition of cognitive

task load levels using single channel EEG and

stacked denoising autoencoder, Chinese Control

Conf. CCC vol 2016 pp 3907–12.

[8] Yu Z and Song J, Multi-class motor imagery clas-

sification by singular value decomposition and

deep Boltzmann machine, 2017 IEEE 3Rd Inf.

Technology Mechatronics Engineering Conf. pp

376–9

[9] Yohanandan S A C, A robust low-cost EEG motor

imagery-based brain–computer interface, 2018

40th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.

(EMBC) pp 5089–5092

[10] Salma Alhagry, Aly Aly Fahmy, Reda A. El-

Khoribi. Emotion Recognition based on EEG us-

ing LSTM Recurrent Neural Network, Interna-

tional Journal of Advanced Computer Science

and Applications (IJACSA), Vol. 8, No. 10, 2017.

[11] Qiao R, Qing C, Zhang T, Xing X and Xu X, A

novel deep-learning based framework for multi-

subject emotion recognition, Int. Conf. Informa-

tion, Cybernetics and Computational Social Sys-

tems, ICCSS 2017 pp 181–5.

[12] Jing-Shan Huang, Yang Li, Bin-Qiang Chen,

Chuang Lin and Bin Yao, An Intelligent EEG

Classification Methodology Based on Sparse

Representation Enhanced Deep Learning Net-

works, Front. Neurosci., 30 September 2020 |

https://doi.org/10.3389/fnins.2020.00808.

[13] Hasan, M.J., Shon, D., Im, K., Choi, H.-K., Yoo,

D.-S.,; Kim, J.-M. Sleep State Classification Us-

ing Power Spectral Density and Residual Neural

Network with Multichannel EEG Signals. Appl.

Sci. 2020, 10, 7639.

[14] Christensen, J.A.E., Munk, E.G.S., Peppard,

P.E., Young, T., Mignot, E., Sorensen, H.B.D.,

Jennum, P. The diagnostic value of power spectra

analysis of the sleep electroencephalography in

narcoleptic patients. Sleep Med. 2015, 16, 1516–

1527.

[15] Rahi, P.K., Mehra, R. Analysis of power spec-

trum estimation using welch method for various

window techniques, Int. J. Emerg. Technol. Eng.

Vol. 2, 2014, 106–109

Conflict of interest disclosure

The authors declare that they have no conflict of in-

terest.

Ethical approval

For this type of study formal consent is not required.

Informed consent

The dataset used in this article is freely available.

Corresponding Author

Maheshi Dissanayake (maheshid@eng.pdn.ac.lk)

Orcid:0000-0001-5209-5441

Contribution of Individual Authors to the

Creation of a Scientific Article (Ghostwriting

Policy)

The authors equally contributed in the present re-

search, at all stages from the formulation of the prob-

lem to the final findings and solution.

Sources of Funding for Research Presented in a

Scientific Article or Scientific Article Itself

No funding was received for conducting this study.

Conflicts of Interest

The authors have no conflicts of interest to

declare that are relevant to the content of this

article.

Creative Commons Attribution License 4.0

(Attribution 4.0 International , CC BY 4.0)

This article is published under the terms of the

Creative Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en

_US

WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE

DOI: 10.37394/23208.2023.20.13

W. M. N. D. Kulasinghe, Maheshi B. Dissanayake

E-ISSN: 2224-2902

139

Volume 20, 2023