Epilepsy Diagnosis Using EEG Image Analysis

Seeba Doddmani, Sana Mulla, Dilipsingh Rajpurohit, Rajashri Khanai and Prema T. Akkasaligar

Department of Computer Science and Engineering, KLE Technological University, Belagavi, India

Keywords:

Epilepsy Detection, EEG Image Analysis, Machine Learning (ML), Convolutional Neural Network (CNN),

Scalogram Images, VGG-16 Model, Deep Learning, Feature Extraction, Transfer Learning, Bern-Barcelona

EEG Dataset, Seizure Classiﬁcation, Signal Processing, Time-Frequency Analysis, Neural Networks,

Supervised Learning, Data Preprocessing, Binary Classiﬁcation, Medical Diagnosis, Graph Spectral Features,

Long Short-Term Memory (LSTM), Real-Time Detection.

Abstract:

Epilepsy is a neurological disorder that affects millions of people around the world and is characterized by

recurrent seizures caused by abnormal brain activity. Electroencephalograms (EEG) are the primary diagnostic

tool, but traditional manual analysis is time-intensive and prone to errors. This project leverages machine

learning techniques to automate epilepsy detection using scalogram images generated from EEG signals. A

custom Convolutional Neural Network (CNN) model was developed and trained on the Bern-Barcelona EEG

dataset, achieving a training precision of 74.07% and a testing precision of 73.22%. The model demonstrates

good training performance and testing accuracy. An implemented VGG-16 gave a training accuracy of 81.13%

and a testing accuracy of 80.04%. This study aims to help clinicians improve diagnostic accuracy and provide

a scalable, real-time solution for epilepsy detection, particularly in underserved regions.

1 INTRODUCTION

Epilepsy, a chronic neurological disorder, affects

more than 50 million people worldwide, causing sig-

niﬁcant health and social challenges. Diagnosis of

epilepsy traditionally relies on manual analysis of

EEG recordings, a process that is not only time-

consuming but also susceptible to errors due to the

complexity of EEG waveforms. With advances in ma-

chine learning, there is growing interest in automating

this process to improve diagnostic accuracy and efﬁ-

ciency. This project focuses on converting EEG sig-

nals into scalogram images, which capture both tem-

poral and frequency domain features, and applying a

custom CNN model for seizure detection. The pro-

posed system aims to address the limitations of man-

ual analysis, offering a reliable and scalable solution

for clinicians, particularly in resource-limited areas

where access to specialized neurologists is scarce.

The high prevalence of epilepsy, coupled with

the challenges of timely and accurate diagnosis, mo-

tivates the need for automation in the detection of

epilepsy. Automated systems can signiﬁcantly reduce

the burden on neurologists, improve diagnostic ac-

curacy, and ensure better access to healthcare in un-

derserved areas. Using machine learning, these sys-

tems can offer reliable and real-time diagnoses, facil-

itating earlier intervention and better management of

epilepsy.

Several approaches have been explored for

epilepsy detection, combining traditional machine

learning and deep learning techniques: Krishnasamy

et al. proposed supervised learning algorithms, in-

cluding Support Vector Machines (SVM) and CNNs,

achieving accuracy rates up to 99.7% but requiring

large datasets and computational resources. Pattnaik

et al. utilized scalogram images and transfer learn-

ing with ResNet50, achieving a classiﬁcation accu-

racy of 95.23% with high sensitivity and speciﬁcity.

Sesha Sai et al. combined CNNs with SVMs for auto-

matic feature extraction, achieving 94.48% accuracy

but heavily dependent on data quality. Other stud-

ies, such as by Wang et al., explored hybrid CNN-

LSTM (Long Short Term Memory) models for cap-

turing both spatial and temporal features in EEG data.

Despite their promising results, these approaches

face several limitations: Heavy reliance on large,

high-quality datasets, making them less effective in

diverse clinical settings. Computational complex-

ity, particularly for deep architectures like Residual

776

Doddmani, S., Mulla, S., Rajpurohit, D., Khanai, R. and T. Akkasaligar, P.

Epilepsy Diagnosis Using EEG Image Analysis.

DOI: 10.5220/0013602300004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 776-783

ISBN: 978-989-758-763-4

Network (ResNet) and hybrid models, which hin-

ders real-time application. Limited generalization to

unseen data, with signiﬁcant performance degrada-

tion when applied to different patient demographics

or new datasets. Challenges in adapting models to

resource-limited environments due to high hardware

requirements.

Epilepsy detection faces several challenges: High

variability in seizure patterns across patients compli-

cates the creation of a universally robust model. EEG

data often contains noise and artifacts, requiring so-

phisticated pre-processing. Balancing model com-

plexity with computational efﬁciency is critical for

real-time applications. Addressing class imbalance in

datasets is essential to improve model generalization.

To address these challenges, this project proposes

a machine learning pipeline that converts EEG signals

into scalogram images to capture time-frequency do-

main features. A custom CNN is designed to classify

these images into seizure and non-seizure categories.

The pipeline incorporates preprocessing techniques

like normalization and leverages training strategies to

improve model generalization.

Developed a novel pipeline to transform EEG sig-

nals into scalogram images for time-frequency anal-

ysis. Designed and implemented a custom CNN ar-

chitecture optimized for binary classiﬁcation. Evalu-

ated model performance on the Bern-Barcelona EEG

dataset, achieving a training accuracy of 97.77%.

Identiﬁed challenges in generalization through testing

accuracy analysis, providing insights for future im-

provement. Highlighted the potential for real-time de-

ployment of the proposed system in resource-limited

settings.

The rest of paper is organized as follows. Sec-

tion II provides a brief review of the recent works.

Section III details the problem statement, background

and the proposed methodology. Section IV describes

the experiment details along with the results and dis-

cussions. Finally, the paper concludes in Section V.

2 LITERATURE SURVEY

The methodology for identifying and selecting stud-

ies for the literature survey involved several stages

to ensure relevance and quality. Initially, 42 papers

were identiﬁed from reputable sources such as IEEE

Conference, Springer, Frontiers, and BioMed Cen-

tral(BMC). During the screening phase, 10 papers

were excluded—6 for being older than ﬁve years and

4 for having incorrect titles. Figure. 1 shows the

ﬂowchart illustrating this process. After screening, 32

papers were assessed for eligibility, and 16 were fur-

ther excluded—3 were review papers, and 13 focused

on signals rather than images. Ultimately, 16 stud-

ies were included in the literature survey, providing a

robust foundation for the research.

Figure 1: Flowchart of literature survey paper selection

(Krishnasamy et al., 2024) explored supervised

learning algorithms for epileptic seizure detection,

achieving high accuracy rates (98.5% with SVM and

99.7% with fuzzy classiﬁers). Their system auto-

mates seizure detection effectively, reducing com-

plexity for experts, but struggles with real-time de-

ployment due to computational demands and limited

generalizability across datasets. Similarly, (Hafeez

and Shakil, 2024) utilized EEG-based brainwave im-

ages to classify stress levels using LSTM (70.67%

accuracy) and CNN (90.46% accuracy), demonstrat-

ing CNN’s capability for spatial feature extraction but

facing challenges with noise in EEG data and small

participant pools. (Pattnaik et al., 2024) achieved

95.23% classiﬁcation accuracy using scalogram im-

ages analyzed with a pre-trained ResNet50 model,

showcasing the utility of transfer learning for EEG

signal analysis.

(Sadam and Nalini, 2024) combined CNN and

SVM to achieve 94.48% accuracy, highlighting the

advantages of hybrid models for automated feature

Epilepsy Diagnosis Using EEG Image Analysis

777

extraction but showing sensitivity to noisy datasets.

(Georgis-Yap et al., 2024) compared supervised and

unsupervised approaches like CNN, CNN-LSTM,

and TCN, ﬁnding comparable results for patient-

speciﬁc seizure prediction. While their methods re-

duce preprocessing needs, they require substantial

preictal data for optimal performance. (Krishnan

et al., 2024) applied GASF to convert EEG signals

into image representations, achieving up to 96% ac-

curacy. However, this approach is computationally

intensive, especially during image transformation and

feature extraction.

(Shankar et al., 2023) utilized CNNs with phase

synchronization matrices to classify seizures with

83.3% accuracy, demonstrating effective spatial fea-

ture extraction. However, intensive preprocessing and

the need for large datasets remain barriers to real-

time applications. (Khasawneh et al., 2022) achieved

99.8% precision using Faster R-CNN and transfer

learning for K-complex detection, with adaptability

across datasets. Nonetheless, overlapping image gen-

eration methods may introduce data leakage, compro-

mising model reliability. (Hu et al., 2020) proposed

a hierarchical neural network (HNN) using transfer

learning, attaining 98.97% accuracy on the CHB-

MIT dataset, though the reliance on large pre-trained

DNNs limits real-time applicability.

(Sharma and Meena, 2024) introduced a model

integrating GFT and DWT for feature extraction,

achieving over 98% accuracy across datasets. While

graph spectral features enhance detection accuracy,

the method’s complexity increases computational de-

mand, complicating real-time deployment. (Kunekar

et al., 2024) used LSTM networks to achieve 97% ac-

curacy, effectively handling temporal dependencies in

EEG data. However, dataset class imbalance could

hinder generalizability. (Jridi et al., 2024) employed

deep ResNet for multi-disorder detection, reaching

100% accuracy for epilepsy detection on the UBonn

dataset, but high computational requirements pose

challenges for implementation in resource-limited en-

vironments.

(Saleem et al., 2023) combined CNN with tra-

ditional classiﬁers, achieving 98.49% accuracy in

seizure detection. The hybrid model effectively iden-

tiﬁes subtle EEG patterns but requires larger, more

diverse datasets to ensure generalizability. (Majzoub

et al., 2023) used AlexNet for multi-channel EEG sig-

nal classiﬁcation, achieving 98.25% accuracy in bi-

nary classiﬁcation. However, its accuracy drops to

92.98% with new patient data, highlighting the need

for diverse training samples. (Wang et al., 2023) de-

veloped a CNN-LSTM hybrid model, achieving 98%

accuracy in ternary classiﬁcation and 100% in binary

classiﬁcation. This model excels in capturing both

spatial and temporal features but demands high com-

putational resources.

Finally, (Supriya et al., 2021) adopted graph-

theory-based methods, such as VG and HVG, for fea-

ture extraction, achieving accuracies above 95%.

These methods are advantageous for their process-

ing speed and classiﬁcation efﬁcacy but are limited by

their sensitivity to dataset size and threshold depen-

dencies. Collectively, while these studies demonstrate

signiﬁcant advancements in EEG-based seizure de-

tection and classiﬁcation, the major limitations across

models include dependency on large datasets, compu-

tational demands, and challenges with generalization

across diverse patient populations.

This literature survey reviews advancements in

EEG-based epileptic seizure detection using super-

vised learning, deep learning, and hybrid models.

Techniques like CNN-SVM hybrids (94.48% accu-

racy), transfer learning with scalograms (95.23%),

and GASF image transformations (96%) demonstrate

high accuracy but face challenges with computational

demands and noisy data. Deep learning models like

CNNs and hybrid CNN-LSTM architectures achieve

exceptional spatial-temporal feature extraction but re-

quire large datasets. Despite progress, issues such

as dataset dependency, generalizability, and real-time

applicability remain signiﬁcant hurdles.

3 PROBLEM STATEMENT AND

SYSTEM MODEL

Epilepsy is a chronic neurological disorder marked by

recurrent seizures resulting from abnormal electrical

activity in the brain. Diagnosing epilepsy relies heav-

ily on EEGs, which capture electrical patterns and

help identify abnormalities. Traditional approaches

involve manual inspection, a labor-intensive task

requiring specialized expertise.

EEG, or electroencephalography, plays a signiﬁ-

cant role in medical diagnosis as a non-invasive and

cost-effective method for recording brain activity. It

is particularly useful in identifying abnormal brain

wave patterns, which can indicate potential seizures

or other neurological disorders. By capturing electri-

cal activity in the brain, EEG provides valuable in-

sights into the functional state of the brain, aiding

clinicians in the diagnosis and management of vari-

ous conditions.

Traditional methods of diagnosing epilepsy

present several challenges. The manual analysis of

long EEG recordings is a time-consuming process, re-

INCOFT 2025 - International Conference on Futuristic Technology

778

quiring signiﬁcant effort from trained specialists. Fur-

thermore, the variability in seizure patterns across in-

dividuals adds complexity to achieving a consistent

and accurate diagnosis. These issues are compounded

by the limited availability of trained neurologists, par-

ticularly in rural or underserved areas, making timely

and effective diagnosis even more difﬁcult.

Epilepsy is a neurological disorder that impacts

over 50 million people worldwide, characterized by

abnormal brain activity leading to seizures. Tradi-

tional epilepsy diagnosis using EEGs involves man-

ually analyzing extensive recordings to detect epilep-

tic events, a process prone to errors and inefﬁcien-

cies. This project aims to automate the detection of

epilepsy by employing Machine Learning (ML) tech-

niques to classify epileptic and non-epileptic events

using 2D scalogram images derived from EEG sig-

nals. The goal is to create a robust, accurate system

that can enhance diagnostic efﬁciency, particularly in

resource-limited areas.

The primary objective of this work is to develop

a machine learning pipeline capable of automati-

cally classifying EEG-derived scalogram images as

”seizure” or ”non-seizure.” This automated approach

aims to improve diagnostic accuracy by increasing

sensitivity and speciﬁcity, thereby assisting clinicians

in making reliable diagnoses. Additionally, the solu-

tion is designed to be deployable in real-time systems,

enabling rapid analysis and decision-making in clini-

cal settings. To ensure versatility, the proposed solu-

tion is scalable and adaptable to diverse datasets and

hardware setups, including mobile and low-power de-

vices.

Figure.2. shows the system model that outlines

a process for classifying EEG signals, into two cate-

gories: seizure (focal) and non-seizure. The process

begins with signal input, followed by data preprocess-

ing to transform the signal into a scalogram image,

which visually represents the frequency content over

time. The dataset is then split into training and test-

ing subsets. Two models are employed for classiﬁca-

tion: a custom CNN and a pre-trained VGG16 model.

Both models undergo training and validation to learn

patterns from the data. Finally, model evaluation is

conducted to assess performance, leading to the clas-

siﬁcation of the input signal as either seizure (focal)

or non-seizure.

4 PROPOSED METHODOLOGY

The proposed work is executed on device is powered

by an AMD Ryzen 7 4800H with Radeon Graphics

processor, operating at 2.90 GHz, and is equipped

Figure 2: System model for EEG detection

with 16 GB of RAM. It runs a 64-bit version of

Windows 11 Home and the NVIDIA GTX 3050

GPU. For implementation, we have used several

Python libraries are used for efﬁcient data analysis

and visualization. NumPy supported numerical

computations, Pandas handled data manipulation.

Matplotlib are employed for creating visualizations,

tensorﬂow used for building the CNN Architecture.

The dataset originally consisted of EEG signals

represented by both X and Y components. These sig-

nals were processed and converted into scalogram im-

ages, with separate scalograms generated for the X

and Y components. Following this transformation,

normalization was applied to the images to standard-

ize pixel intensity values, ensuring they fell within a

consistent range to enhance model training.

4.1 Custom CNN

CNNs are a class of deep learning models speciﬁcally

designed for processing structured data like images.

A CNN consists of multiple layers, including convo-

lutional layers that extract spatial and hierarchical fea-

Epilepsy Diagnosis Using EEG Image Analysis

779

Figure 3: Focal Scalagram image (Seizure)

Figure 4: Non-Focal Scalagram image (Non-Seizure)

tures, pooling layers that reduce dimensionality, and

fully connected layers for ﬁnal classiﬁcation. The

training process involves optimizing a loss function,

typically using stochastic gradient descent and back-

propagation, to adjust weights and minimize classiﬁ-

cation errors. Regularization techniques like dropout

are often applied to prevent overﬁtting. CNNs are

highly effective in capturing spatial patterns, making

them well-suited for tasks such as image recognition,

object detection, and, in this case, epilepsy detection

from EEG-based scalogram images.

The custom CNN for a multi-class classiﬁcation

problem consists of sequential convolutional blocks

with increasing ﬁlters (32, 64, 128, and optionally

256), each containing a Conv2D layer with ReLU ac-

tivation, batch normalization for stability, and Max-

Pooling2D for down-sampling. To prevent overﬁt-

ting, L2 regularization is applied to both the convo-

lutional and dense layers, and dropout is used in the

dense layers for further regularization. The fully con-

nected layers include a ﬂatten layer to convert fea-

ture maps into a 1D vector, followed by dense lay-

ers with ReLU activation to learn complex represen-

tations. The model uses the Adam optimizer, an adap-

tive method suitable for deep learning tasks.

After training the CNN model for 50 epochs with

early stopping at 11, it achieved a training accuracy of

74.07%, indicating that the model effectively learned

patterns from the training data and the testing accu-

racy is 73.22%.

Figure.5. This graph depicts the training and test-

ing accuracy of a model over several epochs (labeled

on the x-axis). The training accuracy (blue line)

starts higher and maintains relatively stable values

with slight ﬂuctuations, while the testing accuracy

(orange line) begins lower but rises sharply, eventu-

Figure 5: Training vs testing accuracy for custom CNN

Figure 6: Training and testing error vs Epoch for Custom

CNN

ally ﬂuctuating and surpassing the training accuracy

in later epochs. This pattern may suggest overﬁtting

at the ﬁnal epochs, as the testing accuracy’s instabil-

ity could indicate sensitivity to the validation set. The

testing accuracy surpassing training at times suggests

randomness in the dataset split. However, the con-

vergence of both accuracies around epoch 4 indicates

balanced model performance during this phase.

Figure.6. The graph shows the training and test-

ing loss of a machine learning model over six epochs.

The training loss (blue line) starts at a moderate level,

gradually decreasing as the model learns from the

training data. The testing loss (orange line), which

measures the model’s performance on unseen data,

begins at a higher value but initially drops sharply, in-

dicating improvement in generalization. After a few

epochs, the testing loss increases slightly, where the

model begins to perform better on training data.

4.2 VGG 16

VGG16 is a pre-trained model and is well-suited for

study due to its ability to effectively extract spatial

features from images. EEG signals converted into vi-

INCOFT 2025 - International Conference on Futuristic Technology

780

sual representations, such as scalogram images, con-

tain patterns indicative of epileptic activity. VGG16’s

architecture, with its small 3x3 convolutional ﬁlters,

is adept at capturing these ﬁne-grained spatial de-

tails. Additionally, using a pre-trained VGG16 model

through transfer learning allows leveraging its general

feature extraction capabilities while ﬁne-tuning it for

the speciﬁc task of epilepsy detection. This approach

is particularly advantageous when working with lim-

ited data, as it reduces the need for extensive training

from scratch and ensures robust performance in clas-

sifying complex patterns in EEG images.

Learning Rate: 0.001, adjusted dynamically using

schedulers. Batch Size: 32 for balanced efﬁciency

and performance. Epochs: 25 with early stopping

to prevent overﬁtting. Optimizer: Adam for faster

convergence and adaptive learning. Image Resolu-

tion: 128x128 for consistent input size. Dropout layer

to prevent overﬁtting and also applied the L2 regu-

larization Loss Function : Binary cross entropy for

binary classiﬁcation Activation Function : We used

both Relu and Sigmoid

Figure 7: Training vs testing accuracy for VGG-16

Figure.7. illustrates the training and validation ac-

curacy over epochs. Initially, both training and val-

idation accuracy increase, indicating that the model

is learning effectively. Around epoch 10, the vali-

dation accuracy starts to ﬂuctuate, showing signs of

overﬁtting as the training accuracy continues to im-

prove steadily while the validation accuracy varies.

Despite these ﬂuctuations, the validation accuracy re-

mains relatively close to the training accuracy, sug-

gesting that the model is performing reasonably well.

Figure.8. shows the training and validation loss

over epochs. Both losses decrease steadily, indicating

that the model is learning and improving its predic-

tions. However, the validation loss consistently re-

mains lower than the training loss after a few epochs,

which might suggest differences in how the training

and validation datasets are handled. The smooth de-

cline in loss for both suggests that the training process

is stable, and there are no signiﬁcant issues like over-

Figure 8: Training vs testing loss for VGG-16

Figure 9: The architecture of VGG16

ﬁtting or divergence in this range of epochs.

Figure.9. illustrates the convolutional neural net-

work (CNN) architecture used for binary classiﬁca-

tion of scalogram images into focal and non-focal

categories. The model processes input scalograms

through a series of convolutional layers, each extract-

ing increasingly complex features, with max-pooling

layers reducing the spatial dimensions. The extracted

features are then ﬂattened and passed through fully

connected layers, culminating in a sigmoid-activated

output layer for binary classiﬁcation. This archi-

tecture is adapted from the sea ice classiﬁcation of

SAR imagery presented by Khaleghian et al. (2021)

(Khaleghian et al., 2021).

Figure 10: ROC curve of VGG-16

Epilepsy Diagnosis Using EEG Image Analysis

781

Figure.10. shows the ROC curve demonstrates the

performance of the model in distinguishing between

classes, with an Area Under the Curve (AUC) of

0.87. This indicates that the model has a high ability

to discriminate between focal and non-focal classes,

performing signiﬁcantly better than random guessing

(represented by the diagonal line). The curve’s prox-

imity to the top-left corner suggests a good balance

between the true positive rate (sensitivity) and false

positive rate, making the model reliable for classiﬁca-

tion tasks.

Table 1: Comparison results of custom CNN and VGG16

Parameter VGG16 Custom CNN

Training Accuracy 81.13% 74.07%

Testing Accuracy 80.04% 73.22%

5 CONCLUSIONS

This study highlights the potential of machine learn-

ing in automating epilepsy detection using EEG-

based scalogram images. The proposed custom CNN

model achieved a training accuracy of 74.07% and a

testing accuracy of 73.22%, demonstrating its abil-

ity to learn meaningful patterns from the data. Addi-

tionally, the VGG-16 model outperformed the custom

CNN, achieving a training accuracy of 81.13% and a

testing accuracy of 80.04%.

The CNN model has 9 layers but yet works good

in comparison with the 16 layers of VGG-16. These

results underscore the utility of advanced image-

based analytics in healthcare while also emphasizing

the importance of optimizing models for enhanced

performance and generalization. This project lays

a foundation for developing scalable, real-time sys-

tems for epilepsy diagnosis, with future work focus-

ing on improving model robustness, leveraging larger

and more diverse datasets, and exploring deployment

strategies for real-world applications.

REFERENCES

Georgis-Yap, Z., Popovic, M. R., and Khan, S. S.

(2024). Supervised and unsupervised deep learning

approaches for eeg seizure prediction. Journal of

Healthcare Informatics Research, 8(2):286–312.

Hafeez, M. A. and Shakil, S. (2024). Eeg-based stress iden-

tiﬁcation and classiﬁcation using deep learning. Mul-

timedia Tools and Applications, 83(14):42703–42719.

Hu, D., Cao, J., Lai, X., Wang, Y., Wang, S., and Ding, Y.

(2020). Epileptic state classiﬁcation by fusing hand-

crafted and deep learning eeg features. IEEE Trans-

actions on Circuits and Systems II: Express Briefs,

68(4):1542–1546.

Jridi, A., Djemal, R., and Belwaﬁ, K. (2024). Neurologi-

cal disorder diagnosis through deep residual network-

based eeg signal analysis. In 2024 IEEE 7th Interna-

tional Conference on Advanced Technologies, Signal

and Image Processing (ATSIP), volume 1, pages 144–

149. IEEE.

Khaleghian, S., Ullah, H., Kræmer, T., Hughes, N., Eltoft,

T., and Marinoni, A. (2021). Sea ice classiﬁcation

of sar imagery based on convolution neural networks.

Remote Sensing, 13(9):1734.

Khasawneh, N., Fraiwan, M., and Fraiwan, L. (2022). De-

tection of k-complexes in eeg waveform images using

faster r-cnn and deep transfer learning. BMC Medical

Informatics and Decision Making, 22(1):297.

Krishnan, P. T., Erramchetty, S. K., and Balusa, B. C.

(2024). Advanced framework for epilepsy detection

through image-based eeg signal analysis. Frontiers in

Human Neuroscience, 18:1336157.

Krishnasamy, L., Sriwastav, Y. K., Bharat, S. P., and

Ganachar, S. R. (2024). Epilepsy detection using su-

pervised learning algorithms. In 2024 IEEE Interna-

tional Conference on Contemporary Computing and

Communications (InC4), volume 1, pages 1–6. IEEE.

Kunekar, P., Gupta, M. K., and Gaur, P. (2024). Detection of

epileptic seizure in eeg signals using machine learning

and deep learning techniques. Journal of Engineering

and Applied Science, 71(1):21.

Majzoub, S., Fahmy, A., Sibai, F., Diab, M., and Mahmoud,

S. (2023). Epilepsy detection with multi-channel eeg

signals utilizing alexnet. Circuits, Systems, and Signal

Processing, 42(11):6780–6797.

Pattnaik, S., Rao, B. N., Rout, N. K., and Sabut, S. K.

(2024). Transfer learning based epileptic seizure clas-

siﬁcation using scalogram images of eeg signals. Mul-

timedia Tools and Applications, pages 1–15.

Sadam, S. S. P. and Nalini, N. (2024). Epileptic seizure de-

tection using scalogram-based hybrid cnn model on

eeg signals. Signal, Image and Video Processing,

18(2):1577–1588.

Saleem, A., Khan, M. A., and Yousaf, H. M. (2023). Ad-

vancing epilepsy disease classiﬁcation through ma-

chine learning and deep learning models utilizing eeg

data. In 2023 17th International Conference on Open

Source Systems and Technologies (ICOSST), pages 1–

8. IEEE.

Shankar, A., Chakraborty, D., Saikia, M. J., Dandapat, S.,

and Barma, S. (2023). Seizure type detection using

eeg signals based on phase synchronization and deep

learning. In 2023 IEEE 19th International Conference

on Body Sensor Networks (BSN), pages 1–5. IEEE.

Sharma, R. and Meena, H. K. (2024). Enhanced epilep-

tic seizure detection through graph spectral analysis of

eeg signals. Circuits, Systems, and Signal Processing,

pages 1–21.

Supriya, S., Siuly, S., Wang, H., and Zhang, Y. (2021).

Epilepsy detection from eeg using complex network

techniques: A review. IEEE Reviews in Biomedical

Engineering, 16:292–306.

INCOFT 2025 - International Conference on Futuristic Technology

782

Wang, X., Wang, Y., Liu, D., Wang, Y., and Wang, Z.

(2023). Automated recognition of epilepsy from eeg

signals using a combining space–time algorithm of

cnn-lstm. Scientiﬁc Reports, 13(1):14876.

Epilepsy Diagnosis Using EEG Image Analysis

783