Pith. sign in

REVIEW 4 major objections 7 minor 44 references

DiffEEG shows that diffusion pre-training plus F1-maximizing RL can detect rare seizures under real clinical imbalance with little labeled data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 04:38 UTC pith:MNN5XPOL

load-bearing objection Solid engineering baseline for imbalanced clinical EEG: real patient-wise numbers, but the RL story is unablated and the 97.6% figure is not a fair capacity claim. the 4 major comments →

arxiv 2607.11578 v1 pith:MNN5XPOL submitted 2026-07-13 cs.LG cs.AIcs.CVeess.SP

DiffEEG: A Self-Supervised Denoising Diffusion Model for Learning EEG Generic Representations

Abdulkader Helwan , Lina Abou-Abbas , Hussein El Amouri , Belkacem Chikhaoui , Khadidja Henni This is my paper
classification cs.LG cs.AIcs.CVeess.SP
keywords EEGdenoising diffusionfoundation modelreinforcement learningseizure detectionclass imbalanceself-supervised learningpatient-wise evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Clinical EEG seizure detection is bottlenecked by scarce annotations and extreme class imbalance: seizures often occupy well under 10% of continuous recordings. DiffEEG is a 9.6-million-parameter foundation model that first learns generic brain-signal structure by training a 1D U-Net with self-attention to reverse a diffusion noising process on 1.3 million unlabeled Temple University Hospital segments. Downstream, a reinforced decision layer uses policy gradients to optimize the F1 score directly, so the classifier is rewarded for catching rare ictal events rather than maximizing overall accuracy. On a strict patient-wise leave-one-fold-out protocol over 279 patients the model reaches 81% accuracy and 85% weighted F1 for binary detection (59% seizure recall at 6.7% prevalence) and 61% accuracy / 59% F1 for four-class seizure subtyping; a segment-level upper bound of 97.6% accuracy confirms the architecture can discriminate when patient identity is not held out. The paper therefore claims that diffusion-based self-supervision plus metric-aware reinforcement learning yields representations that support clinically usable monitoring with far fewer labels than fully supervised pipelines.

Core claim

Denoising-diffusion pre-training on large unlabeled EEG, followed by a reinforced decision layer that directly maximizes F1 via policy gradients, produces patient-generalizable seizure detectors that remain sensitive to rare ictal events under the severe class imbalance of continuous clinical monitoring.

What carries the argument

The reinforced decision layer (RDL): a small policy network that adds an input-conditioned logit adjustment and is trained by REINFORCE on batch-level seizure F1, so the model is explicitly rewarded for sensitivity rather than majority-class accuracy.

Load-bearing premise

The multi-timestep encoder features extracted at five fixed noise levels stay similar enough across patients that an F1-maximizing policy trained on some patients still works on completely unseen patients.

What would settle it

A fresh patient-wise hold-out on a larger multi-center EEG corpus (or an independent seizure corpus with the same 5-second windowing) that drops seizure recall well below 50% or collapses weighted F1 under severe imbalance would falsify the claimed clinical viability.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. DiffEEG proposes a 9.6M-parameter EEG foundation model pretrained by denoising diffusion on 1.3M unlabeled TUHSZ segments with a 1D U-Net (FiLM residual blocks, multi-head attention), then adapted via multi-timestep encoder probing and a reinforced decision layer that uses REINFORCE to maximize batch F1. Under patient-wise LOFO on 279 patients the model reports 61% accuracy / 59% weighted F1 for 4-class seizure subtyping; on the official TUHSZ evaluation split it reports 81% accuracy, 85% weighted F1, and 59% seizure recall for binary detection at 6.7% prevalence. A segment-wise 80/20 split yields 97.6% accuracy, framed as an architectural upper bound. The paper also reports few-shot subtype results and a TUAB abnormality-detection transfer experiment (AUROC 0.922).

Significance. If the joint contribution of diffusion pretraining and metric-aware RL is substantiated, the work is a useful, parameter-efficient contribution to EEG foundation models under realistic imbalance. Strengths include honest reporting of prevalence, use of the official TUHSZ evaluation split and strict patient-wise LOFO for subtyping, a compact 9.6M architecture relative to larger EEG FMs, and explicit multi-timestep probing. The clinical-viability narrative and the claim that diffusion+RL together enable deployable monitoring would be more persuasive with matched ablations and tempered language around 59% seizure recall.

major comments (4)
  1. The central claim that diffusion pretraining combined with the reinforced decision layer produces the reported patient-wise binary metrics is not isolated. Sections 3.2.2–3.2.3 introduce the policy-gradient term (Eqs. 16–22, λ=0.1/0.15) and asymmetric logit adjustment; Section 4.3.2 and Figure 5 only show that the RL reward rises after the warmup. No matched patient-wise run with λ=0 (or RDL removed) is reported for binary detection or LOFO subtyping. Without that ablation it is impossible to attribute the 59% seizure recall (Table 2) to the pretrained encoder, inverse-frequency CE, or the RL head. A λ=0 control under the same protocol is load-bearing for the foundation-model-plus-RL narrative.
  2. The abstract and Section 4.2.2 present the segment-wise 97.66% accuracy / 97.67% weighted F1 (Table 3) as an “upper bound” confirming architectural capacity. That split allows segments from the same patient in train and test (explicitly noted as leakage). Framing a leakage-contaminated number as capacity evidence overstates what Table 3 shows relative to the patient-wise LOFO mean of 61%/59% (Table 4). Either reframe Table 3 strictly as a non-generalizing ceiling under leakage, or add a patient-wise full-data upper bound for fair comparison.
  3. “Clinically viable” / “clinically deployable” language (abstract, §5.1, conclusion) rests on 59% seizure recall at 6.7% prevalence (Table 2) and 59% LOFO F1 for 4-class subtyping (Table 4). Missing 41% of seizures is a serious clinical cost; patient-wise subtype F1 is also below several supervised baselines on smaller cohorts (Table 6: MBMD 73.9%, Wavelet2Vec 72%). Either provide clinical justification (e.g., acceptable false-negative rates in continuous monitoring, operating-point analysis, precision–recall curves) or temper the deployability claim to match the reported sensitivity and cross-patient variance (Folds 1 and 5 F1 ≈0.48).
  4. There is no from-scratch or non-diffusion pretrained control under the same patient-wise protocol. The multi-timestep probe T={50,250,500,750,950} (Eqs. 23–25, §3.4.2) and the partial-freeze partition (§3.2.4, §3.4.4) are free design choices without sensitivity analysis. At minimum, report (i) encoder trained from scratch with the same head/RL setup and (ii) single-timestep vs multi-timestep probing on the LOFO or official eval split, so the value of diffusion pretraining itself can be assessed.
minor comments (7)
  1. Section 4.2: “full supervised sine-tuning regime” is a typo for “fine-tuning”.
  2. Table numbering is inconsistent: body text refers to Table I / II / III / IV while captions use Arabic numerals (Table 1–8). Unify.
  3. RL weight is λ=0.1 in Eq. (22) but λ=0.15 in Table 1; state which value was used for the reported runs.
  4. Figure 4 is referenced for few-shot results but numerical values beyond K=342 (W-F1 0.6578) are not tabulated; a small table would aid reproducibility.
  5. Binary detection is described as patient-wise / LOFO in places (abstract, Table 2 header) while §3.2.1 and Table 7 indicate the predefined TUHSZ evaluation set; clarify that binary results use the official eval split, not 5-fold LOFO.
  6. Pretraining is said to use “non-seizure” recordings (§3.1) while also “1.3M unlabeled segments from TUHSZ”; confirm whether any ictal segments entered pretraining and, if not, discuss possible domain shift for seizure morphology.
  7. Citation [32] for REINFORCE points to an auxiliary-loss search paper rather than Williams (1992) or a standard policy-gradient reference; fix the primary citation.

Circularity Check

0 steps flagged

No load-bearing circularity: empirical ML results on held-out patient splits; diffusion and REINFORCE are external standards; minor author-overlap citations are background only.

full rationale

DiffEEG is an empirical deep-learning paper whose central claims are measured performance numbers (patient-wise LOFO accuracy/F1 on TUHSZ binary detection and 4-class subtyping, plus segment-wise upper bound). The pre-training objective is the standard DDPM noise-prediction MSE of Ho et al. (Eq. 14) with a cosine schedule from Nichol & Dhariwal; the reinforced decision layer is ordinary REINFORCE on batch F1 (Eqs. 19–22) with a moving-average baseline. Neither construction defines the reported metrics in terms of themselves. Evaluation uses the official TUHSZ evaluation split and strict 5-fold patient-wise LOFO (279 patients, no segment leakage), so the numbers are not forced by construction from training data. Self-citations to prior EEG work by overlapping authors appear only as background motivation for imbalance and seizure detection; they do not supply uniqueness theorems, ansätze, or fitted constants that the present results reduce to. The large segment-wise vs. patient-wise gap is a genuine generalization limitation, not circularity. Score 1 only for the presence of non-load-bearing author-overlap references; the derivation chain itself is self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard diffusion mathematics, the empirical premise that non-seizure EEG is a useful pre-training manifold, a set of hand-chosen architectural and RL hyper-parameters, and two invented modules (the multi-timestep probe aggregator and the reinforced decision layer) whose value is demonstrated only inside this paper.

free parameters (5)
  • RL loss weight λ = 0.1 / 0.15
    Set to 0.1 in the method text and 0.15 in Table I; controls the trade-off between cross-entropy and F1 policy gradient and is not derived.
  • Probe timesteps T_probe = {50,250,500,750,950}
    Fixed set {50,250,500,750,950} chosen by the authors; different choices would alter the 2400-dimensional feature vector.
  • Partial-freeze partition = early levels frozen, last two + bottleneck trainable
    Which encoder blocks remain frozen versus trainable is a design choice that directly affects reported transfer performance.
  • Policy-network hidden size = max(64, D/4)
    Set to max(64, D/4); another free architectural hyper-parameter of the reinforced decision layer.
  • Baseline momentum α and entropy bonus β = α=0.9, β=0.01
    REINFORCE variance-reduction constants (α=0.9, β=0.01) chosen without ablation.
axioms (4)
  • standard math The forward diffusion process and cosine noise schedule of Ho et al. / Nichol & Dhariwal correctly model the EEG data manifold.
    Equations (1)–(4) are taken unchanged from the DDPM literature.
  • domain assumption Non-seizure EEG segments form a sufficiently rich manifold that denoising them yields transferable features for seizure detection.
    Stated in Section 3.1; the entire pre-training stage rests on this premise.
  • ad hoc to paper Batch-level F1 computed on Bernoulli samples is an unbiased enough reward for REINFORCE to improve patient-wise seizure recall.
    Introduced in Section 3.2.2; no theoretical guarantee is offered that the policy generalizes beyond the mini-batch.
  • domain assumption Five-second windows with 30% overlap and the chosen 1–75 Hz bandpass preserve all clinically relevant seizure morphology.
    Preprocessing choices in Section 3.5.1; different windowing would change both pre-training and evaluation distributions.
invented entities (2)
  • Reinforced Decision Layer (RDL) no independent evidence
    purpose: Asymmetric logit adjustment trained by policy gradient to maximize seizure F1 rather than accuracy.
    Defined in Section 3.2.2; no independent theoretical or empirical existence outside this work.
  • Multi-timestep diffusion feature probe no independent evidence
    purpose: Concatenate global-average-pooled encoder states at five noise levels into a 2400-dim vector for classification.
    Section 3.4.2; the specific aggregation scheme is original to the paper.

pith-pipeline@v1.1.0-grok45 · 21593 in / 3592 out tokens · 32718 ms · 2026-07-14T04:38:17.190734+00:00 · methodology

0 comments
read the original abstract

Deep learning for EEG-based seizure detection faces critical challenges: severe annotation scarcity and extreme class imbalance, where ictal events comprise less than 10\% of clinical recordings. We present DiffEEG, a 9.6M-parameter self-supervised foundation model that addresses both limitations through denoising diffusion pre-training and reinforcement learning (RL)-based fine-tuning. Pre-trained on 1.3M unlabeled segments from the Temple University Hospital Seizure Corpus (TUHSZ), DiffEEG learns generic neural representations via a 1D U-Net with multi-head self-attention. For downstream adaptation, a reinforced decision layer employs policy gradient optimization to directly maximize F1-score, prioritizing sensitivity to rare seizure events over overall accuracy. Under strict patient-wise evaluation (279 patients, Leave-One-Fold-Out), DiffEEG achieves 61\% accuracy and 59\% F1 for 4-class seizure subtyping, and 81\% accuracy with 85\% weighted F1 for binary detection, maintaining clinically viable seizure recall (59\%) despite extreme imbalance (6.7\% prevalence). Segment-level evaluation establishes an upper bound of 97.6\% accuracy, confirming strong architectural capacity. DiffEEG demonstrates that diffusion-based pre-training combined with metric-aware reinforcement learning enables clinically deployable seizure monitoring with minimal labeled data requirements.

Figures

Figures reproduced from arXiv: 2607.11578 by Abdulkader Helwan, Belkacem Chikhaoui, Hussein El Amouri, Khadidja Henni, Lina Abou-Abbas.

Figure 1
Figure 1. Figure 1: Overview of the DiffEEG 1D U-Net architecture, showing the encoder, bottleneck, and decoders pathway used for feature extraction. where ¯αt is derived from: α¯t = cos2  t/T +s 1+s · π 2  cos2  s 1+s · π 2  (3) with s = 0.008 as a small offset to prevent the schedule from reaching zero too early. The cosine schedule provides more stable training dynamics compared to linear alternatives, particularly imp… view at source ↗
Figure 5
Figure 5. Figure 5: Fine-tuning and reinforcement learning dynamics for the proposed model. The vertical dashed line indicates the activation of the Policy Gradient-based decision layer at Epoch 6. 4.3.2 Training Dynamics and RL Contribution The policy gradient (RL) auxiliary loss was activated after epoch 5 (warmup period). As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Model sizes comparison in terms of parameter count. 5.4.2 Binary Seizure Detection Benchmarking (Patient-Wise) The binary detection task evaluates the ability to distinguish ictal segments from interictal background activity. Table VII compares DiffEEG against foundation models and supervised clinical baselines using the primary metrics explicitly reported in each study. Foundational architectures such as … view at source ↗
Figure 2
Figure 2. Figure 2: Detailed structure of the residual block featuring Feature-wise Linear Modulation (FiLM) for timestep-based conditioning. Preprint — 18 [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Patient-wise 5-fold cross-validation for seizure subtype classification [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Few-Shot Seizure 5-Subtypes Classification Performance (Frozen vs. Unfrozen Backbone Strategies), 80/20 stratified split, segment-wise. Preprint — 19 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 7 linked inside Pith

  1. [1]

    Generative ai with wgan-gp for boosting seizure detection accuracy,

    L. Abou-Abbas, K. Henni, I. Jemal, and N. Mezghani, “Generative ai with wgan-gp for boosting seizure detection accuracy,”Frontiers in Artificial Intelligence, vol. 7, p. 1437315, 2024

    2024

  2. [2]

    Deep learning-based electroencephalography analysis: A systematic review,

    Y. Roy, H. Banville, I. Albuquerque, A. Gramfort, T. H. Falk, and J. Faubert, “Deep learning-based electroencephalography analysis: A systematic review,”Journal of Neural Engineering, vol. 16, no. 5, p. 051001, Aug. 2019

    2019

  3. [3]

    Deep learning with convolutional neural networks for eeg decoding and visualization,

    R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, and T. Ball, “Deep learning with convolutional neural networks for eeg decoding and visualization,”Human Brain Mapping, vol. 38, no. 11, pp. 5391–5420, Nov. 2017

    2017

  4. [4]

    Deep convolutional neural network for the automated detection and diagnosis of seizure using eeg signals,

    U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, and H. Adeli, “Deep convolutional neural network for the automated detection and diagnosis of seizure using eeg signals,”Computers in Biology and Medicine, vol. 100, pp. 270–278, Sep. 2018

    2018

  5. [5]

    Eegformer: Towards transferable and interpretable large-scale eeg foundation model,

    Y. Chen, K. Ren, K. Song, Y. Wang, Y. Wang, D. Li, and L. Qiu, “Eegformer: Towards transferable and interpretable large-scale eeg foundation model,”arXiv preprint arXiv:2401.10278, 2024. Preprint — 14

    Pith/arXiv arXiv 2024

  6. [6]

    Deep learning for epileptic seizure prediction from eeg signals: A review,

    X. Huang, H. Meng, and Z. Li, “Deep learning for epileptic seizure prediction from eeg signals: A review,”Biomedical Signal Processing and Control, vol. 117, p. 109518, 2026

    2026

  7. [7]

    Imbalance-aware machine learning for epileptic seizure detection,

    K. Henni, L. Abou-Abbas, I. Jmal, A. Mitiche, and N. Mezghani, “Imbalance-aware machine learning for epileptic seizure detection,” in2024 IEEE 12th International Symposium on Signal, Image, Video and Communications (ISIVC). IEEE, 2024, pp. 1–5

    2024

  8. [8]

    Focal and generalized seizures distinction by rebalancing class data and random forest classification,

    L. Abou-Abbas, I. Jemal, K. Henni, A. Mitiche, and N. Mezghani, “Focal and generalized seizures distinction by rebalancing class data and random forest classification,” in International Conference on Bioengineering and Biomedical Signal and Image Processing. Springer, 2021, pp. 63–70

    2021

  9. [9]

    Eegnet: a compact convolutional neural network for eeg-based brain–computer interfaces,

    V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “Eegnet: a compact convolutional neural network for eeg-based brain–computer interfaces,” Journal of neural engineering, vol. 15, no. 5, p. 056013, 2018

    2018

  10. [10]

    Eeg oscillatory power and complexity for epileptic seizure detection,

    L. Abou-Abbas, I. Jemal, K. Henni, Y. Ouakrim, A. Mitiche, and N. Mezghani, “Eeg oscillatory power and complexity for epileptic seizure detection,”Applied Sciences, vol. 12, no. 9, p. 4181, 2022

    2022

  11. [11]

    Eegwavenet: Multiscale cnn-based spatiotemporal feature extraction for eeg seizure detection,

    P. Thuwajit, P. Rangpong, P. Sawangjai, P. Autthasan, R. Chaisaen, N. Banluesombatkul, P. Boonchit, N. Tatsaringkansakul, T. Sudhawiyangkul, and T. Wilaiprasitporn, “Eegwavenet: Multiscale cnn-based spatiotemporal feature extraction for eeg seizure detection,”IEEE transactions on industrial informatics, vol. 18, no. 8, pp. 5547–5557, 2021

    2021

  12. [12]

    Eegpt: Pretrained transformer for universal and reliable representation of eeg signals,

    G. Wang, W. Liu, Y. He, C. Xu, L. Ma, and H. Li, “Eegpt: Pretrained transformer for universal and reliable representation of eeg signals,”Advances in Neural Information Processing Systems, vol. 37, pp. 39 249–39 280, 2024

    2024

  13. [13]

    Transformer-based spatial-temporal feature learning for eeg decoding,

    Y. Song, X. Jia, L. Yang, and L. Xie, “Transformer-based spatial-temporal feature learning for eeg decoding,”arXiv preprint arXiv:2106.11170, 2021

    Pith/arXiv arXiv 2021

  14. [14]

    Eeg conformer: Convolutional transformer for eeg decoding and visualization,

    Y. Song, Q. Zheng, B. Liu, and X. Gao, “Eeg conformer: Convolutional transformer for eeg decoding and visualization,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 31, pp. 710–719, 2022

    2022

  15. [15]

    Tsd: Transformers for seizure detection,

    Y. Ma, C. Liu, M. S. Ma, Y. Yang, N. D. Truong, K. Kothur, A. Nikpour, and O. Kavehei, “Tsd: Transformers for seizure detection,”bioRxiv, pp. 2023–01, 2023

    2023

  16. [16]

    Toward eeg-free seizure detection: Evidence of eeg-ecg synchronization,

    H. El Amouri, L. Abou Abbas, H. Tfaily, and K. Henni, “Toward eeg-free seizure detection: Evidence of eeg-ecg synchronization,” in2025 IEEE/ACS 22nd International Conference on Computer Systems and Applications (AICCSA). IEEE, 2025, pp. 1–8

    2025

  17. [17]

    The temple university hospital eeg data corpus,

    I. Obeid and J. Picone, “The temple university hospital eeg data corpus,”Frontiers in neuroscience, vol. 10, p. 196, 2016

    2016

  18. [18]

    Eegformer: A transformer–based brain activity classification method using eeg signal,

    Z. Wan, M. Li, S. Liu, J. Huang, H. Tan, and W. Duan, “Eegformer: A transformer–based brain activity classification method using eeg signal,”Frontiers in neuroscience, vol. 17, p. 1148855, 2023

    2023

  19. [19]

    BIOT: Cross-data biosignal learning in the wild,

    C. Yang, M. B. Westover, and J. Sun, “BIOT: Cross-data biosignal learning in the wild,” arXiv preprint arXiv:2305.10351, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Large brain model for learning generic representations with tremendous EEG data in BCI,

    W.-B. Jiang, L.-M. Zhao, and B.-L. Lu, “Large brain model for learning generic representations with tremendous EEG data in BCI,” inProc. 12th Int. Conf. Learn. Representations (ICLR), 2024

    2024

  21. [21]

    Large cognition model: Towards pretrained electroencephalography (EEG) foundation model,

    C.-S. Chen, Y.-J. Chen, and A. H.-W. Tsai, “Large cognition model: Towards pretrained electroencephalography (EEG) foundation model,”arXiv preprint arXiv:2502.17464, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Bioserenity-e1: a self-supervised eeg model for medical applications,

    R. G. Bettinardi, M. Rahmouni, and U. Gimenez, “Bioserenity-e1: a self-supervised eeg model for medical applications,”arXiv preprint arXiv:2503.10362, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Brainbert: Self-supervised representation learning for intracranial recordings,

    C. Wang, V. Subramaniam, A. U. Yaari, G. Kreiman, B. Katz, I. Cases, and A. Barbu, “Brainbert: Self-supervised representation learning for intracranial recordings,”arXiv preprint arXiv:2302.14367, 2023. Preprint — 15

    Pith/arXiv arXiv 2023

  24. [24]

    Wavelet2vec: A filter bank masked autoencoder for eeg-based seizure subtype classification,

    R. Peng, C. Zhao, Y. Xu, J. Jiang, G. Kuang, J. Shao, and D. Wu, “Wavelet2vec: A filter bank masked autoencoder for eeg-based seizure subtype classification,” inProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  25. [25]

    Tie-eegnet: Temporal information enhanced eegnet for seizure subtype classification,

    R. Peng, C. Zhao, J. Jiang, G. Kuang, Y. Cui, Y. Xu, H. Du, J. Shao, and D. Wu, “Tie-eegnet: Temporal information enhanced eegnet for seizure subtype classification,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 30, pp. 2567–2576, 2022

    2022

  26. [26]

    Eeg foundation models: A critical review of current progress and future directions,

    G. Kuruppu, N. Wagh, V. Kremen, and Y. Varatharajah, “Eeg foundation models: A critical review of current progress and future directions,”Journal of Neural Engineering, vol. 23, no. 2, p. 021001, 2026

    2026

  27. [27]

    CHB-MIT Scalp EEG Database,

    J. Guttag, “CHB-MIT Scalp EEG Database,”PhysioNet, Jun. 2010, version 1.0.0. [Online]. Available: https://doi.org/10.13026/C2K01R

    doi:10.13026/c2k01r 2010

  28. [28]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 6840–6851

    2020

  29. [29]

    Improved denoising diffusion probabilistic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” in Proceedings of the International Conference on Machine Learning, 2021, pp. 8162–8171

    2021

  30. [30]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. D. Vries, V. Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, 2018

    2018

  31. [31]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  32. [32]

    Reinforcement learning with automated auxiliary loss search,

    T. He, Y. Zhang, K. Ren, M. Liu, C. Wang, W. Zhang, Y. Yang, and D. Li, “Reinforcement learning with automated auxiliary loss search,”Advances in neural information processing systems, vol. 35, pp. 1820–1834, 2022

    2022

  33. [33]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,”arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  34. [34]

    Wavelet-based multi-class seizure type classification system,

    H. Albaqami, G. M. Hassan, and A. Datta, “Wavelet-based multi-class seizure type classification system,”Applied Sciences, vol. 12, no. 11, p. 5702, 2022

    2022

  35. [35]

    Residual and bidirectional lstm for epileptic seizure detection,

    W. Zhao, W.-F. Wang, L. M. Patnaik, B.-C. Zhang, S.-J. Weng, S.-X. Xiao, D.-Z. Wei, and H.-F. Zhou, “Residual and bidirectional lstm for epileptic seizure detection,”Frontiers in Computational Neuroscience, vol. 18, 2024

    2024

  36. [36]

    Multiband seizure type classification based on 3d convolution with attention mechanisms,

    H. Huang, P. Chen, J. Wen, X. Lu, and N. Zhang, “Multiband seizure type classification based on 3d convolution with attention mechanisms,”Computers in Biology and Medicine, vol. 166, p. 107517, 2023

    2023

  37. [37]

    Seizurenet: Multi-spectral deep feature learning for seizure type classification,

    U. Asif, S. Roy, J. Tang, and S. Harrer, “Seizurenet: Multi-spectral deep feature learning for seizure type classification,” inMachine Learning in Clinical Neuroimaging and Radiogenomics in Neuro-Oncology, 2020, pp. 77–87

    2020

  38. [38]

    Variable weight algorithm for convolutional neural networks and its applications to classification of seizure phases and types,

    G. Jia, H. K. Lam, and K. Althoefer, “Variable weight algorithm for convolutional neural networks and its applications to classification of seizure phases and types,”Pattern Recognition, vol. 121, p. 108226, 2022

    2022

  39. [39]

    Epileptic seizure detection in eeg signals using a unified temporal-spectral squeeze-and-excitation network,

    Y. Li, Y. Liu, W.-G. Cui, Y.-Z. Guo, H. Huang, and Z.-Y. Hu, “Epileptic seizure detection in eeg signals using a unified temporal-spectral squeeze-and-excitation network,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 28, no. 4, pp. 782–794, 2020

    2020

  40. [40]

    A combination of statistical parameters for epileptic seizure detection and classification using vmd and nltwsvm,

    S. Zhang, G. Liu, R. Xiao, W. Cui, J. Cai, and X. Hu, “A combination of statistical parameters for epileptic seizure detection and classification using vmd and nltwsvm,” Biocybernetics and Biomedical Engineering, vol. 42, no. 1, pp. 258–272, 2022

    2022

  41. [41]

    Seizure type classification using eeg signals and machine learning: Setting a benchmark,

    S. Roy, U. Asif, J. Tang, and S. Harrer, “Seizure type classification using eeg signals and machine learning: Setting a benchmark,” inProceedings of the IEEE Signal Processing in Medicine and Biology Symposium, 2020, pp. 1–6. Preprint — 16

    2020

  42. [42]

    Multi-branch mutual-distillation transformer for eeg-based seizure subtype classification,

    R. Peng, Z. Du, W. Liu, D. Wu, C. Zhao, J. Luo, and X. Chen, “Multi-branch mutual-distillation transformer for eeg-based seizure subtype classification,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2024

    2024

  43. [43]

    Adversarial representation learning for robust patient-independent epileptic seizure detection,

    X. Zhang, L. Yao, M. Dong, Z. Liu, Y. Zhang, and Y. Li, “Adversarial representation learning for robust patient-independent epileptic seizure detection,”IEEE Journal of Biomedical and Health Informatics, vol. 24, no. 10, pp. 2852–2859, 2020

    2020

  44. [44]

    Weak supervision of deep learning models for seizure onset detection,

    K. Saab, J. Dunnmon, D. Rubin, and C. Lee-Messer, “Weak supervision of deep learning models for seizure onset detection,”npj Digital Medicine, 2020. Preprint — 17 Time Embedding (512-d) Time MLP SiLU + Linear γ,β Inputh InstanceNorm1d SiLU Conv1d (3×3) FiLM Layer h·(1 +γ) +β InstanceNorm1d SiLU Dropout (0.1) Conv1d (3×3) Multi-Head Attn* (8 heads) Add Res...

    2020