A Comparative Analysis of Classifier Performance for Epileptic

Seizure Detection Using EEG Signals

Nida Alyas, David P. Hastings and Abbas Mehrabidavoodabadi

Department of Computer & Information Sciences, Northumbria University, Newcastle-upon-Tyne, U.K.

Keywords: Machine Learning, Epilepsy, Seizure Detection, Signal Processing, EEG, Classification.

Abstract: In middle and low-income countries, epilepsy remains undiagnosed in many instances because of an

insufficient number of medical specialists and expensive EEG recording devices. In previous studies, many

machine learning (ML) based methods were proposed to investigate and classify the EEG signals. However,

little work has been performed with EEG data recorded with consumer-grade devices. The extraction of the

most discriminating set of features and high misclassification rate is another challenge. To address these

problems, this study empirically investigates several data segment sizes and chooses the optimal window size

to segment the Guinea-Bissau dataset. Several statistical and spectral feature extraction methods were

investigated to obtain useful sets of features from segmented epochs in combination with conventional ML

algorithms and ensemble methods. The proposed framework is then implemented on a comparable dataset

collected from Nigeria to validate the reliability of the framework. A comparative analysis is performed with

conventional ML models and with existing techniques to prove the effectiveness of the proposed methodology.

The obtained results demonstrate that XGBoost and LightGBM achieved the highest levels of performance in

terms of F1 score and AUC.

1 INTRODUCTION

Epilepsy is a chronic neurological illness that affects

1-2% of the population worldwide (Panayiotopoulos,

2010), with nearly 2.4 million people newly

diagnosed per year (Megiddo, et al., 2016).

Unfortunately, 30-40% of epileptic patients have

uncontrolled seizures and are left without any proper

treatment, or their seizures do not respond to

medication (Cook, et al., 2013). A brain test called an

electroencephalogram (EEG) is used to spot

anomalies in the electrical brain signals. EEG signals

are one of the frequently used methods to categorise

and predict neurological diseases and disorders.

Usually, the visual representation of EEG signals

must be analysed and monitored by experienced

medical professionals (Hu et al., 2020) a time-

demanding process and one where a clinician’s

fatigue can cause less accurate outcomes (San-

Segundo et al., 2019). Therefore, an automatic

epilepsy seizure detection system that can assist

healthcare experts and staff in analyzing the EEG

signals quickly and efficiently would assist in

providing more precise and reliable diagnoses.

Historically, epilepsy seizures are diagnosed by

classifying electroencephalography (EEG) electrical

signals into epilepsy or control classes and require

expensive devices to record the EEG signals. The

detection of epileptic seizures by classifying EEG

signals is a demanding and challenging task, as it

identifies the seizure and seizure-free states from non-

linear and non-stationary data. Previous research has

seen many machine learning based approaches

introduced to analyze and interpret EEG signals for

accurate classification. However, the nature of EEG

data (non-linear and non-stationary) makes it difficult

to extract proper information regarding these

dynamic biomedical signals, while the extraction of

the most relevant features set from EEG recordings is

also challenging. Another issue is the potential for

high misclassification rates due to the oscillatory and

fractal characteristics of EEG signals (epilepsy and

control) possessing a high resemblance. To address

these problems, this study empirically investigates

several data segment sizes and selects the optimal

window size to segment the Guinea-Bissau dataset, (a

dataset collected with consumer-grade devices,

mimicking the conventional way to record these

signals in many parts of the word). Several statistical

and spectral feature extraction methods are

investigated to obtain useful features from segmented

epochs, in combination with conventional ML

Alyas, N., Hastings, D. and Mehrabidavoodabadi, A.

A Comparative Analysis of Classiﬁer Performance for Epileptic Seizure Detection Using EEG Signals.

DOI: 10.5220/0011631600003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 237-244

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

237

algorithms and ensemble methods. The proposed

framework is then implemented on a dataset of

similar quality (the Nigeria dataset, collected with

same protocol as Guinea-Bissau dataset), to validate

the reliability of the framework. A comparative

analysis is performed to demonstrate the

effectiveness of the proposed framework. The

obtained results demonstrate that XGBoost and

LightGBM achieved the highest levels of

performance, in terms of F1 score and AUC. The

main contributions of this research project address

relevant feature extraction problems, low

performance levels, and detection using recordings

from low-cost devices, are stated as follows:

• An effective ensemble method for epileptic

seizure detection is implemented that classifies EEG

signals obtained using a low-cost device into epilepsy

and control classes with enhanced performance.

• The best window size is determined empirically

to get the optimal segments of the long EEG signals

for better interpretation.

• A combination of spectral and temporal feature

extraction techniques is investigated and extracts the

most useful set of features by implementing a TFD-

based statistical analysis of segmented EEG signals.

• The comparison analysis of state-of-the-art and

ensemble classifiers is performed using several

evaluation metrics.

2 LITERATURE REVIEW

Data can be collected using an EEG monitoring

device that records the brain activity in form of brain

signals through different channels/electrodes

connected to the scalp. It records the signals with

voltage and spatial information. EEGs are non-

stationary data, as the statistical characteristics of data

change over time (Azami et al., 2011). To allow for

the data to be used by a range of classifiers, data

segmentation divides the signals into segments that

are expected to contain the same temporal and

spectral features (Hassanpour & Shahiri 2007).

Dividing the data in such a way helps to generate

more training samples and a more appropriate way to

train classification algorithms. Feature extraction is a

fundamental component of developing an efficient

epileptic seizure detection system. The most

frequently used feature extraction techniques for EEG

data include Time-domain (TD), Frequency-domain

(FD), Time-Frequency domain (TFD), Fourier

transform (FT), Short-time Fourier Transform

(STFT) and Continuous wavelet transform (CWT)

(Usman et al., 2019). Moreover, Empirical mode

decomposition (EMD), Kalman filter (KF), Singular

spectrum analysis (SSA), Discrete wavelet transform

(DWT), and Savitzgy-Golay (SG) filtering are also

used to enhance the performance of ML and deep

learning (DL) techniques (Azami & Sanei 2014).

Regarding the algorithmic approach, different

classifiers such as Support Vector Machine (SVM),

Random Forest (RF), Decision Tree (DT), and K-

Nearest Neighbour (KNN) have been seen to produce

high levels of performance, especially in brain signal

processing, and many previous studies preferred

hybrid models for automatic seizure detection

systems. Shoeb & Guttag (2010) presented a machine

learning-based classifier to develop a patient-specific

model for onset detection of an epileptic seizure and

took a step to explore and solve the automatic

epileptic seizure detection problem. They used the

data from the CHB-MIT database that contains a total

of 163 seizure episodes and EEG signals recordings

of 844 hours. They used SVM with feature extraction,

using time and frequency domains (FD). The

evaluation was performed using performance

measures such as sensitivity, detection delay, and

false alarms per hour. The model achieved good

results with an average sensitivity of around 96%.

However, the study highlights the main challenges

that are intrinsic to this problem, including data

quality issues. Wang et al., (2015) implemented and

compared ML algorithms such as DT based algorithm

named C4.5, RF, SVM, and SVM-based random

forest (SVM+RF), and DT-based SVM (SVM+C4.5)

for seizure detection. A RF outperformed all other

implemented models in this study, yielding the

highest accuracy among all algorithms (Wang et al.,

2015). Dash et al., (2020) proposed a novel approach

that extracted sub-components from EEG signals

using an iterative filtering decomposition approach

and implemented Hidden Markov Model (HMM) as

a classifier. They evaluated the models using a private

EEG dataset from the All India Institute of Medical

Science (AIIMS), Patna, and a publicly available

database (CHB-MIT). The final class was decided

based on the maximum score from HMM classifier.

The proposed novel approach using decomposition of

EEG signals and HMM attained 99.60% and 99.74%

accuracy for the online CHB-MIT and AIIMS Patna

EEG datasets, respectively. ML was found effective

for epilepsy detection by Kavitha et al., (2022), who

implemented KNN, DT, Naive Bayes (NB), and

SVM for EEG signal classifications. A University of

Bonn database and real-time private dataset obtained

from the Senthil Multispecialty Hospital, India, were

used. The EEG signals were extracted from both

datasets and broken up into six frequency sub-bands

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

238

using DWT and six statistical features were extracted,

allowing for the combining of different features and

classifiers. Van Hees et al. (2018) collected EEG

signals for 5-minutes with an EMOTIVE device for

epilepsy identification from two low-income

countries in rural areas: Guinea-Bissau and Nigeria.

They achieved an accuracy of 83% and 70% in

Guinea-Bissau and Nigeria respectively. The key

results from the a number of key studies in this

domain are shown in Table 1.

Table 1: Existing machine learning approaches for seizure

detection.

3 PROPOSED FRAMEWORK

This study proposes an improved framework to

automate epileptic seizure detection with the aim of

classifying the input EEG signals into epilepsy and

control classes, as shown in Figure 1. The proposed

framework consists of 6 sub-modules: (1) EEG

dataset acquisition using the consumer-grade device,

that describes the EEG data and its parameters in

detail; (2) pre-processing of data, which deals with

the cleaning of data from noise and artefacts; (3)

signal segmentation, related to dividing the long

signal into segments using a window size; (4) feature

extraction to extract the most relevant features from

the data; (5) classifier building, including building a

classifier using ensemble method to compare with

conventional ML models; and (6) evaluation, which

deals with the performance evaluation of models

using different metrics.

3.1 Data Collection

The datasets used in this research were acquired by

van Hees et al. (2018), and are available for public

access. They collected EEG signals from epileptic

(N=51) and healthy individuals (N=46) in the low-

income country of Guinea-Bissau. A low-cost,

portable, and consumer-grade device (EMOTIVE)

was used to acquire the 5-minutes of 14 channels

includes: AF3, AF4, F3, F4, F7, F8, FC5, FC6, O1,

O2, P7, P8, T7, T8) resting-state EEG data from rural

areas with the 128 Hz sampling frequency.

Figure 1: Proposed framework for epilepsy detection.

Figure 2: Emotive-EPOC device/headset 14 channels scalp

placement (Mehmoud & Lee, 2016).

A Comparative Analysis of Classiﬁer Performance for Epileptic Seizure Detection Using EEG Signals

239

Table 2: Description of Dataset.

The electrodes were placed on the scalp at

anterofrontal (AF3, AF4, F3, F4, F7, F8), parietal (P7,

P8), occipital (O1, O2), frontocentral (FC5, FC6), and

temporal sites (T7, T8), according to the International

standard 1020 as shown in Table 2. The Guinea-

Bissau dataset consisted of 97 subjects in total while

Nigeria dataset consist of 204 subjects.

3.2 Data Pre-processing

Data processing was performed to prepare the data for

use. The datasets contain non-EEG parameters: for

this research they can be considered superfluous,

hence they are removed, leaving the 14 EEG

channels. Data is filtered using a band-pass filter

range of 0.1Hz-45Hz: removal of the high

frequencies eliminates the effects of a few artifacts as

well as line noise, while the suppression of low

frequencies below 0.1 Hz eliminates the effects of

slow voltage shifts due to the skin potentials. Re-

referencing was then used to normalize the signal.

3.3 Signal Segmentation

To segment the signals, the full EEG signal of around

300 seconds time series is segmented into small

epochs to make a better interpretation and

classification. The appropriate segment size is

selected by evaluating different split sizes. Firstly,

different splits were investigated with overlapping of

1-s. The selected EEG epochs are of 8-s window size

with 1-s overlapping.

3.4 Feature Extraction

3.4.1 Power Spectral

Power spectral density using the Welch method is

performed to calculate the spectral features. The EEG

signal Y is decomposed to evaluate the power

distribution across the neurological frequency

spectrum. The Welch method is a Power Spectral

Density (PSD) estimation technique that uses a STFT

to calculate the periodogram for segmenting the EEG

data (TD signal to FD). Overlapping segments are

windowed with a discrete Fourier transform applied

to calculate the periodogram, then the data is squared

and each periodogram is averaged to obtain the power

measure.

3.4.2 Statistical Methods

Statistical methods were used to extract the different

temporal features such as average, standard deviation

(SD), peak to peak (PTP), variance, minimum,

maximum, index of minimum value, index of

maximum value, root mean square, and absolute

difference of signal, skewness, and kurtosis.

3.5 Classifier Building and Evaluation

In this study, four conventional machine learning

(ML) algorithms, Logistic Regression (LR), K-

Nearest Neighbour (KNN), Decision Tree (DT),

Support Vector Machine (SVM), and ensemble

methods as Random Forest (RF), Bagging, Majority

Voting, Extra Tree (ET), AdaBoost, Gradient

Boosting, XGBoost, LightGBM are implemented to

build classifiers for epileptic seizure detection using

EEG.

For the experiments, the chosen dataset consists

of a total of 4245 EEG epochs. For each classifier, the

dataset is spliced into train, validation, and test

subsets as shown in Table 3. Grid search is used for

hyper-parameters tuning and to overcome the

overfitting problem using k-fold cross validation. The

training subset is used to develop the classifier, the

validation set to tune the algorithms and test subsets

are used to evaluate the performance of classifiers.

Table 3: Train Test Split.

The classifiers are evaluated through a range of

performance metrics, namely accuracy, precision,

recall (both used to calculate the F1 score), and AUC,

along with the macro and weighted averages, to

ensure robustness in the evaluative process.

4 RESULTS AND DISCUSSION

This subsection provides an analysis and discussion

of the results obtained by implementing conventional

ML and ensemble methods with a different set of

features. All the stated results were achieved after

model-specific hyperparameter tuning to optimise

model performance. F1 score and AUC-ROC have

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

240

been chosen as the principal methods of evaluation,

as in combination they provide a robust means

evaluating predictive performance.

4.1 Results of Conventional ML

Algorithms

Tables 4 and 5 show the F1 scores for the

conventional ML algorithms. The results demonstrate

that highest level of performance among these models

comes from k-NN using spectral features (with a

weighted average F1 score of 0.833). SVM achieves

a similar level of performance, with a weighted

average F1 score of 0.830.

The experimental results are also evaluated in

terms of AUC-ROC, presented in Figures 3 and 4,

which demonstrate the performance of conventional

ML algorithms with spectral features achieve better

performance than those using temporal features.

Table 4: F1 scores for conventional ML algorithms (ML +

statistical features).

Table 5: F1 scores for conventional ML algorithms (ML +

spectral features).

The AUC plots confirm the performance discrepancy

between models, with K-NN and SVM considerably

outperforming the decision tree and logistic

regression models.

4.2 Results of Ensemble Methods

Nine ensemble methods were used in this study:

bagging, RF, extra tree (ET), hard majority voting,

soft majority voting, AdaBoost, gradient boosting,

(XGBoost), and LightBGM.

Figure 3: AUC plot for conventional ML models with

statistical features.

Figure 4: AUC plot for conventional ML models with

spectral features.

The F1 score results of the ensemble classifiers

using statistical features are shown in Table 6. In

these results, XGBoost demonstrates the best

performance with a weighted average of 0.921.

LightGBM also performed well when compared to

with the other methods, with a weighted average of

0.900.

A Comparative Analysis of Classiﬁer Performance for Epileptic Seizure Detection Using EEG Signals

241

Table 6: F1 scores for ensemble methods + statistical

features.

Table 7: F1 scores for ensemble methods + spectral features.

Table 7 shows the performance of ensemble

methods with the set of spectral features. The results

demonstrate that ET achieved the highest levels of

performance when compared to other ensemble

algorithms, with a weighted average of 0.879.

The experimental results of the ensemble methods

are evaluated in terms of AUC-ROC, the graphical

demonstration of which is in Figures 5 and 6. When

using the statistical features, XGBoost, LightGBM

and ET all demonstrate high levels of discriminant

ability, with AUC values of 0.98, 0.97 and 0.95

respectively, reinforcing the findings generated

through the F1 score. The same models achieve the

highest levels of performance when using the spectral

features, each achieving AUC values of 0.93; lower

than those achieved when using the statistical

features. This pattern is repeated for the majority of

ensemble models, which differs from the

conventional ML models, where the individual

highest levels of performance were achieved by using

the spectral features.

Figure 5: AUC plot for ensemble methods with statistical

features.

Figure 6: AUC plot for ensemble methods with spectral

features

From the comparative analysis of classification

results obtained from the Guinea-Bissau dataset it can

be seen that the classification accuracies of KNN and

SVM with spectral features are the best among all the

conventional ML algorithms implemented in this

study. Equally, for ensemble classifiers, XGBoost

with statistical features achieved the best overall

performance: this combination of model and features

was also the optimal approach when considering all

other approaches used in the study.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

242

4.3 Results with Nigeria Dataset

We used the Nigeria EEG dataset that was recorded

using the same protocols and standards as the Guinea-

Bissau dataset to validate the performance of the

proposed framework. The results achieved with this

dataset were satisfactory and prove the reliability of

the model. All the conventional ML algorithms and

ensemble methods along with feature extraction

techniques were implemented. The results gathered

from this data demonstrate similar outcomes to those

achieved using the Guinea-Bissau data: the highest

performing model was XGBoost with a set of

statistical features, with 79.45% accuracy and a

weighted F1 score of 0.793. While the results for the

Nigeria dataset are lower than those achieved when

using the Guinea-Bissau dataset, this mirrors the

findings of van Hees et al. (2018) and Anwar et al.,

(2021), who also document reduced levels of

performance when using the data collected from

Nigeria.

5 CONCLUSION

Epileptic seizures cause abnormalities of the brain

and physical activities of epileptic patients,

considered a chronic disease with an increased

number of patients and sudden deaths every year. As

earlier indicated, a better approach for epilepsy

detection uses EEG data recorded using a consumer-

grade device, and this study demonstrates that the

optimal performance for an epilepsy detection model

using such data can be achieved through ensemble

machine learning methods using statistical features

derived from the data. Accommodating the low-

quality data using low-cost devices has not frequently

been an approach used in previous research.

However, the use of such data in the development of

a system to detect epileptic seizures is better able to

replicate the real-world data that can be collected

from patients in much of the world and opens an

avenue to increase the diagnosis rate of this disorder

in low-income countries. However, additional factors

may be considered that remain unaddressed within

the study, such as geographical location of the

patients and patient genetics that may affect the

results. Further work will address this limitation to aid

in the development of more generalisable findings.

Moreover, when building the automatic seizure

detection system, the potential effectiveness of deep

learning methods should be investigated. Future work

will identify whether deep learning algorithms can be

implemented to further improve the development of

accurate and reliable detection systems, along with

attempting to optimise the datasets themselves,

through the use of combined statistical and spectral

features.

REFERENCES

Megiddo, I. et al., 2016. Health and economic benefits of

public financing of epilepsy treatment in India: An

agent-based simulation. Epilepsia, 57(3), pp. 1-11.

Panayiotopoulos, C. P., 2010. A Clinical Guide to Epileptic

Syndromes and their Treatment. NewYork: Springer.

Cook, M. J. et al., 2013. Prediction of seizure likelihood

with a long-term, implanted seizure advisory system in

patients with drug-resistant epilepsy: a first-in-man

study. Lancet Neurol, 12(6), pp. 563-571.

Hu, X. et al., 2020. Scalp EEG classification using deep Bi-

LSTM network for seizure detection. Computers in

Biology and Medicine, Volume 124.

San-Segundo, R. et al., 2019. Classification of epileptic

EEG recordings using signal transforms and

convolutional neural networks. Computers in Biology

and Medicine, Volume 109, pp. 148-158.

Azami, H., Mohammadi, K. & Hassanpour, H., 2011. An

Improved Signal Segmentation Method using Genetic

Algorithm. International Journal of Computer

Applications, 29(8).

Hassanpour, H. & Shahiri, M., 2007. Adaptive

Segmentation Using Wavelet Transform. Lahore,

Pakistan, IEEE.

Rashed-Al-Mahfuz, M. et al., 2021. A Deep Convolutional

Neural Network Method to Detect Seizures and

Characteristic Frequencies Using Epileptic

Electroencephalogram (EEG) Data. IEEE Journal of

Translational Engineering in Health and Medicine,

Volume 9.

Usman, S. M., Latif, S. & Beg, A., 2019. Principle

components analysis for seizures prediction using

wavelet transform. International Journal of Advances

in Applied Sciences, 6(3), p. 50–55.

Azami, H. & Sanei, S., 2014. Spike detection approaches

for noisy neuronal data: Assessment and comparison.

Neuro computing, Volume 133, pp. 491-506.

Shoeb, A. H. & Guttag, J. V., 2010. Application of Machine

Learning To Epileptic Seizure Detection. s.l., ICML

2010.

Wang, G., Deng, Z. & Choi, K.-S., 2015. Detection of

epileptic seizures in EEG signals with rule-based

interpretation by random forest approach. s.l.,

Advanced Intelligent Computing Theories and

Applications.

Dash, D. P., Kolekar, M. H. & Jha, K., 2020. Multi-channel

EEG based automatic epileptic seizure detection using

iterative filtering decomposition and Hidden Markov

Model. Computers in Biology and Medicine.

Kavitha, K. N. et al., 2022. On the Use of Wavelet Domain

and Machine Learning for the Analysis of Epileptic

A Comparative Analysis of Classiﬁer Performance for Epileptic Seizure Detection Using EEG Signals

243

Seizure Detection from EEG Signals. Journal of

Healthcare Engineering.

Anwar, T., Rehmat, N., & Naveed, H. (2021). A Generic

Approach for Classification of Psychological Disorders

Diagnosis using EEG. Mexico: IEEE.

Birjandtalab, J., Pouyana, M. B., Cogana, D., Nourani, M.,

& Harvey, J. (2017). Automated seizure detection using

limited-channel EEG and non-linear dimension

reduction. Computers in Biology and Medicine, 82, 49-

58.

Donos, C., Dümpelmann, M., & Schulze-Bonhage, A.

(2015). Early Seizure Detection Algorithm Based on

Intracranial EEG and Random Forest Classification.

International Journal of Neural Systems, 25(5).

Hees, V. T., Diessen, E. v., Sinke, M. R., Buitenhuis, J. W.,

Maas, F. v., Ridder, L., & Otte, W. M. (2018). Reliable

and automatic epilepsy classification with affordable,

consumer-grade electroencephalography in rural sub-

Saharan Africa. bioRxiv .

Mehmood, R. M. & Lee, H. J., 2016. Towards human brain

signal preprocessing and artifact rejection methods.

Lasvegas, USA, Int'l Conf. Biomedical Engineering

and Sciences (BIOENG'16).

Pinto-Orellana, M. A., & Cerqueira, F. R. (2016). Patient-

dependent epilepsy seizure detection using random

forest classification over one-dimension transformed

EEG data. Elsevier.

Wang, X., Gong, G., Li, N., & Qiu, S. (2019). Detection

Analysis of Epileptic EEG Using a Novel Random Forest

Model Combined With Grid Search Optimization.

Frontiers in Human Neuroscience, 13, 12.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

244