Journal of Robotics and Control (JRC) Volume 6.
Issue 5, 2025 ISSN: 2715-5072.
DOI: 10.
18196/jrc.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Yuri Pamungkas 1*.
Riva Satya Radiansyah 2.
Stralen Pratasik 3.
Made Krisnanda 4.
Natan Derek 5 Department of Medical Technology.
Institut Teknologi Sepuluh Nopember.
Surabaya.
Indonesia Department of Medicine.
Institut Teknologi Sepuluh Nopember.
Surabaya.
Indonesia Department of Informatics Engineering.
Universitas Negeri Manado.
Manado.
Indonesia School of Information and Physical Sciences.
University of Newcastle.
Callaghan.
Australia Department of Neurology and Neurological Sciences.
Stanford School of Medicine.
Palo Alto.
United States Email: 1 yuri@its.
id, 2 riva.
satya@its.
id, 3 stralente@unima.
id, 4 made.
krisnanda@uon.
au, 5 nderek@stanford.
*Corresponding Author AbstractAiIdentifying epileptogenic zones (EZ.
is a crucial step in the pre-surgical evaluation of drug-resistant epilepsy Conventional methods, including EEG/SEEG visual inspection and neurofunctional imaging, often face challenges in accuracy, reproducibility, and subjectivity.
The rapid development of artificial intelligence (AI) technologies in signal processing and neuroscience has enabled their growing use in detecting epileptogenic zones.
This systematic review aims to explore recent developments in AI applications for localizing epileptogenic zones, focusing on algorithm types, dataset characteristics, and performance outcomes.
A comprehensive literature search was conducted in 2025 across databases such as ScienceDirect.
Springer Nature, and Ie Xplore using relevant keyword combinations.
The study selection followed PRISMA guidelines, resulting in 34 scientific articles published between 2020 and 2024.
Extracted data included AI methods, algorithm types, dataset modalities, and performance metrics .
AUC, sensitivity, and F1-scor.
Results showed that deep learning was the most used approach .
%), followed by machine learning .
%), multi-methods .
%), and knowledgebased systems .
%).
CNN and ANN were the most commonly applied algorithms, particularly in scalp EEG and SEEG-based Datasets ranged from public sources (Bonn.
CHB-MIT) to high-resolution clinical SEEG recordings.
Multimodal and hybrid models demonstrated superior performance, with several studies achieving accuracy rates above 98%.
This review confirms that AI .
specially deep learning with SEEG and multimodal integratio.
has strong potential to improve the precision, efficiency, and scalability of EZ detection.
facilitate clinical adoption, future research should focus on standardizing data pipelines, validating AI models in real-world settings, and developing explainable, ethically responsible AI KeywordsAiEpileptogenic Zone.
Artificial Intelligence.
Deep Learning.
Machine Learning.
Stereo-EEG.
INTRODUCTION
Epilepsy is a neurological disorder characterised by a tendency to experience recurrent seizures, caused by abnormal electrical activity in the brain.
One important approach in the management of refractory epilepsy, which is epilepsy that does not respond to pharmacological treatment, is to accurately determine the epileptogenic zone (EZ) through clinical and electrophysiological evaluation .
Accurate identification of the EZ is crucial in determining the success of surgical intervention, which aims to significantly reduce or eliminate seizure frequency.
Currently, approximately 30% of epilepsy patients experience refractory conditions, where the quality of life of patients heavily depends on the accuracy of the medical or surgical interventions applied .
Data from the World Health Organisation (WHO) indicate that over 50 million people worldwide have epilepsy, with approximately 30-40% of them experiencing refractory epilepsy .
This number is projected to continue rising, given the various diagnostic and therapeutic challenges still faced in the field of epilepsy neurology .
, .
Conventional electroencephalography (EEG), magnetic resonance imaging (MRI), positron emission tomography (PET), and single photon emission computed tomography (SPECT) have been widely used in clinical practice to identify EZ.
However, these methods have limitations in terms of inconsistent sensitivity and specificity among patients.
Additionally, interpreting the results of these techniques is often complex and highly dependent on the clinician's experience.
These challenges lead to variations in diagnostic outcomes and may result in suboptimal identification of EZ.
These challenges underscore the need for more advanced and accurate diagnostic approaches.
The development of alternative methods based on cutting-edge technology, particularly AI, is expected to address the limitations of conventional diagnostic methods .
Artificial intelligence (AI) offers potential and transformative solutions to address various challenges in the field of neurology, particularly in identifying epileptogenic zones (EZ) in patients with drug-resistant epilepsy.
AI is broadly defined as a branch of computer science focused on developing algorithms and systems capable of mimicking human intelligent behaviour, including the ability to understand, learn, and make data-driven decisions.
In the medical context.
AI has advantages in handling large and complex datasets, such as electroencephalography (EEG), magnetoencephalography (MEG), and various neuroimaging modalities like MRI and PET scans .
Ae.
Through Journal Web site: http://journal.
id/index.
php/jrc Journal Email: jrc@umy.
Journal of Robotics and Control (JRC) ISSN: 2715-5072 processes such as prediction, classification, segmentation, and feature extraction.
AI algorithms can improve diagnostic accuracy, accelerate clinical decision-making processes, and reduce reliance on subjective interpretations by medical professionals .
Ae.
A range of artificial intelligence (AI) techniques has been applied in epilepsy research, including machine learning (ML), deep learning (DL), expert systems (ES), and integrated multimodal frameworks that combine multiple data types and analytical methods .
Ae.
ML and DL, in particular, have shown strong capabilities in detecting intricate patterns within electrophysiological recordings and neuroimaging data, patterns that are often challenging to discern through manual analysis.
Techniques such as ANN.
SVM.
Decision Trees.
CNN, and ensemble models like random forest and gradient boosting have achieved notable accuracy in localizing epileptic foci .
Ae.
These strengths highlight the potential of AI to enhance and streamline the pre-surgical assessment of epilepsy, offering more precise and efficient diagnostic support.
Furthermore.
AI is not only used for diagnosis but also plays a growing role in guiding therapeutic decisions, such as identifying suitable candidates for surgical resection, neuromodulation, or laser ablation.
Accurate identification of EZs can lower the failure rate of invasive interventions and significantly enhance patientsAo quality of life.
However, despite its promise, the application of AI in this domain still faces several challenges, including variability in data quality, limited interpretability, and inconsistent validation across clinical settings.
Given these gaps, it is essential to conduct a systematic review of existing AI methods used for EZ This includes analyzing the types of algorithms implemented, the nature and sources of datasets, and reported performance metrics.
A clearer understanding of these components will help formulate strategic recommendations for real-world clinical integration and guide future development of more robust, interpretable, and clinically reliable AI technologies for managing refractory epilepsy.
II.
METHODOLOGY
Research Question This study aims to identify the application of artificial intelligence (AI) in determining the epileptogenic zone (EZ).
We reviewed various scientific articles that have reported AI methods and techniques specifically applied to identify the EZ in epilepsy patients.
The population in this study consists of individuals with epilepsy.
The intervention of focus is the use of various AI algorithms and methods to support the accurate identification of EZ.
In this study, no comparative analysis was conducted, as the review's focus was on a comprehensive exploration of available AI methods and The main goal of this review is to deliver an indepth overview of the AI methodologies applied, including machine learning (ML), deep learning (DL), and multimodal strategies, while assessing their performance in localizing the epileptogenic zone (EZ).
Search Strategy This systematic review was conducted in 2025, employing a structured literature search across major scientific databases such as ScienceDirect.
Springer Nature, and Taylor & Francis.
The search strategy utilized combinations of relevant keywords derived from MeSH, as outlined in Table I.
The entire process followed the PRISMA guidelines to ensure methodological rigor.
To minimize selection bias, two independent reviewers performed the screening and selection of articles.
In instances where discrepancies arose between the reviewers, a third independent evaluator was consulted to reach a consensus.
The inclusion criteria were limited to English-language publications from the past five years .
0Ae2.
TABLE I.
SEARCH STRATEGY OF THE RESEARCH
Database Limits Data Search Query Search strategy ScienceDirect.
Springer Nature, and Taylor & Francis .
Inclusion criteria included English-language sources and studies in human populations.
January 1, 2020 to December 31, 2024 ("Epileptogenic Zone") AND ("Detection" OR "Diagnosis") AND ("AI" OR "Artificial Intelligence" OR "Machine Learning" OR "Deep Learning") Inclusion and Exclusion Criteria The inclusion criteria for this review include original research articles, experimental studies, and meta-analysis reports discussing the application of artificial intelligence (AI) for the identification of epileptogenic zones (EZ) in epilepsy patients.
Only articles reporting evaluations of AI model performance, such as accuracy, sensitivity, specificity, precision.
F1 score, or area under the curve (AUC), were included in the analysis.
Selected studies must have used AIbased methods, including machine learning, deep learning, or other computational techniques applied for the classification or prediction of EZ locations.
Exclusion criteria included articles not written in English, articles that did not provide full-text access, and non-original research publications such as narrative reviews, comments, opinions, letters to the editor, brief communications, and conference proceedings Additionally, studies not conducted on human subjects or those that did not present quantitative data on AI model performance were also excluded from this review.
Selection Process The article selection process in this review followed the PRISMA guidelines as shown in Fig.
After screening the titles, abstracts, and full texts, 34 articles were finally selected for further analysis.
The entire selection and quality evaluation process was conducted independently by two researchers to ensure objectivity and avoid selection bias.
there were differences of opinion between the two researchers, the final decision was made through discussion with a third independent reviewer.
For data analysis purposes, each article that met the inclusion criteria was extracted using a standard form covering seven main categories, namely .
author name, .
year of publication, .
artificial intelligence (AI) method applied, .
type of algorithm used, .
type of data used to identify epileptogenic zones, including EEG data.
MRI images.
PET, or multimodal combinations, .
characteristics of the study population or sample, and .
best model performance based on evaluation metrics such as accuracy, sensitivity, specificity.
F1 score.
Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Journal of Robotics and Control (JRC) ISSN: 2715-5072 and AUC.
All successfully extracted data were systematically analysed and synthesised, then presented in the form of tables and graphical visualisations to illustrate the main findings of this review.
RESULTS
Based on the study search terms, 34 articles were reviewed in detail and presented in Table II.
Meanwhile.
Fig.
2 shows the distribution of articles published in the 2020Ae2024 period, related to the topic of epileptogenic zone Based on the data, it is evident that 2021 and 2024 were the years with the highest number of publications, each contributing 26%, or approximately 9 articles per year.
Meanwhile, 2022 and 2023 each contributed 18% .
articles per yea.
, indicating a relatively stable publication trend but not as intense as the two peak years prior.
The year 2020 had the fewest publications, at only 12%, or equivalent to 4 articles, likely influenced by the initial impact of the COVID19 pandemic on clinical research activities.
This trend indicates an increasing interest and need for approaches to identify epileptogenic zones in clinical and research contexts, particularly in recent years .
1 and 2.
, which may be earch y eywor s in cience irect linked to the development of technologies such as SEEG.
HFO analysis, and the application of AI in neurodiagnostics.
Fig.
3 shows the distribution of artificial intelligence (AI) methods used in studies identifying epileptogenic zones.
the total 34 articles described.
Deep Learning was the most dominant method, used in 44% of publications .
This indicates that Deep Learning is increasingly relied upon due to its ability to extract complex patterns from brain signals such as EEG and SEEG.
Additionally, classical Machine Learning was used in 35% of studies .
, indicating that this approach remains relevant, particularly for smaller datasets or those based on manual features.
Multi Methods .
ombining two or more AI technique.
were used in 18% of publications .
, reflecting an integrative trend in epileptogenic research.
Meanwhile.
KnowledgeBased AI, such as expert systems or inference based on brain network theory, was only used in 1 article .
%), indicating that symbolic approaches are increasingly rare in modern AI These findings reflect a shift in research from conventional approaches toward more automated, highprecision, and measurable deep learning to support clinical decision-making in identifying epileptogenic zones.
earch y eywor s in pringer ature otal apers ligi ility apers eeting general criteria recent years, full te t, hu ans, nglish, ournal article, clinical trials an eta analysis inclu e otal apers aper e clu e apers eeting general criteria recent years, full te t, hu ans, nglish, ournal article, clinical trials an eta analysis inclu e aper e clu e apers were relevant y title an a stract to su ect inclu e aper e clu e apers eeting general criteria recent years, full te t, hu ans, nglish, ournal article, clinical trials an eta analysis inclu e aper e clu e apers were relevant y title an a stract to su ect inclu e aper e clu e earch y eywor s in otal apers aper e clu e apers ha perfor ance assess ent criteria aper e clu e apers were relevant y title an a stract to su ect inclu e aper e clu e apers ha perfor ance assess ent criteria aper e clu e apers ha perfor ance assess ent criteria Fig.
PRISMA process for data collection Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Journal of Robotics and Control (JRC) ISSN: 2715-5072 achine earning eep earning nowle ge Fig.
Number of articles published in the period 2020-2024 related to the identification of Epileptogenic Zones Fig.
Frequency of AI methods used in the identification of Epileptogenic Zones TABLE II.
SELECTED PAPERS ACCORDING TO THE SPECIFIED CRITERIA
Authors & Year AI methods Roger et al.
Machine Learning Algorithm Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoos.
Dataset The "original" dataset of 57 unilateral mTLE patients and the "reduced and working" dataset of 46 patients Characteristics of Drug-resistant presurgical evaluation (NPE.
EEG.
MRI).
divided into LmTLE and RmTLE Personalized, connectome-based data derived from non-invasive imaging (MRI.
DTI).
used to simulate brain The best Performance XGBoost AUROC: 90.
Accuracy: 77.
No-U-Turn
Sampler
(NUTS)
Accuracy: 100% Hashemi et al.
KnowledgeBased AI
No-U-Turn
Sampler
(NUTS),
Automatic Differentiation Variational Inference (ADVI) Guo et al.
, 2020
Deep Learning Attention Neural Network (AttNN) MEG data from 20 epilepsy patients .
ripples and 50 fast MEG 306-channel, 4000 Hz frequency sampling.
Manual
labeling by experts Attention
Neural
Network
(AttNN) Deep
Learning EMS-Net
(CNN
multiview: 1D 2D feature MEG data from 20 epilepsy patients recorded at Sanbo Hospital.
Beijing 306-channel, 1000
Hz, 300 ms
epochs, spike & non-spike, data
EMS-Net
Machine Learning Simple neural network
MLPNN), least-square support vector
(LS-SVM)
Bonn University EEG dataset
500 EEG
segments, 5 subsets (AAeE), 100 segments per set, 61 Hz sampling, 0.
5Ae40 Hz filtered GDA sMLPNN Accuracy: 100%.
Sensitivity: 100%.
Specificity: 100%.
Precision: 100%.
AUC: 1 CNN-based STFT CWT Accuracy: 91.
LSTM with feature (CCP) Accuracy: 99%.
Precision: 95%.
Recall: 100%.
F1score: 98% Zheng et al.
Nkengfack et , 2021 .
Simulated synthetic data using The Virtual Brain (TVB).
patient-specific MRI and DTI SEEG data Xia et al.
, 2021
Deep
Learning Convolutional neural network
(CNN)
Bern-Barcelona EEG Database .
0 pairs of Focal and NonFocal signals from 5 epilepsy Aliyu et al.
Deep Learning Long shortterm memory (LSTM) EEG Bonn University intracranial EEG, 512 Hz, 20 seconds, 10240 data points per from epileptogenic & nonepileptogenic Data has been normalized & de-noise 5 EEG subsets, 100 segments each, duration 23.
seconds, sampling 61 Hz Accuracy: 89.
AUC: 0.
Sensitivity: 84.
Specificity: 92.
F1 Score: 88.
Accuracy: 99.
Precision: 99.
Sensitivity: 99.
Specificity: 99.
F1 Score: 99.
AUC: 0.
Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Journal of Robotics and Control (JRC) Garcya et al.
Deep
Learning neural network
D CNN)
Nkengfack et , 2021 .
Machine Learning Least Squares Support Vector
Machine (LSSVM) dengan RBF kernel
Machine Learning Multilayer Perceptron
Neural
Network
(MLPNN),
Support Vector Machine (Linear SVM dan RBF SVM) Torabi et al.
ISSN: 2715-5072
EPISURG (Postoperative MRI
dataset of refractory epilepsy brain from various institutions, total of 430 postoperative Publicly available EEG dataset
from the University of Bonn
Mean Square Error
y AA
Hypergraph
SEEG HFOs
(HSO)
Accuracy: 90.
Sensitivity: 80.
Specificity: 96.
Epileptorbased model using Bayesian Precision: 80%.
Recall: 85% Temple University Hospital EEG Seizure Corpus (TUSZ) EEG of patients with epilepsy, multi-channel scalp EEG, sampling rate .
Ae512 H.
Random Forest Accuracy: 97.
AUC: 99.
Sensitivity (Recall/TPR):
Specificity (TNR): 99.
EEG dataset (Sets AAeE), total 500 segmen Each set consisted of 100 fixedduration EEG segments, from healthy and epileptic subjects Neural Architecture Search (NAS) Accuracy: 76.
F1-score: 76.
Kappa coefficient:
EEG dataset from University of Bonn (Set AAeE, total 500 trial.
3-class EEG .
ormal, interictal, samples per trial.
sampling rate 61 Hz Gradient (GBM) fusion genetic algorithm (GA) for feature Accuracy = 100% Interictal ECoG from 7 patients with refractory epilepsy (Focal Cortical Dysplasi.
1 hour of interactive ECoG recording, 2000 Hz sampling rate, adult and pediatric SVM with RBF Kernel AUC: 0.
EEG and MRI of epilepsy patients .
wo subject.
Guo et al.
, 2021
Machine Learning Hypergraph Learning SEEG data from 19 refractory focal epilepsy patients .
otal 4000 segment signals: 1640 HFO and 2360 baseline control.
Machine Learning Epileptor ynamical syste.
Hierarchical Bayesian Machine Learning Random Forest.
Support Vector Machine (SVM).
MultiLayer Perceptron
(MLP)
Deep Learning Deep neural network using NAS and EEGNet Sunaryono et , 2022 .
Machine Learning Miao et al.
Multi Gradient Boosting Machines (GBM) fusion Genetic Algorithm (GA) for SVM (Linear & RBF Kerne.
LightGBM, 2D Accuracy: 88.
75% 100%.
AUC: 0.
- 1.
Spiking Neural Network Bonn University EEG Dataset Spiking Neural Network Liu et al.
, 2022
Dice Score (DSC):
Multilayer Perceptron
Neural
Network
(MLPNN)
Deep Learning Wang et al.
3D MRI T1weighted (T1.
, resolusi isotropik 1
3D CNN
193 y 229 y 193 5 sets (AAeE), each of 100 EEG data of 23.
LS-SVM
Condition:
berbasis Jacobi healthy .
ith eyes Polynomial open/close.
, epileptic Transform patients without (JPT) seizures, and epileptic patients during seizures Accuracy ABCD/E:
Accuracy AB/CD/E: 98.
Accuracy A/D/E:
Accuracy A/E: 100%.
Accuracy D/E:
Saeedinia et al.
Vattikonda et , 2021 .
Retrospective SEEG data from 25 focal epilepsy patients.
includes synthetic and empirical patient data Epilepsy EEG, 5 sets (A-E), 100 segments each, seconds, sampling Multi-channel EEG .
, personal MRI, data duration Ou hours patient and 40 minutes .
SEEG 256channel, sampling rate 2000 Hz, segmentation per 1000 ms, includes interictal and ictal data, gold standard determined by SEEG data, synthetic seizure generation, real patient outcome labels (Engel IAeIV) Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review
Journal of Robotics and Control (JRC) ISSN: 2715-5072
neural network
(CNN)
Machine Learning Artificial Neural Networks (ANN.
MultiLayer Perceptron
(MLP)
Scalp EEG recordings of 21 adult patients with focal epilepsy .
ata from Baptist Hospital of Miami, 19 10-20 syste.
Wang et al.
Deep
Learning Multiscale neural network
(MSCNN),
Bidirectional LSTM (BiLSTMAM).
GradCAM Public multicenter SEEG dataset (MAYO.
FNUSA) and Private clinical SEEG dataset Mohsen et al.
Multi Methods LSTM.
SVM .
ith Fast WalshHadamard Transfor.
EEG dari University of Bonn Sun et al.
, 2023
Deep Learning Deep Source Imaging Framework (PDeepSIF) MEG data from 29 focal epilepsy patients Dou et al.
, 2023
Deep
Learning Autoencoder Adaptive Graph
Convolutional Network
(GCN)
SEEG
(Stereoelectroencephalograph.
& CCEP (Cortico-Cortical Evoked Potential.
Li et al.
, 2023
Machine Learning RUSBoost (Random Under Sampling Boostin.
HFO data from 26 epilepsy patients .
hospitals: Tiantan & Fengtai Hospital.
Beijin.
Ilias et al.
, 2023
Deep Learning EfficientNetB7.
CNN.
Gated Multimodal Unit Weiss et al.
Multi Mohammed et , 2022 .
Kim et al.
, 2024
Murugan et al.
Multi Methods Deep Learning Graph metrics (FR ratedistance radius.
Convolutional neural network, random forest.
SVM.
XGBoost
Convolutional Neural
Network
(CNN)
EEG database dari University of Bonn iEEG from 23 epilepsy patients (UCLA & Thomas Jefferson Univ.
EEG from 150 patients .
NCSE, 50 ME, 50 BI), 19 channel, 20 seconds epoch The public EEG dataset, consisting of 500 EEG recordings, each 23.
6 seconds long, is divided into 23 segments per recording patients, number of channels 36Ae76 3-second EEG segments (IED vs NIED), band-pass 5Ae70 Hz, sampling rates of 200/256/512 Hz.
PCA ICA annotated IEDs Multicenter SEEG data from epilepsy patients, including physiological, and artifact signals.
high-frequency SEEG data.
crosubject variation.
500 EEG signals, single-channel, 6 s.
Hz.
only class C&D .
on-seizure and seizur.
is MEG interictal head model from MRI.
validation with iEEG/surgical Time-frequency SEEG data from 18 patients, 3 behavioral states .
wake, sleep, 113,457 HFO .
raining: 89,844 pathoHFO 23,613 phyHFO), testing: 12,695 pathoHFO 5,599 phyHFO Single-channel EEG, consisting of healthy, interictal, and ictal classes, is processed with STFT to produce 3-channel images .
pectrogram, delta, etc.
SEEG, non-REM sleep, fast ripples >350 Hz, 2 kHz EEG 32-channel:
19 channel, 200 Hz, bandpass 0.
1Ae 70 Hz, 20s optimal EEG 1D signals .
data points/segment.
, consisting of seizure and nonseizure classes, recorded from various individuals FC-NNPruned .
pruned neural network using features from 4 subband.
ROC-AUC: 0.
SEEG-Net
(MSCNN
BiLSTM-AM FDG-los.
Accuracy: 93.
TPR: 87.
TNR:
FPR:
LSTM
Accuracy: 99.
Precision: 99.
Recall: 99.
F1score: 99.
PDeepSIF Sensitivity: 77%.
Specificity: 99% Adaptive Graph Convolutional Network Accuracy: 83.
F1-score: 76.
RUSBoost AUC: 0.
Accuracy: 0.
Sensitivity: 0.
Specificity: 0.
F1-score: 0.
Multimodal CNN EfficientNetB7 Gated Multimodal Unit Accuracy: 95.
33% 98.
FR ratedistance radius AUC: 0.
Accuracy: 78.
Sensitivity: 100%.
Specificity: 61.
NPV: 100%
CNN uses FC AUC = 0.
Convolutional Neural
Network
(CNN)
Accuracy: 98.
Precision: 0.
Recall (Sensitivit.
F1 Score: 0.
Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Journal of Robotics and Control (JRC) ISSN: 2715-5072 Accuracy: 82.
, 75.
4% .
ast Convolutional Neural
Network
(CNN)
Precision: 90.
Recall: 89.
Own EEG dataset .
ediatric patients with epileps.
Bonn EEG dataset multi-channel EEG .
wn dataset, 500 Hz sampling rate, 2-second singlechannel (Bonn, 61 Hz, 23.
sec segmen.
TCN-SA Self-dataset accuracy: 95.
Sensitivity: 91.
Specificity: 98.
AUC: 0.
95 and Bonn A-E dataset accuracy: 97.
Sensitivity: 94.
Specificity: 99.
F1 Score: 97.
Bonn EEG dataset & CHB-MIT EEG dataset EEG signals classified as normal or seizure, benchmark sets GBSOTAENN Accuracy: 99.
Specificity: 99.
Sensitivity: 99% SVM with RBF kernel using TF-IDF F1-score: 85.
emporal vs extratemporal ANN Accuracy: 64.
Sensitivity: 76.
Specificity: 43.
DCNN
Genetic Algorithm (GA) CrossValidation (CV) Accuracy: 93.
Precision: 90%.
Recall/Sensitivity:
Specificity:
F1-score:
AUC: 0.
Machine Learning Logistic Regression iEEG recordings from 20 MRE patients .
umber: CRCNS.
Payman et al.
Deep
Learning Convolutional Neural
Network
(CNN)
3,560 annotated skull base images .
rom 34 dry human Kantipudi et al.
Mora et al.
Mercier et al.
Krishnamoorthy et al.
, 2024 .
Intracranial EEG .
ubdural & depth interictal sleep sampling window per night, freely Multi-angle, highresolution images of 10 types of annotated with bounding boxes Logistic Regression Stergiadis et al.
Huang et al.
Deep
Learning Temporal Convolutional Neural
Network
dengan SelfAttention Layer
(TCN-SA)
Multi Methods GBSO-TAENN (Gradientbased Spider Optimization Temporal Aware Ensemble Neural Networ.
Multi Methods Logistic Regression.
SVM .
RBF.
NLP
536 seizure descriptions from 122 patients .
etrospective from Italian epilepsy surgery cente.
Machine Learning Artificial Neural
Network
(ANN),
Logistic Regression (LR) 123 paediatric patients.
EEG
akefulness & slee.
, 246
1-minute interictal scalp EEG
Deep
Learning Optimized Deep
Convolutional Neural
Network
(DCNN)
Bonn EEG dataset.
New Delhi EEG dataset Based on a systematic review of 34 scientific articles, various artificial intelligence (AI) approaches have been used to detect epileptogenic zones (EZ.
with varying performance and characteristics.
These methods include deep learning, conventional machine learning, multi-method .
approaches, and knowledge-based AI systems.
Each has its own advantages in specific aspects, such as accuracy, scalability, interpretability, and robustness to data variability.
However, each approach also faces limitations, ranging from data requirements, modelling complexity, to clinical validity.
Table i summarises the types of AI methods used in studies identifying epileptogenic zones, along with the main advantages and limitations of each approach.
Text-based seizure descriptions in Italian, labeled by EZ side & region.
highly curated clinical EMR Clean EEGs without artefacts or discharges, using standard 10Ae20 montage, sampling rate 256/512 Hz Bonn EEG dataset is categorized into 3 (Normal.
Interictal.
Seizur.
and New Delhi dataset is categorized into 3 classes .
ctal, preictal, intericta.
Fig.
4 shows the frequency of AI algorithm use in identifying epileptogenic zones based on all articles reviewed .
The two most dominant algorithms are convolutional neural network (CNN) and artificial neural network (ANN), each used in 26% of studies .
This reflects researchers' tendency to rely on neural network models, both in the form of convolutional networks for extracting spatial-temporal features from brain signals and classical feedforward networks for classification.
Furthermore.
Support Vector Machine (SVM) and ensemble algorithms .
uch as XGBoost or GBM) were used in 9% of studies .
each, indicating that conventional machine learning approaches remain relevant, especially in situations with limited data.
Long short-term memory (LSTM), which Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Journal of Robotics and Control (JRC) ISSN: 2715-5072 focuses on temporal dynamics, was used in 6% .
of studies, while random forest and logistic regression were each used in only 3% of studies .
Another 18% of studies .
used other methods such as NUTS.
NAS, or adaptive graph algorithms that were not explicitly This distribution reflects the trend that while neural network algorithms dominate, combination or alternative approaches remain necessary to address the complexity of SEEG/EEG data in the context of epilepsy.
of multimodal datasets .
uch as SEEG.
MRI, and MEG) to achieve more precise identification of epileptogenic zones, particularly in the context of refractory epilepsy requiring surgical intervention.
linical escription linical escription Fig.
Frequency of datasets used in the identification of Epileptogenic Zones IV.
Fig.
Frequency of algorithms used in the identification of Epileptogenic Zones In addition.
Fig.
5 presents the frequency distribution of the use of various types of datasets in studies identifying epileptogenic zones, based on an analysis of a number of scientific articles.
It can be seen that scalp EEG is the most widely used dataset, accounting for 53% of all studies.
This reflects that scalp EEG remains the primary and most accessible method in both research and clinical practice, despite its limitations in spatial resolution.
On the other hand.
SEEG (Stereo-EEG), an invasive technique with high spatial and temporal resolution, was used in 12% of studies, followed by MRI .
%) and MEG .
%), both of which play a crucial role in non-invasive mapping of brain anatomy and iEEG (Intracranial EEG) datasets were also used in 6% of studies, indicating a trend toward increased use of data from subdural or intracortical electrodes.
Meanwhile, highfrequency oscillation (HFO).
Electrocorticography (ECoG), and clinical descriptions were each used in only 3% of studies, indicating that although highly informative, such data remain limited in use due to access constraints, costs, and the need for invasive procedures.
This pattern highlights that while scalp EEG remains dominant due to its non-invasive and easily accessible nature, there is an increasing utilisation DISCUSSION This systematic review shows a growing number of publications between 2020 and 2024 discussing the application of artificial intelligence (AI) in identifying epileptogenic zones, a crucial area in the management of refractory epilepsy.
A total of 34 articles were analysed, with a significant increase in publications in 2021 and 2024.
This reflects advancements in AI technology and the urgency to improve the accuracy of epileptogenic zone identification, particularly in the context of pre-surgical evaluation.
The review also highlights a shift in approach from conventional methods toward deep learning techniques, as well as increased use of high-resolution Stereo-EEG (SEEG) data.
Deep learning methods are the most dominant AI approach used, appearing in 44% of studies.
Architectures such as CNN.
LSTM, and TCN have proven highly effective in extracting spatio-temporal features from EEG/SEEG signals.
For example, the SEEG-Net model (Wang et al.
, 2.
which combines CNN.
BiLSTM, and Grad-CAM , achieved an accuracy of 93.
85% on a multicenter dataset .
Similarly, the TCN-SA model (Huang et al.
, 2.
demonstrated high performance on both internal and external datasets .
These results reinforce the potential value of deep learning in mapping epileptogenic zones with high TABLE i.
SUMMARY OF TYPES OF AI METHODS.
THEIR ADVANTAGES AND LIMITATIONS IN THE CONTEXT OF EPILEPTOGENIC ZONE IDENTIFICATION
AI Method Deep Learning Machine Learning Multi-method (Hybri.
Knowledge-Based AI Advantages in Epileptogenic Zone Identification Automatically extracts spatial-temporal features from EEG/SEEG Highly accurate .
ccuracy >90%) with SEEG.
Suitable for large and complex datasets.
Suitable for small to medium-sized datasets.
Faster to train and more interpretable.
Can be optimized through feature selection or transformation.
Combines strengths of multiple models .
CNN SVM).
More robust and adaptive.
Can enhance generalization to noisy or varied data.
Can integrate clinical knowledge and brain network theory.
More transparent and interpretable .
ule-based system.
Limitations in Epileptogenic Zone Identification Requires large, well-annotated datasets.
Low interpretability .
lack-bo.
Prone to overfitting on small datasets.
Relies on manually extracted features.
Less effective for raw EEG signal.
Lower performance than DL overall.
More complex to design and validate.
Harder to reproduce without detailed Requires tuning many parameters.
Less flexible for real-world data variability.
Not suitable for raw EEG signals.
Rarely used and less scalable.
Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Journal of Robotics and Control (JRC) ISSN: 2715-5072 On the other hand, conventional machine learning methods such as SVM.
Random Forest, and XGBoost are still used in certain conditions, particularly when the dataset size is limited or when better interpretability is required.
A study by Roger et al.
used a combination of SVM and XGBoost to classify mTLE laterality, achieving an AUROC 2% .
Although simpler, this approach remains highly relevant, especially in clinical settings with data limitations or the need for transparent interpretation.
Some studies also explore multi-method and hybrid approaches, such as Miao et al.
, who combined SVM.
LightGBM, and CNN .
, and Kantipudi et al.
with the GBSOTAENN model, which integrates spider algorithm optimisation (GBSO) with temporal neural networks .
Such hybrid models can leverage the strengths of each algorithm and demonstrate superior performance in handling the complexity of brain signals.
This review also highlights trends in dataset usage.
Scalp EEG is the most commonly used dataset type .
%), primarily due to its non-invasive nature and availability in open-access formats .
Bonn EEG.
CHB-MIT).
However, the most accurate results are generally found in studies using SEEG or iEEG datasets .
%), which have high spatial and temporal resolution.
This is evident in the study by Zheng et .
using MEG, which achieved an accuracy of 48% and nearly perfect AUC .
Studies using SEEG data produce very high performance because they are able to record brain electrical activity in depth and with Guo et al.
using an HSO Detector based on hypergraph learning achieved an accuracy of 90.
7% and specificity of 96.
9% .
SEEG data offers advantages over scalp EEG in revealing hidden epileptogenic activity patterns within brain structures, particularly in patients with complex focal epilepsy.
Recent approaches also demonstrate the use of graph models, such as graph convolutional networks (GCN) in the study by Dou et al.
, which model the relationships between SEEG channels as an adaptive graph .
This approach treats the brain as a complex interconnected network, reflecting the new perspective that epileptogenic zones are not fixed locations but part of the brain's dynamic network system.
Although the models generally perform well, many studies still face limitations, particularly in terms of small sample sizes, reliance on synthetic data .
uch as TVB), and lack of validation on external datasets or clinical outcomes .
uch as post-surgical Engel classificatio.
This limits the application of these models in real-world contexts and needs to be addressed in future studies.
Another challenge is the lack of standardisation in EEG/SEEG preprocessing across Filtering, segmentation, and HFO annotation techniques vary, making it difficult to replicate or compare Differences in HFO definitions .
athological vs.
also add complexity.
Therefore, clinical consensus and open standards for invasive EEG signal processing are needed.
From a clinical perspective, these findings hold great potential.
AI can accelerate the identification of epileptogenic zones, reduce SEEG monitoring time, and assist in electrode placement and epilepsy surgery planning.
Models such as SEEG-Net and HSO Detector have the potential to be integrated into clinical decision support systems (CDSS) at epilepsy surgery centres .
However, to date, the application of AI in clinical settings remains limited.
Most models have not been prospectively tested in real-world practice, and there are still ethical, regulatory, and interpretability challenges .
Transparent and explainable AI models are crucial in the context of high-stakes surgical decision-making .
For future development, research should focus on integrating multimodal data (SEEG MRI DTI clinica.
, tracking long-term outcomes, and conducting prospective multicentre clinical trials.
Collaboration between neurologists, epileptologists, and AI scientists is essential to develop robust, clinically valid systems ready for implementation .
Additionally, future research should also emphasize transparent documentation of the AI model structure, preprocessing procedures, and patient demographic details.
Moreover, it is important to report postsurgical clinical outcomes, such as those measured by the Engel scale, to assess the practical efficacy of EZ predictions beyond statistical metrics alone.
CONCLUSIONS
This review confirms that artificial intelligence (AI), particularly deep learning methods based on SEEG and multimodal approaches, has great potential to revolutionise the process of identifying epileptogenic zones.
Models such as CNN.
LSTM, and hybrid networks that combine spatialtemporal features have demonstrated high accuracy in detecting abnormal patterns in EEG and SEEG data.
This advantage is reinforced by consistent results across various studies, especially those using high-quality data such as SEEG and MEG, as well as validation against clinical AI enables more objective, efficient, and scalable exploration of epileptogenic zones across diverse patient However, for AI to be widely implemented in clinical practice, a series of important prerequisites are required, including: multi-centre validation, standardisation of preprocessing and data annotation processes, and transparent reporting of model structures and clinical Key challenges also include the need for explainable AI models, integration with clinical decision support systems (CDSS), and attention to ethical and medical data privacy aspects.
Bridging the gap between research and clinical practice requires coordinated efforts from scientists, clinicians, and tech developers.
With the right direction of development.
AI has the potential to become a cornerstone of future precision epileptology practice.
The combination of AI's capabilities in large-scale and complex data analysis, along with clinicians' expertise in contextual interpretation and medical decision-making, will create a strong synergy in the management of refractory epilepsy.
In the long term.
AI systems integrated into clinical workflows can contribute to improved diagnostic accuracy, pre-surgical evaluation efficiency, surgical success, and overall quality of life for epilepsy patients.
ACKNOWLEDGMENT
The authors would like to acknowledge the Department of Medical Technology.
Institut Teknologi Sepuluh Nopember, for the facilities and support in this research.
The Yuri Pamungkas.
The Emerging Role of Artificial Intelligence in Identifying Epileptogenic Zone: A Systematic Literature Review Journal of Robotics and Control (JRC) ISSN: 2715-5072 authors also gratefully acknowledge financial support from the Institut Teknologi Sepuluh Nopember for this work, under project scheme of the Publication Writing and IPR Incentive Program (PPHKI) 2025.
REFERENCES