Performance Analysis of Epileptic Seizure

Detection from EEG Signal by Employing

Machine Learning
Dr.Raghu K Suchith Reddy S Suhas H
Assitant Professor,School of ECE School of ECE School of ECE
REVA University, Bangalore REVA University, Bangalore REVA University, Bangalore
[email protected] [email protected] [email protected]
Supreeth U Tejas S
School of ECE School of ECE
REVA University,Bangalore REVA University,Bangalore
[email protected] [email protected]

I. ABSTRACT scalp electrodes that show distinct patterns that influence

seizures. However, the task of analyzing EEGs recorded for
Epileptic seizure detection from electroencephalogram
up to 72 hours is challenging and carries the risk of human
(EEG) signals is a critical task in neuroscience and medical
error. To overcome these limitations, methods for
diagnostics, requiring accurate and efficient methods for diagnosing diabetes based on signal processing, machine
timely intervention. This paper presents a comprehensive learning, and image analysis have been developed. The
performance analysis of machine learning techniques for main methods include methods combining PCA with cell
seizure detection, focusing on the comparison of key EEG counts, which achieve 97.64% accuracy, and wavelet-based
feature extraction methods and classification algorithms. methods such as wavelet denoising and continuous wavelet
The study evaluates four feature domains—time-domain, transform, which achieve 96.72% sensitivity and 96.72%
frequency-domain, time-frequency domain, and discrete specificity. Using image processing equipment to convert
wavelet transform (DWT) features—for their efficacy in EEG signals into images, the accuracy rate is 98.91%. Other
capturing the unique patterns associated with epileptic approaches that include phase space reconstruction with
seizures. For classification, two widely used machine Poincaré maps followed by analysis and classification
learning models, Support Vector Machines (SVM) and also show exceptional promise. Furthermore, machine
Random Forest, are employed and compared in terms of learning algorithms such as support vector machines (SVM)
accuracy, sensitivity, specificity, and computational and recurrent convolutional neural networks (RCNN) have
efficiency. Experimental results highlight the strengths and achieved good performance, with SVM detecting up to 96%
limitations of each feature domain and classifier, offering of seizures in CHB-MIT dataset. Bayesian models and
insights into their suitability for seizure detection in real- additional appropriate techniques increase sensitivity while
world scenarios. This work provides valuable guidance for reducing bias. However, despite these advances, challenges
remain in ensuring robustness, computational efficiency,
the development of robust EEG-based seizure detection
and ease of use for clinical applications. On this basis, this
systems and underscores the importance of feature
research focuses on improving dementia diagnosis by
selection and model choice in achieving high- combining automatic channel selection, spatiotemporal
performance outcomes. feature extraction, and random forest classifiers. This
approach reduces the number of EEG channels analyzed
Keywords—component, formatting, style, styling, using Kaggle epilepsy detection data, thus increasing the
insert (key words) efficiency of computations without loss of accuracy. This
II. INTRODUCTION work not only addresses the limitations of existing methods,
but also provides clinicians with powerful, effective, and
Epilepsy is a rare disease that affects about 1% of the world's user-friendly tools to improve the accuracy and accessibility
population. It is characterized by abnormal electrical of epilepsy diagnosis and management.
activity in the brain that causes seizures. Medications can
control symptoms in two-thirds of patients, but almost a
third respond to medication, and only 7 to 8 percent of
III. THE MODEL OF EPILEPTIC SEIZURE DETECTION FROM
patients can treat their pain with surgery. Early detection
and treatment of seizures in such patients can improve EEG SIGNALS
quality of life. The most common test used to diagnose
epilepsy is an electroencephalogram, or Main Contribution of the paper:
electroencephalogram. It captures electrical activity from
1.This study has shown that classifications based on study statistics and machine learning. We present a
decision tree clusters (e.g., decision forest and random comparative analysis to highlight the similarities between
forest) outperform other machine learning methods, these methods.
including those using neural networks (ANNs), nearest
neighbors (KNNs), support vector machines, and self- In the section “Epilepsy Localization”, we delve into the
determination trees. This finding demonstrates the research to identify specific lobes of the brain affected
effectiveness and robustness of the integrated approach to during epilepsy by using machine learning to improve local
research activities. reality.

2. We believe that decision forest algorithms have great “Problems found in the existing literature” addresses the
potential for widespread use in knowledge discovery, in limitations and problems found in previous studies. This
addition to their success in epilepsy detection. These section highlights gaps in systems, materials, and processes
algorithms can be adapted to solve a variety of clinical and lays the foundation for future success.
problems, such as identifying patterns in data or uncovering
relationships that can help make clinical decisions beyond The "Evaluation of Current Systems and Operating
epilepsy research. Systems" section presents our findings regarding the best
products and statistical functions for epilepsy diagnosis
3. This study provides researchers with general guidance based on our review.
through data analysis and analysis of the most appropriate
statistical and machine learning methods to improve the Research Guidelines in Diabetes Research - Important
process of diagnosing dementia. By understanding the issues for future research, including new methods,
strengths and limitations of these systems, researchers can techniques, technology, and improved translation standards
develop strategies to improve performance. are discussed.

4. This study can also serve as a resource for readers The final section, "Conclusion", summarizes the main
looking for information about epilepsy in the public findings and contributions of this study and highlights their
domain. These documents are important for the importance for current and future research practices.
advancement of research in the field, and our discussion
provides a better understanding of their function, scope, This method allows readers to easily follow the article and
and how machine learning can be used for practical find the most important information that suits their needs.
purposes in epilepsy diagnosis.

5. Finally, we present the findings and insights from the

analysis, as well as recommendations for solving problems
in the context of current research. Our recommendations
aim to encourage further innovation and advancement in
epilepsy detection techniques and to provide more reliable
and effective solutions.

6. This document is intended to provide the following

useful information and insights:
BLOCK DAIGRAM
Also known as “The role of data scientists in visual search
Role of data scientist in epileptic seizure detection:
epilepsy”, this document explores the involvement of
machine learning experts in the analysis of brain learning
Health and biological data have been significantly
on electrographic datasets and highlights the important role
influenced by machine learning techniques that have
that data scientists play in the restoration of the region.
yielded better results [15,15] In order to improve seizure
detection, scientists and researchers are working hard to
In “Epilepsy Detection Framework”, we propose a
develop new solutions, especially in data mining and
generalized epilepsy detection model that describes each
machining. In order to find significant and valuable items
step from data preprocessing to evaluation. The framework
in data from various fields, machine learning techniques
provides a blueprint for researchers.
have been used very often. be essential in solving difficult
problems in healthcare [17, 19–25].
The Publicly Available Datasets section provides detailed
information about data used in epilepsy research. It
In engineering, machine learning has been widely applied
includes the description of these files, their characteristics,
to brain-related data for tasks such as seizure localization,
and their importance in machine learning design.
seizure dеtесtion, diffееrіlеrsy lateralization seіzurе [26-
29]. A number of machine learning classifiers such as
“Diabetes Disease Research Based on Advanced Models
Random Forests, Artificial Neural Networks (ANN),
and Machine Learning Classifiers” provides a
Surort Vеstor Mасhinеs (SVM), and possibilities of use, hd
comprehensive literature review focusing on multi-use
applications [26, 28].
 where:
Several reviews have previously gone through classifiers, xi=signal amplitude at i
extended features and labeled seizure levels, 3–37, 30. N = total number of samples
However, the difficulties that data scientists face when
working with data related to neurological disorders have Variance (σ2𝝈𝟐):
been largely overlooked in these cases. The current expert Indicates the spread of the signal values around the mean.
fills this knowledge gap by offering a thorough analysis of
machine learning methods in еpileptic sеizurе . 𝑁
1
𝜎 = ∑(𝑥𝑖 − 𝜇)2
2
The participants in this review were carefully selected from 𝑁
𝑖=1
rеsресtаblе journals listed in the Wеb of Sсiеnсе (WOS)
and SCOPUS. High security conference attendees were Entropy (H):
also brought into the discussion. ехtensive ехamination of Measures the randomness in the signal using Shannon's
various functions аnd classifiers аррliеd to EEG dаtаsеts entropy.
for seizure detection is рrovidеd thе lеrаture,3.53 сurrеntly 𝑁
in. The implementation of classifiers and extraction 𝐻 = − ∑ 𝑝𝑖 log 2 𝑝𝑖
procedures remains very difficult.
𝑖=1

In recent years, a growing number of реорlе аrе interested where:

in using machine learning classifiers to gather valuable pi = probability of occurrence of the i th amplitude.
information from EEG signals. These techniques have
shown potential in identifying seizure sites in the brain, Correlation Coefficient (rxy𝒓𝒙𝒚):
detecting seizures, and assisting with associated knowledge Measures the linear relationship between two EEG
[28,37] channels.
𝐶𝑜𝑣(𝑥, 𝑦)
Jеаn Gotmаn [6, 38–40] is notable for being the first to 𝑟𝑥𝑦 =
develop models using EEG signals for automatic seizure 𝜎𝑥 𝜎𝑦
detection thirty years ago. In his work, he used statistical
and computational methods to achieve certain results. Since where:
then, much has changed in the field as sophisticated science Cov(x,y) = covariance between channels x and y
and signal techniques are used to improve outcomes [34,4– σx,σy = standard deviations of x and y
7]
Referen Feature Classifi Sensitiv Specific
Remarks
IV. SURVEY OF FEATURES AND FEATURE EXTRACTION ce s Used er ity ity
ALGORITHMS IN EEG SIGNAL PROCESSING Achieved
Bayesia
Saab & Mean, short
n
Gotman variance, 76% 64% detection
Classifi
A. Time Domain Features: (2005) entropy delay
er
Time domain features are important for analyzing EEG (9.8s).
signals because they represent temporal changes in signal Incorporat
amplitude. These features are well calculated and capture ed
Correlati
important features related to epilepsy. The extracted Kuhlma additional
on Decisio
features are basic features including measures such as mean nn et al. 81% 70% features to
coefficie n Trees
value, variability, entropy, and correlation coefficient. (2009) improve
nts
detection
1. Time Domain Feature Description accuracy.
Time-domain features involve statistical analysis of raw Reduced
EEG signals over specific time windows. These features channel
are calculated from the amplitude values of EEG signals Variance
Rando count for
and include: Hills ,
m 91.33% 94.02% computati
Mean(μ): (2014) correlati
Forest onal
Represents the average amplitude of the signal over a given on
efficiency.
period.
𝑁
1
µ = ∑ 𝑥𝑖 Time domain features are very important in the analysis of
𝑁
𝑖=1 EEG signals because they represent the time changes in
signal amplitude. These features were carefully calculated
to capture the essential features associated with epilepsy.
The extracted features are the main features and include
measures such as mean, variance, entropy, and correlation networks (ANN) are used for task classification. This mod
coefficient. el shows improvement in detection accuracy but requires
more field data for training.
B. Time-Frequency Domain Features: Patient-
specific solutions: A recent trend proposes personalized pat
D. Zabish et al(2013) reported a patientspecific epilepsy st terns learned from the patient's profile. This model takes in
udy using longterm EEG data in patients with refractory ep to account differences in seizure incidence but requires sig
ilepsy. This study highlights the limitations of universal epi nificant services.
lepsy detection systems, which often fail to adapt to the spe
cific characteristics of patients. The authors focused on dev The differences mentioned by Zabihi et al. (2013).
eloping a method that uses signal processing and machine l
earning techniques to detect epilepsy patients with high an The recognition of EEG symptoms and seizure characterist
d low accuracy using the highest signal and machine learni ics in children that differ from adults makes this work even
ng methods to improve the system. more important to the field.

Core services include:

STFT:
Specialty medicine: Custom-
designed trials including patient-specific seizure patterns. 𝑋(𝑡, 𝑓) = ∫𝑥(𝜏)𝑤(𝜏 − 𝑡)𝑒 − 𝑗2𝜋𝑓𝜏𝑑𝜏

Rovust feature extraction: Use of features in EEG signals t Variance:

𝑁
o improve epilepsy diagnosis.
𝜎2 = 𝑁1𝑖 = 1 ∑ (𝑥𝑖 − 𝜇)2
Machine Learning Model: Based on supervised learning to
distinguish epileptic from non-epileptic conditions. Entropy:

Pediatric Data Validation: Evaluation of a method for EEG 𝐻(𝑋) = −𝑖 ∑ 𝑝(𝑥𝑖) log 𝑝 (𝑥𝑖)
recording in children, focusing on unsuspected diabetes, a
nd demonstration of clinical validity of this method. Power spectral density:
Background and History
𝑃(𝑓) =∣ 𝑋(𝑓) ∣2
Epilepsy diagnosis is an important area of
research in neuroscience and biomedical engineering. Lon
g- Hyperplane eqution:
term EEG monitoring is the gold standard for epilepsy diag 𝑓(𝑥) = 𝑤𝑇𝑥 + 𝑏
nosis and treatment planning. However, manual analysis of
EEG data is time- Activation function:
consuming and requires the use of advanced systems for an 𝑎 = 1 + 𝑒 − 𝑧1
alysis.

Advanced Studies C. Discrete Wavelet Transform Features:

Automatic EEG analysis can support neurologists in diagn
Universal Epilepsy Detection System: Early research has f osing epilepsy by processing largescale data. The system i
ocused on developing a universal standard for diagnosing e dentifies each EEG time point itself using multistage discre
pilepsy in different patient populations. Although these tec te wavelet transform (DWT) for feature extraction, princip
hniques have been successful, they often yield poor results al component analysis (PCA) for noise reduction, and supp
when applied to patient EEG data. ort vector machine (SVM) for classification. The self-
improvement system allows doctors to correct any classific
Disadvantages: No change in the patient's risk of developi ation errors, allowing the system to adapt over time and inc
ng epilepsy. rease its accuracy. This approach aims to reduce dependen
cy on professionals and increase diagnostic accuracy. Preli
Feature- minary testing of data from 21 patients at the Punjab Healt
based approach: Researchers have investigated EEG- h Department (PIMH) and 21 patients at the Boston Childr
based features such as frequency, amplitude, and wavelet c en’s Hospital (CHB) showed robust, diverse systems acros
oefficients to identify seizures. Custom options are also pla s different phones and models, proving their effectiveness
nned to improve the performance of the computation. and adaptability to real-world situations.

Challenge: It is difficult to determine the best features for e Features

ach patient.
Epoch Size:
Machine learning for epilepsy detection: Trace learning su Eps are sequences of EEG signals that are processed separ
ch as support vector machines (SVM) and artificial neural ately.According to the findings of Seng et al., the best dura
tion was determined as 1 second. Summary-Box Window: Provides an overview of the
entire EEG signal of the selected channel.
Discrete Wavelet Transform (DWT):
Corrective Markings

Perform multilayer DWT using Daubechies • Introduces a user-adaptive mechanism to improve

4 (db4) as the mother wavelet. classification performance over time.
Adjust the level of expansion coefficients according to the • Users can select incorrect classifications and log
sample frequency, keep the coefficients in the range of 0- details (e.g., user name, channel, epileptic pattern,
30 Hz. epochs) for retraining the classifier.
• Allows collaboration among neurologists by
Statistical features: comparing markings across different users.
Retraining -The retraining process incorporates
Extract features such as mean, standard deviation, and pow previously logged and new training examples to enhance
er from the selected DWT coefficients. classifier performance.
Use Zscore normalization to normalize the features.
Data Management
Principal Component Analysis (PCA):
1. PCA for Reducing Training Examples: - PCA was
applied to retain training examples that defined 98% of the
PCA is used to reduce redundancy and noise in the features variance.
.
The custom set was reduced from 21 items to 9 items, pres 2. Training Data Storage: -After every retraining, PCA is
erving 93% of all variation. applied to stored data to retain components covering 90% of
the mixture model while discarding the rest to manage
Classifier storage limitation.

Support Vector Machine (SVM) classifier is used for initia

l training in rows and with a box limit of 50 samples, whic EEG Databases
h makes all classifiers equal to the number of channels mul
CHB-MIT: 916 hours of EEG data from 24 subjects with
tiplied by 10 types of epilepsy models.
intractable seizures. 23 channels, sampled at 256 Hz, 16-bit
resolution.
PIMH: 33-channel EEG, sampled at 500 Hz.
Features
Built 220 classifiers for CHB-MIT and 330 classifiers for
PIMH.
Focused on 0.3-30 Hz frequencies, using DWT
decomposition up to the 7th level to identify Beta, Alpha,
Theta, and Delta patterns.
SVM Classifier /VM Classifier
Used 10-fold cross-validation with 8736 epochs.
Best SVM kernel (box constraint 50) achieved 94.8%
accuracy, 95.7% specificity, and 91.7% sensitivity.
Independent channel processing improved accuracy by
4%.
Retraining mechanism adapted classifiers, improving
accuracy from 94.8% to 95.12% after correcting 269
epochs.
Classification-Box Window: Used 10-fold cross-validation with 8736 epochs. Best SVM
kernel (box constraint 50) achieved 94.8% accuracy, 95.7%
• Allows users to view EEG data classified as specificity, and 91.7% sensitivity. Independent channel
epileptic or non-epileptic. processing improved accuracy by 4%. Retraining
• Features zoom and pan options for detailed signal mechanism adapted classifiers, improving accuracy from
analysis. 94.8% to 95.12% after correcting 269 epochs.

Zoom-Box Window: Enables comparison of different

parts of the data using adjustable sliders or manual input for D. Frequency Domain Features
axis values.
Overview of EEG Signal Analysis:
Electroencephalograms (EEGs) are critical for monitoring These metrics highlight the enhancement in seizure
brain electrical activity. This document explores methods detection accuracy compared to established traditional
for detecting epileptic seizures through advanced analysis methods, advocating for the integration of this hybrid model
techniques, specifically Fast Fourier Transform (FFT) and in clinical tools for better epilepsy management.
Discrete Wavelet Transform (DWT), complemented by
statistical methods like Information Gain (InfoGain) for
enhanced feature selection.
V. MACHINE LEARNING CLASSIFIERS FOR
Fast Fourier Transform (FFT) in EEG Analysis: SEIZURE DETECTION
FFT allows for decomposition of signals into their
frequency components, facilitating the examination of Black-box classifier: Definition: Black-box classifier,
specific brain wave patterns associated with different including support vector machine (SVM), artificial neural
mental states. The formula for FFT is: network (ANN), and k nearest neighbor (KNN), are very
good, but no translation.
[X(f) = \sum_{n=0}^{N-1} x(n) e^{-j 2 \pi \frac{fn}{N}}]
where (X(f)) is the output spectrum, (x(n)) is the input Application: Widely used to identify epileptic events
signal, and (N) is the number of samples. This (ictal, interictal) and to distinguish between epileptic and
transformation is useful for identifying frequencies non-epileptic data.
characteristic of cognitive states, exemplified by analyzing
theta waves (4-8 Hz) and alpha waves (8-12 Hz) during Limitations: Failure to provide the reasons or logic
calm and alert conditions. behind the predictions makes them unsuitable for
cognitive exploration.
Importance of Information Gain in Feature Selection: Non-black-box classifiers
InfoGain measures the effectiveness of different features in
predicting outcomes, which significantly enhances model Definition: Non-black-box classifiers, such as decision
accuracy. The formula for InfoGain is: trees and decision forests (e.g., random forests), generate
[IG(T, A) = H(T) - \sum_{v \in Values(A)} interpretable and human-readable code.
\frac{|T_v|}{|T|} H(T_v)]
Application: Widely and locally used in the diagnosis of
Where the entropy (H(T)) is calculated as: Diabetes Mellitus. This process provides greater accuracy
[H(T) = -\sum_{c} p(c) \log_2 p(c)] and allows researchers to understand the decision-making
process.
This process prioritizes the most informative attributes,
facilitating the identification of relevant EEG features that Advantages: Performs well on large data sets and
enhance the model's ability to differentiate between normal generates multiple decisions to achieve better classification
and seizure states.
and prediction.

Application of Discrete Wavelet Transform (DWT): Statistical features in epilepsy diagnosis

DWT offers time-frequency analysis, essential for non- Time domain features: including mean and variance,
stationary signals like EEG. It can be defined skewness, kurtosis and line length. The "Length" feature in
mathematically as: particular is often found to be very effective.
[W(a, b) = \frac{1}{\sqrt{a}} \int_{-\infty}^{\infty}
x(t) \psi\left(\frac{t-b}{a}\right) dt] Frequency domain features: Features such as spectral
power and entropy help determine the frequency of EEG
In this equation, (W(a, b)) provides the wavelet coefficients, data.
where (\psi) is a wavelet function (e.g., Morlet or
Daubechies). DWT excels in pinpointing sudden changes in
signal amplitude, crucial for detecting seizures which may Time-Frequency Features: Provides time and frequency
not be evident in the frequency domain alone. data to capture changing patterns.

Wavelet transform function: discrete and continuous

System Performance and Evaluation Metrics: wavelet transform, extracts local information to provide
better performance.
The effectiveness of the combined FFT, DWT, and Info
Gain methods was evaluated based on specific performance VI. PERFORMANCE ANALYSIS AND EVALUATION
metrics. Sensitivity and specificity metrics achieved were:
Sensitivity: 95%, indicating the model's ability to correctly A. Evaluation Method: Category Accuracy is an
identify true positives (actual seizures). important measure for epilepsy diagnosis and is used to
evaluate the
Specificity: 90%, reflecting the ability to correctly identify
true negatives (non-seizure events).
 proportion of each category. the number of events is 4. Future directions:
excluded from all events. In the study of epilepsy, there are
generally two groups: epileptic and non-epileptic. Develop hybrid classifiers that combine black-box and
Fact: Most studies have achieved high accuracy with non-black-box models to improve accuracy and
advanced techniques like decision forests and random interpretation. Improve model to see faster time to EEG
forest recording without inaccuracy. Use of classifiers to identify
classifiers, reaching almost *100% accurate on some data, epileptogenic regions to aid surgical intervention.
like the BONN dataset.
This performance review highlights advances in epilepsy
Precision (Recall): A measure of the ability to identify the detection while also highlighting areas for further research
nature of seizures. Achievements include: and improvement.

99.38% precision was achieved using SVM with DWT on VII. CONCLUSION
the BONN dataset.
In conclusion, epilepsy presents a significant challenge due
The AUC (area under the curve) of the random forest to the complexity of its detection, driven by the large
classifier on the Kaggle dataset is 96.94%. volume, non-linear, and noisy nature of EEG signals.
Machine learning has proven to be an effective approach
Specificity: Measures how well a model predicts non- for addressing these challenges, offering high accuracy in
epileptic seizures. Specific results often add precision to seizure detection and opportunities for knowledge
performance metrics. discovery. Among the reviewed methodologies, non-
black-box classifiers like decision forests stand out for
B. Evaluation their ability to provide interpretable, explanatory logic rules
alongside high predictive performance. This attribute is
1. Classifier Performance: critical for advancing both accurate seizure detection and
related applications, such as seizure localization and
Black-box classifiers: SVM, ANN, and KNN exhibit high classification of seizure types.
prediction accuracy (i.e., SVM succeeds in 96% of the The study emphasizes the importance of carefully selecting
CHB-MIT dataset). However, their inability to provide relevant statistical features, such as entropy, line length,
detailed explanations limits their usefulness in tasks that and energy, which contribute to classifier performance
require detailed instructions. without introducing unnecessary computational overhead.
Furthermore, balancing dataset dimensions and focusing on
Non-black-box classifiers: Random Forests and Decision interpretable patterns ensure both computational efficiency
Forests are superior in terms of performance, achieve high and practical applicability.
accuracy and provide easily interpretable rules. Random Future research should address challenges like dataset
forests are particularly popular due to the combination of imbalance, efficient feature selection, and real-time seizure
features that provide greater accuracy and robustness. detection on extended EEG recordings. These efforts will
contribute to more robust, clinically viable solutions,
2. Important feature: enhancing the quality of life for individuals with epilepsy.

Time-recording features such as “Line Length” have VIII. REFERENCES

proven to be very effective and very useful for accurate
classification. Various frequency domain features such as • Ahmad, I., Mumtaz, R., Rashid, M. M., &
spectral power and entropy capture specific seizure Khoso, M. A. (2018). A hybrid approach for
patterns. epileptic seizure detection using DWT and SVM.
• Zabihi, M., Asl, B. M., Mohammadi, M. R., &
Wavelet-based features (e.g. DWT and CWT coefficients) Mohseni, H. (2018). Patient-specific epileptic
improve epilepsy diagnosis by representing local signal seizure detection using time-frequency features
changes. and SVM classifier.
• Truong, N. D., Kuhlmann, L., Bonyadi, M. R.,
Challenges 3: Yang, J., Ippolito, S., & Kavehei, O. (2018).
Toward seizure detection using EEG signals: A
Insufficient class: The duration of seizures is shorter than novel channel selection method. IEEE
the duration of seizures, resulting in conflicting data. Transactions on Biomedical Engineering, 65(7),
Techniques such as weighted pricing and SMOTE have • Al Ghayab, H. R., Subasi, A., & Abdulwahab, A.
been partially successful in solving this problem. (2019). Detection of epileptic seizures using LS-
SVM and entropy feature extraction from EEG
Scalability: Despite the efficiency of the data measurement signals. Journal of King Saud University -
system, the application faces problems such as Computer and Information Sciences, 31(4), 496–
computational efficiency and new transfer to the data. 506.
• Lahmiri, S. (2018). An accurate system to
distinguish between normal and abnormal
electroencephalogram records with epileptic
seizure-free intervals. Biomedical Signal
Processing and Control, 40, 312–317.
• Shoeb, A., & Guttag, J. (2010). Application of
machine learning to epileptic seizure detection. In
Proceedings of the 27th International Conference
on Machine Learning (ICML).
• 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(05), 1550023.
• Tzallas, A. T., Tsipouras, M. G., & Fotiadis, D. I.
(2007). Automatic seizure detection based on
time-frequency analysis and artificial neural
networks. Computational Intelligence and
Neuroscience, 2007, 80510.
• Polat, K., & Güneş, S. (2007). Classification of
epileptiform EEG using a hybrid system based on
decision tree classifier and fast Fourier transform.
Applied Mathematics and Computation, 187(2),
1017–1026.