Generative Adversarial Networks for Performance Enhancement in EEG-Based Brain-Computer Interfaces: A Comprehensive Review

Generative Adversarial Networks for Performance Enhancement in EEG-Based Brain-Computer Interfaces: A Comprehensive Review Student's Name Institution Course Code: Course Name Professor's Name Submission Date Abstract EEG-based BCIs enable individuals to interact with and control devices using brain signals, offering significant potential for assistive technology and neurorehabilitation applications. GANs, derived from game theory, employ adversarial training between generator and discriminator networks to generate synthetic data that closely mimic real EEG data distributions. Using literature studies as the data source, this paper thematically analyzed the findings from major key studies. Key findings highlighted that GANs's effectiveness is crucial in improving BCI performance metrics such as accuracy, signal-to-noise ratio, and Inception Score. GANs are used in motor imagery decoding in BCI systems, motor rehabilitation, and communication aids. Despite advancements, challenges persist, including the need for robust training protocols and addressing the variability of EEG signals across individuals. Future research directions emphasize refining GAN architectures tailored to EEG characteristics and exploring novel applications beyond motor imagery tasks. This review contributes to understanding how GANs can advance EEG-based BCI technology, offering insights for researchers and practitioners aiming to optimize BCI performance through innovative machine-learning techniques. The implications of these findings suggest that GAN-enhanced BCIs hold significant potential for improving assistive technologies, neurorehabilitation, and human-computer interaction. The paper concludes by proposing future research directions to address the persistent challenges in integrating GANs with BCIs. Keywords: Brain-computer interface (BCI), Generative Adversarial Networks (GANs), motor imagery (MI), data augmentation, signal enhancement, feature extraction, deep learning, epilepsy Table of Contents TOC \o "1-3" \h \z \u 1. INTRODUCTION PAGEREF _Toc175766264 \h 41.1 Background of the Study PAGEREF _Toc175766265 \h 41.2 Introduction to Generative Adversarial Networks (GANs) PAGEREF _Toc175766266 \h 71.3 Purpose of the Review PAGEREF _Toc175766267 \h 112. METHODOLOGY PAGEREF _Toc175766268 \h 122.1 Literature Search Strategy PAGEREF _Toc175766269 \h 122.2 Guidelines for Selection Criteria PAGEREF _Toc175766270 \h 142.3 Review Process PAGEREF _Toc175766271 \h 162.4 Ethical Considerations PAGEREF _Toc175766272 \h 173. RESEARCH RESULTS PAGEREF _Toc175766273 \h 183.1 Applications of GANs in EEG-Based BCIs PAGEREF _Toc175766274 \h 354. PERFORMANCE METRICS AND EVALUATION PAGEREF _Toc175766275 \h 444.1 Evaluation Metrics PAGEREF _Toc175766276 \h 444.2 Comparative Analysis PAGEREF _Toc175766277 \h 465. DISCUSSION OF KEY FINDINGS AND TRENDS PAGEREF _Toc175766278 \h 475.1 Recent Advances PAGEREF _Toc175766279 \h 475.2 Challenges and Limitations PAGEREF _Toc175766280 \h 495.3 Research Gaps PAGEREF _Toc175766281 \h 516. FUTURE DIRECTIONS PAGEREF _Toc175766282 \h 526.1 Innovative Approaches PAGEREF _Toc175766283 \h 526.2. Practical Implications PAGEREF _Toc175766284 \h 527. CONCLUSION PAGEREF _Toc175766285 \h 53References PAGEREF _Toc175766286 \h 54 1. INTRODUCTION 1.1 Background of the Study 1.1.1 Background on EEG-Based BCIs Until recently, the ability to manipulate one's surroundings via thoughts was considered fictional. Nevertheless, the progression of technology has introduced a novel paradigm: Presently, physicians can utilize the electrical impulses generated by brain function to engage with, exert influence upon, or modify their surroundings (Karakas & Yildiz, 2020). Kübler (2019) states that the developing domain of Brain-Computer Interfaces (BCIs) technology can enable individuals who are unable to talk or use their limbs to regain the ability to communicate and control assistive devices for walking and moving objects. Without depending on the typical pathways for communication between the brain and the body's motor processes, a BCI enables users to command and interact (Padfield et al., 2019). BCI users' brain activity is frequently measured using electroencephalographic (EEG). According to Abiri et al. (2019), BCI normally functions through four processes: logging brain activity, extracting characteristics, gathering pertinent data, and combining data for valuable applications. BCIs establish a connection between the brain and an external device, such as a computer, for various uses like motor rehabilitation and cognitive training. EEG detection is the most used non-invasive BCI brainwave capture method. In 1924, Hans Berger discovered how to monitor brain electrical impulses through the scalp, making EEG possible (Värbu et al., 2022). According to Padfield et al. (2019), EEG signals are less expensive and easier to obtain compared to functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG); moreover, their enhanced capacity to record temporal variations makes them better for BCI applications. They also travel better than MEG or fMRI. Wang et al. (2019) divided BCIs into evoked and spontaneous groups using EEG. Visual, auditory, or sensory cues must be presented to evoked systems. By measuring brain reactions to stimuli, the BCI system determines volition. Spontaneous BCIs use mental processes to function without external stimulation. 1.1.2 Significance, Applications of EEG-Based BCIs EEG is non-invasive, as opposed to other BCI modalities, and it is utilized in assistive technology, neurorehabilitation, and medical diagnostics. Since it is a portable device, EEG recording is vital within numerous disciplines for real-time monitoring of connected brain activities with therapeutic implications (Orban et al., 2022). Support for EEG-based BCI systems has been increased by machine learning (ML), among other recent technologies like wireless recording, that might better the lives of disabled people. The temporal and spatial resolution, which are reliable, make it possible to diagnose epilepsy, insomnia, and Alzheimer’s using EEG. EEG is relatively cheaper to record when assessed against fMRI, MEG, Near-infrared (NIR) spectroscopy and other methods for capturing brain waves non-invasively. The BCI utilizes the cerebral cortex dynamics of the EEG to enhance human comprehension of cognitive tasks and contribute to computer science and engineering progress. The EEG-based BCI is extensively utilized in two prominent application areas: enhancing human-computer interaction (HCI) and entertainment games and playing a crucial part in pattern identification and machine learning (ML) of brain dynamics. Conventional HCI mainly relies on manual operations and communications, which limits its effectiveness in promoting HCI. Nevertheless, the BCI has significantly altered the approach to HCI in intricate and challenging operational settings, potentially transforming HCI into a groundbreaking strategy. In addition, the tiny wireless BCI headset, primarily designed for the gaming sector, is both wearable and adaptable and may be fitted without any prior configuration. While they may not possess the same efficiency level as other cutting-edge BCI technologies utilized in clinical trials, they are appropriate for entertainment. They are employed in the development of interactive games. 1.1.3 Current Challenges of EEG-Based BCIs While BCI technology opens doors for those with severe motor disabilities by facilitating communication and control, a significant challenge still lies in reliably and correctly identifying tasks by employing EEG data. Developing a BCI system employing MI brain signals has faced usability and technological challenges. Multiple sources of noise make EEG data extraction problematic (Chaddad et al., 2023). Increased BCI performance requires signal quality and viable noise-reduction strategies. EEG signal-to-noise ratios are attenuated, especially in real life. EEG-based motor imagery brain-computer interface (MI-BCI) systems have successfully used several signal-processing methods. However, Dhiman (2023) notes that usability, signal quality and noise, real-time processing, artifact removal, training, and non-linearity remain issues. Usually, users want an accessible system. Dhiman (2023) attributes these problems to training challenges in group differentiation. Mumtaz et al. (2021) imply that artifact removal may be difficult due to the complexity of the techniques or nonlinear noise in the EEG data. Due of aberration nonlinearity, Mumtaz et al. say it's difficult to extract artifacts without losing neuronal data. Anomalies, including cardiac variations, muscular activity, and eye blinks disrupt EEG readings. Artifact removal involves recognizing and removing specific items (Jiang et al., 2019). Thus, finding a practical artifact removal method that meets all application requirements is difficult. The training procedure conducted by the user is a time-consuming activity for brain-computer interface systems. In addition to handling the electrical impulses from the EEG, the task entails teaching and guiding users on effectively using the BCI system (Rashid et al., 2020). A key area of research is to make BCIs more intuitive and accessible to a broader range of people, reducing the time it takes to learn how to use them. A key research solution to address the issue of time consumption is to employ single trials instead of conducting multi-trial testing. Padfield et al. (2022) state that significant training is necessary for subjects to control MI-based BCIs effectively. The human brain is a prime illustration of a non-linear process that produces chaotic biological electrical EEG signals. Developing effective classifiers using non-linear signals requires extensive research. 1.2 Introduction to Generative Adversarial Networks (GANs) 1.2.1 Basic Architecture of GANs and General Capabilities GANs are ML techniques that are based on game theory. Occasionally, researchers may encounter limitations in the availability of real-world EEG data, which can provide a weakness when training ML. In order to address this issue, a GAN is utilized. The "generator networks" and "discriminator networks," two neural networks trained simultaneously, comprise the GAN. This structure aims to enhance the data (Song et al., 2021). See Fig 1; the “generator” creates synthetic data that closely mimics accurate data by matching the original distribution. However, the discriminator's role is to determine whether the created data is real or fake (Cao, 2020). Song et al. (2021) argue that via repeated rounds of adversarial training, the two modules eventually reach a state of equilibrium, where the generator can produce very realistic data that the discriminator cannot differentiate. Figure 1 Basic Structure of Discriminator and Generator GANS Adopted from Cho and Yoon (2020). Both the Auxiliary Classifier GANs advocated by Odena et al. (Cho & Yoon, 2020) and Mirza et al. conditional GANs (Song et al., 2021) integrated category information as one more pre-existing supportive condition to both the “generator” and the “discriminator models.” Hence, samples that can be categorized can be produced. A conditional GAN (cGAN) uses latent vectors to create data on given information (Cho & Yoon, 2020). The most popular among various cGANs is the auxiliary classifier GAN (ACGAN) (He et al., 2019). However, in ACGAN, if natural and created datasets have the same probability distributions, both the auxiliary subjects that classify the generator and the discriminator can be understood as an assemblage of many GANs. Each GAN in this cluster undergoes cross-entropy adversarial reduction training with all shared hidden layers. In the modified ACGAN, the generated data classification loss of the ACGAN discriminator is eliminated to prevent interference with the training of each GAN, treating the ACGAN as a collection of GANs. Each GAN can only undergo training when the distributions of natural and produced data are identical. Consequently, a potential issue may occur where each GAN may not receive training at the first stages of ACGAN training. The adversarial process facilitates the enhancement of the generator's outputs to deceive the discriminator, generating exceptionally lifelike data. Once convinced of the ability of GANs to produce EEG and non-stationary time series, various applications have been explored. GANs have been widely utilized in many domains, such as “image synthesis, video production, and data augmentation.” Jiang et al. (2021) employed a conditional deep convolutional GAN to enhance data by applying wavelet transform. Scholars have attempted to create unprocessed EEG waves for more extensive applications and create EEG characteristics. Song et al.'s (2021) study demonstrated the extraction of a feature vector from the target subject's data as a prerequisite for GANs. This process resulted in acquiring multi-channel EEG signals that inherited the subject's characteristics. Their capacity to produce superior synthetic data renders them especially pertinent for activities that involve difficulties or significant costs in acquiring authentic data. 1.2.2 Importance of GANs for EEG-Based BCIs Motor imagery (MI) describes the activation of brain regions relevant to motor function when a person visualizes moving a particular body component. One of the main focuses of BCI research is interpreting the MI EEG data. By generating synthetic EEG data, Habashi et al. (2023) have demonstrated how GAI techniques can improve the sparse EEG data utilized for BCI device calibration. BCI devices use multiple EEG patterns in their operation. Eldawlatly (2024) first described the P300 pattern, which has been applied widely in many different contexts. A critical component of event-related potentials (ERP) is P300 evoked potentials. The EEG's P300 signal is predominantly seen as a positive waveform in response to erratic inputs that are somatosensory, visual, or aural. This signal appears 300–400 ms after the person directs their attention from several frequent stimuli to a rare target stimulus. Applying P300 recognition has created remarkable devices and communication aids for those suffering from motor neuron diseases. P300-based BCIs can offer economically viable, portable, and non-intrusive communication solutions to needy individuals, improving their overall quality of life. Epilepsy is a persistent neurological condition characterized by recurrent seizures experienced by individuals. Epileptic seizures manifest in patients through a range of symptoms, including involuntary spasms, convulsions, impaired consciousness, and sensory auras (Feyissa & Bower, 2022). Consequently, epilepsy hinders the overall well-being of patients and raises the death rate among those who experience frequent seizures. Drugs may not always be efficacious for a significant number of individuals with epilepsy (Pascual et al., 2021). Therefore, detecting and monitoring seizures have a significant role in diagnosing patients, enhancing their quality of life, and deepening our understanding of seizures. Intelligent methods for detecting and forecasting seizures have demonstrated potential in machine learning. However, these systems often necessitate a substantial amount of training data. Regrettably, the acquisition of EEG signals during epileptic convulsions might be deemed a costly and time-intensive procedure for both patients and physicians (Sheykhivand et al., 2020). Therefore, generating artificial seizure-like EEG data is proposed as a potential way to train algorithms for detecting and predicting seizures. GAN, or Generative Adversarial Network, was introduced as a superior way among other generative data methods for creating generated data on epileptic seizures. This artificial data is used to train seizure algorithms. Emotion recognition, a crucial component of emotion computing, plays a significant role in uncovering individuals' mental states and comprehending their actions. Artificial emotional intelligence aims to produce devices and systems that accurately identify human emotions (Olider et al., 2024). Compared to facial expression and gesture-based approaches, EEG-based emotion identification is reliable and accurate (Li et al., 2019). Emotion identification methods assess valence, which ranges from negative to positive experiences (Itkes & Kron, 2019), and arousal, which ranges from calm to excited/activated states (Singh & Singh, 2021). 1.3 Purpose of the Review This review explores how GANs can enhance BCI systems based on EEG. This comprehensive review will focus on how GANs affect feature extraction, classification methods, signal enhancement, data augmentation, and employed assessment metrics. Additionally, this comprehensive review aims to provide a succinct summary of EEG-based BCI systems, allowing the reader(s) to choose the most suitable approach for a particular BCI system under several BCI applications, whether medical or non-medical. Upon reviewing recent studies, the research will highlight the linked research gaps that require additional investigation. Significant issues related to EEG-based BCI systems are categorized based on their uses. The outline of this paper is: Chapter 1, Introduction, offers a comprehensive explanation of GANs and their significance in EEG-based BCI technology. Chapter 2 will contain the methodology of the review; this will be discussed in detail regarding how the articles were selected and how they will be analyzed. The third chapter will dwell on research results, provide key studies selected, and discuss the main findings from the selected studies. The fourth, fifth, sixth, and seventh chapters will contain the performance metrics and evaluation strategies, a discussion of the essential findings and trends, future directions, and the conclusion, respectively. Each chapter will be discussed in depth, highlighting the key points. The aim is to present a well-structured and detailed review that analyzes how GANs enhance the performance of EEG-based BCIs. 2. METHODOLOGY 2.1 Literature Search Strategy To acquire the required data, the researcher will concentrate on different categories of EEG signals recognized for their significance in BCIs. This encompasses signals that have both beneficial and detrimental effects on BCI performance, particularly in improving the quality of signals and the accuracy of categorization through GANs. Applying the appropriate formal technique is essential for gathering pertinent research. The researcher will utilize several databases such as SCOPUS, ScienceDirect, and Medline to record all the articles about this study. Moreover, other recognizable databases, such as PubMed, Google Scholar, and IEEE Xplore, will be used to collect pertinent literature about GANs and EEG-based BCIs. The search approach will employ relevant keywords listed in Table 1 below. Data acquisition will not be constrained to a particular location. The primary screening phase will focus on the introduction and summary of the study. The next phase will thoroughly evaluate each chosen study to gather the essential and credible details. This study will exclusively incorporate scholarly articles that are complete in content and freely accessible. Table 1 Keywords: Articles Selection S. No Keywords 1 "Generative Adversarial Networks" 2 "EEG-based Brain-Computer Interface" 3 "GANs in EEG Signal Processing" 4 "Data Augmentation with GANs" 5 "Signal Enhancement using GANs" 6 "Feature Extraction in BCIs" 7 "Classification Improvement in BCIs" 8 "EEG Data Synthesis" 9 "GANs for BCI Performance" 10 "EEG Artifact Removal with GANs" 11 "GANs in Neurotechnology" 12 "Machine Learning in BCIs" 13 "Deep Learning in EEG Analysis" 14 "Synthetic EEG Data for BCIs" 15 "Advanced GAN Architectures in BCIs" The role of GANs in increasing EEG-based BCI performance will be assessed in both retrospective and prospective research, in conjunction with other strategies employed to enhance EEG data processing and improve classification accuracy. Research involving study-specific data will be thoroughly analyzed, including accuracy rates, signal-to-noise ratios, and computational efficiency measures with enough data to provide equivalent results. The study setting will not pose any limitations. Once the appropriate articles have been selected, two contributors will independently thoroughly examine the ‘full-text articles’ according on the predetermined selection criteria. A flow diagram will be created to enhance transparency and clearly explain this process. After choosing appropriate articles, they will undergo additional assessment based on a particular set of criteria. Only studies that meet the inclusion criteria will be considered for the comprehensive review. 2.2 Guidelines for Selection Criteria This comprehensive review will examine every paper that explores the utilization of GANs to improve EEG-based BCIs. The analysis will focus on many aspects, such as augmenting EEG data, enhancing signals, extracting features, and classifying data. It is necessary to establish a comprehensive criterion to address the challenge posed by the scarcity of published publications on the topic. 2.2.1. Inclusion Criteria * Collecting articles about applying GANs in enhancing EEG-based BCI performance. * Studies relating to the role of GANs in augmenting and improving EEG signal quality for BCI systems. * Different GAN architectures and their mechanisms of action on EEG data processing and classification. * Studies about traditional and GAN-enhanced methods for improving accuracy, reliability, and adaptability in EEG-based BCIs. 2.2.2. Exclusion Criteria * Articles published before 2019 * Articles written in languages other than English. * Submission of manuscripts to editors for the purpose of review, including case studies and investigations that involve subjects that are not human. * Replicated studies. * Acceptable sources include credible and reliable sources, with the exception of opinion articles, editorials, and non-peer-reviewed sources. * Studies that did not have full accessibility. Figure 2 Flow Diagram for Literature Selection on GAN-Enhanced EEG-Based BCIs Template Adopted from Page et al. (2021) 2.2.3. Data Collection and Management The authors' data will primarily be stored in individual detailed data extraction sheets. EndNote will input all identified citations once the data has been collected, and duplicates will be removed. The researcher will examine the selected studies’ author details, publication year, research scope, specifics of the EEG data (such as sample size and participant demographics like age, gender, and ethnicity), the application or context of the BCI (such as assistive technologies or healthcare applications), the type of GAN model used, and the critical performance metrics assessed (such as accuracy, signal-to-noise ratio, and Inception Score). If the data cannot be accessed, the study will be excluded. 2.3 Review Process The review process employed a deductive approach known as framework synthesis (Toglia & Goverover, 2022). This method determined a preliminary structure of topics by applying GANs to improve EEG-based BCI systems. The scoping exercise focused on four key issues, one being the applications of GANs in EEG-Based BCIs. (1.1) Data augmentation involves increasing the magnitude of data available for analysis by leveraging diverse approaches. (1.2) Signal enhancement refers to improving the quality and clarity of signals. (2) Feature extraction and transformation involves extracting relevant information from the data and transforming it into a more suitable format for analysis. (2.1) The impact of GANs on extracting and transforming features; and (3) classifying and predicting. The previous topic examined was Performance Metrics and Evaluation. Using the previously developed framework, the reviewers consistently sought information to support the identified themes. New themes were added, and some old themes were combined as coding progressed. Research that has previously been analyzed was reassessed utilizing the updated approach. The researcher then assigned key themes to each study and checked their accuracy. Inconsistencies were discussed and fixed. Studies were classified as primary or secondary themes when appropriate. Still, research was categorized according to its central theme. 2.4 Ethical Considerations The ethical components involve carefully protecting participants and adhering to ethical standards during the research process. This study utilized secondary data studies and did not directly include human participants in the data collection process. Regardless, it is essential to recognize ethical considerations. Ruggiano and Perry (2019) assert that researchers must consistently adhere to ethical research standards, even when utilizing secondary data. The procedure commences with the initial formulation of the study, which should give utmost importance to the general population's welfare and, at the minimum, refrain from generating any negative consequences. It continues until the distribution of the results, which should ensure transparency, accessibility, and the reproducibility of the findings. Researchers must verify with the data sources that their data use matches participant consent. Researchers can fill out the Data Management section of the ethical review application form or design a data management strategy if the data is susceptible or the source requires it. When seeking ethical permission for research that involves secondary data, researchers should take into account the following factors: corresponds with the participants' initial consent, The data should ensure the genuineness of participants' initial consent, strategies for managing data, the measures that will be taken to guarantee confidentiality and data security in the presence of identifiable, personal, or sensitive data, research and the methods of data utilization, management, and storage adhere to the requirements specified by the data sources and the essential documentation been completed and the required authorizations obtained. Moreover the the researchers should ensure the data source be appropriately acknowledged and referenced. Any copyright issues associated with the data and participants de-anonymization by aggregating multiple data sources. Before conducting the research they researchers should check whether there a risk of bias or the development of profiles for a certain group when this data is combined or incorporated with other data. Moreover they should consider key questions such as Which methodology will they employ to communicate the data or analysis? Does this ensure the upholding of participants' privacy and confidentiality? This study was based on the principles of academic integrity, which emphasized the importance of accuracy and trustworthiness in synthesizing knowledge. The researchers conducted a thorough assessment of the quality and validity of the sources, taking into account the methodological rigor and relevance of the papers included in the review. 3. RESEARCH RESULTS Out of the 3432 papers found, only 33 were selected for this review. (Figure 2 describes the inclusion and exclusion criteria.) The items were grouped into three main categories based on common themes. (1) Applications of GANs in EEG-Based BCIs(n=13). (2) Feature extraction and transformation(n=10). (3) Twelve studies have focused on classification and prediction. It should be noted that various research studies are classified into several themes. See Table 2 below. Table 2 Summary Description of the 33 Included Studies Study Purpose GAN Type Method of Research Findings Theme George et al. (2022) Enhance motor imagery decoding performance through data augmentation. RGANs Two public datasets, six augmentation techniques Synthesized data exhibited similar characteristics to actual data, gaining up to 3% and 12% increases in mean accuracies. Data Augmentation Motamed et al. (2021) Improve classification accuracy for pneumonia and COVID-19 using chest X-ray images. Not Specified Semi-supervised detection, comparison with traditional methods GAN-based augmentation surpasses traditional methods for detecting anomalies in X-ray images. Data Augmentation Wickramaratne and Mahmud (2021) Improve task classification accuracy using fNIRS data through data augmentation. CGANs fNIRS data with CGAN and CNN Achieved a task classification accuracy of over 95% with CGAN-CNN combination. Data Augmentation Fahimi et al. (2020) Improve BCI classifier performance through artificial EEG generation. DCGANs EEG data with deep convolutional GANs Significant accuracy improvements for diverted (7.32%) and focused attention (5.45%) tasks. Data Augmentation Vahid et al. (2022) Inform about the inter-relation of antagonistic behaviors on a neural level using EEG data. cGANs EEG recordings with cGAN cGANs can infer amplitude and timing of scalp potentials during response inhibition. Signal Enhancement An et al. (2022) Automatically denoise multichannel EEG signals using GANs. Not Specified EEG signal denoising with new normalization method Achieves comparable performance to manual denoising methods with reduced processing time. Signal Enhancement Brophy et al. (2022) Denoise EEG signals for real-world BCI applications using GANs. Not Specified EEG signal denoising with supervised learning Demonstrates competitive performance in denoising EEG with various artefacts. Signal Enhancement Sumiya et al. (2019) Reduce noise in mice EEG signals using GANs. NR-GANs EEG noise reduction NR-GAN successfully reduces noise in biological EEG signals without requiring separate noise records. Signal Enhancement Sekhar et al. (2023) Classify Alzheimer's disease using EEG with GANs and Marine Predators Algorithm. Not Specified EEG data with GAN and MPA GANs with MPA enhance classification precision, sensitivity, and specificity for AD. Signal Enhancement Chaddad et al. (2023) Review and analyze EEG signal processing methods and techniques. Not Specified Comprehensive literature review Summarizes preprocessing, feature extraction, and classification techniques, and identifies future trends. Signal Enhancement Dong et al. (2023) To present a DL framework that utilizes GANs for the purpose of denoising EEG signals. AR-WGANs Comparison of AR-WGAN with Denoised AutoEncoder, Wiener Filter, and Empirical Mode Decomposition using public and self-collected datasets. AR-WGAN effectively removes artifacts. Achieves high correlation and low error rates. Promises real-time applications in clinical settings. Signal Enhancement Sawangjai et al. (2022) To develop a framework for removing ocular artifacts from EEG signals using GANs. Not Specified Tested on EEG eye artifact dataset and BCI applications without relying on EOG channels. Comparable to state-of-the-art methods. Effective in BCI applications without EOG signals. Shows potential for multivariate bio-signal processing. Signal Enhancement Song et al. (2024) To enhance EEG signal classification using a GAN-based data augmentation model. EEGAN CGAN EEGGAN-Net model with data augmentation, cropped training strategy, and SE attention mechanism. Achieves high classification accuracy on BCI Competition datasets. Surpasses CNN-based models. Effective feature extraction and classification. Signal Enhancement Värbu et al. (2022) To review the evolution and current state of EEG-based BCI applications. Not Specified Systematic review Comprehensive overview of EEG-based BCI applications. Analysis of medical and non-medical domains. Identification of current challenges and future directions. Feature Extraction and Transformation Ma et al. (2019) To review the application of deep learning in remote sensing, including GANs. Not Specified Meta-analysis of over 200 publications on DL in remote sensing. Overview of GANs in remote sensing image analysis. Discussion of challenges and future research directions. Insight into various DL applications like image fusion and classification. Feature Extraction and Transformation Pan, B., & Zheng, W. (2021). To propose a method for emotion recognition using GANs and CNNs. Not Specified Sample generation method using GANs to address data imbalance and improve emotion recognition. Effective in improving emotion recognition performance. Frequency band correlation model is more conducive to recognition. Enhances deep learning model performance with data augmentation. Feature Extraction and Transformation Goodfellow et al. (2020) To discuss the principles and applications of Generative Adversarial Networks. Not Specified Overview of GANs and their applications in generative modeling. GANs are effective in generating high-resolution images. Feature Extraction and Transformation Habashi at al. (2023) To provide an overview of GAN techniques in EEG analysis. Not Specified Review of various GAN-based EEG augmentation techniques. GANs improve the quality of EEG features and classification performance. Focus on BCI, emotion recognition, and seizure prediction. - Highlights limitations and future research directions. Feature Extraction and Transformation Classification and Prediction Panachakel and Ramakrishnan (2021) To review systems for decoding covert speech from EEG. Not specified Comprehensive review of EEG-based covert speech decoding methods. Detailed comparison of different speech decoding systems. Review of data acquisition, feature extraction, and classifiers. Discussion of future directions for real-time BCI systems. Feature Extraction and Transformation Luo et al. (2020) To propose a GAN-based method for EEG signal reconstruction. WGAN Reconstructi...