Content area
Epilepsy is a neurological disorder characterized by abnormal neuronal discharges in the brain. As a rich source of biometric information, electroencephalography (EEG) provides favorable conditions for automated detection. Traditional algorithms and manual analysis possess solid theoretical foundations and good interpretability, however, these methods predominantly require extensive domain expertise and involve lengthy processing pipelines for complex data. The advent of artificial intelligence (AI) has facilitated the application of neural networks in the detection and prediction of epilepsy. Although such approaches heavily rely on high-quality annotated data, suffer from limited model interpretability, and involve complex training and parameter tuning, these efficient, real-time, end-to-end models still demonstrate significant potential in epilepsy analysis. This review systematically analyzes and summarizes the neural network technologies used in 341 papers published in the past three years, employing the PRISMA standard procedure. To facilitate readers’ related research, the review also summarizes the basic information of 16 publicly available datasets, common features, and metrics. Specifically, this review offers a comprehensive evaluation of diverse neural network architectures, concluding that convolutional neural networks have become a prevalent choice as classic neural networks. Furthermore, graph neural networks and transformers are experiencing a marked surge in popularity. The application of hybrid neural networks to fully extract information from EEG is also a growing trend. The review concludes with a comprehensive discussion and summary of the technical characteristics, research directions, and limitations of current methods, including patient-to-patient identification, explainable AI, dataset bias, and zone location.
Introduction
Epilepsy is a chronic neurological condition marked by recurrent episodes of sensory disturbance in the human brain (Gotman 2011). An epileptic seizure (ES) is described by the international league against epilepsy (ILAE) as a passing event characterized by indications and/or symptoms resulting from abnormal, excessive, or synchronous neuronal activity in the brain. Epilepsy affects approximately 1% of the world’s population (Khurshid et al. 2024), with the primary age group being 10–20 years old (Soliman et al. 2022).
According to the World Health Organization (WHO), 70% of epileptic patients can be treated with medication, making early detection and prediction of epilepsy particularly important. Causes of epilepsy can be divided into structural, genetic, metabolic, or unknown factors (Falco-Walter et al. 2018). During a seizure, patients may experience loss of consciousness and uncontrolled tonic-clonic convulsions, preventing them from performing self-care activities. Furthermore, epilepsy is associated with life-threatening complications, including sudden unexpected death in epilepsy, a significant risk that cannot be overlooked (Tomson et al. 2005; Supriya et al. 2021).
Electroencephalogram (EEG) is a recording technique that is used to measure electrical signals in the brain. Due to its rich dynamics and high temporal resolution, it provides biological signals for the recognition of epilepsy (Llorente-Vidrio et al. 2022; Fathima et al. 2024). In the clinical setting, physicians primarily use the visual interpretation of EEG to determine ES (Ibrahim et al. 2022). This approach is widely considered the current gold standard for epilepsy detection (Noachtar and Rémi 2009).
However, this method has significant limitations, as its accuracy is heavily dependent on the judgment of personnel and the expertise of physicians, increasing the risk of misdiagnosis and missed diagnoses (Noachtar and Rémi 2009). In addition, the background noise recorded by EEG also interferes with the diagnostic process. Epilepsy is characterized by numerous subclassifications [focal seizures (Gill et al. 2024), generalized seizures, West syndrome, etc.] that are challenging to evaluate manually (Samee et al. 2022). The recognition of epilepsy is a time-consuming and tedious task that places a significant burden on clinicians (Shanmugam and Dharmar 2023). Since Cook et al. (2013) experimentally demonstrated the feasibility of using EEG signals for epilepsy detection, research in this field has gained significant attention.
Although traditional machine learning (ML) methods, such as support vector machines (SVMs) (Atlam et al. 2025), linear discriminant analysis (LDA) (Hasnaoui and Djebbari 2024), and decision trees (DT) (Hasnaoui and Djebbari 2025), remain effective and offer advantages such as lower complexity, interpretability, and suitability for smaller datasets, this review has chosen to focus on deep learning (DL) approaches due to their growing prominence and proven ability to model complex, nonlinear patterns in EEG data. Today, DL methods are playing an increasingly significant role in EEG recognition, as they excel in the extraction of discriminative features (Cui et al. 2025). These models enable end-to-end learning and have demonstrated superior performance in recent studies, particularly in tasks involving high-dimensional and unstructured data.
Neural networks consist of a series of interconnected nodes, each connection being assigned a weight to facilitate the learning and classification processes of the network. Traditional ML methods, such as SVMs, have several inherent limitations. First, they are unable to handle nonlinear data, which is a prominent characteristic of EEG signals (Huang and Duan 2023). Second, their generalization capabilities are limited. Third, they struggle with high-dimensional data. Finally, their ability to process large-scale datasets is limited. In contrast, neural network-based DL methods show promising learning capabilities, with potential applications for seizure prediction and detection.
However, these methods cannot be simply transferred to perform epilepsy detection/prediction tasks. EEG signals are inherently nonlinear and chaotic (Ilakiyaselvan et al. 2022). Most ES detection and prediction methods overlook the nonlinear and implicit properties of EEG, which can influence accuracy (Huang and Duan 2023). Furthermore, EEG recordings are susceptible to artifacts, leading to misleading data analysis and even significantly impacting the interpretation of results (Jin et al. 2023; Narmada and Shukla 2023).
[See PDF for image]
Fig. 1
Division of epilepsy into different periods. The seizure prediction horizon (SPH) defines the minimum lead time between the alarm () and seizure onset (), ensuring an actionable warning. The seizure occurrence period (SOP) specifies the subsequent time window during which a seizure is expected to occur. Here, defines the transition time between SPH and SOP determined by EEG signals, which does not necessarily coincide with the clinical symptom, i.e., epileptic seizure. This discrepancy exists because pre-ictal changes are often reflected in EEG signals before any clinical symptoms appear (Fathima et al. 2024; Nemati and Meshgini 2022)
EEG recordings can be divided into four stages: pre-ictal, ictal, post-ictal, and interictal (Fathima et al. 2024), which are illustrated in Fig. 1. The pre-ictal stage occurs before the onset of ES, the ictal stage is when the ES takes place, the post-ictal stage follows ES, and the interictal stage refers to the time outside these specific periods. Predicting ES before their onset is crucial, making the pre-ictal stage particularly significant for research.
[See PDF for image]
Fig. 2
Epilepsy recognition tasks by number of classifications. “Non-ictal" refers to other periods outside of ES. “Normal" indicates non-epileptic individuals. “Pre-I" and “Pre-II" denote different prediction timeframes, such as 5 min and 30 min before seizure onset. “Subclasses" represent more specific conditions, such as focal seizures
Epilepsy recognition tasks, including detection and prediction, can be categorized by the number of distinct classifications, as illustrated in Fig. 2. Most prior research has concentrated on distinguishing between ictal and interictal states, the most fundamental binary classification task (i.e., seizure detection). In contrast, epilepsy prediction typically requires differentiating the pre-ictal phase from the interictal phase, with some studies further incorporating the ictal phase as an additional class. Specifically, this review focuses on epilepsy detection and prediction tasks, with particular attention to their advancements within neural network-based approaches.
Comparative analysis with existing reviews in this field
Before describing our work in detail, we conducted a comparative analysis of recent reviews in the field of automatic epilepsy detection/prediction. The main purpose was to explain the focus and limitations of these reviews and highlight the focus of our review.
Fathima et al. (2024) classified the features into four domains, including the time domain, frequency domain, time-frequency (TF) domain, and nonlinear domain. In addition, the authors listed four public datasets and discussed the main metrics. They also proposed a comprehensive recognition framework that included feature selection, classification, and performance verification. However, there is still a lack of comprehensive summaries of datasets, features, and performance metrics.
Supriya et al. (2021) focused on graph theory-based detection methods, revealing the limitations of the above four types of features and highlighting the significant advantages of graph theory in extracting channel features from EEG. The study provides a detailed explanation of several graph parameters used in the EEG feature vector analysis. In their paper, EEG-based graph networks are categorized into five types: visibility graph methods, horizontal visibility graph methods, time series of complex network methods, traditional graph theory-based methods, and weighted graph-based methods. This approach primarily focuses on graph theory and is therefore not comprehensive for other features.
Ren et al. (2023) divided the epilepsy framework into three subproblems: signal acquisition, feature extraction, and classifier design. They provided a detailed explanation of signal acquisition, with particular emphasis on key components such as electrodes, preprocessing circuits, filters, as well as system-level chips. The article focused on trends in the development of low-power signal acquisition devices, but there was relatively little discussion of feature extraction and classifier design.
Jahan et al. (2023) focused on the combination of the Internet of Things (IoT) framework and artificial intelligence (AI). The paper delves into IoT-based epilepsy prediction and detection models and outlines future challenges and directions for development. To support their arguments, the authors carefully selected six datasets for description and revealed their review process. In addition, the review describes features, architectures, classification algorithms, and evaluation metrics. However, it lacks a comprehensive analysis of AI algorithms used.
Huang et al. (2023) divided the ES detection framework into three core aspects: data pre-processing, feature extraction, and neural network detection algorithms. They also conducted in-depth research on epilepsy prediction models and specifically mentioned the exploration of spiking neural networks (SNNs) in this field. The authors particularly discussed the limitations of existing technologies, arguing that the results are limited due to single evaluation indicators and insufficient data, but they did not analyze the technical details of existing technologies in sufficient depth.
Naganur et al. (2022) conducted a review and meta-analysis of portable devices to detect ES. Their work focused more on statistical analysis of the sensitivity of different devices or protocols, but lacked analysis of the algorithms involved.
Acharya et al. (2018) review more on the limitations and prospects of epilepsy prediction technology, including EEG data bias, biomarker extraction, noise interference, and improving prediction accuracy with large amounts of data. Furthermore, the review explores the potential application of cloud computing in the prediction of epilepsy. However, it should be noted that the coverage may not reflect latest breakthroughs.
Spikes are considered a reliable feature for epilepsy detection, and El-Samie et al. (2018) focus on a review of spike detection. Spikes are often generated by ES and are considered markers of seizures or potential seizures. The authors classify ES into three dimensions: time domain, spectral domain, and wavelet domain. At the end of this review, they provide suggestions for improving spike detection algorithms. However, there is insufficient understanding of the new epilepsy feature extraction methods that have emerged in recent years.
A comparative analysis of these existing reviews reveals that they offer unique and insightful perspectives within their respective fields. However, due to limitations in the time frame covered and the number of documents surveyed, they lack a thorough analysis of the latest advances in EEG for epilepsy detection and prediction. In view of this, the present review focuses on technical breakthroughs in neural networks for epilepsy detection/prediction, including new trends like transformers and graph neural networks (GNNs). In addition, based on an extensive review of the existing literature, it also includes a comprehensive statistical analysis of key components such as datasets, metrics, and features. The objective of this endeavor is to address the gaps in existing reviews.
This study was conducted according to the preferred reporting items for systematic reviews and meta-analyses (PRISMA) framework. PRISMA is a set of guidelines designed to improve the quality of reporting in systematic reviews and meta-analyses (Page et al. 2021). Its primary goal is to ensure that such reports are transparent, comprehensive, and easily understandable and reproducible. PRISMA provides researchers with a standardized structure, facilitating the normalization of content in their literature reviews and meta-analyzes.
Chapter arrangement
There are three main research questions driving this review paper:
What datasets, evaluation metrics, and feature extraction methods have been commonly used in epilepsy detection/prediction studies recently?
What are the most influential advances in neural network architectures for epilepsy detection/prediction recently?
What are the current applications, technological advances, and key limitations of neural networks in epilepsy detection/prediction systems?
Systematic review methodology
This systematic review follows the guidelines provided by the PRISMA statement (Rethlefsen et al. 2021; Page et al. 2021). The following subsections provide the necessary information related to the identification, selection, and interpretation of the studies included in this review. Figure 3 illustrates the literature search process with a formal PRISMA flowchart.
The eligibility criteria for the papers are presented in Sect. 2.1, which served as the basis for screening to identify the literature of interest for this review. Studies were initially collected by exploiting scientific search engines and applying the keyword search filters mentioned in Sect. 2.2. In Sect. 2.3, the methodology and strategy for examining each of the literature obtained from the initial screening are presented. Section 2.4 presents the detailed results of the screening and describes some of the non-compliance cases to explain why some papers were not included in this study.
[See PDF for image]
Fig. 3
PRISMA-based literature review flowchart
Eligibility criteria
According to the following criteria, this systematic review includes research on ES with EEG signals and neural networks.
The present study is limited to the written literature in English, a choice based on the desire to ensure a high degree of generalizability.
Selection is limited to journal papers, whereas conference proceedings, reviews, book chapters, posters, master’s theses, and Ph.D. dissertations were excluded.
The temporal restriction on papers published after January 2022 was implemented with the objective of examining State-Of-The-Art (SOTA) methods and research trends, including preprints.
Papers that did not utilize at least one publicly available dataset to evaluate their methodology were excluded.
Papers that did not clearly demonstrate the validity of their methodology were excluded (required to be validated at least using metrics such as accuracy, sensitivity, false positives, etc.).
Papers that did not use neural networks (at least one layer of neural network) for classification, comparison, or evaluation were excluded.
Papers that did not present sufficient information to reproduce their results (e.g., a model architecture or a valid model evaluation strategy) were excluded.
At least one of the methods in the paper (those applying neural networks) had an advanced level of accuracy. After screening, our research considers that the accuracy should be greater than or equal to 85%.
Papers whose accessibility has been restricted in such a way that their information cannot be fully accessed were excluded (e.g., retracted, database inaccessible, paper being protected, etc.).
Search strategy
The papers to be screened were collected from Scopus and IEEE Xplore. For Scopus, the papers were last accessed on December 18th, 2024. For IEEE Xplore, the papers were last accessed on July 1st, 2025. These studies were extracted by querying
Scopus as follows: TITLE-ABS-KEY ( ( seizure OR epilepsy OR ( epileptic AND seizure ) ) AND ( eeg OR electroencephalogram ) AND ( ( deep AND learning ) OR dl OR ( neural AND network ) ) ) AND LANGUAGE ( english ) AND PUBYEAR >2021 AND PUBYEAR < 2026 AND ( LIMIT-TO ( DOCTYPE, “ar" ) )
IEEE Xplore as follows: ( ( “All Metadata": “epilepsy" OR “seizure" OR “epileptic" ) AND ( “All Metadata": “EEG" OR “electroencephalogram" OR “eeg" ) AND ( “Abstract": “deep learning" OR “neural network" OR “ANN" ) )
Screening process
This section elaborates on the eligibility criteria outlined in Sect. 2.1. As noted previously, papers were firstly filtered using keyword restrictions to identify relevant literature. Each retained paper was then screened against the inclusion criteria. Since the automatic search has already completed Criteria 1–3, our screening is mainly based on Criteria 4–9.
The screening process for the 702 selected papers proceeded as follows. One author conducted a thorough manual review while a second author recorded detailed information from each paper. A third author cross-checked the results to ensure compliance with the screening criteria and verify that no data were omitted. In cases of disagreement or ambiguity, the team revisited the contested papers for reevaluation. A detailed description of our review process is provided below.
First, we screened the abstracts of the selected papers. Criterion 7 was prioritized as its required information was explicitly verifiable from the abstracts. We also examined the titles and abstracts for the use of neural networks by Criterion 6.
Next, since many papers did not specify their datasets in the abstracts, we performed a full-text skim to locate this information, typically found in the methods or conclusions sections. Papers with restricted access or unavailable through our institution were excluded according to Criteria 9.
During this review, we documented the neural network frameworks and performance metrics used in each study. Papers that did not meet Criteria 5 or 8 were excluded from further consideration.
Screening outcome
This subsection details the final results of this review. After applying the review criteria in Sect. 2.1 to each article, 341 of the 702 journal papers were retained. 102 articles that were not retained, partly because they did not use a publicly available dataset or did not explicitly use a dataset, and partly because they did not fit the topic of seizure diagnosis. 158 papers were excluded because they did not report an explicit accuracy rate or did not use neural networks. 56 papers were excluded because the accuracy was not SOTA or not accessible. Statistical information on the use of datasets, networks, and feature extraction methods obtained during the review process is presented in Sects. 3, 4, and 5, respectively.
In order to provide a more visual representation of the current research trends in the field, the keywords appearing in the title are plotted as a cloud diagram in Fig. 4.
[See PDF for image]
Fig. 4
Word cloud maps based on the investigated literature
Datasets, metrics and framework
Datasets
This section presents and elaborates on the 16 publicly available datasets mentioned in the 341 papers. Table 1 provides a brief overview of all datasets, including their original attribution, EEG types, task types and links. Table 2 offers the detailed information about these datasets.
Table 1. Original attribution, EEG type, task type and link to the EEG public dataset on ES
Dataset | Original attribution | EEG type | Task type | Link |
|---|---|---|---|---|
Bonn(UoB) (Andrzejak et al. 2001) | Bonn University | iEEG | Detection | Link |
CHB-MIT (Shoeb 2010) | Children’s Hospital Boston | Scalp EEG | Both detection and prediction | Link |
UCI (Kode et al. 2024) | UCI Epileptic Seizure Recognition dataset | iEEG | Detection | Link |
Kaggle (Abderrahim et al. 2024) | American Epilepsy Society Seizure Prediction Challenge database | iEEG | Prediction | Link |
TUH(TUEG) (Obeid and Picone 2016) | Temple University Hospital EEG Seizure Corpus | Scalp EEG | Both Detection and prediction | Link |
Siena (Detti et al. 2020) | Unit of Neurology and Neurophysiology at University of Siena | Scalp EEG | Both detection and prediction | Link |
Helsinki (Stevenson et al. 2019) | The NICU at Helsinki University Hospital | Scalp EEG | Detection | Link |
BMC (Nasreddine 2021) | Epilepsy monitoring unit of American University of Beirut Medical Center | Scalp EEG | Both detection and prediction | Link |
Hauz Khas (Swami et al. 2016) | Neurology and sleep Center, Hauz Khas, New Delhi | Scalp EEG | Detection | Link |
SWEC-ETHZ (Burrello et al. 2019) | Sleep-Wake-Epilepsy-Center at ETH Zurich | iEEG | Both detection andprediction | Link |
Bern-Barcelona (Varli and Yilmaz 2023) | Bern-Barcelona Database | iEEG | Prediction | Link |
Freiburg (Malekzadeh et al. 2021) | Freiburg hospital, Germany | iEEG | Prediction | Link |
Epilepsiae (Alvarado-Rojas et al. 2012) | European Union funded project EPILEPSIAE | iEEG/ Scalp EEG | Both detection and prediction | Link |
Melbourne (Massoud et al. 2023) | Melbourne University AES/MathWorks/NIH Seizure Prediction | iEEG | Prediction | Link |
UFP (Pereira and Fiel 2019) | Universidade Federal do Para, Brazil | Scalp EEG | Detection | Link |
SeizeIT2 (ICASSP) (Vieira et al. 2023) | Seizure challenge dataset from ICASSP 2023 | Scalp EEG | Detection | Link |
A significant proportion of the papers we reviewed used several publicly available datasets simultaneously. Of all the papers, CHB-MIT was the most widely used dataset, possibly because it contained multiple patients and was recorded over a long period of time. The distribution of the datasets is demonstrated by a pie chart shown in Fig. 5.
[See PDF for image]
Fig. 5
Statistical pie charts of the usage frequency for different datasets, where the datasets used less than 10 times are given separately in the right pie chart for better presentation
EEG can be categorized into scalp EEG (sEEG) and intracranial EEG (iEEG). sEEG involves placing electrodes on the scalp to record the electrical activity of the brain. It is the most common EEG technique in neuroscience and is widely used for the diagnosis of epilepsy, sleep disorders (Fathima and Ahmed 2025), encephalopathy, and other neurological problems. The primary method for electrode placement is the 10–20 system (El-Samie et al. 2018). This recording method is susceptible to power-frequency interference and artifacts from subject movements and eye activities, resulting in sEEG signals that are often noisy and of low intensity.
Table 2. Complement information of 16 public datasets
Dataset | Number of subjects | Ages | Sampling rate | Number of segments | Number of channels |
|---|---|---|---|---|---|
Bonn(UoB) | 10 | NA | 173.61Hz | 500 | Single |
CHB-MIT | 22 (5 M 17F) | 1.5–22 | 256Hz | 664 | 23 (22, 18, 28) |
UCI | 10 | NA | 173.61Hz | 11,500 | Single |
Kaggle | 2 & 5 cannies | NA | 5000Hz | 48 seizures | 16 (cannie) 24 (people) |
TUH(TUEG) | 10,874 | 1–90 | 250 (256 400 512)Hz | 16,986 | 31 |
Siena | 14 | 20–71 | 512Hz | 47 seizures | 29 |
Helsinki | 79 | weeks | 256Hz | NA | 19 |
BMC | 6 | NA | 400Hz | NA | 21 |
Hauz Khas | 10 | NA | 200Hz | 51,200 | 57 |
SWEC-ETHZ | 18 | NA | 256Hz | 116 seizures | 24–128 |
Bern-Barcelona | 5 | NA | 512Hz | NA | 64 |
Freiburg | 21 | 10–50 | 256Hz | 84 seizures | 128 |
Epilepsiae | 93 | 35.60 ± 14.21 | 256Hz | 518 seizures | NA |
Melbourne | 3 | 22–51 | 400Hz | 1139 seizures | 16 |
UFP | 14 | 17-45.9 | 256Hz | NA | 20 |
SeizeIT2(ICASSP) | 42 | NA | 250Hz | NA | 21 |
However, iEEG signals are recorded through surgically implanted intracranial electrodes, allowing direct measurement of electrical activity in deep brain structures. Compared to sEEG, iEEG offers higher spatial resolution and lower noise interference, along with the ability to detect subtle epileptic discharges and neural features that are often obscured in non-invasive recordings. Despite these advantages, the clinical applications of iEEG remains limited due to its invasive nature, restricting its use compared to sEEG techniques (Fathima et al. 2024).
Metrics
Performance metrics are essential for evaluating neural network-based classification tasks, as they quantify the accuracy and generalization ability of a model, reflecting the effectiveness of the classification approach. Although accuracy is the most widely used metric, it can be misleading in cases of class imbalance. For example, if 90% of the samples belong to a single class, a model that always predicts that class will achieve high accuracy, despite poor discriminative performance. To ensure a more comprehensive and reliable assessment, we summarize the key evaluation metrics in Table 3, which guides readers in selecting the most appropriate measures for their applications.
The confusion matrix is a fundamental tool for evaluating classification models, which sorts predictions into four key categories, i.e., true positives (TP), where the model correctly predicts the positive class; true negatives (TN), where it correctly predicts the negative class; false positives (FP or Type I errors), where negative cases are erroneously classified as positive; and false negatives (FN or Type II errors), where positive cases are misclassified as negative. TP and TN represent correct classifications, while FP and FN reveal model errors. Figure 6 illustrates the confusion matrix and a typical example.
[See PDF for image]
Fig. 6
Confusion matrix. a Four categories of classification results; b An example of confusion matrix under four signal phases
Table 3. Evaluation Metrics. SPH is illustrated in Fig. 1
Metric name | Equation | Description |
|---|---|---|
Accuracy | Overall classification correctness | |
Sensitivity (recall) | Ability to detect positive cases (avoid FNs) | |
Specificity | Ability to exclude negative cases (avoid FPs) | |
Precision | Proportion of true positives among predicted positives | |
FPR (false positive rate) | Proportion of false alarms (Type I errors) | |
FNR (false negative rate) | Proportion of missed detections (Type II errors) | |
G-mean | Geometric mean of sensitivity and specificity | |
F1-score | Harmonic mean of precision and recall | |
MCC (matthews correlation coefficient) | Balanced measure for all confusion matrix entries. (values belonging to ) | |
AUC-ROC (receiver operating characteristic) | - | Overall discrimination ability. (where one stands for optimal performance) |
SPH (seizure prediction horizon) | Minimum lead time between the alarm () and seizure onset () |
Framework
Given that ES detection/prediction tasks are often end-to-end, it is crucial to develop a comprehensive framework (Li et al. 2024; Wu et al. 2023; Kumar et al. 2023). Generally, the seizure process encompasses three aspects: pre-processing, feature extraction, and classification. The objective of pre-processing is to remove noise from EEG signals. The aim of feature extraction is to highlight the epileptic biomarkers in the EEG signals, thereby facilitating recognition by the classifier. Classification aims to determine whether an unknown segment represents a seizure or to identify the type of seizure, thus achieving the detection or prediction objective. In prediction tasks, post-processing is often involved. For example, the model is applied to a long recording of EEG data to evaluate whether it correctly issues a warning. Furthermore, some studies consider brain region localization as part of the post-processing phase (Yaqing et al. 2025; Zhao et al. 2018). The basic processing framework for seizure detection and prediction is illustrated in Fig. 7.
[See PDF for image]
Fig. 7
Basic framework for seizure detection and prediction
Considering that papers focusing on post-processing constitute a minority in our surveyed scope, post-processing is not included in this section. Instead, the processing effects and application scenarios of interest in some papers will be discussed in Sect. 5.
Pre-processing
Pre-processing aims to extract clean, reliable, and consistent EEG signals, thereby avoiding interference from noise and artifacts that could affect the conclusions of subsequent analyses. The noise sources in EEG signals are primarily categorized into three components: physiological noise, electrode noise, and environmental noise. Physiological noise encompasses ocular movements (such as eye blinks), muscle activity (as indicated by EMG artifacts), and more complex neural current interference. Electrode noise arises from powerline interference in acquisition devices as well as artifacts caused by improper electrode placement or crosstalk between electrodes. Environmental noise encompasses amplifier-related disturbances and external electromagnetic interference. Table 4 details the different types of noise, their sources, frequency ranges, and the corresponding denoising methods.
Table 4. EEG artifacts, sources, frequencies, and denoising methods
Artifact type | Source | Frequency | Denoising methods |
|---|---|---|---|
Ocular | Eye blinks | 0.1–5 Hz | ICA, regression, high-pass filtering |
Muscle | Facial EMG | 20–300 Hz | ICA, low-pass filtering |
Power noise | Power line | 50/60 Hz | Notch filtering |
Electrode pop | Bad contact | Broadband | Channel interpolation |
Amplifier noise | Circuitry | Above 1 kHz | Low-pass filtering |
Electromagnetic interference | Wireless | Above 1 kHz | Faraday cage |
Several methods are commonly used to remove noise from EEG signals, including filtering, independent component analysis (ICA), and regression techniques. Notch filters are particularly effective for eliminating power-line interference (Einizade et al. 2023). Aslam et al. (2022) implemented a preprocessing pipeline that went beyond basic notch filtering, incorporating regression-based techniques to further improve the signal-to-noise ratio. Similarly, the Kalman filter has been applied to suppress noise in raw EEG signals (Borhade et al. 2024), while bandpass filtering helps retain low-frequency biomarkers by discarding higher-frequency noise (Chung et al. 2024; Hermawan et al. 2024). Additionally, Golla and Maloji (2023) utilized a finite Haar wavelet transform (WT) to decompose, filter, and remove artifacts from raw EEG data.
Despite their benefits, conventional filtering methods can introduce signal distortions. Since unwanted noise often overlaps spectrally with EEG activity, aggressive filtering may result in information loss, ultimately limiting classifier performance. Various denoising strategies have been explored in recent literature. Shi et al. (2023) introduced a preprocessing module based on stochastic resonance, employing parallel non-symmetric overdamped bistable systems. By adjusting the asymmetry of these units, intrinsic noise is redistributed across different spectral components of the EEG signal.
ICA is another widely used approach, operating under the assumption that EEG signals are linear mixtures of independent sources. ICA separates these mixed signals into statistically independent components for further analysis. Sunkara (2024) combined ICA with principal component analysis (PCA) and high-pass filtering to pre-process EEG data, while Sivasaravana Babu et al. (2023) applied heuristic ICA to extract independent components. Additionally, Varnosfaderani et al. (2024) employed an automatic pre-processing technique based on common average referencing to remove artifacts.
In addition to filtering, common pre-processing methods include data augmentation, normalization, and segmentation, as shown in Table 5. Due to the imbalance in sample size between seizure and non-seizure periods, data augmentation is required to compensate for the uneven distribution of data available at different stages of the EEG signal. In general, it is necessary to augment the data for seizure periods to balance the number of samples with different labels.
To mitigate class imbalance in EEG datasets, the synthetic minority oversampling technique (SMOTE) is widely employed (Alshaya and Hussain 2023; Patel et al. 2024). SMOTE generates synthetic samples for minority classes using linear interpolation and other ML techniques, thereby balancing the class distribution. Recent studies have refined SMOTE by integrating clustering and enhancement methods. Ahmad et al. (2024) combined K-means clustering with SMOTE to generate synthetic samples that better reflect the underlying data structure. Similarly, Du et al. (2023) proposed a mixture of experts model to pre-train on imbalanced data before joint training, improving learning efficiency. Although SMOTE augments the data, the quality of the generated data is questionable. In addition, SMOTE is more suitable for one-dimensional data, and the generated data may not reflect TF features.
Alternative approaches include Co-MixUp (Hu et al. 2024), which mixes samples to enhance data diversity, and positional/random data augmentation (Kamakshi and Rengaraj 2024), which synthesizes seizures of varying sizes and shapes. Ru et al. (2024) further improved data diversity using adversarial and mixed augmentation techniques, significantly expanding training samples. Shu et al. (2024) introduced DiffEEG, a novel diffusion-based data augmentation method designed to model EEG data distributions in a comprehensive way and generate highly diverse synthetic samples.
Table 5. EEG signal pre-processing techniques
Operation | Purpose | Methods | Equation |
|---|---|---|---|
Filtering | Noise removal | Notch filter | |
Kalman filter | |||
Bandpass filter | |||
IIR filter | |||
Butterworth filter | |||
Data augmentation | Balance samples | SMOTE | |
GAN | |||
Segmentation | Reduce complexity | Segmentation | Split x(t) into K epochs: |
Windowing | Apply window function w(t): | ||
Normalization | Reduce complexity | Min-Max | |
Z-score | |||
Robust scaling |
Segmentation is a method for simplifying EEG signal analysis by breaking long-term recordings into smaller, manageable segments based on time or event windows. Proper segmentation strikes a balance, where too long segments increase computational burden, while too short segments may lose critical ES biomarkers. An appropriately chosen segment length (typically 1 to 10 s) maintains detection accuracy while reducing processing demands, as supported by recent studies (Albaqami et al. 2023; Caffarini et al. 2022; Chung et al. 2024; Dutta et al. 2023).
Normalization scales data of varying magnitudes to a common range, ensuring uniformity and reducing the influence of outliers, an essential step when processing large datasets. In Omar’s study (Omar and Abd El-Hafeez 2024), standard scaling (Z-score normalization) outperformed MinMax scaling, significantly enhancing the accuracy of the models.
Downsampling is a pre-processing step for EEG data, which enhances training efficiency by reducing redundant signal information. Pan et al. (2023) explored downsampling as a novel approach to lower computational complexity in DL models. They introduced three methods, including direct, compressed, and convolutional downsampling, and directly fed the downsampled signals into neural networks for seizure detection. Their results showed significant computational savings without compromising detection accuracy.
Under moderate downsampling factors, several proposed methods not only reduced computational load but also improved seizure detection/prediction accuracy. For dimensionality reduction, Das et al. (2024) applied local linear embedding, multidimensional scaling, and t-distributed stochastic neighbor embedding (t-SNE), while Golla and Maloji (2023) used the reduced-dimensionality butterfly optimization technique to enhance classification performance.
Feature extraction
The feature extraction process of EEG signals is considered a critical stage in ES tasks. Properly selected features can effectively capture biomarkers associated with epileptic ictal or pre-ictal states while reducing the complexity of neural networks. Based on the implementation methodology, feature extraction approaches can be categorized into two types: manually engineered features and automated feature extraction via neural networks. Compared with manual extraction, automated feature extraction represents a distinctive capability of neural networks. However, due to the prevalent "black-box" nature and limited interpretability of neural network systems, this section emphasizes manually engineered features appeared in recent papers.
Recent studies in EEG-based epileptic detection/prediction leverage neural networks through three main technical frameworks.
End-to-end DL, where networks automatically extract and classify features from raw EEG signals.
Hybrid approaches that combine manually engineered features with neural network classifiers.
A multi-stage pipeline integrating automated feature extraction, ML-based dimensionality reduction, and subsequent classification.
Table 6. Summary of features under different categories
Feature category | Extraction methods |
|---|---|
Time-domain | Raw EEG (Abdallah et al. 2023), Phase space (Feizbakhsh and Omranpour 2023), event-related potential, peak amplitude, signal slope, mean absolute deviation |
Frequency-domain | Short time Fourier transform, continuous Fourier transform (Albaqami et al. 2023), discrete wavelet transform (Khalfallah et al. 2025), Fourier transform (Amer and Belhaouari 2024), empirical wavelet transform (Anita and Meena Kowshalya 2024), spike-and-wave discharges (Baser et al. 2022), empirical mode decomposition (Bhandari and Huchaiah (2022), Mel frequency cepstral coefficients (Dissanayake et al. 2022), Fast Fourier transform, variational mode decomposition (Wu et al. 2022), hilbert-huang transform (Sharma 2024) |
Spatial-domain | Recurrence plot (Goel et al. 2024), common spatial pattern (Amiri et al. 2023), independent component analysis (Salafian et al. 2023), granger causality (Rijnders et al. 2022), mutual information (Salafian et al. 2023), coherence, spatial entropy |
Nonlinear metrics | Sample entropy (Birajadar et al. 2024), Hjorth (Bhadra et al. 2024), spectral entropy, fuzzy entropy, multiscale entropy, fractal dimension, largest lyapunov exponent, kolmogorov complexity |
Statistical features | Mean value, variance, Kurtosis, Skewness, range, interquartile range, zero-crossing rate, covariance matrix, coefficient of variation |
[See PDF for image]
Fig. 8
Different types of features in epilepsy detection/prediction
Our review classifies epileptic features into six key categories as shown in Fig. 8, i.e., time, frequency, spatial, nonlinear, statistical, and spike-related features. Notably, time and frequency features dominate in applications because of their interpretability and effectiveness, while spike features remain underrepresented in the literature. Statistical features, typically derived as secondary measures from other feature sets, serve as a manually engineered input rather than as standalone descriptors. In contrast, spatial features, which increasingly leverage graph-based connectivity to model interactions between EEG channels, have emerged as a promising area, capitalizing on the inherent topological structure of brain networks.
1. Time-domain features
Characterizing the raw signal or its transformations is the most straightforward approach. The raw signal data are fed directly into the networks, with no signal processing expertise required to extract the signals (Abdallah et al. 2023). However, due to the limited characterization of raw signals, particularly in the domains of epilepsy detection/prediction, most of these methods demonstrate high accuracy in epilepsy detection but perform below the chance level in prediction (Shafiezadeh et al. 2024). Furthermore, given the one-dimensional nature of the time series, most papers employ sequence-sensitive models, such as 1-Dimensional convolutional neural network (1D-CNN) and long short-term memory (LSTM), for signal processing (Srinivasan et al. 2023; Quadri et al. 2024; Poorani and Balasubramanie 2023).
Due to the low discriminative power of raw EEG signals, some studies employ transformations to project them into higher-dimensional representations. Phase space reconstruction (PSR), for instance, converts one-dimensional time series into geometric trajectories, revealing underlying dynamical patterns. Recent innovations extend this approach, Feizbakhsh and Omranpour (2023) replaced traditional ellipse-based clustering with shape-agnostic clustering to enhance feature extraction, while (Goel et al. 2024) leveraged recurrence plots to map EEG signals into 2D recursive structures, subsequently applying transfer learning (TL) to derive discriminative features.
2. Frequency-domain features
The non-stationarity of an EEG signal means that its statistical properties (e.g., frequency components and amplitude) vary over time rather than being constant. Specifically, the EEG signal exhibits different frequency activities over time. This non-stationarity makes it difficult to fully capture the features by simple time-domain analysis or frequency-domain analysis. Therefore, the frequency-domain features that we discussed are basically TF ones. Our statistics show that most papers based on frequency domain features use the short-time Fourier transform (STFT) as a means of feature processing. Besides, WT and empirical modal decomposition (EMD) are also commonly used methods.
The STFT is the application of the Fourier transform to each frame after the signal is sub-framed. The basic idea is to consider the signal as approximately smooth over a short period of time so that its spectral properties can be analyzed (Abdulwahhab et al. 2024). Amer and Belhaouari (2024) proposed a TF transform called the forward-backward Fourier transform and used convolutional neural networks (CNNs) to extract meaningful features from TF images and classify brain diseases. Varlı and Yılmaz (2023) proposed a combined model using raw EEG signals and TF image transform of time-dependent EEG signals. In their paper, continuous WT and STFT methods were used to convert signals into images.
The continuous WT is a localized signal processing method particularly effective for analyzing transient events, such as those occurring during ES. By capturing non-stationary EEG characteristics, continuous WT provides a robust theoretical foundation for multiscale feature extraction, as demonstrated by Kaur and Gandhi (2023). Their approach leverages WT at varying scales and time intervals, enhancing discriminative feature construction. The discrete WT is a multiresolution analysis technique that provides TF localization by decomposing signals into subbands of varying frequencies using orthogonal wavelet basis functions. At its core, discrete WT employs filter banks to iteratively downsample and decompose the signal, enabling step-by-step multiscale feature extraction (Bhandari and Huchaiah 2022; Fawad Hussain and Mian Qaisar 2022; Halawa et al. 2022).
The wavelet packet decomposition (WPD) is an extended form of the discrete WT, which not only decomposes the low-frequency approximation coefficients, but also further decomposes the high-frequency detail coefficients to form a finer-grained band division and overcome the defect of insufficient high-frequency resolution of the discrete WT (Einizade et al. 2023).
The core concept of EMD is to decompose the original signal into multiple physically meaningful intrinsic modal functions (IMFs) through an iterative screening process (Bhandari and Huchaiah 2022). Each IMF satisfies the following conditions: the difference between the number of signal extremes and the number of zeros is not greater than one; the local mean at each point is zero; and the upper and lower envelopes have zero mean.
Using EMD decomposition, a complex signal can be decoupled into a finite number of IMFs, each of which represents vibration modes in the signal at different time scales with explicit TF localization properties (Srinath and Gayathri 2022). Das et al. (2024a, 2024b) decomposed the EEG signal into six IMF functions by EMD and three different features were manually extracted, that is, volatility index, variance and ellipse area of the second-order difference plot. And they were arranged in 1D and 2D form to be fed into the CNNs, respectively.
3. Spatial-domain features
Recently, graph theory has emerged as a promising tool for biomedical signal analysis, where signals are transformed into graph networks and represented as adjacency matrices or Laplacian matrices. However, as the size of time series increases, the dimensionality of transformation matrices expands accordingly, leading to heightened computational demands for analysis. Consequently, there is urgent need for efficient feature extraction methods with low computational time. Patel et al. (2024) introduced a novel feature extraction technique based on the Gershgorin circle theorem (GCT) for biomedical signals, termed Gershgorin circle feature extraction (GCFE). In the GCFE method, features are extracted from specially modified weighted Laplacian matrices in visibility graphs.
Mutual Information (MI) can also be used as biomarkers for ES. Hassan et al. (2022) selected relevant features by using an MI-based estimator to reduce the curse of dimensionality and improve accuracy. Most existing multi-modal approaches rely on feature fusion or decision fusion techniques but fail to incorporate cross-correlation learning. To address this limitation, Kumbam and Mary (2023) demonstrated that modeling cross-lead MI significantly improves the model’s performance. Their proposed cross-modal attention mechanism extracts discriminative features from multimodal EEG data to enhance classification accuracy while mitigating false positives.
Salafian et al. (2023) proposed a MI-based CNN-assisted factor graph for ES detection. Recognizing that soft estimates from neural MI estimators and CNNs fail to inherently model temporal dependencies between EEG blocks when used as combinatorial features, they instead employed these estimates to compute function nodes in a factor graph. This approach enables structured inference, leveraging probabilistic relationships between signal segments to enhance detection accuracy.
Rijnders et al. (2022) applied CNN algorithms to directional granger causality (GC) connectivity measurements. This also represents a connection-based feature approach.
Common spatial patterns (CSP) is a spatial filtering feature extraction algorithm that is capable of extracting the spatially distributed components of each class from within multichannel EEG data. The basic principle of the common spatial pattern algorithm is to find a set of optimal spatial filters for projection using diagonalization of matrices to maximize the difference between the variance values of the two classes of signals, thus obtaining a more discriminative feature vector. Amiri et al. (2023) proposed an algorithm that utilizes the sparse CSP and adaptive STFT based compressive transform to extract features from signals.
4. Nonlinear metrics
Nonlinear methods derive statistical entropy features from EEG signals, including sample entropy, spectral entropy, permutation entropy, and Shannon entropy. These features can be fed into an LSTM-based fully connected neural network to classify EEG signals into ictal and interictal states, as demonstrated in Goel et al. (2023), Assali et al. (2023).
Sharma (2023) proposed a hybrid approach combining higher-order statistics (HOS), sensitivity analysis, and residual WT. HOS explores nonlinear dynamics in high-dimensional spaces without loss of generality.
In the second module by Sun et al. (2024), a grayscale recurrence plot (GRP) was constructed to incorporate more nonlinear dynamic features than conventional recurrence plots. This GRP was then input into a residual-connected convolutional module to effectively learn nonlinear dynamic characteristics.
Assali et al. (2023) introduced a stability Index (SI) derived from a multivariate autoregressive model, quantifying dynamic phenomena during transitions between epileptic states. SI reflects the stability status of epileptic neural systems and was incorporated into a CNN when combined with conventional features. The integration of SI enhanced the model’s ability to learn discriminative EEG patterns, improving classification accuracy for epileptic states.
5. Statistical features
Some papers use statistical features extracted from multiple domains as model inputs. These features are more conducive to neural network modeling because they reflect the state of an EEG segment in multiple aspects. However, it is undeniable that the extraction of statistical features causes information loss, which may lead to a decrease in accuracy. Table 7 describes the main statistical features in detail.
Table 7. Statistical features and their equations
Feature | Equation |
|---|---|
Mean value | |
Variance | |
Kurtosis | |
Skewness | |
Range | |
Inter quartile range (IQR) | |
Zero-crossing rate (ZCR) | |
Covariance matrix | |
Coefficient of variation |
6. Spikes
Spikes are a prominent feature of ES, appearing as small peaks in EEG signals. Kumar et al. (2024) proposed an innovative DL approach for epileptic spike detection using spiking and non-spiking deep CNNs. They utilized CNNs and an adaptive Layered Adaptive Moments optimizer to efficiently extract relevant features from the time-domain) and frequency-domain representations of spiking and non-spiking signals.
In recent papers, multimodal data inputs are becoming increasingly common, where these inputs include features from multiple domains. Working with multimodal inputs, researchers can develop customized models to enhance the performance Cui et al. (2025).
Classification
A classifier is a concept in ML used to categorize input data into different classes or labels. For seizure detection/prediction frameworks that employ neural networks, the classifier is continuously trained through the backpropagation mechanism. However, traditional classifiers, such as SVMs, can also be used in the final decision-making stage Basha et al. (2024), which will be detailed in Sect. 4.8.3.
The specific implementation of classification processes through neural networks will be comprehensively detailed in Sect. 4.
Neural networks
In ES detection and prediction with EEG signals, neural network models offer significant advantages by overcoming the limitations of traditional methods. Through nonlinear feature learning and hierarchical representation, they effectively capture complex spatiotemporal EEG patterns. These systems often incorporate spatiotemporal feature separation modules, cross-modal attention mechanisms, or multi-branch feature aggregation strategies to build versatile, high-performance networks.
[See PDF for image]
Fig. 9
Nested donut chart of neural networks applied in epilepsy detection and prediction
This study systematically categorizes existing methodologies into four primary network architecture paradigms, including CNNs, recurrent neural networks (RNNs), AutoEncoders (AEs), and generative adversarial networks (GANs). Additionally, emerging architectures, such as transformer-based models, GNNs, and selected feedforward neural networks (FFNNs), have also gained significant research interest as shown in Fig. 9. The findings and comparative analyses of these approaches are detailed in dedicated subsections within this review.
Due to the intrinsic differences in feature scales and semantic interpretations of EEG data across temporal, spectral, and spatial domains, significant structural alignments exist between domain-specific features and particular neural network architectures: one-dimensional temporal features demonstrate high compatibility with 1D-CNN or RNN architectures; two-dimensional spectral representations (e.g., TF spectrograms) inherently align with 2D-CNN processing paradigms; spatial-topological features possess intrinsic compatibility mechanisms with GNNs.
Furthermore, this study synthesizes the key advantages demonstrated by various hybrid neural network architectures. A particular focus is placed on examining the fundamental trade-offs between interpretability and computational efficiency in these systems. These discussions are systematically structured into dedicated subsections to facilitate comprehensive analysis.
Finally, a subset of studies has leveraged cutting-edge advancements in neuro computing for seizure detection/prediction. These approaches are elaborated in Sect. 4.9.
CNN
The network architecture of CNNs typically consists of multiple convolutional layers, pooling layers, and fully connected layers. Convolutional layers capture local features by applying convolutional operations to input data and automatically learn hierarchical feature representations through multiple filters, a process particularly effective in detecting epileptic waveforms. Subsequent pooling layers reduce dimensionality via downsampling while preserving essential features, enhancing broader spatial feature extraction and alleviating model complexity and overfitting risks. As these layers process the data sequentially, deeper network layers progressively extract more abstract features, crucial for distinguishing normal brainwave patterns from epileptiform discharges. Finally, the fully connected layer integrates these high-level features to perform classification, determining the presence or absence of ES. Additionally, studies have confirmed that CNNs can autonomously extract and learn features from multichannel, time-sequenced EEG signals (Chetana et al. 2023). Figure 10 illustrates the accuracy distribution of CNNs across different datasets.
[See PDF for image]
Fig. 10
Violin plot of accuracy distribution for CNN-based seizure detection and prediction
1D-CNN
The 1D-CNN architecture is of interest due to its inherent applicability. The architecture simplifies the data processing pipeline by eliminating the need for pre-processing (Sameer and Gupta 2022). In addition, 1D convolutional layers are capable of extracting localized patterns in the input sequence. Their key strengths are the parameter sharing mechanism and the hierarchical feature learning capability. While preserving temporal dependence, they are also effective in capturing spatial correlations between EEG channels.
32 articles involved 1D-CNNs, including (Ahmad et al. 2024; Alshaya and Hussain 2023; He et al. 2024; Handa et al. 2024; Li et al. 2023; Poorani and Balasubramanie 2023; Qiu et al. 2023; Rivera et al. 2024; Sateesh Kumar Reddy and Suchetha 2023; Wang et al. 2023; Zaid et al. 2023). To enhance processing efficacy, researchers have designed diverse 1D-CNN models. Ahmad et al. (2024) developed a five-layer 1D-CNN capable of extracting both temporal and spectral features, combined with a multi-view forest-based ML classifier for final predictions. Alshaya and Hussain (2023) proposed two lightweight deep network models: a 1D multiscale neural network and an LSTM-integrated compact CNN. He et al. (2024) introduced an intelligent learning classifier incorporating heuristic strategies, which employs tunable Q-factor WT and 1D-CNNs for feature extraction, followed by parameter optimization to generate weighted fusion features for classification.
Researchers continue to develop innovative 1D-CNN architectures for both single-channel and multichannel time series analysis, with significant applications in EEG signal processing. Notable advances include a pyramidal 1D-CNN that enhances model generalizability through augmented EEG training data, and an end-to-end 1D atrous Conv-Net proposed by Handa et al. (2024), which integrates atrous convolutions with LSTM layers for automated epilepsy detection. Li et al. (2023) improved seizure prediction by combining 1D convolutions for dimensionality reduction with capsule networks to model spatial dependencies and instantiated features. Additionally, Wang et al. (2023) introduced a channel-incremental 1D-CNN framework optimized for iEEG-based seizure prediction. These advancements demonstrate the versatility and effectiveness of 1D-CNNs in extracting temporal and spatial features from complex biomedical signals.
Recent advances in seizure detection highlight the adaptability of 1D-CNN architectures for EEG analysis. Poorani and Balasubramanie (2023) implemented a pure 1D-CNN on the CHB-MIT database, demonstrating robust temporal feature extraction. For resource-constrained applications, Qiu et al. (2023) proposed LightSeizureNet, a lightweight model combining dilated 1D convolutions, global average pooling, and kernel-wise pruning to optimize efficiency for wearable devices. Rivera et al. (2024) introduced two spectrogram-compatible architectures: Network 1D Raw (1D convolutions on raw EEG) and Network 2D Conv (2D convolutions in spectrograms), both leveraging separable convolutions for computational efficiency. Further improving precision, Sateesh Kumar Reddy and Suchetha (2023) integrated fuzzy C-means clustering with CNNs to refine seizure signal classification. Zaid et al. (2023) demonstrated that modified EEG signal combinations yield superior classification performance, with lightweight-to-moderate DL models achieving competitive accuracy while minimizing computational resource consumption. Taking together, these approaches underscore the trade-offs between accuracy, computational cost, and deployability across clinical and ambulatory settings.
Despite their computational efficiency in seizure detection/prediction, 1D-CNNs face inherent constraints due to their reliance on raw time-series data. Performance is highly sensitive to input signal quality, making them vulnerable to artifacts and noise common in clinical EEG recordings. Additionally, their ability to detect subtle biomarkers is limited by the need for extended signal segments, as many pathological patterns only emerge over longer durations. This requirement complicates real-time deployment where low latency is critical.
TCN
The temporal convolutional networks (TCNs) are specifically designed for long-sequence modeling, utilizing stacked dilated causal convolutions to build a deep temporal feature extractor. Four papers involve TCNs, including (Darvishi-Bayazi et al. 2024; Georgis-Yap et al. 2024; Huang et al. 2024; Rao et al. 2024). Its architecture exhibits three key characteristics:
Causal structure The model enforces strict temporal causality by ensuring each timestep’s output depends solely on past inputs, effectively preventing future information leakage while enabling real-time online detection/prediction capabilities.
Exponentially expanding receptive field Through strategically designed dilation rates that grow exponentially across network layers, the architecture achieves efficient long-range dependency modeling while maintaining parameter efficiency.
Residual learning Each convolutional block incorporates skip connections to facilitate gradient flow throughout the deep network, simultaneously mitigating vanishing gradient issues and ensuring stable training dynamics.
Researchers continue to develop more sophisticated implementations of TCNs for EEG analysis. Georgis-Yap et al. (2024) introduced an advanced framework that combines temporal convolution modules with robust feature normalization. These architectures integrate multiple processing layers while maintaining stable training dynamics through careful regularization.
Recent work has produced specialized versions of TCNs that address specific challenges in seizure detection. Huang et al. (2024) enhanced the standard architecture with attention mechanisms that automatically highlight diagnostically relevant segments of EEG signals. Meanwhile, Rao et al. (2024) created a patient-adaptive version that not only processes temporal sequences efficiently but also personalizes its predictions based on individual patient information.
Current research focuses on further improving model efficiency and generalization capabilities to make them practical in diverse medical environments and patient populations. If the latency and computational resource adjustments associated with complex models can be resolved, it will be possible to develop more powerful diagnostic systems.
2D-CNN
The 2D-CNN is a deep learning model specifically designed for image or gridded two-dimensional data, utilizing convolutional kernels to extract local spatial features by sliding along the spatial dimension. Raw EEG signals can be transformed into two-dimensional time-frequency domain representations, such as STFT. These representations are then processed by 2D-CNN to capture spatial distribution features. The multi-level feature fusion capability of 2D-CNNs effectively models global evolutionary patterns in signal spectra while improving robustness to complex noise backgrounds. In contrast to deep CNN architectures, this subsection focuses on simple and shallow neural networks Aslam et al. (2022). Methods based on 2D-CNNs are the most popular, with a total of 163 papers included in our research.
Numerous studies have validated the rationality of the model through comparative experiments involving feature extraction and classification (Abderrahim et al. 2024; Sarvi Zargar et al. 2023). Comparisons covered include different CNN architectures and the effect of different feature inputs and prediction intervals on model performance.
Collaborative CNN architectures have improved recognition capabilities through multi-branch, multi-feature and multi-scale frameworks. Experiments have been conducted with different kernel sizes (Chung et al. 2024), different feature representations (Liu et al. 2024; Sabarivani and Ramadevi 2022), and different numbers of branches (Ouichka et al. 2022).
Attention mechanisms in the design of CNNs can improve the ability to discriminate between key inter-frequency information to recognize specific seizure patterns. These implementations have included multi-head attention (Ding et al. 2023; Yang et al. 2024), dilation convolution (Gao et al. 2022) and additive convolution based attention mechanisms (Huang et al. 2023).
Multi-view frameworks classify from diverse perspectives to improve seizure fragment identification. Einizade et al. (2023) developed fAttNet, fusing temporal multichannel EEG signals, WPD, and manually extracted features into three key views. To resolve dimensional conflicts between 1D EEG data and 2D-CNNs, Ibrahim et al. (2022) employed PSR as an alternative to spectrograms, preserving dominant signal trends via direct time-domain projections. Patel and Yildirim (2023) rasterized non-stationary 1D physiological signals into 2D images using Bresenham’s line algorithm for training 2D-CNNs, classifying neuro-spikes and seizure/non-seizure EEG with shape-based criteria. Kalita et al. (2024) designed EpiNET, integrating computationally efficient statistical, spectral, and temporal features to predict seizures 2 h before onset.
Studies addressing feature weight allocation across receptive fields. Jia et al. (2022) presented a variable-weight CNN by utilizing dynamic rather than static weights in convolutional and fully connected layers. The network adapts to input characteristics, outperforming standard CNNs in noisy conditions through superior classification accuracy, generalization, and robustness.
Architectural module modifications to enhance recognition rates remain a research focus. Peng et al. (2022) developed a streamlined EEGNet variant for seizure classification, augmenting its first convolutional layer with sinusoidal time-encoding and automated hyperparameter selection strategies. Sadam and Nalini (2024) proposed a hybrid CNN-SVM model employing continuous WT for signal-to-image conversion, with SVM replacement improving classification accuracy. Clinical applicability considerations have driven the development of models, such as the EEGWaveNet proposed by Thuwajit et al. (2022), which is an end-to-end framework that utilizes trainable deep convolution techniques for multi-scale feature extraction across EEG channels, eliminating the need for manual parameter tuning.
As the most classic pipeline, 2D-CNN has unique advantages in epilepsy detection/prediction tasks. However, efforts are still needed in network parameter optimization, multi-angle extraction, and inter-channel relationships.
3D-CNN
The 3D-CNN has outstanding advantages in processing multi-channel bioelectric signals. Compared with 1D-CNNs and 2D-CNNs, the increase in dimensions brings the model’s ability to perceive spatial relationships. Three-dimensional convolutional kernels better capture coupled spatiotemporal patterns (Liu et al. 2024). Here three papers, including (Liu et al. 2024; Qi et al. 2023; Xu et al. 2024), used 3D-CNN.
Lu et al. (2023) proposed a CBAM-3D CNN-LSTM model for seizure prediction. They combined Bi-LSTM with 3D-CNN for classification. Qi et al. (2023) used a 3D-2D hybrid CNN to extract spatial depth features from the EEG data, and the model can automatically use inter-channel correlation to improve seizure prediction. Xu et al. (2024) proposed a framework to shorten the detection delay through probabilistic prediction. They combined the STFT and 3D-CNN architectures to accurately capture the prediction probabilities. In addition, a correction weighting strategy is introduced to improve the prediction probability, and a cumulative decision rule is used to significantly reduce the detection delay.
Currently, there is limited research on 3D-CNNs, which is largely due to their high complexity. If 3D-CNNs could be incorporated as a small part of the model, such as the front-end extraction layer, it might enhance the model’s capabilities.
Deep CNN
Deep CNNs effectively model the multi-scale characteristics of epileptic EEG signals through their hierarchical architecture, where shallow layers capture high-frequency details like local signal variations while deep layers integrate low-frequency global patterns such as seizure periodicity. This inherent multi-scale processing capability aligns with the dynamic time-frequency-spatial evolution of epilepsy, enabling robust feature extraction even in challenging scenarios involving electrode displacement or subtle seizure manifestations. The networks’ deep architecture (typically comprising dozens to hundreds of layers) provides the necessary complexity to process large-scale inputs and learn discriminative patterns at varying temporal resolutions. Among the 341 papers we reviewed, 32 gave detailed descriptions of deep CNNs.
Deep CNNs face challenges such as high parameter dimensionality and low training efficiency in EEG-based seizure detection/prediction. To address these issues, researchers have proposed various optimization strategies, including lightweight architectures and advanced training techniques. For instance, Baser et al. (2022) introduced a sparse wavelet-detectable framework using CNNs, which uses non-invasive EEG with reduced channel counts to lower computational and memory requirements while maintaining detection performance.
Strategies have been explored to prevent gradient vanishing and overfitting. Alshaya and Hussain (2023) utilized residual (ResNet) modules to enable deeper training without vanishing gradients while improving generalization. These residual connections have proven particularly useful for analyzing long-term EEG signals, as they allow ultra-deep network structures without degradation. Several studies have adopted hybrid architectures, integrating well-established CNNs such as ResNet50 and VGG19 (Buldu et al. 2024; Cao et al. 2022; Narin 2022). Further enhancements include adding specialized layers to deep networks (Mekruksavanich and Jitpattanakul 2023) or combining multiple architectures (Borhade et al. 2024). Given the complexity of such models, transfer learning has been widely employed (Pattnaik et al. 2024; Thara et al. 2023; Song et al. 2022).
Multi-branch convolutional architectures have been used to capture cross-scale temporal-spectral features, accommodating the nonstationary nature of epileptic EEG. Input representations include nonlinear metrics (Ilakiyaselvan et al. 2022), time-frequency spectra (Tripathi et al. 2023), and statistical features (Islam et al. 2023). Attention mechanisms (particularly multi-head self-attention) have been integrated to enhance training efficiency and classification accuracy. For example, Gill et al. (2024) combined attention with a Deep CNN to classify multiple seizure subtypes using spectral entropy and wavelet coefficients.
To improve computational efficiency, researchers have developed compact models suitable for real-world deployment. Karnati et al. (2024) proposed a multiscale dilated CNN for seizure prediction under limited training data, while Li et al. (2023) introduced a lightweight CNN with uncertainty-aware training. Signal processing techniques such as discrete WT have been combined with medium-weight architectures (Nemati and Meshgini 2022), and Rout et al. (2022) implemented a model based on field programmable gate array (FPGA) using empirical WT for real-time diagnosis.
Recent work emphasizes hybrid and modular Deep CNN designs to optimize performance. Shafiqul et al. (2022) introduced a dynamic seizure detection framework integrating dense blocks, residual connections, and attention modules. Similarly, Zhu et al. (2024) developed a multiscale dense network with cascaded feature fusion for enhanced epilepsy detection. These approaches demonstrate the ongoing evolution of efficient and interpretable deep CNN architectures for EEG analysis.
Neural networks are evolving toward deeper and more complex architectures. In recent years, classical deep CNN models have shown success in epilepsy detection/prediction, reflecting this trend. However, like 3D-CNNs, deep CNNs still face challenges related to computational efficiency and real-time performance. By employing advanced techniques, such as novel modules, attention mechanisms, or optimized training schemes, researchers can enhance model performance in this domain, thereby making them more robust and competitive.
RNN
CNNs have been utilized to accurately identify irregular interictal discharges as non-ES; however, they may fail to detect ictal states or slower oscillations (Palanichamy and Basheer Ahamed 2022). RNNs are specifically designed to modal sequential data, capturing temporal dynamics through their cyclic structure to enable implicit memory and contextual associations. They serve as critical tools for time-dependent tasks such as analyzing epileptic EEG signals. Figure 11 shows the accuracy distribution of different RNN networks, and it can be seen that Bi-LSTM performs better.
Simple RNNs are not frequently mentioned in epilepsy detection/prediction tasks, with only 12 papers discussing them, including (Bhanja et al. 2023; Jibon et al. 2024; Jusseaume and Valova 2022; Li et al. 2023; Palanichamy and Basheer Ahamed 2022; Patro et al. 2024; Swami et al. 2024; Vidyaratne et al. 2022; Kabir et al. 2023). Most articles that utilize simple RNNs for recognition leverage their relatively lower complexity to address the identification challenges of multidimensional time-series data.
[See PDF for image]
Fig. 11
Violin plot of accuracy distribution for RNN-based seizure detection and prediction
Kabir et al. (2023) employed RNNs to evaluate the importance of individual EEG channels in epilepsy detection. Through sequential omission of each of the 14 channels, their study identified the most discriminative electrode configurations. Swami et al. (2024) proposed an expert system that combines a non-dominated sorting genetic algorithm with RNNs to detect ictal patterns. Their hybrid feature set integrated statistical metrics from PSR, singular value decompositions, and time-frequency energy distributions of wavelet packet coefficients.
To handle spatiotemporal EEG complexity, Vidyaratne et al. (2022) introduced a deep cellular RNN. This architecture used cellular automata-inspired parallelism to process multisource signals with spatial awareness, while parameter-sharing mechanisms ensured computational efficiency for high-dimensional inputs.
One problem with simple RNNs is that their structure is somewhat outdated. Additionally, due to their limited memory characteristics, they do not perform well when handling long-term EEG recordings.
LSTM
LSTM, as an advanced variant of RNN, introduces a core innovation through its cell state and three gating mechanisms (input gate, forget gate, and output gate), enabling precise control over critical information (Kannan et al. 2022; Shekokar and Dour 2022). LSTM, which is well-suited for time series analysis, has demonstrated excellent performance in EEG analysis. Kunekar et al. (2024) presented a comparative analysis of seizure detection using ML and DL techniques. Their findings demonstrated that conventional ML and DL algorithms achieved optimal performance when integrated with LSTM-based models, a key rationale for LSTM’s prominence in ES detection tasks.
70 studies have used the LSTM architecture to enhance its effectiveness in seizure detection/prediction. Key advancements include integrating convolutional operations to improve feature representation in EEG signals (Chalaki and Omranpour 2023), employing multiple stacked LSTM layers for hierarchical temporal modeling (Dash et al. 2024, 2025), and applying Bayesian optimization for hyperparameter tuning (Yadav et al. 2024).
Beyond processing sequential data efficiently, the LSTM’s gating mechanism mitigates gradient instability, enabling robust classification after feature extraction. To optimize classification performance, Mandal et al. (2022) conducted a systematic comparison of feature selection methods and classifier models. Further improving LSTM’s accuracy, Rani and Kavitha (2024) applied swarm decomposition to denoise input signals, isolating diagnostically relevant frequency components.
The adaptability of LSTMs has also facilitated end-to-end seizure prediction frameworks. Wu et al. (2023) developed an LSTM-based model that directly processes gamma-band EEG signals, eliminating the need for manual feature extraction. Similarly, Yu et al. (2022) combined handcrafted expert-defined features with deep-learned representations through a dedicated fusion module.
Given the subtlety of epileptic biomarkers, advanced architectures have been proposed to capture fine-grained anomalies. For instance, Zhang et al. (2024) introduced a position-aware multi-length and mutual-attention network, which integrates positional encoding of transient EEG abnormalities with multi-scale temporal dependencies using a hierarchical Residual Dilated LSTM. Meanwhile, Kulasinghe and Dissanayake (2023) evaluated various ML and DL approaches, including LSTM under low-data and high-noise conditions, benchmarking their performance against traditional spectral-based techniques.
Bi-LSTM
Bi-LSTM introduces bidirectional temporal modeling capability based on the standard LSTM, capturing global contextual dependencies in EEG series through parallel forward and reverse LSTM layers, enhancing contextual sensitivity and feature integration efficacy for epileptic events. 35 papers we reviewed clearly used Bi-LSTM.
Bi-LSTMs are often combined with CNNs to improve their feature acceptance. Chetana et al. (2023) compared the performance of Conv1D, Conv1D+LSTM, and Conv1D+Bi-LSTM. Zhao et al. (2024) connected a customized 1D-CNN and a Bi-LSTM layer to model temporal correlation. Furthermore, owing to the limited bidirectional capturing capability of Bi-LSTM, attention mechanisms can effectively enhance training speed. Dutta et al. (2024) designed a Bi-LSTM network integrated with multi-head self-attention mechanisms. Their results demonstrated the superior efficacy of this technique in detecting seizures from multichannel EEG recordings across multiple patients. Additionally, different strategies can be fused to meet the requirements of reinforcement learning, and these strategies can be used together with attentional mechanisms to enhance model performance (Tang et al. 2024; Prabhakar and Lee 2022). Bi-LSTM can also be combined with adversarial networks to enhance model generalization. Li et al. (2024) designed a semi-supervised model integrated GANs with Bi-LSTMs to enhance seizure prediction.
Due to the significant impact of the Bi-LSTM parameters on model training speed, Tuncer and Doğru Bolat (2022) proposed a classification method for EEG data using a deep neural network architecture. In the proposed method, a RNN-Bi-LSTM model was employed. Additionally, the study emphasized the importance of hyperparameter selection in Bi-LSTM models (optimization algorithm, initial learning rate, neuron count) and evaluated their influence on accuracy.
The LSTM-based pipelines including LSTM and Bi-LSTM have outstanding advantages in modeling time series. However, this type of time modeling is highly dependent on signal quality. In addition, the number of parameters introduced by the Bi-LSTM network structure also needs to be optimized, which requires further exploration by researchers.
GRU
The gated recurrent unit (GRU) is an efficient variant of RNNs, designed to achieve long-term dependency modeling capabilities comparable to LSTM with a more simplified architecture. Through the synergistic operation of the update gate and reset gate, the GRU balances computational efficiency and model performance in EEG analysis, particularly in real-time monitoring and resource-constrained scenarios. 17 papers mentioned GRU, including (Zhu et al. 2024; Basha et al. 2024; Bhadra et al. 2024; Cherukuvada and Kayalvizhi 2023; Kalita et al. 2024; Mekruksavanich and Jitpattanakul 2023; Omar and Abd El-Hafeez 2024; Prakash and Kumar 2023; Ru et al. 2024; Sadam and Nj 2023; Varun and Bollu 2024; Xu et al. 2024, 2023; Zhang et al. 2024; Lucasius et al. 2024).
The introduction of GRU aims to enhance the model’s temporal sensitivity. Xu’s study re-evaluated the temporal scope of the pre-ictal phase and subdivided it into multiple time windows (Xu et al. 2023). Lucasius et al. (2024) introduced a model combining modular Volterra kernel convolutional networks with Bi-GRU, augmented by phase-amplitude cross-frequency coupling features derived from scalp EEG.
Furthermore, ML integration has been shown to enhance GRU performance. A modified gorilla troops optimization-based feature selection algorithm was employed to identify optimal feature subsets, followed by a GRU model for the prediction of epilepsy (Cherukuvada and Kayalvizhi 2023). Golla and Maloji (2023) adopted a gated term memory unit recurrent network model to classify EEG signals as healthy or epileptic with high precision. This process leverages the recently developed ladybug beetle optimization algorithm to compute a solution-based logistic sigmoid function.
Recent improvements in GRU architectures have primarily focused on optimizing gate structures and enhancing integration with convolutional blocks. Prakash and Kumar (2023) proposed a modified GRU with an adaptive update gate mechanism that dynamically adjusts based on reset gate outputs, improving temporal dependency learning. For wearable EEG systems, Varun and Bollu (2024) developed WearEuroNet, which automatically optimizes the parameters of the convolutional layer (including depth, kernel size, and 1D convolution) while employing GRU layers connected without gaps to better capture the temporal dynamics of the EEG. More comprehensive approaches have also emerged, such as the multi-class feature fusion framework, which combines a hybrid CNN-GRU architecture with attention mechanisms for simultaneous spatial-temporal feature extraction in seizure detection and prediction (Zhang et al. 2024). This demonstrates how GRU enhancements contribute to end-to-end clinical analysis systems.
GRUs have difficulty capturing long-term EEG recording features and cannot adapt to different time windows, which limits its application. Improving the architecture to enhance channel perception capabilities could also enhance its competitiveness.
AutoEncoders
AEs are unsupervised learning models that reconstruct input data by learning low-dimensional representations of the data. In epilepsy detection/prediction tasks, AEs are widely employed to extract features from EEG signals and identify seizure events through reconstruction errors or generated features. As a variant of classical FFNNs, AEs align with the core definition of FFNNs in terms of their encoder-decoder architecture, unidirectional data flow, and training methodologies. Their unsupervised nature and diverse application scenarios further expand the scope of FFNNs. 11 studies mentioned AEs, including (Bhanja et al. 2023; Ghanimi et al. 2024; Caffarini et al. 2022; Lin et al. 2025; He et al. 2023; Hilal et al. 2022; Huang et al. 2022; Liang et al. 2024; Parija et al. 2024; Prabhakar and Lee 2022; Ramkumar et al. 2024; Wang et al. 2023, 2024).
Early studies primarily utilized basic AE models to extract EEG signal features, providing robust inputs for subsequent classification models. Bhanja et al. (2023) employed an autoencoder to reconstruct the signals, fusing the derived features with those of a 1D-CNN, and applied a Radial Basis RNN for epilepsy detection. Hilal et al. (2022) proposed a model based on an intelligent deep-canonical sparse autoencoder. Huang et al. (2022) developed a time domain feature extraction method based on AEs. They both elaborated on the architecture and mechanism, defining and computing specified features through AE-based signal reconstruction quantification, and validated the efficacy of the proposed detection framework via EEG recognition.
To enhance feature extraction and improve model robustness, researchers have developed diverse AEs for seizure detection and prediction. Ghanimi et al. (2024) proposed an innovative reconstruction approach using a generative adaptive adaptive encoder network incorporating the cramer distance and a temporal-spatial frequency loss function to measure divergence between probability distributions. He et al. (2023) introduced a data augmentation method based on a random temporal shift strategy within a residual temporal shift convolutional variational AE to mitigate indistinct feature extraction and overfitting while preserving information during unsupervised training. Further advancing representation learning, Liang et al. (2024) developed a novel double discrete variational AE network to achieve deep discretization of EEG signals. For early seizure detection, Caffarini et al. (2022) engineered two deep neural networks to encode 16 intracranial EEG features, while Cherukuvada and Kayalvizhi (2023) automated seizure detection through an improved crystal structure algorithm combined with a stacked autoencoder, optimizing feature selection and applying the arithmetic optimization algorithm for hyperparameter tuning.
Integrating AEs with novel network layers has improved recognition rates, particularly through convolutional features (Parija et al. 2024). Lin et al. (2025) applied an AE-based singular attractor reconstruction method to encode EEG signals, manually extracted statistical features from reconstructed signals, and employed K-nearest neighbors for seizure detection. Srinivasan et al. (2023) leveraged the automatic feature learning capability of a 3D deep convolutional AE to construct an ensemble framework. The label hybrid convolutional AE model exhibited optimal efficiency, integrating a Bi-LSTM classifier with 4-second EEG segments. Wang et al. (2024) introduced a semi-supervised hybrid architecture that integrates feature dataset with a convolutional autoencoder.
Researchers should pay attention to the computational resource consumption during the encoding-decoding process. Integration with other networks can also further enhance the performance of AEs.
GAN
GAN is a powerful deep learning model that synthesizes high-quality synthetic data through adversarial training between a generator and a discriminator. In epilepsy detection/prediction, GANs have been widely applied to address data imbalance issues, enhance model generalization capabilities, and boost performance. 8 studies applied GAN, including (Ganti et al. 2022; Gao et al. 2022; Ghanimi et al. 2024; Li et al. 2024; Narmada and Shukla 2023; Wang et al. 2024; Xu et al. 2022; Zhang et al. 2024).
One study investigated whether the synthetic data generated by temporal GANs (TGANs) could expand the sample sizes to improve the performance of deep learning classifiers based on sEEG with limited samples (Ganti et al. 2022). The results demonstrated that synthetic data generated by TGANs significantly enhanced the performance of Bi-LSTM networks in classifying thalamic seizure and baseline states, with the detection model’s accuracy improving by 18.5%.
Additionally, Gao et al. (2022) constructed a GAN to modify the imbalanced EEG data distributions by generating seizure-phase EEG samples to form a more balanced training set. Xu et al. (2022) further proposed a method for EEG-based epilepsy recognition using a triple GAN, which independently processes time, frequency, and TF domain features.
Since GANs can generate high-quality signals, they are also used for data augmentation. Zhang et al. (2024) investigated two approaches to address the challenge of cross-patient generalization in deep learning methods. First, they proposed a novel data augmentation technique, spatiotemporal EEG augmentation, to generate synthetic seizure-phase training data with spatiotemporal dependencies. Second, they developed a patient adversarial neural network that employs adversarial optimization between a feature extractor and an identity discriminator to retain only shared seizure-related features.
These applications not only improve model accuracy and robustness, but also effectively mitigate challenges posed by small sample sizes and imbalanced data through synthetic data generation and adversarial training mechanisms, enhancing adaptability and generalization capabilities in processing complex EEG signals. However, GANs require a large amount of resources for training. Whether it is suitable for clinical application still needs further exploration.
Regarding the hybrid approaches combining this architecture with other neural networks, we mentioned them in Sect. 4.8.
Transformer
The transformer is a sequence modeling architecture based on the self-attention mechanism, which completely discards traditional recurrent structures, such as RNN, LSTM, and GRU. Through global parallel modeling and hierarchical feature abstraction, it exhibits robust capabilities in capturing long-range dependencies and uncovering cross-temporal dynamic correlations in EEG time-series analysis. Notably, it demonstrates exceptional competence in processing ultra-long sequences and fusing multi-lead signals. 26 papers concerning transformers are included in our review.
One of the main features of transformers is its parallel computing capability, which ensures fast processing speeds and meets the real-time requirements of epilepsy detection/prediction. Busia et al. (2024) proposed a compact transformer architecture called EEGformer, optimized for efficient execution on a low-power MicroController unit (MCU). Their most responsive method achieves a response time of 13.7 ms. Rukhsar and Tiwari (2023) proposed a deep learning architecture based on a lightweight convolutional transformer. Their approach addresses the limitations of the vision transformer (ViT) in terms of translational isotropy and localization through convolutional tokens.
Currently, the method of enhancing transformers involves combining them with CNNs. Generally, convolutional layers are used to extract features and transformers are used for further classification. Deng et al. (2023) proposed a novel hybrid ViT architecture combining CNNs with transformers. Furthermore, they developed a learnable method for optimizing constraint coefficients in patient-specific loss functions and an EEG signal continuity-based uncertainty quantification approach for alarm evaluation in k-of-n continuous prediction strategies. Hu et al. (2023, 2024) designed hybrid transformer models. The enhanced transformer structure replaces conventional feedforward layers with dual half-step feedforward layers to strengthen nonlinear representations. Holguin-Garcia et al. (2024) conducted comparative analysis among CNNs, CNN-transformer hybrids, and traditional ML algorithms, demonstrating the effectiveness of CNN-transformer frameworks.
The transformer’s global search capability provides an effective means for comprehensive seizure feature discovery. Global spatial interactions across channels and long-term temporal dependencies play crucial roles in seizure prediction, necessitating extensive feature space exploration to learn generalized representations. Shi and Liu (2024) proposed B2-ViT-Net, a bi-level programming model for learning novel generalized spatiotemporal long-range correlation features that characterize global inter-channel interactions and extended temporal dependencies critical for seizure prediction. Yan et al. (2022) developed a transformer-based seizure prediction model incorporating STFT for time-frequency feature extraction, followed by a triple transformer tower architecture for feature fusion and classification, addressing sequence length limitations.
Several papers have made attempts at long-range EEG analysis and clinical applications (Tong et al. 2025; Sun et al. 2022; Wu et al. 2022). These attempts include the study of inter-channel relationships (Sun et al. 2022) and the identification of ES patterns (Tong et al. 2025).
As a novel architecture, transformers still need continuous improvement. Existing methods can be effectively enhanced by introducing breakthrough in the field. Regarding the hybrid approaches combining transformers with other neural networks, we mentioned them in Sect. 4.8.
GNN
The fundamental limitation of current deep learning approaches lies in their inability to effectively represent physiological EEG recordings, as their inherently irregular and unstructured nature resists adaptation to conventional matrix-based formats (Jibon et al. 2024). Furthermore, most ES detection/prediction methods neglect the nonlinear and implicit characteristics of EEG signals, which adversely impacts detection accuracy (Huang and Duan 2023). GNNs, however, can comprehensively leverage inter-channel relationships.
GNNs present a promising solution by inherently encapsulating relational data among variables. Representing interactive nodes as edge-connected entities with weights determined by temporal correlations or anatomical linkages, GNNs have garnered significant attention for their potential in modeling cerebral anatomical systems. Centered on graph-structured data modeling, GNNs explicitly capture spatial topologies and dynamic functional network properties of multi-channel EEG through nodes, edges, and message-passing mechanisms. This framework enables quantitative analysis of cross-regional neural synchronization anomalies and epileptic propagation pathways, offering a unique value for epileptic network localization and seizure propagation pattern inference. 24 papers applied GNN, and some of the papers are summarized in Table 8.
Table 8. Comparative analysis of recent papers in GNN-based seizure detection and prediction
Refs | Dataset | Feature extraction methods | Classifiers | Accuracy(%) |
|---|---|---|---|---|
Guo et al. (2023) | CHB-MIT | STFT | GCN | 96.7 |
Huang et al. (2022) | TUH | Graph regularization | Graph-regularized fuzzy broad learning system, GCN | 92.2 |
Jia et al. (2022) | CHB-MIT | Band energy, Hjorth parameters, Higher order crossings, 1-order difference, 2-order difference, Differential entropy, Fractal dimension (node), Correlation (edge) | GCN | 96.51±4.02 |
Jibon et al. (2024) | CHB-MIT, TUH | Line length, Auto-covariance and Auto-correlation, Periodogram | Sequential GCN, Deep RNN | 99.007 |
Jibon et al. (2024) | CHB-MIT | PSD | GAT, RBFN | 98.74 |
Li et al. (2022) | TUH | Spatial-temporal connectivity | GNN, CNN, Transformer | 91 |
Liu et al. (2025) | CHB-MIT | Connectivity (Cosine similarity, Gaussian kernel) | GCN, GNN | 95.52±0.47 |
Patel et al. (2024) | Bonn | Gershgorin circle | GNN | 98.333 |
Raeisi et al. (2022) | Helsinki | Mean squared coherence, phase-locking value, Euclidean distance | GCN | 99.1 (AUC) |
Wang et al. (2024) | Bonn | EEG signal graph representation | GNN | 99.9 |
Wang et al. (2023) | CHB-MIT | Cauchy principal value, Hilbert transform | GAT | 98.74 |
Wang et al. (2023) | CHB-MIT | Pearson correlation | GAT, Transformer | 98.25 |
Xu et al. (2024) | CHB-MIT | Pearson correlation | GCN, GRU | 97.35 |
Xu et al. (2024) | CHB-MIT | Pearson correlation | GCN, CNN | 88 |
A subset of studies explores static graph structures based on Euclidean grids. Dissanayake et al. (2022) developed a deep learning framework comprising two components: a classification GNN and a graph synthesis network. Feature extraction at the input layer employs Chebyshev graph convolution. Jia et al. (2022) proposed a universal Graph Convolutional Network (GCN) architecture for seizure prediction that addresses model complexity by framing the task as graph classification. The network includes graph convolution layers for single-hop node feature extraction, pooling layers for feature abstraction, and fully connected layers for classification. Raeisi et al. (2022) introduced a GCN-based model for automated neonatal seizure detection, embedding temporal information as graph signals in EEG representations alongside time-frequency features. Spatial information is encoded through functional connectivity metrics or ED graphs between EEG channels.
Low signal-to-noise ratio challenges in brain connectivity modeling hinder effective model learning. To address this, Liu et al. (2025) proposed a Graph Contrastive Denoising (GCD) module that enhances graph structure learning for small, densely weighted EEG graphs via combined pre-training and fine-tuning strategies. Wang et al. (2024) developed a sparse spectral GCN for EEG-based seizure classification. Their weighted neighborhood field graph representation reduces redundant inter-node edges while optimizing graph generation time and memory usage.
As research progresses, modeling dynamic functional connectivity has emerged as a critical frontier. Wang et al. (2023) proposed a channel-weighted transformer feature fusion network with multi-branch dynamic multi-graph convolution. This framework combines multi-branch feature extraction for joint temporal-spatial-spectral EEG characterization, dynamic multi-graph convolutional networks for high-level feature extraction, and channel-weighted transformers for multi-domain graph feature fusion. Guo et al. (2023) exploited inter-dependencies among multi-domain features to enhance seizure prediction, implementing a novel triplet attention layer for cross-dimensional interaction and a spatial dynamic GCN to dynamically model electrode spatial relationships.
The dynamic complexity of cerebral functional connectivity impedes conventional signal processing and ML methods in neurological diagnostics. Li et al. (2022) addressed this issue by proposing a graph generation network that dynamically identifies functional connectivity through sEEG analysis, thereby generating time-frequency brain connectivity graphs for diverse seizure detection. Recognizing insufficient attention to EEG complexity in existing methods, Xu et al. (2024) developed a dynamic functional connectivity neural network (DynFCNet) for seizure prediction. DynFCNet discovers dynamic brain connectivity patterns, generates dynamic functional connectivity graphs, and extracts features via GCNs while incorporating CNNs for feature extraction.
Novel GNN architectures continue to penetrate epilepsy detection/prediction research. Huang and Duan (2023) developed a hybrid approach using GCNs for feature extraction from undirected EEG-derived graphs, followed by graph fuzzy broad learning systems for ES detection. Jibon et al. (2024) leveraged GNNs’ capacity to exploit implicit neuroanatomical information through temporally or anatomically weighted node interactions, proposing a hybrid sequential GCN-Deep RNN framework for seizure detection. Li et al. (2022) introduced a subject-specific spatiotemporal-spectral hierarchical GCN with active preictal-interval learning. This framework constructs hierarchical graphs to characterize epileptic cortices across rhythms, extracts spatiotemporal dependencies via spectral-temporal CNNs and variable self-gating mechanisms, and integrates intrarhythm features through hierarchical graph convolution.
Graph attention networks (GATs) implement attention mechanisms to dynamically weight the contributions of neighboring nodes. Wang et al. (2023) proposed a synchronization-based spatiotemporal GAT, employing phase-locking values to model EEG channel connectivity to capture time-frequency correlations. Experimental validation demonstrates superior performance on benchmark datasets. To address the challenges of underutilized time-frequency information and cross-patient detection, Zhao et al. (2023) developed a hybrid attention network combining front-end GATs for spatial feature extraction and transformers for temporal modeling, augmented by focal loss to handle class imbalance.
GCNs leverage adjacency and degree matrices for neighbor information aggregation, effectively capturing local topological dependencies. Xu et al. (2024) designed an end-to-end GCN-Bi-GRU architecture for seizure detection. Wavelet-preprocessed pearson correlation matrices inform GCN-based spatial feature extraction before modeling via Bi-GRUs. Ge et al. (2024) proposed a multi-hybrid attention convolutional network with hierarchical spatiotemporal-spectral multi-hybrid attention mechanisms for cross-channel correlation analysis. Spatial dynamic GCNs complement this to model the multi-band spatiotemporal dynamics in epileptic cortices to capture optimal seizure biomarkers.
In recent years, research on epilepsy detection/prediction based on GNNs has become a hot topic. However, current research still needs more thorough argumentation to demonstrate that these methods are not simply a pile of technologies. Regarding the hybrid approaches combining GNNs with other neural networks, we mentioned them in Sect. 4.8.
FFNN
The core feature of the FFNNs is the unidirectional flow of data from the input layer to the output layer with no loops or feedback connections. 12 papers researched and analyzed multi-layer perception (MLP) (Benzaid et al. 2024; Das et al. 2024; He et al. 2023; Li et al. 2023; Srinivasan et al. 2023; Affes et al. 2022), extreme learning machine (ELM) (Zhang et al. 2023; Xue et al. 2022), and deep belief network (DBN) (Cherukuvada and Kayalvizhi 2023; Karunakar Reddy and Kumar AV 2022; Visalini et al. 2023; Saminu et al. 2024).
MLP, as a simple neural network, can intuitively and quickly access the importance of channels. A novel cascade algorithm integrating attention mechanisms and MLP architectures has been proposed for optimal EEG signal channel selection (Affes et al. 2022). Li et al. (2023) developed an end-to-end seizure prediction methodology based on MLP networks.
ELMs have been used for EEG detection to reduce parameters and training due to their single hidden layer structure, strict forward transfer of data, fixed or random hidden layer parameters, output layer computed by parsing methods, and no feedback mechanism. Zhang et al. (2023) proposed a deep learning based ELM on the principle of stacked generalization. Enhanced EEG knowledge is considered as a complementary component, which is then mapped to the next module.
DBNs are constructed through the stacking of multiple restricted Boltzmann machines (RBMs). During the pre-training phase, each layer undergoes unsupervised learning sequentially, while the fine-tuning phase may employ backpropagation. Cherukuvada and Kayalvizhi (2023) introduced a feature selection technique based on the oppositional aquila optimizer and DBN for ES detection using EEG signals. Karunakar Reddy and Kumar (2022) extract features from signals is performed through Haar DWT and spike detection. These features are subsequently processed by an enhanced DBN incorporating RBM layers, where parameter tuning of the DBN is executed via an Adam-based coyote optimization Algorithm for signal classification. Visalini et al. (2023) demonstrated a DBN-based ML architecture for binary classification of seizure and non-seizure episodes.
Hybrid neural networks
Traditional single-architecture deep learning models often face limitations in incomprehensive feature representation or insufficient temporal dependency modeling when processing such complex bioelectrical signals. By organically integrating neural network modules with complementary advantages, hybrid architectures demonstrate unique value, as they enable multiscale feature abstraction hierarchies that simultaneously capture the morphological characteristics of local EEG waveforms, global temporal evolution patterns, and inter-channel synergistic relationships. In particular for tasks such as seizure prediction that require joint attention to transient spike patterns and long-term rhythmic transitions, hybrid networks provide a multidimensional analytical framework that better aligns with clinical decision-making mechanisms.
This section reviews recent epilepsy detection/prediction methodologies based on hybrid neural networks, which is a supplement to the literature in the previous section. The section is organized as follows. Section 4.8.1 critically analyzes the widely adopted CNN and LSTM architecture, elucidating its time-frequency feature co-encoding mechanism. Section 4.8.2 explores innovative combinations of CNNs with other networks (e.g., transformers, GNNs) and their breakthroughs in multi-scale bioelectrical correlation modeling. Section 4.8.3 introduces fusion strategies between deep networks and traditional machine learning methods, dissecting how hybrid feature engineering enhances model interpretability and few-sample adaptation capabilities.
CNN and LSTM combination
The integration of CNNs and LSTM networks can effectively exploit time-frequency features in EEG recordings. Each model possesses unique strengths: CNNs excel in spatial feature extraction, whereas LSTMs are adept at capturing temporal dynamics, and hybrid models combine these complementary advantages (Hermawan et al. 2024).
In feature extraction, both models can comprehensively extract time-domain and frequency-domain features. Specifically, CNNs focus on frequency-domain features (or spatial correlations), while LSTMs extract time-domain patterns (Lu et al. 2022; Abdulwahhab et al. 2024). Ahmad et al. (2023) integrated a 1D-CNN with a Bi-LSTM based on truncated backpropagation through time to efficiently extract time-frequency information while lowering complexity. Ma et al. (2023) proposed a multi-channel feature fusion model that requires only minimal pre-processing. Here, CNN extracts spatial features, Bi-LSTM captures temporal dependencies, and an attention mechanism assigns channel weights to filter electrode channels with higher contributions for classification. Qiu et al. (2023) proposed a differential attention ResNet LSTM network, which employs ResNet and LSTM to capture spatial correlations and temporal dependencies, respectively.
From a dimensionality perspective, such hybrid models effectively leverage multimodal data inputs. Most current methods either convert EEG signals into spectrograms for 2D-CNN processing or segment 1D features into fragments for 1D-CNNs. These approaches are further constrained by neglecting the temporal relationships between sequential segments or spectrogram frames. To address this, Wang et al. (2023) proposed a dual-modal information bottleneck network for EEG seizure detection. This architecture extracts EEG features from both time-series and spectrogram dimensions, allowing information from different modalities to pass through the dual-modal IB, which compresses the most relevant features from each modality and shares only essential information.
Given the complexity of multi-layer architectures, attention mechanisms can enhance training efficiency. Abdallah et al. (2023) proposed a hybrid CNN-LSTM-SAT network incorporating a self-attention (SAT) mechanism. The study also documented parameter tuning processes for CNN and LSTM dimensions and optimizers. Their final network comprised four CNN layers and one LSTM layer. Lebal et al. (2023) introduced and evaluated epilepsy net, a suite of deep learning EEG signal processing tools for detecting seizure events without manual feature extraction. The model is represented by ResNet, inception, GRU, and convolutional block attention modules, respectively. Zhou et al. (2024) proposed a novel end-to-end model integrating attention mechanisms, TCN, and Bi-GRU for seizure detection. The model requires only artifact-filtered raw EEG signals, thereby eliminating the need for time-consuming feature extraction.
Hybrid models typically stack diverse network layers to achieve synergistic effects (Vaithilingam and Regulagedda 2024; Anita and Meena Kowshalya 2024; Pandey et al. 2023). Bhadra et al. (2024) combined CNNs with two GRU layers. Kaur et al. (2022) designed a 14-layer CNN combined with a 16-layer stacked CNN-LSTM architecture. Quadri et al. (2024) proposed a framework for seizure prediction by stacking CNNs and Bi-LSTM layers. The design employs 1D convolutional layers with exponentially varying filter lengths, followed by deep Bi-LSTM layers to form a densely connected feedforward structure. Mallick and Baths (2024) utilized 1D convolutional layers integrated with Bi-LSTM, GRU, and average pooling as a recurrent unit for feature extraction, followed by dense layers for EEG waveform classification. Sadam and Nj (2023) fused time-frequency features from CNN-GRU models with Hilbert marginal spectrum and EMD outputs using canonical correlation analysis for multi-channel EEG seizure detection.
Some studies apply CNNs and LSTMs to distinct framework components, typically using CNNs for feature extraction and LSTMs for classification. Liu et al. (2022) proposed a patient-independent seizure detection method in which a CNN extracts features, a Bi-LSTM captures temporal variations, and channel perturbation enhances generalization during training. Prasanna et al. (2023) trained a custom CNN on pre-processed EEG data for precise feature extraction, optimized features via exhaustive random forest selection, and employed Bi-LSTM for seizure detection. Rao and Sridevi (2023) developed a model using 1D-stacked CNN for feature extraction, fisher discriminant analysis for feature selection, and a tuned hybrid fuzzy Bi-LSTM optimized via a probabilistic dingo coyote optimization algorithm.
Other hybrid networks
In addition to CNNs and LSTMs, numerous research frameworks integrate hybrid networks. These hybrid architectures compensate for individual model limitations, improving overall robustness. He et al. (2022) combined GAT as the front-end to extract spatial features by using the topological relationships between EEG channels and a Bi-LSTM network as the back-end to mine temporal dynamics, enabling final decisions based on contextual states before and after the current moment. Lian and Xu (2023) combined a GNN with a transformer architecture. Their model comprises a graph-based component to uncover intrinsic relationships in multi-channel signals and a transformer to reveal heterogeneous correlations across channels. Wang et al. (2023) introduced a model, Gatformer, for prediction. By fusing GAT and transformer, this model synthesizes TF attention to extract EEG information through spatiotemporal interactions.
To holistically address long-term and short-term dependencies in EEG signals, many studies integrate CNNs with transformers. Ma et al. (2024) proposed a parallel dual-branch fusion network, combining the complementary strengths of CNNs and transformers. Yuan et al. (2024) introduced an architecture merging DenseNet and ViT with an attention fusion layer for seizure prediction. DenseNet extracts hierarchical features with parameter efficiency, while ViT provides global feature representation via self-attention. Zhang et al. (2023) noted that single-convolution models for seizure prediction are limited by convolutional operations’ focus on local features and inability to capture long-range EEG dependencies. They thus proposed a hybrid swin transformer-CNN model. The fusion of these blocks optimizes seizure-related information utilization, improving prediction performance. Capitalizing on CNNs’ proficiency in local features and transformers’ capacity for global context, another method was proposed by Zhao et al. (2023) based on interactive local–global feature coupling. Convolutional operations and self-attention separately extract local features and global representations, while a feature coupling block fuses these interactively.
Xia et al. (2024) proposed a hybrid LSTM-transformer model for seizure prediction using sEEG data. The model combined transformers’ long-range dependency modeling with LSTM’s strength in processing variable-length sequences. Raeisi et al. (2023) developed a class imbalance-aware and interpretable CNN-GAT model for seizure detection. The model combines temporal EEG information with spatial channel relationships via graph representations of multi-channel EEG segments. A 1D-CNN automatically generates features differentiating seizure and non-seizure periods in the time domain, while GAT employs attention mechanisms to highlight critical inter-channel connections and information flow.
Incorporating auxiliary networks into the core module can further improve model performance. Lopes et al. (2024) used a convolutional AE as a feature extractor and a Bi-LSTM layer as a classification layer for classification. In the work of Ilias et al. (2023), they added GRU modules to the CNN architecture for integration to weight the modal importance. The use of AEs enables gated units to better capture latent features. Khan et al. (2023) proposed a model integrating time-frequency features derived from EEG signals with key statistical attributes (mean, median, variance). These statistics are fused with compressed continuous WT images processed by an AE, then fed into an LSTM for classification. Mir et al. (2023) developed aBi-LSTM model using a deep convolutional AE for automated seizure detection.
Integrating GANs enables classifiers to learn broader feature representations. Wang et al. (2024) proposed a cross-subject seizure detection method through unsupervised domain adaptation. The feature extractor learns domain-invariant representations, improving generalization across new patients.
Neural networks with ML techniques
The integration of traditional ML classifiers into deep learning networks can enhance model performance. These incorporated structures include optimizers (Jaishankar et al. 2023), architectural designs (Pandey et al. 2022), feature selection mechanisms (Prabhakar et al. 2024), and training strategies (Wang et al. 2023). Relevant ML algorithms not only effectively improve a model’s autonomous decision-making capabilities but also avoid introducing highly complex network architectures. Compared to pure DL networks, such hybrid methods demonstrate improvements in computational efficiency.
In terms of feature selection, automated algorithmic feature selection can substantially reduce model inputs and computational loads. For example, Jaffar (2024) designed a novel hybrid optimization model called the alpha bat customized squirrel optimizer that selects optimal structures from extracted features while combining the standard jellyfish search algorithm with particle swarm optimization for epilepsy detection. Jain et al. (2023) employed particle swarm optimization algorithms for optimal feature selection combined with WT for classification. Jaishankar et al. (2023) proposed an adaptive grey wolf optimizer to learn and enhance discriminative features. Kapoor and Nagpal (2024) developed a hybrid Cuckoo-drongo optimization method that employs a tuned deep CNN classifier to identify and predict ES.
In addition, traditional gradient descent optimizers in neural networks suffer from a number of limitations, including susceptibility to local minima and reliance on static learning rates (Dhull and Singh 2022). New algorithms are proposed to address these issues to enhance model training efficiency (Pandey et al. 2022) and accurate prediction (Salini and Sowmy 2024).
In collective decision-making processes involving multiple neural networks, ML algorithms enable effective model selection. Ensemble learning, as a ML paradigm, enhances overall performance by aggregating predictions from multiple models. Hosseinzadeh et al. (2024) combined neural networks with ensemble learning to improve diagnostic accuracy for ES. Furthermore, the integration of ML algorithms into classification layers enhances model’s predictive performance (Sonawane and Helonde 2024).
Novel architecture
Certain neural networks (Sharma and Srivastava 2025a, b, 2024; Sharma 2024; Khan et al. 2024; Mittal et al. 2024), such as deep fuzzy networks, SNNs, encoder–decoder networks, etc., have gained prominence due to their distinctive architectures and notable advantages.
Fuzzy DL (FDL) synthesizes fuzzy logic and deep learning technologies into a hybrid learning framework. This architecture aims to enhance the performance of DL models in handling complex and imprecise data by capitalizing on fuzzy logic’s capacity to manage uncertainty and ambiguity. Khan et al. (2024) implemented a specially designed FDL architecture. Their methodology incorporated critical steps such as pre-processing and feature extraction, and the constructed FDL model exhibited promising results, achieving an accuracy of 92.57%. Additionally, to evaluate the efficacy of their proposed FDL, Khan et al. conducted comparative performance analyses against LSTMs and 1D-CNNs.
The hyperplane neural network (HNN) constitutes a specialized neural network type centered on the utilization of hyperplanes for executing classification and regression tasks. Hyperplanes, defined as subspaces that partition data into distinct categories or regions within high-dimensional space, form the core of HNN. This architecture combines the efficient learning capabilities of traditional neural networks with hyperplane concepts to enhance model transparency and processing efficiency. Ultimately, through the deployment of a hinged HNN, Mittal’s model demonstrated the capacity to detect diverse neurological brain conditions, including neurotypical states, epilepsy, and autism spectrum disorder (Mittal et al. 2024).
Discussions
This section primarily discusses the key technical issues emphasized in recent literature within this field, particularly focusing on how to achieve more accurate and robust epilepsy detection and prediction. Given the characteristics of EEG signals, it is essential to adopt appropriate processing methods to enhance network performance, evaluate models, and advance medical applications. Notably, as existing studies have prioritized overcoming specific practical challenges through continuous experimentation, a systematic summary of these efforts is warranted.
Channel selection
Since EEG signals are long-term multi-channel recordings, reducing sample complexity without compromising recognition accuracy remains a challenge. Channel selection poses a significant technical difficulty, yet lowering the number of channels to a few critical ones could substantially reduce computational costs. Smaller data volumes also facilitate the effectiveness of customized classifiers. Some researchers have suggested using merely two or three EEG channels to predict ES (Affes et al. 2022). A priori channel selection can be avoided by implementing predefined channel selection strategies. Hartmann et al. (2023) demonstrated the feasibility of automated seizure detection using dual-channel EEG data, which enables efficient and effective review of ultra-long-term recordings.
Amer and Belhaouari (2024) employed a channel selection approach based on mean correlation coefficients to further improve classification accuracy. Mutual information from information theory is also used to evaluate the discriminatory power of features for specific classes (Fawad Hussain and Mian Qaisar 2022).
For n channels, channel combinations can theoretically be evaluated. However, as n increases, computational costs grow exponentially. Thus, it is critical to explore pragmatic channel selection strategies that balance computational efficiency with high classification accuracy. To address this, a novel method combining dual CNN-LSTM networks with a channel reordering strategy was proposed (Wang et al. 2024).
Electrode positions in certain datasets exhibit specificity. For instance, in the CHB-MIT dataset, channel 21 may be selected for single-channel identification, channels 2 and 3 for dual-channel detection, and channels 1, 2, and 3 for triple-channel seizure recognition. The method proposed by Xiong et al. (2024) can integrate information from diverse EEG channels, supporting recognition tasks even with variable channel counts.
Feature fusion
EEG signals encompass diverse features, each revealing distinct aspects of brain activity. Examples include spike-and-wave discharges in the time domain, energy concentration within specific frequency bands in spectrograms, and inter-channel correlation dynamics. Leveraging multidimensional features has become a critical strategy to enhance model accuracy. Some researchers have combined heterogeneous features as inputs to networks, while others have adopted multi-view processing frameworks.
ES patterns may manifest across multiple data representations (or “views"), yet subtle anomalies within a single view might remain undetected. Multi-view feature fusion (Cao et al. 2022; Pan et al. 2022; Cui et al. 2022; Gao et al. 2022; Zhu et al. 2024; Zhang et al. 2024) synthesizes information from different perspectives, enhancing the detection of subtle pathological signals and improving model robustness. For EEG signals, these views typically originate from diverse feature descriptors, including time-domain, frequency-domain, inter-channel relationships, and nonlinear characteristics. Although multi-view fusion initially requires additional computational time to process and analyze disparate datasets, its long-term benefits, through reduced false alarm rates and missed detection cases, outweigh these costs by delivering more accurate and comprehensive diagnostic insights.
XAI (explainable AI)
XAI, an abbreviation for explainable artificial intelligence, represents a research direction in AI that aims to develop AI systems capable of providing transparent, clear, and comprehensible explanations for their decision-making processes. Traditional AI models, are often regarded as black boxes due to their opaque internal mechanisms, which hinder interpretability. This limitation restricts their application in epilepsy detection tasks, where high interpretability and trustworthiness are critical. The benefits of XAI lie in enabling systematic evaluation of model processes, with its primary contributions including improved understanding of feature structures, localization of epileptogenic zones, and quantification of feature importance (Ahmad et al. 2024).
t-SNE is a widely used nonlinear dimensionality reduction technique primarily employed for the visualization of high-dimensional data. By projecting high-dimensional data into lower-dimensional spaces (typically 2D or 3D), it facilitates enhanced comprehension of data structures (Ahmad et al. 2024; Duan et al. 2022).
Shapley additive explanations, a widely adopted model interpretation method rooted in the Shapley value from game theory, quantifies the contribution of individual features to ML model predictions. A key characteristic of Shapley additive explanations values is their capacity to provide additive explanations, wherein the contributions of all features collectively sum to the final prediction outcome (Amrani et al. 2024; Kaur and Gandhi 2023; Sánchez-Hernández et al. 2024; Huang et al. 2022; Murugan and Kameswaran 2024; Ding and Zhao 2024). These interpretive methods work to predict the importance of specific segments or brain regions and elaboration of specific features.
In addition the parameters of the framework were also investigated for the importance of explanatory channels and features, including the weight matrix of the attention mechanism (Baghdadi et al. 2023), the saliency heat map (Einizade et al. 2023), and the parameter variations of the time-frequency transformations (Gireesh et al. 2023).
Gao et al. (2023) developed a framework that measures similarity between input data and prototypes learned during training, utilizing these comparisons as evidence for predictions. Furthermore, NN’s interpretability arises from its layer-wise visualizability and interpretability, achieved through an architecture where learned weights correspond to signal-processing-derived features (e.g., frequency, subband, and spatial filters; Jemal et al. 2022).
To address challenges in localizing drug-resistant epileptic foci, Yang et al. (2023) proposed a transfer learning approach leveraging multimodal EEG. Their method revealed significant feature distinctions between epileptic and non-epileptic channels.
Table 9. Comparison of recent papers concerning lightweight and hardware implementation
Refs | Features | Neutral Network | Hardware or network size | Accuracy(%) | Latency |
|---|---|---|---|---|---|
Li et al. (2025) | DWT | CNN, Transformer | 0.87M | 98.54 | - |
Song et al. (2023) | Raw EEG | CNN (Recursive State-Space Neural Network) | 0.68M | 99.79 | less than 25 ms |
Wang et al. (2024) | Raw EEG | CNN, Bi-LSTM | 9371 | 98.52 | - |
Wang et al. (2023) | Raw EEG | CNN | 88K | 98.7 | - |
Zhao et al. (2022) | Raw EEG | CNN | 45K | 99.81 | - |
Zhao et al. (2022) | Raw EEG | CNN (AddNet-SCL) | 0.12M | 94.9 | 138.6M CPU Cycles |
Zhou et al. (2024) | Welch power spectral density estimation | CNN (LMA-EEGNet) | 2471 | 95.71 | - |
Beeraka et al. (2022) | STFT | CNN, Bi-LSTM | 904 LUTs(FPGA) | 97.2 | - |
Chung et al. (2024) | Raw EEG | CNN | - | 97.05–100 | 2.1–3.4s |
Indira Priyadarshini and Reddy (2023) | CWT | CNN | 898 LUTs(FPGA) | 99.48 | 6.1ms |
Li et al. (2022) | Statistical parameters | Parallel CNN | (CMOS) | 99.01 | 1.13 |
Saidi et al. (2023) | Raw EEG | CNN (VGG-16, ResNet-50) | Zynq-7000 | 97.75 | 100.64ms |
Lightweight and hardware implementation
To meet the requirements of local deployment and rapid detection, lightweight design has become an inevitable trend for detection tasks. Lightweight neural networks can be easily implemented on hardware devices for real-time epileptic EEG detection. Our survey reveals that most models are based on CNNs, probably due to their network simplicity and relatively few parameters. These networks are particularly suitable for use on low-cost, computationally constrained hardware devices, including wearable technology and clinical applications, due to their few training parameters. Methods to reduce network parameters include the use of adder-accelerated convolution (Zhao et al. 2022) and depthwise dilated separable convolution (Zhou et al. 2024).
To enable clinical applications and bridge the gap between models and medical implementations, several scholars have investigated hardware-oriented model realizations, as is shown in Table 9. Given the limited programming languages supported by hardware devices and inherent computational resource constraints, localized deployment challenges inevitably arise during hardware implementation. Current research primarily addresses two critical issues: achieving high-accuracy recognition under wearable device constraints and reducing computational complexity.
Hardware implementations include CMOS-based architectures (Li et al. 2022) and FPGA embedding (Beeraka et al. 2022; Indira Priyadarshini and Reddy 2023). CMOS designs benefit from lower power consumption and faster response due to closer proximity to the underlying layers. Customised designs for FPGAs offer greater flexibility compared to computers.
Framework migration
The generalizability of the framework facilitates the achievement of more general epilepsy recognition. Alharthi et al. (2022) aimed to investigate the feasibility of integrating locally acquired EEG signals from the Epilepsy Center of King Abdulaziz University Hospital into the CHB-MIT dataset by applying a novel data integration compatible framework. Assim et al. proposed a method to achieve high precision using EEG signals acquired from any device method for generalized epilepsy diagnosis. Due to the variability of the data, commonly used methods of independent customization for individual patients often perform poorly in other subjects. Multi-subject modeling extends the scope of patient-specific modeling to incorporate data from specialized patient cohorts, thereby providing a useful, albeit relatively low, level of generalization (Jemal et al. 2024). In addition, knowledge distillation helps the framework migrate between datasets (Singh et al. 2024). Baghersalimi et al. (2024) enhanced the generalizability of the system by implementating decentralized learning strategies and the utilization of adaptive ensemble learning techniques. In light of the application scenario, knowledge distillation techniques were employed during the deployment phase.
IoT related researches
The IoT facilitates the management of remote patient healthcare monitoring systems. IoT devices have been developed to collect patient data and transmit it to computer programs for review by medical professionals. Intelligent seizure detection models should be compatible with an IoMT (internet of medical things)-cloud-based smart healthcare framework (Patro et al. 2024). The advent of cloud and edge computing has given rise to an interface between local detection and cloud-based identification, thereby fostering the development of portable EEG detection and diagnostics (Qiu et al. 2022).
In the contemporary era, the advent of IoT-assisted technologies has precipitated the exploration of the potential of cloud and fog computing services. These technologies have been shown to leverage deep learning to provide solutions for neurological disorders (Singh and Malhotra 2022).
Epileptogenic zone localization
The epileptogenic zone (EZ) localization task identifies the specific brain region where seizures originate, with high-frequency oscillations (HFOs) serving as its biomarker. Research demonstrates that removing the brain region containing HFOs correlates with post-operative seizure freedom (Weiss et al. 2018; Zweiphenning et al. 2022). Currently, in clinical practice, EZ localization still primarily relies on physicians’ visual interpretation using imaging techniques like EEG. This approach is relatively inefficient and highly dependent on the physician’s individual expertise (Yaqing et al. 2025).
In the epilepsy detection task, EZ localization serves as a component for evaluating the effectiveness of their network. Due to its high correlation with brain regions, its evaluation metrics include spatial specificity, dispersion, and coverage (Sun et al. 2024). In recent years, researchers have proposed novel methods for EZ localization. These measures include utilizing the shapley method to visualize EZ (Amrani et al. 2024), employing feature extraction methods such as higher-order statistics (Sharma 2023), maximum-mean discrepancy (Amirshahi et al. 2023), adopting novel network paradigms like geometric deep learning (Dissanayake et al. 2022), and multimodal learning (iEEG and sEEG) (Yang et al. 2023).
However, in epilepsy prediction/detection tasks, the EZ localization task is more often treated as a secondary objective. Therefore, developing neural networks that can model the brain’s spatial patterns for more precise detection of EZ holds greater potential for practical applications.
Patient-independent detection/prediction
Patient-independent seizure detection represents a significant challenge in the current research landscape. Although the majority of existing methodologies fall under the category of patient-specific approaches, these methods typically exhibit strong reliance on individual patient data. Due to the pronounced heterogeneity observed in EEG signals across patients and the distinct variability in epileptogenic zone localization, existing models applied directly across patients often necessitate recalibration. This process requires labelled, personalized datasets and proves labor-intensive and resource-demanding (Cui et al. 2023). Consequently, the development of robust patient-independent seizure detection methodologies holds substantial potential for enhancing the generalisability and practical utility of detection systems.
To address this challenge, researchers are actively exploring novel solutions. For detection tasks, given the significant differences in features between interictal and ictal periods, robust pipelines can be achieved by selecting sufficiently prominent features through channel selection (Singh and Malhotra 2022). Additionally, hybrid neural networks (Zhao et al. 2023) and channel perturbation techniques (Liu et al. 2022) can enhance generalization capabilities. For prediction tasks, transfer learning (Sarvi Zargar et al. 2023; Hu et al. 2023) is considered a universal strategy for boosting model generalization. Since brain channels represent prior signal features to varying degrees, calibration mechanisms (Shafiezadeh et al. 2024) and knowledge distillation (Wu et al. 2022) are also applied to patient-independent approaches.
While these studies provide valuable contributions, the advancement of current patient-independent detection methods remains subject to several critical limitations. These include the scarcity of annotated datasets, the high degree of heterogeneity in patient-specific seizure foci and ictal expression patterns, and the complexity associated with evaluating and validating model generalisability. Consequently, focused research efforts are urgently needed to devise comprehensive solutions to these persistent challenges.
Formal approaches for verifying AI pipelines
Due to the lack of clear, unified datasets or testing standards, there is doubt as to whether AI pipelines truly achieve the stated level of model performance. In order to rigorously prove the superiority of the stated methods, many papers have used strict mathematical methods and training strategies, as illustrated in Fig. 12.
[See PDF for image]
Fig. 12
Workflow for mathematical verification and cross-validation strategy in model evaluation
In terms of training strategies, common ones include holdout (Alshaya and Hussain 2023; Kode et al. 2024), K-fold (Abdallah et al. 2023; Abdulwahhab et al. 2024; Chung et al. 2024), and leave-one-out cross validation (LOOCV) (Esmaeilpour et al. 2024; Gao et al. 2022; Pandey et al. 2023; Salafian et al. 2023). The holdout method splits data into fixed training and test sets (e.g., separated by 70% and 30%) and is efficient for large datasets, but sensitive to random sampling bias. The K-fold mitigates this by partitioning data into K subsets, iteratively using one fold for testing and the rest for training, providing a robust performance estimate. LOOCV, an extreme case of K-fold where each sample serves as the test set once, minimizes bias but is computationally expensive, making it suitable only for very small datasets.
For rigorous AI validation, methods that enhance persuasiveness include ablation experiments, training curves, and t-SNE. For rigorous validation of AI models, ablation experiments systematically remove/modify components to isolate their contributions, revealing the true drivers of performance. Training curves track loss and metrics across epochs to detect overfitting, convergence issues, or insufficient data. t-SNE visualizes high-dimensional embeddings, exposing whether learned representations align with clinical distinctions.
Others
Hyperparameter optimization can further improve the performance of the proposed model. Anita and Meena Kowshalya (2024) proposed a hybrid algorithm called adaptive spider monkey black widow optimization (ASMBWO). Experimental results show that ASMBWO improves the average F1 score in the epilepsy classification task by 12.3% compared to a single optimization method, highlighting its adaptive superiority.
The sensitivity of medical data necessitates rigorous privacy preservation. To address vulnerabilities in EEG signal confidentiality against theft and tampering in open-network environments, Ein Shoka et al. (2023) designed a dual-phase encryption framework. Phase I employs chaotic Baker mapping for pixel-level scrambling of EEG signals to disrupt spatiotemporal correlations. Phase II utilizes Arnold transform for periodic diffusion in the frequency domain to enhance ciphertext irreversibility. He et al. (2023) further proposed anonymous synthetic data generation (ASDG), leveraging GANs to produce distributionally equivalent but untraceable synthetic data. This approach reduces patient re-identification risks by 83.5% in data-sharing scenarios.
In addressing the challenge posed by the scarcity of labelled data, scholars have proposed a range of strategies to facilitate small-sample learning. These include the correction of missing values (Liu et al. 2024), the utilization of inter-channel mutual information (Sami Nafea and Ismail 2024), multimodal integration (Yang et al. 2023), and fuzzy classification (Yang and Li 2024). These methods effectively utilize EEG information in cases where a small number of labelled samples are available.
Conclusion
Significant advancements have been achieved in automated epileptic seizure detection and prediction technologies recently, with performance metrics of SOTA algorithms exceeding 95% accuracy (229 of 341 papers). The application of neural networks has established novel methodological frameworks for end-to-end seizure detection and prediction research. Continued innovations in neural network theory and architecture are anticipated to enhance the interpretability of such methods, thereby facilitating more rigorous investigation into the neurodynamic mechanisms underlying epileptic seizure.
Research findings The present review focuses on systematically reporting neural network applications and their trends in epilepsy detection and prediction over the past 3 years. In this systematic review, 341 papers are summarized, compared, and discussed using the PRISMA framework. In particular, the review hopes to provide readers new to the field with inspiration to get started. Our research findings are summarized as follows.
Epilepsy detection/prediction has become a prominent application platform for neural networks, demonstrating strong potential as a research hotspot.
CNNs remain the most frequently used architecture. However, transformer and GNN-based approaches are gaining increasing popularity. Convolutional layers can be effectively integrated with diverse neural networks to enhance task performance.
Different features can be processed through different network types. Traditional manual feature extraction methods are being progressively replaced by automated approaches like CNNs or AEs.
Integrating the advantages of different neural networks can further improve pipeline performance. Multi-branch networks for multi-domain feature extraction represent a promising research direction, with CNN-LSTM hybrids being the most widely adopted combination.
Emerging network architectures frequently employ epilepsy detection/prediction as downstream tasks to evaluate their capabilities. Customized networks have become a preferred approach for technical innovation, marking a shift in mainstream methodologies from machine learning to deep learning.
Accuracy is no longer the sole evaluation metric, as researchers now optimize systems through multidimensional assessments.
Complex frameworks inherently suffer from overfitting and training challenges (Amiri et al. 2023). Simplifying computational processes remains a key research focus, including linear classifiers (Liu et al. 2025), reduced network complexity (Li et al. 2025), novel EEG representations (Wang et al. 2024), sparse electrode usage (Varnosfaderani et al. 2024), cascaded algorithms (Li et al. 2024), fixed-band approaches (Karnati et al. 2024), and weight learning (Cui et al. 2025).
Noise in the dataset itself can also cause problems. These include data labeling errors (Liu et al. 2025) and differences between patients and electrode placement (Zhao et al. 2024). Processing uncertainty in EEG signals using fuzzy logic can also enhance feature learning (Sharma and Srivastava 2025a, 2025b).
Dataset bias can also make the method less convincing, primarily referring to representativeness and scale. For example, the CHB-MIT dataset consists solely of pediatric patients (Zhang et al. 2024). Some validation conclusions were drawn from smaller datasets and require further evaluation in larger datasets (Ramkumar et al. 2024).
Cross-patient seizure detection will be a key approach (Zhao et al. 2022). Data differences between patients often pose a major challenge for cross-patient seizure detection (Li et al. 2025), particularly when based on raw EEG data (Zhu et al. 2024). Of course, if the training process itself is easily achievable or end-to-end, personalized networks would also be more precise (Yang et al. 2024). To address patient-specific issues, transfer learning strategies have proven effective (Darvishi-Bayazi et al. 2024; Shi and Liu 2024), as they can fully leverage external data (Lopes et al. 2024) and allow model parameters to be updated in real-time without the need for repeated training (Meng et al. 2025).
Due to the large amount of neural network parameters, many of the models are difficult to deploy on wearable devices, thus it is necessary to investigate lightweight networks. Research into reducing the complexity of neural network models through channel pruning will help to achieve applications on wearable devices (Ramkumar et al. 2024). For example, EEGformer has been deployed and verified on Apollo4 MCU and two processors (Busia et al. 2024).
Cloud-based systems facilitate the implementation of new models for predicting epileptic seizure and enhance the real-time application of healthcare systems (Ahmad et al. 2024). Patro et al. (2024) envisioned a process where an automated detection system on the cloud assists doctors in decision-making. Combining epilepsy framework conclusions with electronic health records is also a direction for application (Khurshid et al. 2024).
Information loss during feature extraction makes it more difficult to capture complex spatiotemporal patterns (Cheng et al. 2024). Additionally, networks that focus on small abnormal features are susceptible to changes in EEG recording settings or patient-specific factors (Zhang et al. 2024). Therefore, mining long-term EEG dependencies can help better capture seizure patterns (Tang et al. 2024). Increasing the number of network branches that recognize different features also enhances the model’s feature extraction capability (Ma et al. 2024).
Most deep learning models have an inherent “black box” nature, limiting their interpretability and robustness (Li et al. 2025; Wang et al. 2023). Network coefficients help explain epilepsy features. Wang et al. (2024) visualized learnable weight vectors at different frequencies and found that weights were higher in the 60–100 Hz frequency band, consistent with medical research. Shi et al. visualized the attention weights between channels and found that channels in the epileptic focus may exhibit strong abnormal connections with other channels (Shi and Liu 2024). The coefficients of GNNs are also used as important evidence of the degree of inter-channel association (Raeisi et al. 2023). Methods to enhance interpretability also include Shapley (Khan et al. 2024). It is worth noting that clinical doctors’ opinions should also be included in interpretability assessments (Sánchez-Hernández et al. 2024).
The performance metrics of the network largely depend on the number of patients and individual differences (Xu et al. 2024), meaning that there is no unified assessment standard for evaluating the effectiveness of the network. If authoritative institutions or journals could publish a challenging standard, it would make subsequent research more convincing. In addition, the real-time performance of deep learning is rarely evaluated, which is crucial for medical applications (Omar and Abd El-Hafeez 2024).
Author contributions
Y.Wu and L.Lu contributed equally to this work. Y.Wu contributed to the conceptualization, performed literature search and review, conducted statistical analysis, and wrote the original draft. L.Lu. was involved in manuscript revision, performed statistical analysis, and was responsible for visualization and typesetting. A.Xu carried out literature screening and assisted in statistical analysis. Y.Wang contributed to conceptualization, supervised the research direction, participated in manuscript review and editing, and performed validation. Z.Li conducted manuscript review and editing. Z.Yang wrote the medical applications section. L.Zeng. wrote the AI methodology and related descriptions. Q.Li was responsible for project administration. All authors have reviewed and approved the final manuscript.
Funding
This work was supported by the National Natural Science Foundation of China (Grant Nos. U23A20322, 62304254, and 62404253). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Data availability
No datasets were generated or analysed during the current study.
Code availability
Not applicable.
Materials availability
Not applicable.
Declarations
Conflict of interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Ethical approval and consent to participate
Not applicable
Consent for publication
All authors have reviewed the final manuscript and approved its submission for publication.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
Abdallah, T; Jrad, N; Abdallah, F; Humeau-Heurtier, A; Van Bogaert, P. A self-attention model for cross-subject seizure detection. Comput Biol Med; 2023; [DOI: https://dx.doi.org/10.1016/j.compbiomed.2023.107427]
Abderrahim, N; Echtioui, A; Khemakhem, R; Zouch, W; Ghorbel, M; Ben Hamida, A. Epileptic seizures detection using iEEG signals and deep learning models. Circuits Systems Signal Process; 2024; 43,
Abdulwahhab, AH; Abdulaal, AH; Thary Al-Ghrairi, AH; Mohammed, AA; Valizadeh, M. Detection of epileptic seizure using EEG signals analysis based on deep learning techniques. Chaos Solitons Fractals; 2024; [DOI: https://dx.doi.org/10.1016/j.chaos.2024.114700]
Acharya, UR; Hagiwara, Y; Adeli, H. Automated seizure prediction. Epilepsy Behav; 2018; 88, pp. 251-261. [DOI: https://dx.doi.org/10.1016/j.yebeh.2018.09.030]
Affes, A; Mdhaffar, A; Triki, C; Jmaiel, M; Freisleben, B. Personalized attention-based EEG channel selection for epileptic seizure prediction. Expert Syst Appl; 2022; [DOI: https://dx.doi.org/10.1016/j.eswa.2022.117733]
Ahmad, I; Wang, X; Javeed, D; Kumar, P; Samuel, OW; Chen, S. A hybrid deep learning approach for epileptic seizure detection in EEG signals. IEEE J Biomed Health Inf; 2023; [DOI: https://dx.doi.org/10.1109/JBHI.2023.3265983]
Ahmad, I; Zhu, M; Liu, Z; Shabaz, M; Ullah, I; Tong, MCF; Sambas, A; Men, L; Chen, Y; Chen, S. Multi-feature fusion based convolutional neural networks for EEG epileptic seizure prediction in consumer internet of things. IEEE Trans Consumer Electron; 2024; [DOI: https://dx.doi.org/10.1109/TCE.2024.3363166]
Ahmad, I; Liu, Z; Li, L; Ullah, I; Aboyeji, ST; Wang, X; Samuel, OW; Li, G; Tao, Y; Chen, Y; Chen, S. Robust epileptic seizure detection based on biomedical signals using an advanced multi-view deep feature learning approach. IEEE J Biomed Health Inform; 2024; 28,
Albaqami, H; Hassan, GM; Datta, A. MP-SeizNet: a multi-path CNN bi-LSTM network for seizure-type classification using EEG. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.104780]
Alharthi, MK; Moria, KM; Alghazzawi, DM; Tayeb, HO. Epileptic disorder detection of seizures using EEG signals. Sensors; 2022; [DOI: https://dx.doi.org/10.3390/s22176592]
Alshaya, H; Hussain, M. Classification of epileptic seizure types using multiscale convolutional neural network and long short-term memory. Mathematics; 2023; [DOI: https://dx.doi.org/10.3390/math11173656]
Alshaya, H; Hussain, M. EEG-based classification of epileptic seizure types using deep network model. Mathematics; 2023; [DOI: https://dx.doi.org/10.3390/math11102286]
Alvarado-Rojas, C; Schulze-Bonhage, A; Navarro, V; Valderrama, M; Sales, F; Klatt, J; Ihle, M; Adam, C; Witon, A; Dourado, A; Schelter, B; Neufang, M; Timmer, J; Feldwisch-Drentrup, H; LeVanQuyen, M; Teixeira, C. The Epilepsiae database: an extensive electroencephalography database of epilepsy patients. Epilepsia; 2012; [DOI: https://dx.doi.org/10.1111/j.1528-1167.2012.03564.x]
Amer, NS; Belhaouari, SB. Exploring new horizons in neuroscience disease detection through innovative visual signal analysis. Sci Rep; 2024; [DOI: https://dx.doi.org/10.1038/s41598-024-54416-y]
Amiri, M; Aghaeinia, H; Amindavar, HR. Automatic epileptic seizure detection in EEG signals using sparse common spatial pattern and adaptive short-time Fourier transform-based synchrosqueezing transform. Biomed Signal Process Control; 2023; 79, 104022. [DOI: https://dx.doi.org/10.1016/j.bspc.2022.104022]
Amirshahi, A; Thomas, A; Aminifar, A; Rosing, T; Atienza, D. M2d2: Maximum-mean-discrepancy decoder for temporal localization of epileptic brain activities. IEEE J Biomed Health Inform; 2023; 27,
Amrani, G; Adadi, A; Berrada, M. An explainable hybrid DNN model for seizure vs. non-seizure classification and seizure localization using multi-dimensional EEG signals. Biomed Signal Process Control; 2024; [DOI: https://dx.doi.org/10.1016/j.bspc.2024.106322]
Andrzejak RG, Lehnertz K, Mormann F, Rieke C, David P, Elger CE (2001) Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: dependence on recording region and brain state. E Stat Nonlinear Soft Matter Phys Phys Rev https://doi.org/10.1103/PhysRevE.64.061907
Anita, M; Meena Kowshalya, A. Automatic epileptic seizure detection using MSA-DCNN and LSTM techniques with EEG signals. Expert Syst Appl; 2024; [DOI: https://dx.doi.org/10.1016/j.eswa.2023.121727]
Aslam, MH; Usman, SM; Khalid, S; Anwar, A; Alroobaea, R; Hussain, S; Almotiri, J; Ullah, SS; Yasin, A. Classification of EEG signals for prediction of epileptic seizures. Appl Sci (Switzerland); 2022; [DOI: https://dx.doi.org/10.3390/app12147251]
Assali, I; Ghazi Blaiech, A; Ben Abdallah, A; Ben Khalifa, K; Carrère, M; Hédi Bedoui, M. CNN-based classification of epileptic states for seizure prediction using combined temporal and spectral features. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2022.104519]
Atlam, HF; Aderibigbe, GE; Nadeem, MS. Effective epileptic seizure detection with hybrid feature selection and smote-based data balancing using SVM classifier. Appl Sci; 2025; [DOI: https://dx.doi.org/10.3390/app15094690]
Baghdadi, A; Fourati, R; Aribi, Y; Daoud, S; Dammak, M; Mhiri, C; Chabchoub, H; Siarry, P; Alimi, A. A channel-wise attention-based representation learning method for epileptic seizure detection and type classification. J Ambient Intell Humaniz Comput; 2023; 14,
Baghersalimi, S; Teijeiro, T; Aminifar, A; Atienza, D. Decentralized federated learning for epileptic seizures detection in low-power wearable systems. IEEE Trans Mob Comput; 2024; 23,
Baser, O; Yavuz, M; Ugurlu, K; Onat, F; Demirel, BU. Automatic detection of the spike-and-wave discharges in absence epilepsy for humans and rats using deep learning. Biomed Signal Process Control; 2022; [DOI: https://dx.doi.org/10.1016/j.bspc.2022.103726]
Basha, NK; Surendiran, B; Benzikar, A; Joyal, S. Hybrid approach for the detection of epileptic seizure using electroencephalography input. Int J Inf Technol (Singapore); 2024; 16,
Beeraka, SM; Kumar, A; Sameer, M; Ghosh, S; Gupta, B. Accuracy enhancement of epileptic seizure detection: a deep learning approach with hardware realization of STFT. Circuits Syst Signal Process; 2022; 41,
Benzaid, A; Djemili, R; Arbateni, K. Seizure detection using nonlinear measures over EEG frequency bands and deep learning classifiers. Comput Methods Biomech Biomed Eng; 2024; 10,
Bhadra, R; Singh, PK; Mahmud, M. HyEpiSeiD: a hybrid convolutional neural network and gated recurrent unit model for epileptic seizure detection from electroencephalogram signals. Brain Inf; 2024; [DOI: https://dx.doi.org/10.1186/s40708-024-00234-x]
Bhandari, V; Huchaiah, MD. A new design of epileptic seizure detection using hybrid heuristic-based weighted feature selection and ensemble learning. Int J Intell Robot Appl; 2022; 6,
Bhanja, N; Dhara, SK; Khampariya, P. Heuristic-assisted adaptive hybrid deep learning model with feature selection for epilepsy detection using EEG signals. Biomed Eng Appl Basis Commun; 2023; [DOI: https://dx.doi.org/10.4015/S1016237223500369]
Birajadar, GB; Mulani, AO; Khalaf, OI; Farhah, N; Gawande, PG; Kinage, K; Hamad, AA. Epilepsy identification using hybrid CoPrO-DCNN classifier. Int J Comput Digital Syst; 2024; 16,
Borhade, RR; Barekar, SS; Ohatkar, SN; Mathurkar, PK; Borhade, RH; Bangare, PM. ResneXt-LeNet: a hybrid deep learning for epileptic seizure prediction. Intell Decis Technol; 2024; 18,
Buldu, A; Kaplan, K; Kuncan, M. A hybrid study for epileptic seizure detection based on deep learning using EEG data. J Univ Comput Sci; 2024; 30,
Burrello A, Cavigelli L, Schindler K, Benini L, Rahimi A (2019) Laelaps: An energy-efficient seizure detection algorithm from long-term human iEEG recordings without false alarms. In: 2019 design, automation and test in Europe conference and exhibition. https://doi.org/10.23919/DATE.2019.8715186
Busia, P; Cossettini, A; Ingolfsson, TM; Benatti, S; Burrello, A; Jung, VJB; Scherer, M; Scrugli, MA; Bernini, A; Ducouret, P; Ryvlin, P; Meloni, P; Benini, L. Reducing false alarms in wearable seizure detection with EEGformer: a compact transformer model for MCUs. IEEE Trans Biomed Circuits Syst; 2024; 18,
Caffarini, J; Gjini, K; Sevak, B; Waleffe, R; Kalkach-Aparicio, M; Boly, M; Struck, AF. Engineering nonlinear epileptic biomarkers using deep learning and Benford’s law. Sci Rep; 2022; [DOI: https://dx.doi.org/10.1038/s41598-022-09429-w]
Cao, J; Hu, D; Wang, Y; Wang, J; Lei, B. Epileptic classification with deep-transfer-learning-based feature fusion algorithm. IEEE Trans Cognit Dev Syst; 2022; 14,
Chalaki, M; Omranpour, H. Epileptic seizure classification using ConvLSTM deep classifier and rotation short-time Fourier transform. J Ambient Intell Humaniz Comput; 2023; 14,
Cheng, L; Xiong, J; Duan, J; Zhang, Y; Chen, C; Zhong, J; Zhou, Z; Quan, Y. SaE-GBLS: an effective self-adaptive evolutionary optimized graph-broad model for EEG-based automatic epileptic seizure detection. Front Comput Neurosci; 2024; [DOI: https://dx.doi.org/10.3389/fncom.2024.1379368]
Cherukuvada, S; Kayalvizhi, R. Automated epileptic seizure detection using improved crystal structure algorithm with stacked autoencoder. Int J Adv Comput Sci Appl; 2023; 14,
Cherukuvada, S; Kayalvizhi, R. Feature selection with deep belief network for epileptic seizure detection on EEG signals. Comput Mater Continua; 2023; 75,
Chetana, R; Shubha Rao, A; Mahantesh, K. Application of conv-1D and bi-LSTM to classify and detect epilepsy in EEG data. Int J Adv Comput Sci Appl; 2023; 14,
Chung, YG; Cho, A; Kim, H; Kim, KJ. Single-channel seizure detection with clinical confirmation of seizure locations using CHB-MIT dataset. Front Neurol; 2024; [DOI: https://dx.doi.org/10.3389/fneur.2024.1389731]
Cook, MJ; O’Brien, TJ; Berkovic, SF; Murphy, M; Morokoff, A; Fabinyi, G; D’Souza, W; Yerra, R; Archer, J; Litewka, L; Hosking, S; Lightfoot, P; Ruedebusch, V; Sheffield, WD; Snyder, D; Leyde, K; Himes, D. Prediction of seizure likelihood with a long-term, implanted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study. The Lancet Neurol; 2013; 6, pp. 563-571. [DOI: https://dx.doi.org/10.1093/eurjhf/hfr089]
Cui, X; Hu, D; Lin, P; Cao, J; Lai, X; Wang, T; Jiang, T; Gao, F. Deep feature fusion based childhood epilepsy syndrome classification from electroencephalogram. Neural Netw; 2022; 150, pp. 313-325. [DOI: https://dx.doi.org/10.1016/j.neunet.2022.03.014]
Cui, X; Cao, J; Lai, X; Jiang, T; Gao, F. Cluster embedding joint-probability-discrepancy transfer for cross-subject seizure detection. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 593-605. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3229066]
Cui, W; Xiang, Y; Wang, Y; Yu, T; Liao, X-F; Hu, B; Li, Y. Deep multiview module adaption transfer network for subject-specific EEG recognition. IEEE Trans Neural Netw Learn Syst; 2025; 36,
Darvishi-Bayazi, M-J; Ghaemi, MS; Lesort, T; Arefin, MR; Faubert, J; Rish, I. Amplifying pathological detection in EEG signaling pathways through cross-dataset transfer learning. Comput Biol Med; 2024; [DOI: https://dx.doi.org/10.1016/j.compbiomed.2023.107893]
Das, S; Mumu, SA; Akhand, MAH; Kamal, MAS. Epileptic seizure detection using integrated decomposed features from EEG signal. J Comput Sci; 2024; 20,
Das, S; Mumu, SA; Akhand, MAH; Salam, A; Kamal, MAS. Epileptic seizure detection from decomposed EEG signal through 1D and 2D feature representation and convolutional neural network. Information (Switzerland); 2024; [DOI: https://dx.doi.org/10.3390/info15050256]
Das, A; Nayak, SR; Mohanty, MN; Shukla, PK. Biomedical signal to image conversion and classification using flexible deep learning techniques. Int J Bus Intell Data Min; 2024; 24,
Dash, DP; Kolekar, M; Chakraborty, C; Khosravi, MR. Review of machine and deep learning techniques in epileptic seizure detection using physiological signals and sentiment analysis. ACM Trans Asian Low-Resource Language Inf Process; 2024; [DOI: https://dx.doi.org/10.1145/3552512]
Dash, S; Dash, DK; Tripathy, RK; Pachori, RB. Time-frequency domain machine learning for detection of epilepsy using wearable EEG sensor signals recorded during physical activities. Biomed Signal Process Control; 2025; [DOI: https://dx.doi.org/10.1016/j.bspc.2024.107041]
Deng, Z; Li, C; Song, R; Liu, X; Qian, R; Chen, X. EEG-based seizure prediction via hybrid vision transformer and data uncertainty learning. Eng Appl Artif Intell; 2023; [DOI: https://dx.doi.org/10.1016/j.engappai.2023.106401]
Detti, P; Vatti, G; Lara, GZM. EEG synchronization analysis for seizure prediction: a study on data of noninvasive recordings. Processes; 2020; [DOI: https://dx.doi.org/10.3390/pr8070846]
Dhull, S; Singh, KK. ESIMD: epileptic seizure identification using metaheuristic deep learning technique. Expert Syst; 2022; [DOI: https://dx.doi.org/10.1111/exsy.12897]
Ding, Y; Zhao, W. Channel selection for seizure detection based on explainable AI with Shapley values. IEEE Sens J; 2024; 24,
Ding, X; Nie, W; Liu, X; Wang, X; Yuan, Q. Compact convolutional neural network with multi-headed attention mechanism for seizure prediction. Int J Neural Syst; 2023; [DOI: https://dx.doi.org/10.1142/S0129065723500144]
Dissanayake, T; Fernando, T; Denman, S; Sridharan, S; Fookes, C. Geometric deep learning for subject independent epileptic seizure prediction using scalp EEG signals. IEEE J Biomed Health Inform; 2022; 26,
Du, Z; Peng, R; Liu, W; Li, W; Wu, D. Mixture of experts for EEG-based seizure subtype classification. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 4781-4789. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3337802]
Duan, L; Wang, Z; Qiao, Y; Wang, Y; Huang, Z; Zhang, B. An automatic method for epileptic seizure detection based on deep metric learning. IEEE J Biomed Health Inform; 2022; 26,
Dutta, KK; Manohar, P; Indira, K. Time and frequency domain pre-processing for epileptic seizure classification of epileptic EEG signals. J Intell Fuzzy Syst; 2023; 45,
Dutta, AK; Raparthi, M; Alsaadi, M; Bhatt, MW; Dodda, SB; Prashant, GC; Sandhu, M; Patni, JC. Deep learning-based multi-head self-attention model for human epilepsy identification from EEG signal for biomedical traits. Multimed Tools Appl; 2024; 83,
Ein Shoka, AA; Dessouky, MM; El-Sayed, A; El-Din Hemdan, E. An efficient CNN based epileptic seizures detection framework using encrypted EEG signals for secure telemedicine applications. Alex Eng J; 2023; 65, pp. 399-412. [DOI: https://dx.doi.org/10.1016/j.aej.2022.10.014]
Einizade, A; Nasiri, S; Mozafari, M; Sardouie, SH; Clifford, GD. Explainable automated seizure detection using attentive deep multi-view networks. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2022.104076]
El-Samie, FEA; Alotaiby, TN; Khalid, MI; Alshebeili, SA; Aldosari, SA. A review of EEG and MEG epileptic spike detection algorithms. IEEE Access; 2018; 6, pp. 60673-60688. [DOI: https://dx.doi.org/10.1109/ACCESS.2018.2875487]
Esmaeilpour, A; Tabarestani, SS; Niazi, A. Deep learning-based seizure prediction using EEG signals: a comparative analysis of classification methods on the CHB-MIT dataset. Eng Rep; 2024; [DOI: https://dx.doi.org/10.1002/eng2.12918]
Falco-Walter, JJ; Scheffer, IE; Fisher, RS. The new definition and classification of seizures and epilepsy. Epilepsy Res; 2018; 139, pp. 73-79. [DOI: https://dx.doi.org/10.1016/j.eplepsyres.2017.11.015]
Fathima, S; Ahmed, M. Sleep Apnea detection using EEG: a systematic review of datasets, methods, challenges, and future directions. Ann Biomed Eng; 2025; 53,
Fathima, SN; Rekha, KB; Safinaz, S; Ahmed, ST. Computational techniques, classification, datasets review and way forward with modern analysis of epileptic seizure – a study. Multimed Tools Appl; 2024; 83,
Fawad Hussain, S; Mian Qaisar, S. Epileptic seizure classification using level-crossing EEG sampling and ensemble of sub-problems classifier. Expert Syst Appl; 2022; [DOI: https://dx.doi.org/10.1016/j.eswa.2021.116356]
Feizbakhsh, B; Omranpour, H. Cluster-based phase space density feature in multichannel scalp EEG for seizure prediction by deep learning. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.105276]
Ganti, B; Chaitanya, G; Balamurugan, RS; Nagaraj, N; Balasubramanian, K; Pati, S. Time-series generative adversarial network approach of deep learning improves seizure detection from the human thalamic SEEG. Front Neurol; 2022; [DOI: https://dx.doi.org/10.3389/fneur.2022.755094]
Gao, Y; Chen, X; Liu, A; Liang, D; Wu, L; Qian, R; Xie, H; Zhang, Y. Pediatric seizure prediction in scalp EEG using a multi-scale neural network with dilated convolutions. IEEE J Transl Eng Health Med; 2022; [DOI: https://dx.doi.org/10.1109/JTEHM.2022.3144037]
Gao, B; Zhou, J; Yang, Y; Chi, J; Yuan, Q. Generative adversarial network and convolutional neural network-based EEG imbalanced classification model for seizure detection. Biocybern Biomed Eng; 2022; 42,
Gao, Y; Liu, A; Cui, X; Qian, R; Chen, X. A general sample-weighted framework for epileptic seizure prediction. Comput Biol Med; 2022; [DOI: https://dx.doi.org/10.1016/j.compbiomed.2022.106169]
Gao, Y; Liu, A; Wang, L; Qian, R; Chen, X. A self-interpretable deep learning model for seizure prediction using a multi-scale prototypical part network. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 1847-1856. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3260845]
Ge, S; Wang, W; Shi, X; Wang, M. Hierarchical spatio-temporal-spectral multiple hybrid attention mechanisms for seizure prediction. IAENG Int J Comput Sci; 2024; 51,
Georgis-Yap, Z; Popovic, MR; Khan, SS. Supervised and unsupervised deep learning approaches for EEG seizure prediction. J Healthc Inf Res; 2024; 8,
Ghanimi, HMA; Sri, ST; Sadu, VB; Yellamma, P; Surya, U; Poon, K. Cramer distance: a deep learning approach for better epileptic seizure prediction. J Mach Comput; 2024; 4,
Gill, TS; Zaidi, SSH; Shirazi, MA. Attention-based deep convolutional neural network for classification of generalized and focal epileptic seizures. Epilepsy Behav; 2024; [DOI: https://dx.doi.org/10.1016/j.yebeh.2024.109732]
Gireesh, ED; Skinner, H; Seo, J; Ching, P; Hyeong, LK; Baumgartner, J; Gurupur, V. Deep neural networks and gradient-weighted class activation mapping to classify and analyze EEG. Intell Decis Technol; 2023; 17,
Goel, S; Agrawal, R; Bharti, RK. Epileptic seizure prediction and classification based on statistical features using LSTM fully connected neural network. J Intell Fuzzy Syst; 2023; 44,
Goel, S; Agrawal, R; Bharti, RK. Automated detection of epileptic EEG signals using recurrence plots-based feature extraction with transfer learning. Soft Comput; 2024; 28,
Golla, SK; Maloji, S. A novel finite spectral entropy: Gated term memory unit recursive network integrated with ladybug beetle optimization algorithm for epileptic seizure detection. Int J Numer Methods Biomed Eng; 2023; [DOI: https://dx.doi.org/10.1002/cnm.3769]
Golla, SK; Maloji, S. Maximum overlap discrete transform (MODT)—Gaussian Kernel radial network (GKRN) model for epileptic seizure detection from EEG signals. J Adv Inf Technol; 2023; 14,
Gotman, J. A few thoughts on “What is a seizure?”. Epilepsy Behav; 2011; 22, pp. 2-3. [DOI: https://dx.doi.org/10.1016/j.yebeh.2011.08.025]
Guo, L; Yu, T; Zhao, S; Li, X; Liao, X; Li, Y. CLEP: contrastive learning for epileptic seizure prediction using a spatio-temporal-spectral network. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 3915-3926. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3322275]
Halawa, RI; Youssef, SM; Elagamy, MN. An efficient hybrid model for patient-independent seizure prediction using deep learning. Appl Sci (Switzerland); 2022; [DOI: https://dx.doi.org/10.3390/app12115516]
Handa, P; Gupta, M; Gupta, E; Goel, N. One-dimensional atrous conv-net based architecture for automatic diagnosis of epilepsy using electroencephalography signals and its brain–computer interface applications. Expert Syst; 2024; 41,
Hartmann, M; Koren, J; Baumgartner, C; Duun-Henriksen, J; Gritsch, G; Kluge, T; Perko, H; Fürbass, F. Seizure detection with deep neural networks for review of two-channel electroencephalogram. Epilepsia; 2023; 64, pp. 34-39. [DOI: https://dx.doi.org/10.1111/epi.17259]
Hasnaoui, LH; Djebbari, A. Robust dimensionality-reduced epilepsy detection system using EEG wavelet packets and machine learning. Res Biomed Eng.; 2024; [DOI: https://dx.doi.org/10.1007/s42600-024-00355-6]
Hasnaoui, LH; Djebbari, A. Leveraging epilepsy detection accuracy through burst energy integration in CWT and decision tree classification of EEG signals. Biomed Eng Appl Basis Commun; 2025; [DOI: https://dx.doi.org/10.4015/S1016237224500522]
Hassan, F; Hussain, SF; Qaisar, SM. Epileptic seizure detection using a hybrid 1d CNN-machine learning approach from EEG data. J Healthc Eng; 2022; [DOI: https://dx.doi.org/10.1155/2022/9579422]
He, J; Cui, J; Zhang, G; Xue, M; Chu, D; Zhao, Y. Spatial-temporal seizure detection with graph attention network and bi-directional LSTM architecture. Biomed Signal Process Control; 2022; [DOI: https://dx.doi.org/10.1016/j.bspc.2022.103908]
He, P; Wang, L; Cui, Y; Wang, R; Wu, D. Unsupervised feature learning based on autoencoder for epileptic seizures prediction. Appl Intell; 2023; 53,
He, Z; Yang, J; Alroobaea, R; Yee Por, L. SeizureLSTM: an optimal attention-based trans-LSTM network for epileptic seizure detection using optimal weighted feature integration. Biomed Signal Process Control; 2024; [DOI: https://dx.doi.org/10.1016/j.bspc.2024.106603]
Hermawan, AT; Zaeni, IAE; Wibawa, AP; Hendrawan, WH; Kristian, Y. A multi representation deep learning approach for epileptic seizure detection. J Robot Control; 2024; 5,
Hilal, AM; Albraikan, AA; Dhahbi, S; Nour, MK; Mohamed, A; Motwakel, A; Zamani, AS; Rizwanullah, M. Intelligent epileptic seizure detection and classification model using optimal deep canonical sparse autoencoder. Biology; 2022; [DOI: https://dx.doi.org/10.3390/biology11081220]
Holguin-Garcia, SA; Guevara-Navarro, E; Daza-Chica, AE; Patiño-Claro, MA; Arteaga-Arteaga, HB; Ruz, GA; Tabares-Soto, R; Bravo-Ortiz, MA. A comparative study of CNN-capsule-net, CNN-transformer encoder, and traditional machine learning algorithms to classify epileptic seizure. BMC Med Inf Decis Mak; 2024; [DOI: https://dx.doi.org/10.1186/s12911-024-02460-z]
Hosseinzadeh, M; Khoshvaght, P; Sadeghi, S; Asghari, P; Noroozi Varzeghani, A; Mohammadi, M; Mohammadi, H; Lansky, J; Lee, S-W. A model for epileptic seizure diagnosis using the combination of ensemble learning and deep learning. IEEE Access; 2024; 12, pp. 137132-137143. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3457018]
Hu, S; Liu, J; Yang, R; Wang, Y; Wang, A; Li, K; Liu, W; Yang, C. Exploring the applicability of transfer learning and feature engineering in epilepsy prediction using hybrid transformer model. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 1321-1332. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3244045]
Hu, W; Wang, J; Li, F; Ge, D; Wang, Y; Jia, Q; Yuan, S. A modified transformer network for seizure detection using EEG signals. Int J Neural Syst; 2024; [DOI: https://dx.doi.org/10.1142/S0129065725500030]
Huang, Z; Duan, J. GFBLS: graph-regularized fuzzy broad learning system for detection of interictal epileptic discharges. Eng Appl Artif Intell; 2023; [DOI: https://dx.doi.org/10.1016/j.engappai.2023.106763]
Huang, X; Sun, X; Zhang, L; Zhu, T; Yang, H; Xiong, Q; Feng, L. A novel epilepsy detection method based on feature extraction by deep autoencoder on EEG signal. Int J Environ Res Public Health; 2022; [DOI: https://dx.doi.org/10.3390/ijerph192215110]
Huang, W; Xu, H; Yu, Y. MRP-net: Seizure detection method based on modified recurrence plot and additive attention convolution neural network. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.105165]
Huang, H; Zhang, F; Lv, L; Si, X. Review of application of neural networks in epileptic seizure prediction. J Front Comput Sci Technol; 2023; 11, pp. 2543-2556. [DOI: https://dx.doi.org/10.3778/j.issn.1673-9418.2302001]
Huang, L; Zhou, K; Chen, S; Chen, Y; Zhang, J. Automatic detection of epilepsy from EEGs using a temporal convolutional network with a self-attention layer. BioMed Eng; 2024; [DOI: https://dx.doi.org/10.1186/s12938-024-01244-w]
Obeid, I; Picone, J. The temple university hospital EEG data corpus. Front Neurosci; 2016; [DOI: https://dx.doi.org/10.3389/fnins.2016.00196]
Ibrahim, FE; Emara, HM; El-Shafai, W; Elwekeil, M; Rihan, M; Eldokany, IM; Taha, TE; El-Fishawy, AS; El-Rabaie, E-SM; Abdellatef, E; Abd El-Samie, FE. Deep-learning-based seizure detection and prediction from electroencephalography signals. Int J Numer Methods Biomed Eng; 2022; [DOI: https://dx.doi.org/10.1002/cnm.3573]
Ilakiyaselvan, N; Khan, AN; Shahina, A. Reconstructed phase space portraits for detecting brain diseases using deep learning. Biomed Signal Process Control; 2022; [DOI: https://dx.doi.org/10.1016/j.bspc.2021.103278]
Ilias, L; Askounis, D; Psarras, J. Multimodal detection of epilepsy with deep neural networks. Expert Syst Appl; 2023; [DOI: https://dx.doi.org/10.1016/j.eswa.2022.119010]
Indira Priyadarshini, B; Reddy, DK. Adaptive octopus deep transfer learning based epileptic seizure classification on field programmable gate arrays. Evol Syst; 2023; 14,
Islam, T; Basak, M; Islam, R; Roy, AD. Investigating population-specific epilepsy detection from noisy EEG signals using deep-learning models. Heliyon; 2023; [DOI: https://dx.doi.org/10.1016/j.heliyon.2023.e22208]
Jaffar, AY. IoT-enabled EEG-based epilepsy detection using multilayer deep learning and the evolutionary algorithm approach. Eng Technol Appl Sci Res; 2024; 14,
Jahan, S; Nowsheen, F; Antik, MM; Rahman, MS; Kaiser, MS; Hosen, ASMS; Ra, I-H. AI-based epileptic seizure detection and prediction in internet of healthcare things: a systematic review. IEEE Access; 2023; 11, pp. 30690-30725. [DOI: https://dx.doi.org/10.1109/ACCESS.2023.3251105]
Jain, S; Paul, S; Sharma, K. EEG brain signal processing for epilepsy detection. Recent Adv Electric Electron Eng; 2023; 16,
Jaishankar, B; Ashwini, AM; Vidyabharathi, D; Raja, L. A novel epilepsy seizure prediction model using deep learning and classification. Healthc Anal; 2023; [DOI: https://dx.doi.org/10.1016/j.health.2023.100222]
Jemal, I; Mezghani, N; Abou-Abbas, L; Mitiche, A. An interpretable deep learning classifier for epileptic seizure prediction using EEG data. IEEE Access; 2022; 10, pp. 60141-60150. [DOI: https://dx.doi.org/10.1109/ACCESS.2022.3176367]
Jemal, I; Abou-Abbas, L; Henni, K; Mitiche, A; Mezghani, N. Domain adaptation for EEG-based, cross-subject epileptic seizure prediction. Front Neuroinf; 2024; [DOI: https://dx.doi.org/10.3389/fninf.2024.1303380]
Jia, G; Lam, H-K; Althoefer, K. Variable weight algorithm for convolutional neural networks and its applications to classification of seizure phases and types. Pattern Recognit; 2022; [DOI: https://dx.doi.org/10.1016/j.patcog.2021.108226]
Jia, M; Liu, W; Duan, J; Chen, L; Chen, CLP; Wang, Q; Zhou, Z. Efficient graph convolutional networks for seizure prediction using scalp EEG. Front Neurosci; 2022; [DOI: https://dx.doi.org/10.3389/fnins.2022.967116]
Jibon FA, Tasbir A, Miraz MH, Jin HH, Siddiqui FH, Sakib M, Nishar NH, Thakur H, Khandaker MU (2024) Graph attention network and radial basis function neural network-based hybrid framework for epileptic seizure detection from EEG signal. J Auton Intell. https://doi.org/10.32629/jai.v7i3.1149
Jibon, FA; Jamil Chowdhury, AR; Miraz, MH; Jin, HH; Khandaker, MU; Sultana, S; Nur, S; Siddiqui, FH; Kamal, AHM; Salman, M; Youssef, AAF. Sequential graph convolutional network and DeepRNN based hybrid framework for epileptic seizure detection from EEG signal. Digit Health; 2024; [DOI: https://dx.doi.org/10.1177/20552076241249874]
Jin, X; Wang, J; Liu, L; Lin, Y. Uncertainty-aware denoising network for artifact removal in EEG signals. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 4470-4480. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3330963]
Jusseaume, K; Valova, I. Brain age prediction/classification through recurrent deep learning with electroencephalogram recordings of seizure subjects. Sensors; 2022; [DOI: https://dx.doi.org/10.3390/s22218112]
Kabir, SA; Farhan, F; Siddiquee, AA; Baroi, OL; Marium, T; Rahimi, J. Effect of input channel reduction on EEG seizure detection. Przeglad Elektrotechniczny; 2023; 2023,
Kalita, D; Dash, S; Mirza, KB. EpiNET: an optimized, resource efficient deep GRU-LSTM network for epileptic seizure prediction. Biomed Eng Appl Basis Commun; 2024; [DOI: https://dx.doi.org/10.4015/S1016237224500212]
Kamakshi, K; Rengaraj, A. Early detection of stress and anxiety based seizures in position data augmented EEG signal using hybrid deep learning algorithms. IEEE Access; 2024; 12, pp. 35351-35365. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3365192]
Kannan, S; Premalatha, G; Jamuna Rani, M; Jayakumar, D; Senthil, P; Palanivelrajan, S; Devi, S; Sahile, K. Effective evaluation of medical images using artificial intelligence techniques. Comput Intell Neurosci; 2022; [DOI: https://dx.doi.org/10.1155/2022/8419308]
Kapoor, B; Nagpal, B. Hybrid cuckoo finch optimisation based machine learning classifier for seizure prediction using EEG signals in IoT network. Clust Comput; 2024; 27,
Karnati, M; Sahu, G; Yadav, A; Seal, A; Jaworek-Korjakowska, J; Penhaker, M; Krejcar, O. MD-DCNN: multi-scale dilation-based deep convolution neural network for epilepsy detection using electroencephalogram signals. Knowl-Based Syst; 2024; [DOI: https://dx.doi.org/10.1016/j.knosys.2024.112322]
Karunakar Reddy, V; Kumar, AVR. Multi-channel neuro signal classification using Adam-based coyote optimization enabled deep belief network. Biomed Signal Process Control; 2022; [DOI: https://dx.doi.org/10.1016/j.bspc.2022.103774]
Kaur, T; Gandhi, TK. Automated diagnosis of epileptic seizures using EEG image representations and deep learning. Neurosci Inf; 2023; [DOI: https://dx.doi.org/10.1016/j.neuri.2023.100139]
Kaur, A; Puri, V; Shashvat, K; Maurya, AK. Automated identification of inter-ictal discharges using residual deep learning neural network amidst of various artefacts. Chaos Solitons Fractals; 2022; [DOI: https://dx.doi.org/10.1016/j.chaos.2022.111886]
Khalfallah S, Puech W, Tlija M, Bouallegue K (2025) Exploring the effectiveness of machine learning and deep learning techniques for EEG signal classification in neurological disorders. IEEE Access, 17002–17015
Khan, SU; Jan, SU; Koo, I. Robust epileptic seizure detection using long short-term memory and feature fusion of compressed time-frequency EEG images. Sensors; 2023; [DOI: https://dx.doi.org/10.3390/s23239572]
Khan, FA; Umar, Z; Jolfaei, A; Tariq, M. Explainable fuzzy deep learning for prediction of epileptic seizures using EEG. IEEE Trans Fuzzy Syst; 2024; 32,
Khurshid, D; Wahid, F; Ali, S; Gumaei, AH; Alzanin, SM; Mosleh, MAA. A deep neural network-based approach for seizure activity recognition of epilepsy sufferers. Front Med; 2024; [DOI: https://dx.doi.org/10.3389/fmed.2024.1405848]
Kode, H; Elleithy, K; Almazaydeh, L. Epileptic seizure detection in EEG signals using machine learning and deep learning techniques. IEEE Access; 2024; 12, pp. 80657-80668. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3409581]
Kulasinghe, WMND; Dissanayake, MB. A novel LSTM-based data synthesis approach for performance improvement in detecting epileptic seizures. WSEAS Trans Biol Biomed; 2023; 20, pp. 132-139. [DOI: https://dx.doi.org/10.37394/23208.2023.20.13]
Kumar, PR; Shilpa, B; Jha, RK; Mohanty, SN. A novel end-to-end approach for epileptic seizure classification from scalp EEG data using deep learning technique. Int J Inf Technol (Singapore); 2023; 15,
Kumar, P; Kumar Upadhyay, P; Kumar Panda, M. SNSDeepNet: spike and non-spike detection in epilepsy. Eng Res Express; 2024; [DOI: https://dx.doi.org/10.1088/2631-8695/ad7a4f]
Kumbam, VR; Mary, AVA. Cross model attention based deep learning for multi modal epilepsy detection. Int J Intell Eng Syst; 2023; 16,
Kunekar, P; Gupta, MK; Gaur, P. Detection of epileptic seizure in EEG signals using machine learning and deep learning techniques. J Eng Appl Sci; 2024; [DOI: https://dx.doi.org/10.1186/s44147-023-00353-y]
Lebal, A; Moussaoui, A; Rezgui, A. Epilepsy-net: attention-based 1D-inception network model for epilepsy detection using one-channel and multi-channel EEG signals. Multimed Tools Appl; 2023; 82,
Li, Z; Hwang, K; Li, K; Wu, J; Ji, T. Graph-generative neural network for EEG-based epileptic seizure detection via discovery of dynamic brain functional connectivity. Sci Rep; 2022; [DOI: https://dx.doi.org/10.1038/s41598-022-23656-1]
Li, Y; Liu, Y; Guo, Y-Z; Liao, X-F; Hu, B; Yu, T. Spatio-temporal-spectral hierarchical graph convolutional network with semisupervised active learning for patient-specific seizure prediction. IEEE Trans Cybern; 2022; 52,
Li, C; Lammie, C; Dong, X; Amirsoleimani, A; Azghadi, MR; Genov, R. Seizure detection and prediction by parallel memristive convolutional neural networks. IEEE Trans Biomed Circuits Syst; 2022; 16,
Li, C; Zhao, Y; Song, R; Liu, X; Qian, R; Chen, X. Patient-specific seizure prediction from electroencephalogram signal via multichannel feedback capsule network. IEEE Trans Cognit Dev Syst; 2023; 15,
Li, C; Deng, Z; Song, R; Liu, X; Qian, R; Chen, X. EEG-based seizure prediction via model uncertainty learning. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 180-191. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3217929]
Li, C; Shao, C; Song, R; Xu, G; Liu, X; Qian, R; Chen, X. Spatio-temporal MLP network for seizure prediction using EEG signals. Meas J Int Meas Confed; 2023; [DOI: https://dx.doi.org/10.1016/j.measurement.2022.112278]
Li, H; Liao, J; Wang, H; Zhan, CA; Yang, F. EEG power spectra parameterization and adaptive channel selection towards semi-supervised seizure prediction. Comput Biol Med; 2024; [DOI: https://dx.doi.org/10.1016/j.compbiomed.2024.108510]
Li, H; Dong, X; Zhong, X; Li, C; Cui, H; Zhou, W. End-to-end model for automatic seizure detection using supervised contrastive learning. Eng Appl Artif Intell; 2024; [DOI: https://dx.doi.org/10.1016/j.engappai.2024.108665]
Li, C; Li, H; Dong, X; Zhong, X; Cui, H; Ji, D; He, L; Liu, G; Zhou, W. CNN-informer: a hybrid deep learning model for seizure detection on long-term EEG. Neural Netw; 2025; [DOI: https://dx.doi.org/10.1016/j.neunet.2024.106855]
Lian, J; Xu, F. Epileptic EEG classification via graph transformer network. Int J Neural Syst; 2023; [DOI: https://dx.doi.org/10.1142/S0129065723500429]
Liang, S; Zhang, X; Zhao, H; Dang, Y; Hui, R; Zhang, J. Double discrete variational autoencoder for epileptic EEG signals classification. IEEE Access; 2024; 12, pp. 106567-106578. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3429195]
Lin, Y; Dong, L; Jiang, Y; Lian, J. Epileptic EEG classification via deep learning-based strange attractor. Biomed Signal Process Control; 2025; [DOI: https://dx.doi.org/10.1016/j.bspc.2024.106965]
Liu, G; Tian, L; Zhou, W. Patient-independent seizure detection based on channel-perturbation convolutional neural network and bidirectional long short-term memory. Int J Neural Syst; 2022; [DOI: https://dx.doi.org/10.1142/S0129065721500519]
Liu, C; Chen, W; Li, M. A hybrid EEG classification model using layered cascade deep learning architecture. Med Biol Eng Comput; 2024; 62,
Liu, X; Li, C; Lou, X; Kong, H; Li, X; Li, Z; Zhong, L. Epileptic seizure prediction based on EEG using pseudo-three-dimensional CNN. Front Neuroinf; 2024; [DOI: https://dx.doi.org/10.3389/fninf.2024.1354436]
Liu, J-J; Yao, J-P; Liu, J-H; Wang, Z-Y; Huang, L. Missing data imputation and classification of small sample missing time series data based on gradient penalized adversarial multi-task learning. Appl Intell; 2024; 54,
Liu, G; Yan, Y; Cai, J; Qi, WuE; Fang, S; David Cheok, A; Song, A. GCD: graph contrastive denoising module for GNNs in EEG classification. Expert Syst Appl; 2025; [DOI: https://dx.doi.org/10.1016/j.eswa.2024.126013]
Llorente-Vidrio, D; Ballesteros, M; Salgado, I. Deep learning adapted to differential neural networks used as pattern classification of electrophysiological signals. IEEE Trans Pattern Anal Mach Intell; 2022; [DOI: https://dx.doi.org/10.1109/TPAMI.2021.3066996]
Lopes, F; Pinto, MF; Dourado, A; Schulze-Bonhage, A; Dümpelmann, M; Teixeira, C. Addressing data limitations in seizure prediction through transfer learning. Sci Rep; 2024; [DOI: https://dx.doi.org/10.1038/s41598-024-64802-1]
Lu, L; Zhang, F; Wu, Y; Ma, S; Zhang, X; Ni, G. A multi-frame network model for predicting seizure based on sEEG and iEEG data. Front Comput Neurosci; 2022; [DOI: https://dx.doi.org/10.3389/fncom.2022.1059565]
Lu, X; Wen, A; Sun, L; Wang, H; Guo, Y; Ren, Y. An epileptic seizure prediction method based on CBAM-3d CNN-LSTM model. IEEE J Transl Eng Health Med; 2023; 11, pp. 417-423. [DOI: https://dx.doi.org/10.1109/JTEHM.2023.3290036]
Lucasius, C; Grigorovsky, V; Nariai, H; Galanopoulou, AS; Gursky, J; Moshe, SL; Bardakjian, BL. Biomimetic deep learning networks with applications to epileptic spasms and seizure prediction. IEEE Trans Biomed Eng; 2024; 71,
Ma, Y; Huang, Z; Su, J; Shi, H; Wang, D; Jia, S; Li, W. A multi-channel feature fusion CNN-bi-LSTM epilepsy EEG classification and prediction model based on attention mechanism. IEEE Access; 2023; 11, pp. 62855-62864. [DOI: https://dx.doi.org/10.1109/ACCESS.2023.3287927]
Malekzadeh A, Zare A, Yaghoobi M, Alizadehsani R (2021) Automatic diagnosis of epileptic seizures in EEG signals using fractal dimension features and convolutional autoencoder method. Big Data Cognit Comput. https://doi.org/10.3390/bdcc5040078
Mallick, S; Baths, V. Novel deep learning framework for detection of epileptic seizures using EEG signals. Front Comput Neurosci; 2024; [DOI: https://dx.doi.org/10.3389/fncom.2024.1340251]
Mandal, S; Singh, BK; Thakur, K. Epileptic seizure detection using deep learning based long short-term memory networks and time-frequency analysis: a comparative investigation in machine learning paradigm. Braz Archiv Biol Technol; 2022; [DOI: https://dx.doi.org/10.1590/1678-4324-2022210559]
Massoud, YM; Abdelzaher, M; Kuhlmann, L. General and patient-specific seizure classification using deep neural networks. Analog Integr Circuits Signal Process; 2023; [DOI: https://dx.doi.org/10.1007/s10470-023-02153-z]
Ma H, Wu Y, Tang Y, Chen R, Xu T, Zhang W. Parallel dual-branch fusion network for epileptic seizure prediction. Comput Biol Med 176, 108565. https://doi.org/10.1016/j.compbiomed.2024.108565
Mekruksavanich, S; Jitpattanakul, A. Effective detection of epileptic seizures through EEG signals using deep learning approaches. Mach Learn Knowl Extr; 2023; 5,
Meng, K; Wang, D; Zhang, D; Guo, K; Lu, K; Lu, J; Yu, R; Zhang, L; Hu, Y; Zhang, R; Chen, M. Real-time epileptic seizure prediction method with spatio-temporal information transfer learning. IEEE J Biomed Health Inform; 2025; 29,
Mir, WA; Anjum, M; Shahab, S. Deep-EEG: an optimized and robust framework and method for EEG-based diagnosis of epileptic seizure. Diagnostics; 2023; [DOI: https://dx.doi.org/10.3390/diagnostics13040773]
Mittal, R; John Martin, R; Alshehri, H; Malik, V; Goyal, SB; Swapna, SL; Assiri, H; Duraibi, S. RT-NeuroDDSM: real-time EEG-driven diagnostic decision support model for neurological disorders using deep learning. IEEE Access; 2024; 12, pp. 116711-116726. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3436829]
Murugan, TK; Kameswaran, A. Employing convolutional neural networks and explainable artificial intelligence for the detection of seizures from electroencephalogram signal. Results Eng; 2024; [DOI: https://dx.doi.org/10.1016/j.rineng.2024.103378]
Naganur, V; Sivathamboo, S; Chen, Z; Kusmakar, S; Antonic-Baker, A; O’Brien, TJ; Kwan, P. Automated seizure detection with noninvasive wearable devices: a systematic review and meta-analysis. Epilepsia (Series 4); 2022; 63,
Narin, A. Detection of focal and non-focal epileptic seizure using continuous wavelet transform-based scalogram images and pre-trained deep neural networks. IRBM; 2022; 43,
Narmada, A; Shukla, MK. A novel adaptive artifacts wavelet denoising for EEG artifacts removal using deep learning with meta-heuristic approach. Multimed Tools Appl; 2023; 82,
Nasreddine W. Epileptic EEG Dataset. https://doi.org/10.17632/5pc2j46cbc.1
Nemati, N; Meshgini, S. A medium-weight deep convolutional neural network-based approach for onset epileptic seizures classification in EEG signals. Brain Behav; 2022; [DOI: https://dx.doi.org/10.1002/brb3.2763]
Noachtar, S; Rémi, J. The role of EEG in epilepsy: a critical review. Epilepsy Behav; 2009; 15,
Omar, A; Abd El-Hafeez, T. Optimizing epileptic seizure recognition performance with feature scaling and dropout layers. Neural Comput Appl; 2024; 36,
Ouichka, O; Echtioui, A; Hamam, H. Deep learning models for predicting epileptic seizures using iEEG signals. Electronics (Switzerland); 2022; [DOI: https://dx.doi.org/10.3390/electronics11040605]
Page, MJ; McKenzie, JE; Bossuyt, PM; Boutron, I; Hoffmann, TC; Mulrow, CD; Shamseer, L; Tetzlaff, JM; Akl, EA; Brennan, SE; Chou, R; Glanville, J; Grimshaw, JM; Hróbjartsson, A; Lalu, MM; Li, T; Loder, EW; Mayo-Wilson, E; McDonald, S; McGuinness, LA; Stewart, LA; Thomas, J; Tricco, AC; Welch, VA; Whiting, P; Moher, D. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. Int J Surgery; 2021; 88, [DOI: https://dx.doi.org/10.1016/j.ijsu.2021.105906] 105906.
Palanichamy, I; Basheer Ahamed, FB. Prediction of seizure in the EEG signal with time aware recurrent neural network. Revue d’Intelligence Artificielle; 2022; 36,
Pan, Y; Zhou, X; Dong, F; Wu, J; Xu, Y; Zheng, S. Epileptic seizure detection with hybrid time-frequency EEG input: a deep learning approach. Comput Math Methods Med; 2022; [DOI: https://dx.doi.org/10.1155/2022/8724536]
Pan, Y; Dong, F; Wu, J; Xu, Y. Downsampling of EEG signals for deep learning-based epilepsy detection. IEEE Sens Lett; 2023; 7,
Pandey, A; Singh, SK; Udmale, SS; Shukla, KK. Epileptic seizure classification using battle Royale search and rescue optimization-based deep LSTM. IEEE J Biomed Health Inform; 2022; 26,
Pandey, SK; Janghel, RR; Mishra, PK; Ahirwal, MK. Automated epilepsy seizure detection from EEG signal based on hybrid CNN and LSTM model. Signal Image Video Process; 2023; 17,
Pandey, A; Singh, SK; Udmale, SS; Shukla, KK. An intelligent optimized deep learning model to achieve early prediction of epileptic seizures. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.104798]
Parija, S; Dash, PK; Bisoi, R. Epileptic EEG signal classification using an improved VMD-based convolutional stacked autoencoder. Pattern Anal Appl; 2024; [DOI: https://dx.doi.org/10.1007/s10044-024-01221-y]
Patel, SA; Yildirim, A. Non-stationary neural signal to image conversion framework for image-based deep learning algorithms. Front Neuroinf; 2023; [DOI: https://dx.doi.org/10.3389/fninf.2023.1081160]
Patel, SA; Smith, RJ; Yildirim, A. Gershgorin circle theorem-based feature extraction for biomedical signal analysis. Front Neuroinf; 2024; [DOI: https://dx.doi.org/10.3389/fninf.2024.1395916]
Patel, V; Bhatti, D; Ganatra, A; Tailor, J. EEG-based epileptic seizure state detection using deep learning. Int J Model Ident Control; 2024; 44,
Patro, KK; Prakash, AJ; Sahoo, JP; Routray, S; Baihan, A; Samee, NA; Huang, G. SMARTSeiz: deep learning with attention mechanism for accurate seizure recognition in IoT healthcare devices. IEEE J Biomed Health Inform; 2024; 28,
Pattnaik, S; Rao, BN; Rout, NK; Sabut, SK. Transfer learning based epileptic seizure classification using scalogram images of EEG signals. Multimed Tools Appl; 2024; 83,
Peng, R; Zhao, C; Jiang, J; Kuang, G; Cui, Y; Xu, Y; Du, H; Shao, J; Wu, D. TIE-EEGNet: temporal information enhanced EEGNet for seizure subtype classification. IEEE Trans Neural Syst Rehabil Eng; 2022; 30, pp. 2567-2576. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3204540]
Pereira A, Fiel J. Resting-state interictal EEG recordings of refractory epilepsy patients. https://doi.org/10.17632/6hx2smc7nw.1
Poorani, S; Balasubramanie, P. Deep learning based epileptic seizure detection with EEG data. Int J Syst Assuran Eng Manage; 2023; [DOI: https://dx.doi.org/10.1007/s13198-022-01845-5]
Prabhakar, SK; Lee, S-W. SASDL and RBATQ: sparse autoencoder with swarm based deep learning and reinforcement based q-learning for EEG classification. IEEE Open J Eng Med Biol; 2022; 3, pp. 58-68. [DOI: https://dx.doi.org/10.1109/OJEMB.2022.3161837]
Prabhakar, SK; Lee, JJ; Won, D-O. Ensemble fusion models using various strategies and machine learning for EEG classification. Bioengineering; 2024; [DOI: https://dx.doi.org/10.3390/bioengineering11100986]
Prakash, V; Kumar, D. A modified gated recurrent unit approach for epileptic electroencephalography classification. J Inf Commun Technol; 2023; 22,
Prasanna, CSL; Rahman, MZU; Bayleyegn, MD. Brain epileptic seizure detection using joint CNN and exhaustive feature selection with RNN-BLSTM classifier. IEEE Access; 2023; 11, pp. 97990-98004. [DOI: https://dx.doi.org/10.1109/ACCESS.2023.3312187]
Qi, N; Piao, Y; Yu, P; Tan, B. Predicting epileptic seizures based on EEG signals using spatial depth features of a 3D–2D hybrid CNN. Med Biol Eng Comput; 2023; 61,
Qiu, S; Cheng, K; Zhou, T; Tahir, R; Ting, L. An EEG signal recognition algorithm during epileptic seizure based on distributed edge computing. Int J Interactive Multimed Artific Intell; 2022; 7,
Qiu, S; Wang, W; Jiao, H. LightSeizureNet: a lightweight deep learning model for real-time epileptic seizure detection. IEEE J Biomed Health Inform; 2023; 27,
Qiu, X; Yan, F; Liu, H. A difference attention ResNet-LSTM network for epileptic seizure detection using EEG signal. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.104652]
Quadri, ZF; Saqib Akhoon, M; Loan, SA. Epileptic seizure prediction using stacked CNN-BiLSTM: a novel approach. IEEE Trans Artif Intell; 2024; 5,
Raeisi, K; Khazaei, M; Croce, P; Tamburro, G; Comani, S; Zappasodi, F. A graph convolutional neural network for the automated detection of seizures in the neonatal EEG. Comput Methods Progr Biomed; 2022; [DOI: https://dx.doi.org/10.1016/j.cmpb.2022.106950]
Raeisi, K; Khazaei, M; Tamburro, G; Croce, P; Comani, S; Zappasodi, F. A class-imbalance aware and explainable spatio-temporal graph attention network for neonatal seizure detection. Int J Neural Syst; 2023; [DOI: https://dx.doi.org/10.1142/S0129065723500466]
Ramkumar, M; Jamaesha, SS; Gowtham, MS; Kumar, CS. IoT and cloud computing-based automated epileptic seizure detection using optimized Siamese convolutional sparse autoencoder network. Signal Image Video Process; 2024; 18,
Rani, TJ; Kavitha, D. Effective epileptic seizure detection using enhanced Salp swarm algorithm-based long short-term memory network. IETE J Res; 2024; 70,
Rao, KT; Sridevi, K. Development of tuned hybrid fuzzy and BiLSTM-based epileptic seizure classification model with stacked 1d-CNN-fisher discriminant feature selection. Comput Methods Biomechd Biomed Eng Imaging Vis; 2023; 11,
Rao, S; Liu, M; Huang, Y; Yang, H; Liang, J; Lu, J; Niu, Y; Wang, B. Anchoring temporal convolutional networks for epileptic seizure prediction. J Neural Eng; 2024; [DOI: https://dx.doi.org/10.1088/1741-2552/ad8bf3]
Ren, Q; Sun, X; Fu, X; Zhang, S; Yuan, Y; Wu, H; Li, X; Wang, X; Zhang, F. A review of automatic detection of epilepsy based on EEG signals. J Semicond; 2023; 44,
Rethlefsen, ML; Kirtley, S; Waffenschmidt, S; Ayala, AP; Moher, D; Page, MJ; Koffel, JB. PRISMA-s: an extension to the PRISMA statement for reporting literature searches in systematic reviews. Syst Rev; 2021; 10, 39. [DOI: https://dx.doi.org/10.1186/s13643-020-01542-z]
Rijnders, B; Korkmaz, EE; Yildirim, F. Hybrid machine learning method for a connectivity-based epilepsy diagnosis with resting-state EEG. Med Biol Eng Comput; 2022; 60,
Rivera, MJ; Sanchis, J; Corcho, O; Teruel, MA; Trujillo, J. Evaluating CNN methods for epileptic seizure type classification using EEG data. IEEE Access; 2024; 12, pp. 75483-75495. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3406332]
Rout, SK; Sahani, M; Dora, C; Biswal, PK; Biswal, B. An efficient epileptic seizure classification system using empirical wavelet transform and multi-fuse reduced deep convolutional neural network with digital implementation. Biomed Signal Process Control; 2022; [DOI: https://dx.doi.org/10.1016/j.bspc.2021.103281]
Ru, Y; Wei, Z; An, G; Chen, H. Combining data augmentation and deep learning for improved epilepsy detection. Front Neurol; 2024; [DOI: https://dx.doi.org/10.3389/fneur.2024.1378076]
Rukhsar, S; Tiwari, AK. Lightweight convolution transformer for cross-patient seizure detection in multi-channel EEG signals. Comput Methods Progr Biomed; 2023; [DOI: https://dx.doi.org/10.1016/j.cmpb.2023.107856]
Sabarivani, A; Ramadevi, R. Detection and classification of epilepsy using hybrid convolutional neural network. Concurr Eng Res Appl; 2022; 30,
Sadam, SSP; Nalini, NJ. Epileptic seizure detection using scalogram-based hybrid CNN model on EEG signals. Signal Image Video Process; 2024; 18,
Sadam, SP; Nj, N. Feature fusion of time-frequency and deep learning features for epileptic seizure detection using EEG signals. Int J Electric Electron Res; 2023; 11,
Saidi, A; Othman, SB; Dhouibi, M; Saoud, SB. CNN inference acceleration on limited resources FPGA platforms\_epilepsy detection case study. Int J Inf Commun Technol; 2023; 12,
Salafian, B; Ben-Knaan, EF; Shlezinger, N; De Ribaupierre, S; Farsad, N. MICAL: Mutual information-based CNN-aided learned factor graphs for seizure detection from EEG signals. IEEE Access; 2023; 11, pp. 23085-23096. [DOI: https://dx.doi.org/10.1109/ACCESS.2023.3252897]
Salini, GI; Sowmy, I. Adadelta-CSA: Adadelta-chameleon swarm algorithm for EEG-based epileptic seizure detection. Int J Comput Intell Appl; 2024; [DOI: https://dx.doi.org/10.1142/S1469026824500305]
Samee, NA; Mahmoud, NF; Aldhahri, EA; Rafiq, A; Muthanna, MSA; Ahmad, I. RNN and BiLSTM fusion for accurate automatic epileptic seizure diagnosis using EEG signals. Life; 2022; [DOI: https://dx.doi.org/10.3390/life12121946]
Sameer, M; Gupta, B. CNN based framework for detection of epileptic seizures. Multimed Tools Appl; 2022; 81,
Sami Nafea, M; Ismail, ZH. IMPRESS: informative mutual patch representation for EEG semi-supervised learning in seizure type classification. IEEE Access; 2024; 12, pp. 162251-162266. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3487532]
Saminu, S; Xu, G; Shuai, Z; Abd El Kader, I; Jabire, AH; Ahmed, YK; Karaye, IA; Ahmad, IS. Epileptic EEG signals rhythms analysis in the detection of focal and non-focal seizures based on optimised machine learning and deep neural network architecture. J Mechan Med Biol; 2024; [DOI: https://dx.doi.org/10.1142/S0219519423500653]
Sánchez-Hernández, SE; Torres-Ramos, S; Román-Godínez, I; Salido-Ruiz, RA. Evaluation of the relation between ictal EEG features and XAI explanations. Brain Sci; 2024; [DOI: https://dx.doi.org/10.3390/brainsci14040306]
Sarvi Zargar, B; Karami Mollaei, MR; Ebrahimi, F; Rasekhi, J. Generalizable epileptic seizures prediction based on deep transfer learning. Cogn Neurodyn; 2023; 17,
Sateesh Kumar Reddy, C; Suchetha, M. A 1-d CNN-FCM model for the classification of epileptic seizure disorders. Neural Comput Appl; 2023; 35,
Shafiezadeh, S; Duma, GM; Mento, G; Danieli, A; Antoniazzi, L; Del Popolo, Cristaldi F; Bonanni, P; Testolin, A. Calibrating deep learning classifiers for patient-independent electroencephalogram seizure forecasting. Sensors; 2024; [DOI: https://dx.doi.org/10.3390/s24092863]
Shafiqul, IM; Thapa, K; Yang, S-H. Epileptic-net: an improved epileptic seizure detection system using dense convolutional block with attention network from EEG. Sensors; 2022; [DOI: https://dx.doi.org/10.3390/s22030728]
Shanmugam, S; Dharmar, S. A CNN-LSTM hybrid network for automatic seizure detection in EEG signals. Neural Comput Appl; 2023; 35,
Sharma, R. Localization of epileptic surgical area using automated hybrid approach based on higher-order statistics with sensitivity analysis and residual wavelet transform. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.105192]
Sharma, S. Deep learning for cognitive task and seizure classification with Hilbert-Huang transform and variational mode decomposition. Mach Vis Augment Intell; 2024; [DOI: https://dx.doi.org/10.1007/978-981-97-4359-9_6]
Sharma, S; Srivastava, R. Multi-modality NDE fusion using encoder-decoder networks for identify multiple neurological disorders from EEG signals. Technol Health Care; 2024; [DOI: https://dx.doi.org/10.1177/09287329241291334]
Sharma, S; Srivastava, R. Enhanced EEG decoding with weight freezing in parallel deep neuro-fuzzy networks (PDNFN) for multi-neurological disorder diagnosis. Iran J Comput Sci; 2025; [DOI: https://dx.doi.org/10.1007/s42044-025-00270-8]
Sharma, S; Srivastava, R. Electroencephalogram (EEG) based fuzzy logic and spiking neural networks (FLSNN) for advanced multiple neurological disorder diagnosis. Brain Topography; 2025; [DOI: https://dx.doi.org/10.1007/s10548-025-01106-1]
Shekokar, KS; Dour, S. Automatic epileptic seizure detection using LSTM networks. World J Eng; 2022; 19,
Shi, S; Liu, W. B2-ViT net: broad vision transformer network with broad attention for seizure prediction. IEEE Trans Neural Syst Rehabil Eng; 2024; 32, pp. 178-188. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3346955]
Shi, Z; Liao, Z; Tabata, H. Enhancing performance of convolutional neural network-based epileptic electroencephalogram diagnosis by asymmetric stochastic resonance. IEEE J Biomed Health Inform; 2023; 27,
Shoeb AH (2010) Application of machine learning to epileptic seizure onset detection and treatment. PhD thesis, Massachusetts Institute of Technology
Shu, K; Wu, L; Zhao, Y; Liu, A; Qian, R; Chen, X. Data augmentation for seizure prediction with generative diffusion model. IEEE Trans Cognit Dev Syst; 2024; [DOI: https://dx.doi.org/10.1109/TCDS.2024.3489357]
Singh, K; Malhotra, J. Predicting epileptic seizures from EEG spectral band features using convolutional neural network. Wirel Pers Commun; 2022; 125,
Singh, K; Malhotra, J. Smart neurocare approach for detection of epileptic seizures using deep learning based temporal analysis of EEG patterns. Multimed Tools Appl; 2022; 81,
Singh, S; Jadli, H; Padma Priya, R; Surya Prasath, VB. KDTL: knowledge-distilled transfer learning framework for diagnosing mental disorders using EEG spectrograms. Neural Comput Appl; 2024; 36,
Sivasaravana Babu, S; Prabhu, V; Parthasarathy, V; Saravana Kumar, G. A heuristic-concatenated feature classification algorithm (h-CFCA) for autism and epileptic seizure detection. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.105245]
Soliman S, Fouad AM, Mourad E, Hossam S, Ehab M, Selim S, Darweesh MS (2022) Deep learning approaches for epileptic seizure prediction: a review. In: 2022 4th novel intelligent and leading emerging sciences conference (NILES), pp 01–06. https://doi.org/10.1109/NILES56402.2022.9942420
Sonawane, PS; Helonde, JB. Smart societal optimization-based deep learning convolutional neural network model for epileptic seizure prediction. Comput Methods Biomech Biomed Eng Imaging Vis; 2024; [DOI: https://dx.doi.org/10.1080/21681163.2023.2280551]
Song, Z; Deng, B; Wang, J; Yi, G; Yue, W. Epileptic seizure detection using brain-rhythmic recurrence biomarkers and ONASNet-based transfer learning. IEEE Trans Neural Syst Rehabil Eng; 2022; 30, pp. 979-989. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3165060]
Song, Z; Deng, B; Zhu, Y; Cai, L; Wang, J; Yi, G. Probing epileptic disorders with lightweight neural network and EEG’s intrinsic geometry. Nonlinear Dyn; 2023; 111,
Srinath, R; Gayathri, R. Epilepsy disorder detection and diagnosis using empirical mode decomposition and deep learning architecture. Concurr Comput Pract Exp; 2022; [DOI: https://dx.doi.org/10.1002/cpe.6903]
Srinivasan, S; Dayalane, S; Mathivanan, S; Rajadurai, H; Jayagopal, P; Dalu, GT. Detection and classification of adult epilepsy using hybrid deep learning approach. Sci Rep; 2023; [DOI: https://dx.doi.org/10.1038/s41598-023-44763-7]
Stevenson, NJ; Tapani, K; Lauronen, L; Vanhatalo, S. A dataset of neonatal EEG recordings with seizure annotations. Sci Data; 2019; [DOI: https://dx.doi.org/10.1038/sdata.2019.39]
Sun, Y; Jin, W; Si, X; Zhang, X; Cao, J; Wang, L; Yin, S; Ming, D. Continuous seizure detection based on transformer and long-term iEEG. IEEE J Biomed Health Inform; 2022; 26,
Sun, C; Xu, C; Li, H; Bo, H; Ma, L; Li, H. A novel multi-feature fusion attention neural network for the recognition of epileptic EEG signals. Front Comput Neurosci; 2024; [DOI: https://dx.doi.org/10.3389/fncom.2024.1393122]
Sun, R; Sohrabpour, A; Joseph, B; Worrell, G; He, B. Seizure sources can be imaged from scalp EEG by means of biophysically constrained deep neural networks. Adv Sci; 2024; [DOI: https://dx.doi.org/10.1002/advs.202405246]
Sunkara, M. Tri-SeizureDualNet: A novel multimodal brain seizure detection using triple stream skipped feature extraction module entrenched dual parallel attention transformer. Biomed Signal Process Control; 2024; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.105593]
Supriya, S; Siuly, S; Wang, H; Zhang, Y. Epilepsy detection from EEG using complex network techniques: a review. IEEE Rev Biomed Eng; 2021; 16, pp. 292-306. [DOI: https://dx.doi.org/10.1109/RBME.2021.3055956]
Swami, P; Maheshwari, J; Kumar, M; Bhatia, M. Evolutionary transfer optimization-based approach for automated ictal pattern recognition using brain signals. Front Human Neurosci; 2024; [DOI: https://dx.doi.org/10.3389/fnhum.2024.1386168]
Swami P, Panigrahi B, Nara S, Bhatia M, Gandhi T. EEG Epilepsy datasets. https://doi.org/10.13140/RG.2.2.14280.32006
Tang, Y; Wu, Q; Mao, H; Guo, L. Epileptic seizure detection based on path signature and bi-LSTM network with attention mechanism. IEEE Trans Neural Syst Rehabil Eng; 2024; 32, pp. 304-313. [DOI: https://dx.doi.org/10.1109/TNSRE.2024.3350074]
Thara, DK; Premasudha, BG; Krivic, S. Detection of epileptic seizure events using pre-trained convolutional neural network. VGGNet ResNet Expert Syst; 2023; [DOI: https://dx.doi.org/10.1111/exsy.13447]
Thuwajit, P; Rangpong, P; Sawangjai, P; Autthasan, P; Chaisaen, R; Banluesombatkul, N; Boonchit, P; Tatsaringkansakul, N; Sudhawiyangkul, T; Wilaiprasitporn, T. EEGWaveNet: multiscale CNN-based spatiotemporal feature extraction for EEG seizure detection. IEEE Trans Ind Inf; 2022; 18,
Tomson, T; Walczak, T; Sillanpaa, M; Sander, JWAS. Sudden unexpected death in epilepsy: a review of incidence and risk factors. Epilepsia; 2005; 46, pp. 54-61. [DOI: https://dx.doi.org/10.1111/j.1528-1167.2005.00411.x]
Tong, PF; Dong, B; Zeng, X; Chen, L; Chen, SX. Detection of interictal epileptiform discharges using transformer based deep neural network for patients with self-limited epilepsy with centrotemporal spikes. Biomed Signal Process Control; 2025; [DOI: https://dx.doi.org/10.1016/j.bspc.2024.107238]
Tripathi, PM; Kumar, A; Kumar, M; Komaragiri, RS. Automatic seizure detection and classification using super-resolution superlet transform and deep neural network -a preprocessing-less method. Comput Methods Progr Biomed; 2023; [DOI: https://dx.doi.org/10.1016/j.cmpb.2023.107680]
Tuncer, E; Doğru Bolat, E. Classification of epileptic seizures from electroencephalogram (EEG) data using bidirectional short-term memory (Bi-LSTM) network architecture. Biomed Signal Process Control; 2022; [DOI: https://dx.doi.org/10.1016/j.bspc.2021.103462]
Vaithilingam, SD; Regulagedda, P. The deep learning based epileptic seizure detection using 2-layer convolutional network with long short-term memory. Int J Intell Eng Syst; 2024; 17,
Varlı, M; Yılmaz, H. Multiple classification of EEG signals and epileptic seizure diagnosis with combined deep learning. J Comput Sci; 2023; [DOI: https://dx.doi.org/10.1016/j.jocs.2023.101943]
Varli, M; Yilmaz, H. Multiple classification of EEG signals and epileptic seizure diagnosis with combined deep learning. J Comput Sci; 2023; [DOI: https://dx.doi.org/10.1016/j.jocs.2023.101943]
Varnosfaderani, SM; McNulty, I; Sarhan, NJ; Abood, W; Alhawari, M. An efficient epilepsy prediction model on European dataset with model evaluation considering seizure types. IEEE J Biomed Health Inform; 2024; 28,
Varun, SK; Bollu, TKR. WearNeuroNet: an interpretable light-weight deep learning approach for ictal-interictal classification for limited channel EEG wearables. IEEE Sens Lett; 2024; [DOI: https://dx.doi.org/10.1109/LSENS.2024.3417453]
Vidyaratne, LS; Alam, M; Glandon, AM; Shabalina, A; Tennant, C; Iftekharuddin, KM. Deep cellular recurrent network for efficient analysis of time-series data with spatial information. IEEE Trans Neural Netw Learn Syst; 2022; 33,
Vieira, JC; Guedes, LA; Santos, MR; Sanchez-Gendriz, I. Using explainable artificial intelligence to obtain efficient seizure-detection models based on electroencephalography signals. Sensors (Basel, Switzerland); 2023; [DOI: https://dx.doi.org/10.3390/s23249871]
Visalini, K; Alagarsamy, S; Nagarajan, D. Neonatal seizure detection using deep belief networks from multichannel EEG data. Neural Comput Appl; 2023; 35,
Wang, X; Zhang, C; Karkkainen, T; Chang, Z; Cong, F. Channel increment strategy-based 1d convolutional neural networks for seizure prediction using intracranial EEG. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 316-325. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3222095]
Wang, Y; Shi, Y; He, Z; Chen, Z; Zhou, Y. Combining temporal and spatial attention for seizure prediction. Health Inf Sci Syst; 2023; [DOI: https://dx.doi.org/10.1007/s13755-023-00239-6]
Wang, Y; Shi, Y; Cheng, Y; He, Z; Wei, X; Chen, Z; Zhou, Y. A spatiotemporal graph attention network based on synchronization for epileptic seizure prediction. IEEE J Biomed Health Inform; 2023; 27,
Wang, Y; Cui, W; Yu, T; Li, X; Liao, X; Li, Y. Dynamic multi-graph convolution-based channel-weighted transformer feature fusion network for epileptic seizure prediction. IEEE Trans Neural Syst Rehabil Eng; 2023; 31, pp. 4266-4277. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3321414]
Wang, J; Ge, X; Shi, Y; Sun, M; Gong, Q; Wang, H; Huang, W. Dual-modal information bottleneck network for seizure detection. Int J Neural Syst; 2023; [DOI: https://dx.doi.org/10.1142/S0129065722500617]
Wang, Z; Hou, S; Xiao, T; Zhang, Y; Lv, H; Li, J; Zhao, S; Zhao, Y. Lightweight seizure detection based on multi-scale channel attention. Int J Neural Syst; 2023; [DOI: https://dx.doi.org/10.1142/S0129065723500612]
Wang, S; Feng, H; Lv, H; Nie, C; Feng, W; Peng, H; Zhang, L; Zhao, Y. Cross-subject seizure detection via unsupervised domain-adaptation. Int J Neural Syst; 2024; [DOI: https://dx.doi.org/10.1142/S0129065724500552]
Wang, J; Gao, R; Zheng, H; Zhu, H; Shi, C-JR. SSGCNet: a sparse spectra graph convolutional network for epileptic EEG signal classification. IEEE Trans Neural Netw Learn Syst; 2024; 35,
Wang, X; Gao, Z; Zhang, M; Wang, Y; Yang, L; Lin, J; Karkkainen, T; Cong, F. Combination of channel reordering strategy and dual CNN-LSTM for epileptic seizure prediction using three iEEG datasets. IEEE J Biomed Health Inform; 2024; 28,
Wang, C; Liu, L; Zhuo, W; Xie, Y. An epileptic EEG detection method based on data augmentation and lightweight neural network. IEEE J Transl Eng Health Med; 2024; 12, pp. 22-31. [DOI: https://dx.doi.org/10.1109/JTEHM.2023.3308196]
Weiss, SA; Berry, B; Chervoneva, I; Waldman, Z; Guba, J; Bower, M; Kucewicz, M; Brinkmann, B; Kremen, V; Khadjevand, F; Varatharajah, Y; Guragain, H; Sharan, A; Wu, C; Staba, R; Engel, J; Sperling, M; Worrell, G. Visually validated semi-automatic high-frequency oscillation detection aides the delineation of epileptogenic regions during intra-operative electrocorticography. Clin Neurophysiol; 2018; 129,
Wu, D; Yang, J; Sawan, M. Bridging the gap between patient-specific and patient-independent seizure prediction via knowledge distillation. J Neural Eng; 2022; [DOI: https://dx.doi.org/10.1088/1741-2552/ac73b3]
Wu, X; Zhang, T; Zhang, L; Qiao, L. Epileptic seizure prediction using successive variational mode decomposition and transformers deep learning network. Front Neurosci; 2022; [DOI: https://dx.doi.org/10.3389/fnins.2022.982541]
Wu, X; Yang, Z; Zhang, T; Zhang, L; Qiao, L. An end-to-end seizure prediction approach using long short-term memory network. Front Human Neurosci; 2023; [DOI: https://dx.doi.org/10.3389/fnhum.2023.1187794]
Xia, L; Wang, R; Ye, H; Jiang, B; Li, G; Ma, C; Gao, Z. Hybrid LSTM-transformer model for the prediction of epileptic seizure using scalp EEG. IEEE Sens J; 2024; 24,
Xiong, Z; Liu, Y; Jiang, P. An EEG-based seizure recognition method using dynamic routing. IEEE Access; 2024; 12, pp. 74054-74068. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3402075]
Xu, M; Jie, J; Zhou, W; Zhou, H; Jin, S. Synthetic epileptic brain activities with TripleGAN. Comput Math Methods Med; 2022; [DOI: https://dx.doi.org/10.1155/2022/2841228]
Xu, X; Zhang, Y; Zhang, R; Xu, T. Patient-specific method for predicting epileptic seizures based on DRSN-GRU. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2022.104449]
Xu, Y; Yang, J; Ming, W; Wang, S; Sawan, M. Shorter latency of real-time epileptic seizure detection via probabilistic prediction. Expert Syst Appl; 2024; [DOI: https://dx.doi.org/10.1016/j.eswa.2023.121359]
Xu, J; Yuan, S; Shang, J; Wang, J; Yan, K; Yang, Y. Spatiotemporal network based on GCN and BiGRU for seizure detection. IEEE J Biomed Health Inform; 2024; 28,
Xu, T; Wu, Y; Tang, Y; Zhang, W; Cui, Z. Dynamic functional connectivity neural network for epileptic seizure prediction using multi-channel EEG signal. IEEE Signal Process Lett; 2024; 31, pp. 1499-1503. [DOI: https://dx.doi.org/10.1109/LSP.2024.3400037]
Xue, W; Zhou, T; Cai, J. Horizontal progressive and longitudinal leapfrogging fuzzy classification with feature activity adjustment. Appl Soft Comput; 2022; [DOI: https://dx.doi.org/10.1016/j.asoc.2022.108511]
Yadav, A; Rathee, S; Sheoran, D; Kumar, P. Advanced epileptic seizure detection using deep learning and Bayesian optimization. SN Comput Sci; 2024; [DOI: https://dx.doi.org/10.1007/s42979-024-03201-9]
Yan, J; Li, J; Xu, H; Yu, Y; Xu, T. Seizure prediction based on transformer using scalp electroencephalogram. Appl Sci (Switzerland); 2022; [DOI: https://dx.doi.org/10.3390/app12094158]
Yang, Z; Li, S. A new fuzzy-based ensemble framework based on attention-based deep learning architectures for automated detection of abnormal EEG. Int J Syst Assur Eng Manage; 2024; 15,
Yang, Y; Li, F; Luo, J; Qin, X; Huang, D. Epileptic focus localization using transfer learning on multi-modal EEG. Front Comput Neurosci; 2023; [DOI: https://dx.doi.org/10.3389/fncom.2023.1294770]
Yang, C-Y; Chen, P-C; Huang, W-C. Cross-domain transfer of EEG to EEG or ECG learning for CNN classification models. Sensors; 2023; [DOI: https://dx.doi.org/10.3390/s23052458]
Yang, Y; Luan, T; Yu, Z; Zhang, M; Li, F; Chen, X; Gao, F; Zhang, Z. Technological vanguard: the outstanding performance of the LTY-CNN model for the early prediction of epileptic seizures. J Transl Med; 2024; [DOI: https://dx.doi.org/10.1186/s12967-024-04945-x]
Yaqing L, Yalin W, Tiancheng W (2025) Non-invasive localization of epileptogenic zone in drug-resistant epilepsy based on time–frequency analysis and vgg convolutional neural network. Bioengineering 12(5): https://doi.org/10.3390/bioengineering12050443
Yu, Z; Albera, L; Le Bouquin, Jeannes R; Kachenoura, A; Karfoul, A; Yang, C; Shu, H. Epileptic seizure prediction using deep neural networks via transfer learning and multi-feature fusion. Int J Neural Syst; 2022; [DOI: https://dx.doi.org/10.1142/S0129065722500320]
Yuan, S; Yan, K; Wang, S; Liu, J-X; Wang, J. EEG-based seizure prediction using hybrid DenseNet-ViT network with attention fusion. Brain Sci; 2024; [DOI: https://dx.doi.org/10.3390/brainsci14080839]
Zaid, Y; Sah, M; Direkoglu, C. Pre-processed and combined EEG data for epileptic seizure classification using deep learning. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.104738]
Zhang, X; Dong, S; Shen, Q; Zhou, J; Min, J. Deep extreme learning machine with knowledge augmentation for EEG seizure signal recognition. Front Neuroinf; 2023; [DOI: https://dx.doi.org/10.3389/fninf.2023.1205529]
Zhang, Y; Xiao, T; Wang, Z; Lv, H; Wang, S; Feng, H; Zhao, S; Zhao, Y. Hybrid network for patient-specific seizure prediction from EEG data. Int J Neural Syst; 2023; [DOI: https://dx.doi.org/10.1142/S0129065723500569]
Zhang, G; Zhang, A; Liu, H; Luo, J; Chen, J. Positional multi-length and mutual-attention network for epileptic seizure classification. Front Comput Neurosci; 2024; [DOI: https://dx.doi.org/10.3389/fncom.2024.1358780]
Zhang, J; Zheng, S; Chen, W; Du, G; Fu, Q; Jiang, H. A scheme combining feature fusion and hybrid deep learning models for epileptic seizure detection and prediction. Sci Rep; 2024; [DOI: https://dx.doi.org/10.1038/s41598-024-67855-4]
Zhang, Z; Ji, T; Xiao, M; Wang, W; Yu, G; Lin, T; Jiang, Y; Zhou, X; Lin, Z. Cross-patient automatic epileptic seizure detection using patient-adversarial neural networks with spatio-temporal EEG augmentation. Biomed Signal Process Control; 2024; [DOI: https://dx.doi.org/10.1016/j.bspc.2023.105664]
Zhao, S; Yang, J; Sawan, M. Energy-efficient neural network for epileptic seizure prediction. IEEE Trans Biomed Eng; 2022; 1, pp. 401-411. [DOI: https://dx.doi.org/10.1109/TBME.2021.3095848]
Zhao, Y; Li, C; Liu, X; Qian, R; Song, R; Chen, X. Patient-specific seizure prediction via adder network and supervised contrastive learning. IEEE Trans Neural Syst Rehabil Eng; 2022; 30, pp. 1536-1547. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3180155]
Zhao, Y; He, J; Zhu, F; Xiao, T; Zhang, Y; Wang, Z; Xu, F; Niu, Y. Hybrid attention network for epileptic EEG classification. Int J Neural Syst; 2023; [DOI: https://dx.doi.org/10.1142/S0129065723500314]
Zhao, Y; Chu, D; He, J; Xue, M; Jia, W; Xu, F; Zheng, Y. Interactive local and global feature coupling for EEG-based epileptic seizure detection. Biomed Signal Process Control; 2023; [DOI: https://dx.doi.org/10.1016/j.bspc.2022.104441]
Zhao, W; Wang, W-F; Patnaik, LM; Zhang, B-C; Weng, S-J; Xiao, S-X; Wei, D-Z; Zhou, H-F. Residual and bidirectional LSTM for epileptic seizure detection. Front Comput Neurosci; 2024; [DOI: https://dx.doi.org/10.3389/fncom.2024.1415967]
Zhao X, Zhao Q, Tanaka T, Cao J, Kong W, Sugano H, Yoshida N (2018) Detection of epileptic foci based on interictal iEEG by using convolutional neural network. In: 2018 IEEE 23rd international conference on digital signal processing (DSP), pp 1–5. https://doi.org/10.1109/ICDSP.2018.8631885
Zhou, W; Zheng, W; Feng, Y; Li, X. LMA-EEGNet: a lightweight multi-attention network for neonatal seizure detection using EEG signals. Electronics (Switzerland); 2024; [DOI: https://dx.doi.org/10.3390/electronics13122354]
Zhu, Z; Zhang, M; Cui, Y. Multiscale dense convolutional network for EEG seizure detection. ICIC Express Lett Part B Appl; 2024; 15,
Zhu, P; Zhou, W; Cao, C; Liu, G; Liu, Z; Shang, W. A novel SE-TCN-BiGRU hybrid network for automatic seizure detection. IEEE Access; 2024; 12, pp. 127328-127340. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3406909]
Zweiphenning, W; Klooster, MAVT; Klink, NECV; Leijten, FSS; Ferrier, CH; Gebbink, T; Huiskamp, G; Zandvoort, MJEV; Schooneveld, MMJV; Bourez, M. Intraoperative electrocorticography using high-frequency oscillations or spikes to tailor epilepsy surgery in the netherlands (the hfo trial): a randomised, single-blind, adaptive non-inferiority trial. The Lancet Neurol; 2022; 21,
© The Author(s) 2025. This work is published under http://creativecommons.org/licenses/by-nc-nd/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.