The human brain generates electrical activities measurable through electroencephalography (EEG). These signals are often contaminated by noise, and so its not easy to do an accurate analysis and interpretation which is essential for clinical applications such as epilepsy diagnosis, cognitive neuroscience, and brain-computer interfaces. Traditional denoising techniques frequently fall short in efectively distinguishing between signal and noise, especially when the noise sources exhibit complex and nonlinear characteristics. This paper explores the application of Generative Adversarial Networks (GANs) in denoising EEG signals, ofering a data-driven approach to learn the complex structures of both clean and noisy EEG data. We detail the training of a classifier to distinguish between normal and abnormal EEG signals, the development of an AutoEncoder to compress and reconstruct signals, and the use of a Wasserstein GAN (WGAN) to manipulate abnormal signals towards normality in the latent space. Our results demonstrate the potential of GAN-based methods in enhancing EEG signal quality, paving the way for more accurate clinical analyses.
resolution approach but can be computation-intensive, and also this technique cannot deal with noise that is The human brain is a complex network of connected highly non-stationary in nature. These techniques often neurons that produces electrical activity; this can be mea- tend to delete the main important features of the origisured through EEG. Although EEG signals are quite use- nal EEG signal, especially in cases when noise sources ful to understand brain functioning, they are heavily con- become complex and nonlinear in nature[9, 10]. taminated by numerous sources of noise of very diferent In recent years, machine learning techniques like Genkinds: environmental interference, muscle activity, and erative Adversarial Networks have shown great potential electrical artifacts[1]. For instance, myogenic artifacts in denoising EEG signals. GANs were first introduced by are created in muscle movements around the head and Goodfellow et al. in 2014 [11, 12], where we have two face, while ocular artifacts are generated by eye blinks neural networks, named the generator and the discrimiand movements. Moreover, brain signals are complex nator. The generator tries to generate data samples that and non-stationary; this makes the separation of neural can come from the real data distribution, and the discrimiactivity from noise a challenging task. Denoising EEG nator tries to diferentiate between the real data examples signals is important because it allows us to better ana- and those created by the generator. In the end, the GAN lyze and interpret signals. This is important in various updates its weights so that the generator can fool the settings, such as clinical applications for the diagnosis discriminator. This adversarial training process makes of epilepsy, cognitive neuroscience, and brain-computer GANs capable of learning complex structures of the data, interfaces[2, 3]. which subsequently becomes useful for problems like
To meet these challenges, traditional denoising tech- denoising where typical methods fall short.
niques, such as band-pass filtering, independent compo- Traditional GANs have several training problems, such
nent analysis (ICA) [4, 5, 6], and wavelet transforms and as mode collapse and instability during training. To
hanother domain transforms [
oped to accurately discern between normal and abnormal EEG signals using the provided dataset.
This classifier forms the basis of our subsequent denoising process by ensuring the precise identiifcation of signal types. 2. AutoEncoder Development: Subsequently, an ity to handle multiple artifcat types, having the potential AutoEncoder is constructed and trained exclu- of being used in applications in portable EEG devices. sively on normal EEG data. This network is Similarly we have (Yang An et al. 2022 [19]) approach tasked with compressing and reconstructing the that proposes a new loss function to retain original inforEEG signals, thereby establishing a representa- mation and energy in the filtered signals, demonstrating tive latent space that captures the salient features that the performances are comparable to manual denoisof normal brain activity. This latent space serves ing methods while at the same time, we have a significant as a critical reference for subsequent processing. reduction of the processing time.In this method, we have 3. WGAN Training: In the final stage, a Wasser- also an incorporation of a new normalization method stein Generative Adversarial Network (WGAN) ensures stable generation of EEG signals by the GAN is employed to refine the latent representations model, allowing for automatic denoising across diferent of abnormal signals. The generator within the subjects’ data[20].
WGAN is designed to transform these abnormal Another notable contribution in this domain is prelatent representations so that they more closely sented by (Wang et al. 2022 [21]) in their work titled "An resemble those of normal signals, efectively de- improved Generative Adversarial Network for Denoising ceiving the discriminator. This transformation EEG signals of brain-computer interface systems[22, 23]. facilitates an improved reconstruction of the orig- The authors propose a novel GAN-based framework inal signal, ultimately converting unhealthy sig- that includes a generator with BiLSTM and LSTM layers nals into representations that are consistent with and a discriminator composed of multiple CNN layers. healthy EEG characteristics—a process that holds This architecture aims to address the limitations of presignificant potential for clinical applications. vious models by reducing mode collapse and improving the convergence stability during training. The study
By leveraging the advanced capabilities of GAN-based demonstrates that their improved GAN significantly outarchitectures, the proposed approach addresses several performs traditional methods and other deep learning limitations associated with traditional denoising tech- approaches, particularly in scenarios with high noise levniques. The integration of classifier training, AutoEn- els. Their results show enhanced performance in terms of coder development, and WGAN-based latent space ma- root mean square error (RMSE) and Pearson correlation nipulation provides a comprehensive and robust solution coeficient (CC), making it a robust solution for real-time for EEG signal denoising. This integrated methodology EEG denoising in brain-computer interface (BCI) sysnot only enhances the quality of EEG signals but also tems. increases their utility in clinical diagnostics and cogni- The improved GAN framework not only excels in detive neuroscience research, paving the way for the de- noising accuracy but also enhances the robustness of EEG velopment of more efective brain-computer interface signal processing by efectively handling artifacts such as technologies. eye blinks and muscle movements. The authors utilized the EEGdenoiseNet (Zhang et al. 2022 [24]) dataset to 2. Related Works benchmark their model, which includes a diverse range of artifact-contaminated EEG signals. The proposed There are several recent studies that deals with denois- model’s ability to maintain high performance across varying EEG signals using various methodologies. Among ing signal-to-noise ratios (SNR) underscores its potential these, (Peng Yi et al. 2021 [15]) using transformer, pro- for practical applications in BCI systems, where real-time poses a novel approach integrating non-local and local processing and reliability are crucial [25, 26]. self-similarity of EEG signals through a 1-D EEG signal Moreover, this study show the importance of incorpodenoising network with a 2-D transformer.Using self- rating BiLSTM and LSTM layers in the generator network similarity characteristics, EEGDnet has a improvements to capture temporal dependencies in EEG signals. The in removing ocular and muscle artifacts compared to discriminator network, consisting of multiple CNN layother state-of-the-art models. ers, ensures that the generated clean EEG signals closely
In (Eoin Brophy et al. 2022 [16]), the focus is on utiliz- resemble the true signals, thus enhancing the overall ing Generative Adversarial Networks (GANs) to denoise quality of the denoised output.
EEG time series data. We have to deal with artifacts In addition to the structural improvements, the traininduced in real-world Brain-Computer Interface (BCI) ing strategy employed by Wang et al. achieves superior applications[17, 18], which degrade performance.Using performance. By carefully balancing the learning rates GANs the model is able to, given noisy EEG signals trans- and optimizing the loss functions for both the generator forming it into clean ones,demonstrating promising re- and discriminator, the model is able to avoid the common sults in quantitative metrics such as power spectral den- problems of the GAN, overfitting and non-convergence. sity and signal-to-noise ratio. This study has the capabil- Other interesting works that apply GAN-based models for anomaly detection (Zenati et al. 2018 [27]). In signal from this transformed latent representation will this works, state-of-the-art performances are shown on be an enhanced version of the original abnormal signal. image and network intrusion datasets. The approach con- This enhancement process exploit the combination of sists of using a GAN that learns a latent representation classification, compression, and adversarial training to of normal samples in such a way that the GAN is able improve the quality and normalcy of abnormal data. to detect anomalies by measuring reconstruction and discriminator-based loss. High performance is shown on 3.1. Phase 1: Classifier the MNIST and KDD99 datasets in the method proposed.
The proposed method by (Niu et al. 2020 [28]), is called In the first step, we train a very simple binary classifier. LSTM-based VAE-GAN. This solves the ineficiency of The classifier learns the diference between normal and real-time space-to-latent space mapping when doing abnormal data, in fact it’s role is to accurately identify abanomaly detection. Here we exploit the temporal de- normal signals, which will later be processed to enhance pendencies that an LSTM Network can capture and the their quality. reconstruction and discrimination abilities of VAE and GAN respectively. 3.2. Phase 2: AutoEncoder (Zhang et al. 2023 [29]) is a novel GAN-based model for unsupervised anomaly detection in multivariate time We now train an AutoEncoder (as shown in Figure 2) series (MTS). In this work, the authors use a self-training using only the normal data. The job of this AutoEncoder framework wherein they have a teacher model gen- is to compress and reconstruct each given signal. Once it erating high-quality pseudo-labels iteratively training learns what the typical data distribution is, the AutoEna student model. STAD-GAN involves a generator- coder can then be able to reconstruct signals that reflect discriminator structure with a neural network classifier. normal behavior.
The generator maps the normal data distribution. The discriminator amplifies the reconstruction error of abnor- 3.3. Phase 3: WGAN mal data to enhance recognition performance. It means that the performance of the anomaly classifier will be For the final phase, we train a Wasserstein Generative Adimproved through self-training by iteratively refining the versarial Network (WGAN). The Generator in the WGAN dataset[30]. works on the encoded anomalous latent-space represen
Our approach ofers a diferent contribution. Unlike tation in such a way that the latter becomes more like the existing methodologies that often rely on direct corre- latent representations of normal signals, those obtained spondences between healthy and unhealthy signals for by pre-trained AutoEncoders. The Discriminator’s role denoising, our study uses a comprehensive data-driven is to distinguish between the true normal latent repreapproach without such direct correspondences, by trans- sentations and the manipulated ones produced by the forming various types of unhealthy signals into healthy Generator. The objective is to train the Generator to proones within the same architectural framework. duce enhanced versions of the abnormal signals that fool
Many methods proposed in the literature typically ad- the Discriminator into classifying them as normal. dress only one type of artifact or abnormality, these meth- The objective function for Wasserstein Generative Adods limits versatility and applicability across diferent versarial Networks (WGAN) is given by: scenarios and they often require a specialized design or customization for each specific type of artifact or abnor- min max Ex∼ P [(x)] − Ez∼ P [((z))] (1) mality, which can be both time-consuming and resource- ∈ intensive. Here, P denotes the real data distribution, P denotes
Specifically, our model is designed to handle poten- the prior distribution of the input noise z to the Generator tially a diverse range of signal abnormalities and artifacts, , and is the set of 1-Lipschitz functions which the ensuring that it can be retrained and adapted to improve Discriminator is a part of. any given signal’s quality without the need for significant In particular, we will enforce the 1-Lipschitz constraint, modifications. a gradient penalty term is added to the loss function. The modified objective with gradient penalty (WGAN-GP) is:
Our approach to enhance abnormal data consists of three main phases. The goal of our approach is to transform the latent representation of an abnormal signal to resemble those of normal signals. By doing so, the reconstructed ℒ = Ex∼ P [(x)] − Ez∼ P [((z))] + Ex^∼ Px^ ︀[ (‖∇x^ (xˆ)‖2 − 1)2]︀ (2) where Px^ is the distribution of the interpolated samples between the real and generated data, and is the penalty coeficient.
ular tool used for tracking machine learning experiments.
It’s a strong platform for logging metrics, visualizing reMore specifically, our method is developed in three prin- sults, and collaborating on project implementations. In cipal phases: training of the classifier to diferentiate our implementation, we have integrated wandb to monibetween normal and abnormal signals, training of the tor training and the evaluation processes of our models. AutoEncoder signal compression and its reconstruction, In Table 1, we depict the various hyperparameters of the training of Wasserstein Generative Adversarial Network models. (WGAN) that transform unhealthy into healthy signals.
Then each phase is described in detail with extensive descriptions for data preparation, architectural choices, training processes, and integration steps. Note that in our experiments, we used Weights & Biases (wandb)—a pop
4.1.1. Data Preparation The main dataset (shown in Figure 1) used in our experiments consists of EEG signals recorded under various conditions. The Healthy dataset consists of over 1,500 one- and two-minute EEG recordings obtained from 109 volunteers. The EEG recordings were collected using a 64-channel setup with the BCI2000 system. Each subject performed 14 experimental runs, which included: 1. Baseline, eyes open 2. Baseline, eyes closed 3. Task 1: Open and close left or right fist 4. Task 2: Imagine opening and closing left or right ifst 5. Task 3: Open and close both fists or both feet 6. Task 4: Imagine opening and closing both fists or both feet 7. Task 1 (repeat) 8. Task 2 (repeat) 9. Task 3 (repeat) 10. Task 4 (repeat) 11. Task 1 (repeat) 12. Task 2 (repeat) 13. Task 3 (repeat) 14. Task 4 (repeat)
During these tasks, subjects performed or imagined performing the movement corresponding to the stimuli presented on the screen. All these tasks are designed to evoke diferent types of motor and imagery behaviors and have the event type tag-based indicators: • T0: Rest • T1: Onset of motion (real or imagined) of the left ifst (Tasks 1 and 2) or both fists (Tasks 3 and 4) • T2: Onset of motion (real or imagined) of the right fist (Tasks 1 and 2) or both feet (Tasks 3 and 4) 4.1.2. Unhealthy Dataset
recordings from 7 human participants with orthopedic impairment during motor imagery (MI) tasks. Due to component removal, some subjects have slightly fewer channels (14 or 15) to eliminate noisy channels and improve data quality.
Participants performed a series of MI-related trials across three sessions, each consisting of 40 trials with four diferent MI tasks presented in random order. Each trial included: • 3 seconds of fixation cross • 4 seconds of visual cue • 3 seconds of letters indicating the ready state • 5 seconds of imaginary movement
notch filter) and labels for 10 movement types and a rest state. The corresponding electrode names are provided in TSV files, ensuring a 1:1 mapping with the CSV data.
Note that we have healthy and unhealthy signals, and no correspondences between them ( so given an unhealthy signal, we don’t have the corresponding healthy one) . 4.1.3. Supplementary Dataset
the Epilepsy2 dataset, which contains single-channel EEG measurements from 500 subjects. [31]
Each subject’s brain activity was recorded for 23.6 seconds, resulting in a comprehensive collection of EEG data. To facilitate detailed analysis and model training, the dataset was divided and shufled into 11,500 samples, each representing a 1-second segment of EEG data sampled at 178 Hz. This shufling mitigates sample-subject association and ensures a robust evaluation environment.
The dataset is divided into three groups: 60 samples for training, 20 samples for validation, and 11,420 samples for the test. The validation set was appended directly to the end of the training file to allow easier reproducibility in case a validation set is required. The small size of the training set will actually assist in testing the transfer learning capabilities, and the test set retains the original distribution for a fair evaluation.
The dataset contains five unique classification labels related to diferent conditions or measurement locations:
2. Eyes closed 3. EEG measured in a healthy brain region 4. EEG measured in the region of a tumor 5. Subject experiencing a seizure episode
on this dataset, demonstrating its efectiveness in distinguishing between seizure and non-seizure states. For reference, the Mini Rocket classifier, a well-known model in the field, achieves a test accuracy of 96.25%, highlighting the competitive performance of our approach.
The actual classifier, autoencoder and wgan they will be trained using the main dataset. 4.1.4. Classifier Architecture
work architecture. The classifier consists of multiple fully connected (dense) layers, each followed by a Rectified Linear Unit (ReLU) activation function, and a final Sigmoid activation function to output the probability of the signal being abnormal. The detailed architecture is as follows: • Input Layer: Accepts the preprocessed EEG signal. • Hidden Layers: – First layer: 512 neurons, ReLU activation. – Second layer: 256 neurons, ReLU activation.
– Third layer: 64 neurons, ReLU activation.
• Output Layer: 1 neuron, Sigmoid activation.
4.1.5. Training the Classifier
a binary cross-entropy loss function. The training process involved multiple epochs, where the model parameters were updated iteratively to minimize the loss function. During training, the following steps were performed: • Forward Pass: The input data was passed through the network to obtain the output probabilities. • Loss Calculation: The binary cross-entropy loss between the predicted probabilities and the true labels was calculated. • Backward Pass: Gradients were computed using backpropagation, and the network weights were updated using the Adam optimization algorithm.
4.2.1. AutoEncoder Architecture The AutoEncoder was designed to compress and reconstruct the EEG signals. The AutoEncoder consists of two main parts: the encoder and the decoder. The encoder reduces the dimensionality of the input data to a latent space, capturing the essential features, while the decoder reconstructs the data back to its original form from this compressed representation.
• Encoder: – First layer: 128 neurons, ReLU activation. – Second layer: 64 neurons, ReLU activation. – Third layer: 32 neurons (latent space),
ReLU activation. terms of the reconstruction error. The objective was to minimize the reconstruction error, which should allow the AutoEncoder to reconstruct normal EEG signals. The training loss is described in Figure 6 4.3. Phase 3: WGAN 4.3.1. WGAN Architecture
This architecture allows the AutoEncoder to learn The generator aims to transform the encoded latent space a compressed representation of normal EEG signals, of abnormal signals to resemble the latent space of normal needed for the training of WGAN. signals, while the discriminator evaluates the authenticity of the generated signals. The detailed architectures are as follows: 4.2.2. Training the AutoEncoder The AutoEncoder was trained using the mean squared error (MSE) loss function, which measures the diference between the input and the reconstructed output. The training involved the following steps: • Forward Pass: The normal EEG signals were passed through the encoder to obtain the latent representation, and then through the decoder to reconstruct the signals. • Loss Calculation: The MSE loss was computed
between the original and reconstructed signals. • Backward Pass: Gradients were computed, and the network weights were updated using the
Adam optimizer.
The training was done over several epochs while checking the performance of the model on the validation set in • Discriminator: – First layer: 64 neurons, ReLU activation. – Second layer: 128 neurons, ReLU activa
tion. – Output layer: Same size as latent space,
linear activation. – First layer: 128 neurons, ReLU activation. – Second layer: 64 neurons, ReLU activation. – Output layer: 1 neuron, Sigmoid activa
tion. 4.3.2. Training the WGAN
problems like mode collapse, training of the WGAN was performed with Wasserstein loss and gradient penalty (WGAN-GP). The training process involved alternating between optimizing the discriminator and the generator.
The following steps were performed during training: • Discriminator Training: – Real latent representations (from the encoder) and fake latent representations (from the generator) were fed into the discriminator. – The discriminator’s loss was calculated based on its ability to discern real from fake representations. • Generator Training: – Gradients were computed, and the discrim- signals were performed to make sure that we remove abinator’s weights were updated. normalities but also preserve essential features of the EEG signals.
With this method, we will efectively enhance the qual– The generator produced fake latent repre- ity of EEG signals and transform unhealthy signals to sentations from abnormal signals. healthy ones. The deployment of our model provides a – These representations were fed into the scalable solution for EEG signal denoising applicable in discriminator. many clinical and research applications. – The generator’s loss was calculated based
on its ability to fool the discriminator. – Gradients were computed, and the genera- 5. Results
tor’s weights were updated.
the generator to maintain a balance between the two networks. Thanks to this adversarial training process, the generator learned how to convert signals from unhealthy to healthy. The training loss is described in Figure 7
After training each component, we integrated the models (shown in Figure 8) to form a cohesive system. So, each sub-model work will be: • Classification : The classifier identified abnormal signals from the EEG data. • Encoding: The identified abnormal signals were encoded into latent representations using the autoencoder’s encoder. • Transformation: The WGAN generator transformed these latent representations to match the distribution of normal latent representations. • Reconstruction: The autoencoder’s decoder reconstructed the enhanced signals from these transformed latent representations.
the generator will be the input of the decoder to have the corresponding Healthy signal. The combined system was additionally analyzed with other performance measurements such as: accuracy, RMSE, MPC. mean entropy. Additionally, qualitative assessments of the reconstructed
eral key metrics: Root Mean Square Error (RMSE) calculated in the latent space, Pearson correlation, classifier accuracy, and entropy changes of the samples after manipulation. The results are summarized in Table 2.
The RMSE in the latent space for GAN architecture is very low (0.004166), meaning that the signals reconstructed with the use of GAN are close to the expected latent representations. This further means that the RMSE value is quite low and close to zero, thus suggesting the fact that GAN can successfully capture the structure of signals.
The mean Pearson correlation for the GAN is -0.116, which indicates a slight negative correlation between the generated signals and the true signals. While it is not a strong negative correlation, it still denotes that there is more room of improvements for the model to generate signals that closely resemble the true signal patterns.
The classifier, which identify the healthy samples for the unhealthy one, achieved an accuracy of 0.9047. This high accuracy shows that the classifier is efective in diferentiating between the EEG recordings of normal and abnormal cases, giving a good basis for further processing by the GAN. In an entropy perspective, after manipulation was performed on the 2334 samples, 585 samples showed increased mean entropy while 1728 samples showed decreased mean entropy. Entropy represents a measure of randomness or disorder in the signals. If the entropy has increased, this can be taken as an indication that the signals grew to be more complex, possibly indicating the introduction of noise or artifacts. On the other hand, a decrease in entropy means a decrease in signal complexity, which may be explained by efective signal denoising and enhancement by the GAN. In Figure 9, is that synergistically integrates a classifier, an AutoEnshown an example of transforming an unhealthy signal coder, and a WGAN. Using the unique capabilities of to healthy one, its clear that probably by adding to the each component, the proposed method efectively distinloss also a constraint to the amplitude, we will obtain a guishes between normal and abnormal signals, learns a better signal. compact latent representation of healthy EEG patterns, and transforms noisy or abnormal latent representations towards normality. The experimental results, characterized by a low RMSE in the latent space and robust classifier performance, underscore the potential of the framework to achieve meaningful denoising while preserving critical signal features.
Our integrated approach addresses several of the limitations inherent in traditional signal processing techniques, ofering a data-driven alternative that accommodates the complex, non-stationary, and nonlinear nature of EEG noise. Despite demonstrating promising results, particularly in terms of classifier accuracy and latent space convergence, the study also reveals areas where further renfiement is warranted, such as improving the Pearson correlation between reconstructed and true signals and optimizing entropy measures in the enhanced Figure 9: Example of denoising an unhealthy signal output.
Future work should explore the scalability of the frame
These results show promise for enhancing abnormal work across broader and more diverse datasets, includEEG signals using a GAN-based approach. Particularly ing multi-channel recordings and clinical datasets enencouraging is the very low RMSE and high classifier compassing a wider range of neurological conditions. accuracy, while the entropy results provide further in- Moreover, the real-time implementation of the denoising sight into the nature of signal transformations. Further framework could significantly enhance its applicability improvements in the model could potentially address the in BCI systems. Integrative research eforts that comnegative correlation and refine the signal manipulation bine EEG with other neuroimaging modalities, such as process to achieve even better outcomes. fMRI or MEG, may further enrich the diagnostic precision and clinical relevance of the proposed methodology.
In general, the promising results of this study pave the 6. Conclusions and Future Works way for advanced EEG signal processing paradigms, offering valuable insights into both cognitive neuroscience This study has presented a novel framework for enhanc- research and practical clinical diagnostics. ing abnormal EEG signals using a GAN-based approach
ChatGPT, Grammarly in order to: Grammar and spelling check, Paraphrase and reword. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.