Synthetic ALS-EEG Data Augmentation for ALS Diagnosis Using Conditional WGAN with Weight Clipping

Abdulvahap Mutlu; \c{S}eng\"ul Do\u{g}an; T\"urker Tuncer

REVIEW 4 major objections 5 minor 51 references

Synthetic ALS-EEG Data Augmentation for ALS Diagnosis Using Conditional WGAN with Weight Clipping

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper reports that a conditional Wasserstein GAN can learn the distribution of ALS EEG signals and generate realistic synthetic segments for data augmentation despite severe class imbalance.

desk verdict A reproducible CWGAN application to ALS EEG with honest limitations, but the evidence does not back the central claim that it learned the distribution. read the letter →

arxiv 2506.16243 v1 pith:ZP4B3M3A submitted 2025-06-19 cs.LG cs.AI

Abdulvahap Mutlu , Şengül Doğan , Türker Tuncer This is my paper

classification cs.LGcs.AI

keywords AmyotrophicLateralSclerosis(ALS)Electroencephalography(EEG)DataAugmentationGenerativeAdversarialNetwork(GAN)ConditionalWassersteinGAN(CWGAN)SyntheticBiomedicalSignalsClassimbalanceWeightclipping

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Amyotrophic lateral sclerosis (ALS) is rare, so high-quality EEG recordings from patients are scarce and ALS-versus-control datasets are heavily imbalanced. The authors try to establish that a Conditional Wasserstein Generative Adversarial Network (CWGAN), a GAN variant trained with the Wasserstein loss and weight clipping, can learn the distribution of real ALS EEG segments from a small dataset and generate new, realistic synthetic ALS waveforms. If that is true, the synthetic segments can be added to the minority class to balance training sets, reduce classifier bias toward healthy controls, and be shared without exposing patient identities. The paper's evidence is qualitative: clinicians found many generated traces indistinguishable from real ALS EEG, and the training losses converged stably without visible mode collapse. The authors explicitly note that objective fidelity metrics and a formal diversity analysis have not yet been performed.

What carries the argument

The central object is the Conditional Wasserstein Generative Adversarial Network (CWGAN) with weight clipping: a GAN in which a generator produces synthetic samples from noise plus a class label and a critic scores realness using the Wasserstein (Earth-Mover) loss, with the 1-Lipschitz constraint enforced by clipping critic weights to $[-0.01, +0.01]$ after every update. The conditioning mechanism, an embedding of the class label multiplied element-wise into both generator and critic inputs, is what lets the model generate ALS-class segments on demand by sampling fresh noise and conditioning on the ALS label. The weight-clipping scheme is what the authors credit for stable training on a small dataset, and the label-conditional generation is what makes the output directly usable as minority-class augmentation.

What would settle it

Compute a standard distribution-distance measure (for example, maximum mean discrepancy) between real held-out ALS segments and generated segments, or compare their power spectra across the 32 EEG channels; if synthetic segments are statistically indistinguishable from control segments in ALS-relevant frequency bands, the generator has not captured the disease-specific distribution. Alternatively, train an ALS-versus-control classifier on synthetic ALS data plus real controls and test on real held-out ALS segments: chance-level sensitivity would show the synthetic data carry no usable ALS signal.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims to demonstrate that a CWGAN with weight clipping can be trained on the EEGET-ALS dataset, which contains recordings from six ALS patients and 170 healthy controls, and that the trained generator produces ALS-class EEG segments that look like real patient data. The generator takes a 100-dimensional Gaussian noise vector and a class label, embeds the label, multiplies it into the input, and passes through three dense layers to output a 32-length waveform, while the critic scores realness under the Wasserstein loss with its weights clipped to $[-0.01, +0.01]$ to enforce the 1-Lipschitz constraint. With five critic updates per generator update, RMSprop at $5 \times 10^{-5}$, and 300 epochs, the loss curves stabilized. The authors' central conclusion is that the CWGAN has learned the ALS EEG distribution and can generate new, varied samples that are not simple repeats of training data, making them suitable as augmented training data for ALS-versus-control classifiers.

Load-bearing premise

The whole conclusion rests on the untested assumption that the generator's stable training curves and clinicians' visual approval prove it learned the full variety of real ALS brain-wave patterns rather than a few memorized or repeated shapes.

Editorial extensions

If this is right

The generated ALS segments can be mixed into the minority class of an ALS-versus-control training set, which should reduce the classifier's bias toward healthy-control recordings.
Because generated traces are not recordings of any real patient, they can be shared across institutions with fewer privacy restrictions, enabling larger collaborative benchmarks.
If the synthetic segments truly capture ALS-specific spectral and temporal patterns, downstream classifiers trained with them should generalize better to real held-out ALS data.
The training recipe (weight clipping, five critic updates per generator update, RMSprop) offers a starting point for generating synthetic data in other rare-disease settings with limited EEG cohorts.
The paper positions synthetic data as a complement to real data, not a replacement: disease features the generator misses would remain underrepresented in the augmented set.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the authors assert that generated traces do not correspond to any real individual, but they do not test membership inference; a synthetic dataset could still memorize and leak patient-identifying patterns, so the privacy claim needs its own evaluation.
Editorial inference: a direct frequency-domain comparison (for example, power spectral density in ALS-relevant bands) between real and synthetic segments would turn the qualitative visual assessment into a quantitative one and could reveal whether the generator reproduces the cortical-hyperexcitability signature that motivates EEG-based ALS diagnosis.
Editorial inference: if synthetic-only training transfers to real ALS EEG, the approach could be used to pre-train diagnostic models before fine-tuning on a small real set, reducing the amount of patient data a hospital must collect.
Editorial inference: because the method is not tied to ALS-specific features, the same conditional-generation setup could be applied to other neurological conditions that face the same scarcity-and-imbalance problem.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this paper does one new thing—applying CWGAN to ALS EEG—and it does that clearly, with enough detail to be reproduced. But the evidence for the central claim is thin. The stress-test note is right: stable loss curves and visual inspection do not establish that the generator covered the ALS distribution, especially with only six patients and 32-point MLP outputs.

What's genuinely useful: the problem is real (ALS EEG data is scarce and imbalanced), and the authors are unusually candid. They explicitly state in §5.1 that they have not quantitatively measured fidelity and in §4.2 that they did no formal diversity analysis. They also describe the architecture and hyperparameters in detail—latent dim 100, n_critic=5, RMSprop, weight clipping—which makes the work reproducible. That honesty earns credit.

The soft spots are load-bearing. The conclusion claims CWGAN 'can learn the distribution of ALS EEG signals,' but the evaluation is only qualitative visual review by an unspecified group of clinicians plus smoothing loss curves. A WGAN with MLP critic on 32-point vectors can converge while memorizing the 2,631 training segments or collapsing to a few modes; visual inspection cannot rule that out. No quantitative distributional metrics, no held-out real data, no classifier augmentation experiment, no comparison against simple baselines like random oversampling or SMOTE. The dataset description contains clear inconsistencies: the abstract says 'private EEG dataset', §3.1 says 'public EEGET-ALS', and the class ratio is given as roughly 3.4% in §3.1 but 20% in §1.1. These are not fatal on their own, but they reduce confidence in the report.

The right reading: this is a proof-of-concept that a CWGAN can be trained on this small dataset without obvious divergence. It is not, as the authors suggest, a validated data augmentation strategy. The gap between the evidence and the claim is the paper's core problem.

Who gets value: researchers working on GAN-based EEG augmentation might read it as a starting point, and the detailed training recipe is useful. But as it stands, I would not cite it as evidence that synthetic ALS EEG works.

Recommendation: this deserves a serious referee, but only with the expectation of major revision. The authors should be pushed to add quantitative fidelity and diversity metrics, run a downstream ALS-vs-control classifier with and without augmentation, and fix the dataset inconsistencies. If they can do that, the paper could become a modest but valid application note. As submitted, it is not a supportable contribution.

Referee Report

4 major / 5 minor

Summary. The paper applies a conditional Wasserstein GAN (CWGAN) with weight clipping to generate synthetic EEG segments for ALS patients, motivated by data scarcity and class imbalance in ALS-vs-control classification. Using the EEGET-ALS dataset (six ALS patients, 170 controls), the authors preprocess 32-channel EEG into one-dimensional segments, train an MLP-based CWGAN conditioned on class label, and evaluate the generated ALS-class samples by visual inspection and training-loss stability. The authors claim that the CWGAN can learn the distribution of ALS EEG signals and generate realistic samples, with potential use for data augmentation, but explicitly note that quantitative fidelity metrics and formal diversity analysis are absent.

Significance. If the central claim were substantiated, the paper would provide a potentially useful recipe for GAN-based augmentation in a rare-disease EEG setting, where data are scarce and class imbalance is severe. The intended downstream benefit for ALS diagnosis is clinically relevant. However, the evidence offered is entirely qualitative and the paper itself acknowledges the lack of objective fidelity metrics and diversity analysis. The stable loss curves and clinician inspection reported do not rule out memorization or mode collapse, so the significance at present is only as a preliminary demonstration, not a validated method. A reproducible experimental setup and clearly stated hyperparameters are strengths, but the absence of code and quantitative evaluation limits verifiability.

major comments (4)

[§5.1, §4.2, Conclusion] The central claim that the CWGAN 'can learn the distribution of ALS EEG signals and generate new, realistic samples' rests on qualitative visual inspection and stable loss curves, as the authors explicitly concede in §5.1 ('we have not yet quantified their fidelity using objective metrics') and §4.2 ('we have not performed a formal diversity analysis'). Stable critic loss is not sufficient to guarantee that the generator covers the data distribution; with 2,631 segments of length 32 and weight clipping, the generator could memorize training samples or collapse to a few modes without affecting the loss trajectory. Please add quantitative distributional comparisons (e.g., power spectral density similarity, maximum mean discrepancy, or a classifier-based fidelity test) and a formal diversity analysis (e.g., nearest-neighbor duplicate rates or coverage metrics) to support the central claim.
[Abstract vs. §3.1 and §1.1] There is a direct contradiction between the Abstract and Section 3.1: the Abstract says the CWGAN is trained on 'a private EEG dataset', while §3.1 identifies the public EEGET-ALS dataset. Moreover, §1.1 reports 2,631 ALS segments versus 10,248 control segments (roughly 20% ALS), but §3.1 states 'The resulting ≈3.4% representation of ALS samples' — a large numerical inconsistency. Please clarify which dataset was actually used and correct the imbalance figures, as reproducibility and the motivation for augmentation depend on these facts.
[§1.4, §5.2, §5.3] The paper is framed around data augmentation for ALS diagnosis, but it never demonstrates the downstream benefit. No classifier is trained on augmented data, and any improvement in detection accuracy is deferred to future work (§5.3). The contribution 'provides a strategy to balance class distributions' is therefore not tested. Even a small proof-of-concept experiment, such as training a simple classifier on real-only vs. real-plus-synthetic segments and reporting accuracy/sensitivity/specificity, would be needed to substantiate the augmentation claim.
[§3.4–§3.7] The generator is an MLP producing 32-point vectors, and the training routine (weight clipping, RMSprop, no normalization layers) provides no explicit mechanism against memorization. The claim in §4.1 that synthetic signals 'do not appear as simple repeats of training data' is based on visual inspection, which cannot reliably detect near-duplicates in a 32-dimensional space. Please report quantitative similarity to nearest real training segments (e.g., minimal Euclidean distance or a memorization ratio) to rule out training-set replication, or justify the architecture choice with such an analysis.

minor comments (5)

[§3.2] The preprocessing formula is written in an awkward inline format with spaces and periods ('(𝑥 − 𝑥. 𝑚𝑖𝑛()) / (𝑥. 𝑚𝑎𝑥() − 𝑥. 𝑚𝑖𝑛()) ∗ 2 – 1'); please typeset it properly and define x.min() and x.max() clearly.
[§3.2] The use of 'fake' to denote the non-ALS set is misleading, since 'fake' conventionally denotes generated samples in GAN literature; consider using 'control' or 'alternate class' consistently.
[§4.2 and Figure 3] Figure 3 is referenced as showing training loss curves, but the text does not specify which loss (critic, generator, or both) is shown, nor whether smoothing was applied. Please add a clear caption and axis labels and describe the curves explicitly.
[§1.3] The phrase 'state-of-the-art GAN technique' is vague; please specify which published method or baseline the CWGAN is compared against, or remove the term.
[References] Reference 6 ('Fabbri, C. Conditional Wasserstein Generative Adversarial Networks') lacks venue and year; please provide a complete citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No construct circularity: the paper is an empirical GAN-fitting study whose central claim rests on qualitative inspection and loss convergence; the missing quantitative validation is a correctness gap, not a circular reduction.

full rationale

The paper contains no derivation chain in which an output is defined in terms of its own input. It trains a Conditional Wasserstein GAN on ALS-labeled EEG segments from the EEGET-ALS dataset, generates 32-point synthetic segments, and evaluates them by visual waveform inspection and by the stability of the WGAN loss curves. There is no fitted parameter renamed as a prediction, no equation in which the claimed result is identical to an input by construction, and no load-bearing self-citation: the cited WGAN, CWGAN, and EEG-augmentation methods are external prior work, and the authors do not cite their own prior results to justify the central claim. The Conclusion's assertion that the CWGAN 'can learn the distribution of ALS EEG signals' is supported only by qualitative evidence, and Section 5.1 explicitly concedes that 'we have not yet quantified their fidelity using objective metrics' and that no formal diversity analysis was performed. That is a serious validity limitation: stable critic loss and clinician visual review cannot rule out mode collapse or memorization, and the downstream augmentation benefit is untested. However, missing evidence and weak inference are not circularity. The training data and the generated samples play distinct roles, and no specific reduction from conclusion to input can be exhibited. The appropriate finding is therefore no significant circularity, with the validity concerns noted separately.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's central claim depends on standard WGAN theory, several untested domain assumptions about the EEGET-ALS preprocessing (i.i.d. segments, min-max scaling preserving relevant signal content), and an unverified assumption that visual inspection plus loss convergence demonstrate distributional fidelity. The free parameters are the hand-chosen WGAN hyperparameters; the trained network weights themselves are the fitted object and are not enumerated.

free parameters (6)

weight clipping bound = 0.01
Chosen by hand following the original WGAN (Ref. 5); required for the 1-Lipschitz constraint and directly affects stability and output quality.
critic updates per generator update (n_critic) = 5
Chosen per WGAN recommendation; the training routine depends on it to keep the critic near optimal.
learning rate (RMSprop) = 5e-5
Hand-selected; both networks use this learning rate.
latent dimension = 100
Hand-selected Gaussian noise dimensionality for the generator.
generator/critic widths = G: 256-512-1024; D: 1024-512-256
Hand-selected MLP architecture; no architecture search is reported.
number of training epochs = 300
Hand-chosen; training was monitored for convergence but no early stopping was used.

assumptions (4)

standard math WGAN with weight clipping optimizes an approximation of the Wasserstein distance and enforces the 1-Lipschitz constraint on the critic
Borrowed from Arjovsky et al. (Ref. 5); the entire loss formulation in Section 3.5 depends on this result.
domain assumption Rows of the EEGET-ALS MAT files can be treated as independent, identically distributed EEG segments
Section 3.2 states 'Each segment's rows (or time-series vectors) are treated as independent samples'; this ignores within-patient/session correlation.
domain assumption Per-file min-max scaling to [-1,1] preserves the spectral and temporal features relevant to ALS
Section 3.2 applies this normalization without analyzing its effect on the signal distribution.
ad hoc to paper Converging loss curves and informal clinician inspection are sufficient evidence that the generator has learned the data distribution without mode collapse
Section 4.2 and 5.1; the paper explicitly notes no formal diversity analysis was done, yet uses convergence as evidence of broad learning.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Synthetic ALS-EEG Data Augmentation for ALS Diagnosis Using Conditional WGAN with Weight Clipping." pith.science (2026). https://pith.science/paper/ZP4B3M3A

@misc{pith2026250616243,
  author       = {Pith},
  title        = {Pith review of: Synthetic ALS-EEG Data Augmentation for ALS Diagnosis Using Conditional WGAN with Weight Clipping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZP4B3M3A}},
  note         = {Machine review of arXiv:2506.16243}
}

read the original abstract

Amyotrophic Lateral Sclerosis (ALS) is a rare neurodegenerative disease, and high-quality EEG data from ALS patients are scarce. This data scarcity, coupled with severe class imbalance between ALS and healthy control recordings, poses a challenge for training reliable machine learning classifiers. In this work, we address these issues by generating synthetic EEG signals for ALS patients using a Conditional Wasserstein Generative Adversarial Network (CWGAN). We train CWGAN on a private EEG dataset (ALS vs. non-ALS) to learn the distribution of ALS EEG signals and produce realistic synthetic samples. We preprocess and normalize EEG recordings, and train a CWGAN model to generate synthetic ALS signals. The CWGAN architecture and training routine are detailed, with key hyperparameters chosen for stable training. Qualitative evaluation of generated signals shows that they closely mimic real ALS EEG patterns. The CWGAN training converged with generator and discriminator loss curves stabilizing, indicating successful learning. The synthetic EEG signals appear realistic and have potential use as augmented data for training classifiers, helping to mitigate class imbalance and improve ALS detection accuracy. We discuss how this approach can facilitate data sharing and enhance diagnostic models.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 42 canonical work pages

[1]

Introduction Amyotrophic Lateral Sclerosis (ALS) is a rapidly progressive neuro -degenerative disorder that affects the upper and lower motor neuron s, ultimately leading to muscle weakness, paralysis, and premature death.1,2 Although electro-encephalography (EEG) is non -invasive, cost-effective, and has shown promise for revealing cortical hyper -excita...

work page
[2]

Yu et al

Related Work GAN-based data augmentation for EEG has gained traction recently as a means to boost classification performance in data -scarce scenarios. Yu et al . introduced CWGAN and used it to generate fault data where deep learning models were suffering from data scarcity.11 Zhang et al.22 introduced ERP-WGAN, a GAN framework to augment single-trial EE...

work page
[3]

real” training set, an d non-ALS segments ('0_' files) into the “fake

Methods 3.1. Dataset We used EEGET-ALS Dataset.13,14 It contains raw 32-channel EEG recordings sampled at 256 Hz from six ALS patients—each contributing up to ten sessions over three to five months —and 170 healthy controls (one session each). Every recording comprises nine roughly two-minute blocks of motor imagery, actual movement, eye -tracking–based s...

work page
[4]

Visual Quality of Synthetic EEG: After training, we generated a large number of synthetic ALS EEG segments from the WGAN generator for evaluation

Results 4.1. Visual Quality of Synthetic EEG: After training, we generated a large number of synthetic ALS EEG segments from the WGAN generator for evaluation. Visual inspection of these waveforms suggests that the GAN successfully captured realistic EEG characteristic s of ALS patients. The synthetic signals exhibit oscillatory patterns and amplitude dyn...

work page
[5]

Discussion The present study demonstrates that a Conditional Wasserstein Generative Adversarial Network (CWGAN) with weight clipping can be trained on a small, highly imbalanced set of ALS EEG recordings to synthesize realistic patient -like signals. Qualitative inspection by experienced EEG readers, together with the smoothly convergent generator and cri...

work page
[6]

In the Introduction, we highlighted the rarity of high-quality ALS EEG data and the class imbalance problem that hampers machine learning classifiers

Conclusion We presented a novel study on synthetic EEG data augmentation for ALS diagnosis using WGAN. In the Introduction, we highlighted the rarity of high-quality ALS EEG data and the class imbalance problem that hampers machine learning classifiers. To tackle these issues, our approach utilizes a Conditional Wasserstein GAN to generate realistic EEG s...

work page
[7]

Kiernan, M. C. et al. Amyotrophic lateral sclerosis. The Lancet 377, 942–955 (2011)

work page 2011
[8]

Papaiz, F. et al. Ensemble-imbalance-based classification for amyotrophic lateral sclerosis prognostic prediction: identifying short -survival patients at diagnosis. BMC Med. I nform. Decis. Mak. 24, 80 (2024)

work page 2024

Show all 51 references

[9]

V., Bowyer, K

Chawla, N. V., Bowyer, K. W., Hall, L. O. & Kegelmeyer, W. P. SMOTE: Synthetic Minority Over-sampling Technique. J. Artif. Intell. Res. 16, 321–357 (2002)

2002
[10]

Goodfellow, I. J. et al. Generative Adversarial Networks. Pr eprint at https://doi.org/10.48550/arXiv.1406.2661 (2014)

2014 doi arXiv
[11]

& Bottou, L

Arjovsky, M., Chintala, S. & Bottou, L. Wasserstein GAN. Preprint at https://doi.org/10.48550/arXiv.1701.07875 (2017)

2017 doi arXiv
[12]

Conditional Wasserstein Generative Adversarial Networks

Fabbri, C. Conditional Wasserstein Generative Adversarial Networks
[13]

& Courville, A

Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V. & Courville, A. Improved Training of Wasserstein GANs. Preprint at https://doi.org/10.48550/arXiv.1704.00028 (2017)

2017 doi arXiv
[14]

Creswell, A. et al. Generative Adversarial Networks: An Overview. IEEE Signal Proces s. Mag. 35, 53–65 (2018)

2018
[15]

-Z., Wan, Z

Luo, Y., Zhu, L. -Z., Wan, Z. -Y. & Lu, B. -L. Data Augmentation for Enhancing EEG -based Emotion Recognition with Deep Generative Models. Preprint at https://doi.org/10.48550/ARXIV.2006.05331 (2020)

2020 doi arXiv
[16]

& Osindero, S

Mirza, M. & Osindero, S. C onditional Generative Adversarial Nets. Preprint at https://doi.org/10.48550/arXiv.1411.1784 (2014)

2014 doi arXiv
[17]

Yu, Y. et al. CWGAN: Conditional Wasserstein Generative Adversarial Nets for Fault Data Generation. in 2019 IEEE International Conference on Robotics a nd Biomimetics (ROBIO) 2713–2718 (IEEE, Dali, China, 2019). doi:10.1109/ROBIO49542.2019.8961501

2019
[18]

Arthur, K. C. et al. Projected increase in amyotrophic lateral sclerosis from 2015 to 2040. Nat. Commun. 7, 12408 (2016)

2016
[19]

Ngo, T. D. et al. An EEG & eye-tracking dataset of ALS patients & healthy people during eye- tracking-based spelling system usage. Sci. Data 11, 664 (2024)

2024
[20]

Le, T. H. et al. EEGET-ALS Dataset. 1680831504 Bytes figshare https://doi.org/10.6084/M9.FIGSHARE.24485689.V1 (2024)

2024 doi
[21]

Gelen, M. A. et al. Novel accurate classification system developed using order transition pattern feature engineering technique with physiological signals. Sci. Rep. 15, 15278 (2025)

2025
[22]

Haibo He & Garcia, E. A. Learning from Imbalanced Data. IEEE Trans. Knowl. Data Eng. 21, 1263–1284 (2009)

2009
[23]

Batista, G. E. A. P. A., Prati, R. C. & Monard, M. C. A study of the behavior of several methods for balancing machine learning training data. ACM SIGKDD Explor. Newsl. 6, 20–29 (2004)

2004
[24]

He, C., Liu, J., Zhu, Y . & Du, W. Data Augmentation for Deep Neural Networks Model in EEG Classification Task: A Review. Front. Hum. Neurosci. 15, 765525 (2021)

2021
[25]

Zhang, S., Sun, L., Mao, X., Zhao, M. & Hu, Y. ATGAN: attention -based temporal GAN for EEG data augmentation in personal identification. EURASIP J. Adv. Signal Process. 2024, 94 (2024)

2024
[26]

& Nam, C

Choo, S., Park, H., Jung, J.-Y., Flores, K. & Nam, C. S. Improving classification performance of motor imagery BCI through EEG data augmentation with conditional generative adversarial networks. Neural Netw. 180, 106665 (2024)

2024
[27]

G., Azab, A

Habashi, A. G., Azab, A. M., Eldawlatly, S. & Aly, G. M. Generative adversarial networks i n EEG analysis: an overview. J. NeuroEngineering Rehabil. 20, 40 (2023)

2023
[28]

Zhang, R. et al. ERP-WGAN: A data augmentation method for EEG single -trial detection. J. Neurosci. Methods 376, 109621 (2022)

2022
[29]

Yoon, J. et al. EHR-Safe: generating high-fidelity and privacy-preserving synthetic electronic health records. Npj Digit. Med. 6, 141 (2023)

2023
[30]

K., Mrachacz-Kersting, N

Fahimi, F., Dosen, S., Ang, K. K., Mrachacz-Kersting, N. & Guan, C. Generative Adversarial Networks-Based Data Augmentation for Brain–Computer Interface. IEEE Trans. Neural Netw. Learn. Syst. 32, 4039–4051 (2021)

2021
[31]

& Liu, J

Song, J., Zhai, Q., Wang, C. & Liu, J. EEGGAN -Net: enhancing EEG signal classification through data augmentation. Front. Hum. Neurosci. 18, 1430086 (2024)

2024
[32]

P., Hollenbenders, Y

Carrle, F. P., Hollenbenders, Y. & R eichenbach, A. Generation of synthetic EEG data for training algorithms supporting the diagnosis of major depressive disorder. Front. Neurosci. 17, 1219133 (2023)

2023
[33]

& Shung, D

Giuffrè, M. & Shung, D. L. Harnessing the power of synthetic data in healthcare: innovat ion, application, and privacy. Npj Digit. Med. 6, 186 (2023)

2023
[34]

Du, X. et al. A Data Augmentation Method for Motor Imagery EEG Signals Based on DCGAN-GP Network. Brain Sci. 14, 375 (2024)

2024
[35]

J., Kuhlmann, L

Rasheed, K., Qadir, J., O’Brien, T. J., Kuhlmann, L. & Razi, A. A Generative Model to Synthesize EEG Data for Epileptic Seizure Prediction. IEEE Trans. Neural Syst. Rehabil. Eng. 29, 2322–2332 (2021)

2021
[36]

Luo, Y. & Lu, B. -L. EEG Data Augmentation for Emotion Recognition Using a Conditional Wasserstein GAN. in 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 2535–2538 (IEEE, Honolulu, HI, 2018). doi:10.1109/EMBC.2018.8512865

2018
[37]

Tibermacine, I. E. et al. Adversarial denoising of EEG signals: a comparative analys is of standard GAN and WGAN-GP approaches. Front. Hum. Neurosci. 19, 1583342 (2025)

2025
[38]

Panwar, S., Rad, P., Jung, T. -P. & Huang, Y. Modeling EEG Data Distribution With a Wasserstein Generative Adversarial Network to Predict RSVP Events. IEEE Trans. Neura l Syst. Rehabil. Eng. 28, 1720–1730 (2020)

2020
[39]

& Faria, D

Venugopal, A. & Faria, D. R. Employing WGAN -GP for Synthesizing Biophysical Data: Generating Synthetic EEG for Concentration and Relaxation Level Prediction. in Advances in Information and Communication (ed. Arai, K.) vol. 1284 62–80 (Springer Nature Switzerland, Cham, 2025)

2025
[40]

G., Schirrmeister, R

Hartmann, K. G., Schirrmeister, R. T. & Ball, T. EEG-GAN: Generative adversarial networks for electroencephalograhic (EEG) brain signals. Preprint at https://doi.org/10.48550/arXiv.1806.01875 (2018)

2018 doi arXiv
[41]

& Fang, Y

Chen, Z., Qian, Y., Wang, Y. & Fang, Y. Deep Convolutional Generative Adversarial Network-Based EMG Data Enhancement for Hand Motion Classification. Front. Bioeng. Biotechnol. 10, 909653 (2022)

2022
[42]

& Moscato , F

Berger, L., Haberbusch, M. & Moscato , F. Generative adversarial networks in electrocardiogram synthesis: Recent developments and challenges. Artif. Intell. Med. 143, 102632 (2023)

2023
[43]

Gramfort, A. et al. MNE software for processing MEG and EEG data. NeuroImage 86, 446– 460 (2014)

2014
[44]

& Chintala, S

Radford, A., Metz, L. & Chintala, S. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. Preprint at https://doi.org/10.48550/arXiv.1511.06434 (2016)

2016 doi arXiv
[45]

Smith, K. E. & Smith, A. O. Conditional GAN for timeseries generation. Preprint at https://doi.org/10.48550/arXiv.2006.16477 (2020)

2020 doi arXiv
[46]

L., Hannun, A

Maas, A. L., Hannun, A. Y. & Ng, A. Y. Rectiﬁer Nonlinearities Improve Neural Network Acoustic Models
[47]

(COURSERA: Neural Networks for Machine Learning, 4, 26-31, 2012)

Lecture 6.5 – Rmsprop: Divide the Gradient by a Running Average of Its Recent Magnitude. (COURSERA: Neural Networks for Machine Learning, 4, 26-31, 2012)

2012
[48]

& Khoshgoftaar, T

Sauber-Cole, R. & Khoshgoftaar, T. M. The use of generative adversarial networks to alleviate class imbalance in tabular data: a survey. J. Big Data 9, 98 (2022)

2022
[49]

& Ahamed, S

George, O., Smith, R., Madiraju, P., Yahyasoltani, N. & Ahamed, S. I. Data augmentation strategies for EEG-based motor imagery decoding. Heliyon 8, e10240 (2022)

2022
[50]

& Smith, S

Gonzales, A., Guruswamy, G. & Smith, S. R. Synthetic data in health care: A narrative review. PLOS Digit. Health 2, e0000082 (2023)

2023
[51]

Kingma, D. P. & Welling, M. Auto -Encoding Variational Bayes. Preprint at https://doi.org/10.48550/arXiv.1312.6114 (2022)

2022 doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Introduction Amyotrophic Lateral Sclerosis (ALS) is a rapidly progressive neuro -degenerative disorder that affects the upper and lower motor neuron s, ultimately leading to muscle weakness, paralysis, and premature death.1,2 Although electro-encephalography (EEG) is non -invasive, cost-effective, and has shown promise for revealing cortical hyper -excita...

work page

[2] [2]

Yu et al

Related Work GAN-based data augmentation for EEG has gained traction recently as a means to boost classification performance in data -scarce scenarios. Yu et al . introduced CWGAN and used it to generate fault data where deep learning models were suffering from data scarcity.11 Zhang et al.22 introduced ERP-WGAN, a GAN framework to augment single-trial EE...

work page

[3] [3]

real” training set, an d non-ALS segments ('0_' files) into the “fake

Methods 3.1. Dataset We used EEGET-ALS Dataset.13,14 It contains raw 32-channel EEG recordings sampled at 256 Hz from six ALS patients—each contributing up to ten sessions over three to five months —and 170 healthy controls (one session each). Every recording comprises nine roughly two-minute blocks of motor imagery, actual movement, eye -tracking–based s...

work page

[4] [4]

Visual Quality of Synthetic EEG: After training, we generated a large number of synthetic ALS EEG segments from the WGAN generator for evaluation

Results 4.1. Visual Quality of Synthetic EEG: After training, we generated a large number of synthetic ALS EEG segments from the WGAN generator for evaluation. Visual inspection of these waveforms suggests that the GAN successfully captured realistic EEG characteristic s of ALS patients. The synthetic signals exhibit oscillatory patterns and amplitude dyn...

work page

[5] [5]

Discussion The present study demonstrates that a Conditional Wasserstein Generative Adversarial Network (CWGAN) with weight clipping can be trained on a small, highly imbalanced set of ALS EEG recordings to synthesize realistic patient -like signals. Qualitative inspection by experienced EEG readers, together with the smoothly convergent generator and cri...

work page

[6] [6]

In the Introduction, we highlighted the rarity of high-quality ALS EEG data and the class imbalance problem that hampers machine learning classifiers

Conclusion We presented a novel study on synthetic EEG data augmentation for ALS diagnosis using WGAN. In the Introduction, we highlighted the rarity of high-quality ALS EEG data and the class imbalance problem that hampers machine learning classifiers. To tackle these issues, our approach utilizes a Conditional Wasserstein GAN to generate realistic EEG s...

work page

[7] [7]

Kiernan, M. C. et al. Amyotrophic lateral sclerosis. The Lancet 377, 942–955 (2011)

work page 2011

[8] [8]

Papaiz, F. et al. Ensemble-imbalance-based classification for amyotrophic lateral sclerosis prognostic prediction: identifying short -survival patients at diagnosis. BMC Med. I nform. Decis. Mak. 24, 80 (2024)

work page 2024