REVIEW 4 major objections 4 minor 15 references
Discerning and quantifying high frequency activities in EEG under normal and epileptic conditions
T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read This paper claims that the gamma band of the EEG (40-100 Hz) carries a reliable, quantifiable epilepsy signature, and that isolating two dominant frequencies near 50 Hz and 76 Hz lets a linear/SVM classifier separate healthy from ictal sign
desk verdict The submission is an unrelated federated-learning paper paired with an EEG abstract, so the central 94–95% classification claim is unverifiable; the abstract describes a testable idea, but this manuscript cannot be evaluated as submitted. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the phase-space reconstruction of EEG signals restricted to two narrow frequency bands centered at roughly 50 Hz and 76 Hz, obtained via Welch's transform followed by Fourier reconstruction. The work these bands do is to pinpoint the gamma-band components that change most during epilepsy; the phase-space images of these reconstructed signals then serve as input features for linear-regression and SVM classifiers. The central identity is that these two periodic modulations, when viewed in phase space, show bi-stability and bifurcation-like differences between normal and ictal recordings.
What would settle it
Run the same Welch-transform and phase-space reconstruction pipeline on an independent, multi-subject EEG dataset, selecting each subject's own dominant gamma peaks instead of the fixed 50 and 76 Hz values; if the reported 94-95% separation does not persist, the fixed-frequency assumption fails. A second concrete test is to shift the chosen frequencies by ±5 Hz and check whether the phase-space features and classification accuracy change sharply, which would indicate spectral leakage rather than a stable biological signature.
Extended reading notes
Core claim
The central claim is that the periodic modulations in the gamma band, specifically the dominant components near 50 Hz and 76 Hz, undergo measurable structural changes during epileptic seizures. The authors use Welch's transform to reconstruct the signal from these narrow frequency windows and embed them in phase space. They observe oscillatory motions that display bi-stability and bifurcations, with distinct differences between normal and seizure conditions. These differences, quantified through phase-space features and fed to linear regression and SVM models, yield a classification accuracy of about 94-95% between healthy and ictal states using the publicly available University of Bonn EEG
Load-bearing premise
The paper assumes that the two dominant frequencies around 50 Hz and 76 Hz are stable and physiologically relevant markers for epilepsy, and that Fourier-based spectral methods (Welch's transform) adequately capture the non-stationary dynamics of EEG.
Editorial extensions
If this is right
- If the result holds, the phase-space features derived from the 50 Hz and 76 Hz gamma components can act as a quantitative EEG biomarker for epilepsy, usable without analyzing the full chaotic dynamics.
- The claimed 94-95% classification accuracy indicates that a small set of frequency-resolved features, rather than the full EEG, is sufficient for automatic seizure detection.
- The method suggests that partial spectral reconstruction can isolate relevant pathological dynamics from background noise, a principle transferable to other neurological conditions.
- The use of both linear regression and SVM on the same phase-space features provides a check that the separation is not an artifact of a single classifier family.
- The approach implies that the collective gamma-band synchronization is destabilized during seizures in a way that leaves a measurable footprint in the reconstructed phase space.
Reading between the lines
- The fixed 50 Hz and 76 Hz frequencies may not be universal; a testable extension is to detect each subject's own dominant gamma peaks and see whether classification accuracy improves or degrades, which would reveal whether the fixed-choice assumption is load-bearing.
- A clinically useful biomarker would need to distinguish interictal (between-seizure) from ictal states, not just normal from seizure; the paper's method could be directly tested on that three-way discrimination.
- The provided manuscript body is a different paper (on federated learning), so the abstract's EEG claims could not be checked against methods or results in the text; the summary above follows the abstract.
- The phase-space features are plausibly sensitive to the choice of embedding parameters; a sensitivity analysis over embedding delay and dimension would clarify how robust the 94-95% figure is.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript, as submitted, consists of an abstract claiming a nonlinear-dynamics/EEG study of high-frequency gamma-band activity (40–100 Hz) in normal versus epileptic conditions, with classification accuracy of 94–95% on the University of Bonn EEG dataset using Welch-transform-based spectral analysis, phase-space reconstruction, and linear-regression/SVM classifiers. The full text supplied, however, is an entirely unrelated federated-learning paper titled 'Deploying Models to Non-participating Clients in Federated Learning without Fine-tuning: A Hypernetwork-based Approach' (HyperFedZero). There is no EEG data, no Welch transform, no phase-space reconstruction, no SVM analysis, and no classification protocol anywhere in the body. The abstract's methodological claims are therefore unsupported by any accessible methods, results, or validation details in the manuscript.
Significance. If the abstract's claims were substantiated, the identification of specific gamma-band frequencies (~50 Hz, ~76 Hz) with reproducible phase-space differences between healthy and epileptic EEG could be of interest to the nonlinear-dynamics and clinical neurophysiology communities. However, significance cannot be assessed because the manuscript body contains none of the described analysis. No machine-checked proofs, reproducible code, parameter-free derivations, or falsifiable predictions are present. The paper's contribution as submitted is not evaluable.
major comments (4)
- [Full Text] The submitted full text is a different paper (HyperFedZero, federated learning) and contains none of the methods, results, or validation described in the abstract. All load-bearing details are absent: the preprocessing/artifact rejection, the Welch transform parameters, the procedure for selecting the 50 Hz and 76 Hz peaks, the phase-space embedding (e.g., delay, embedding dimension, reconstruction method), the feature extraction for the classifiers, the training/validation split (within-subject vs. cross-subject), and the error bars or statistical significance of the 94–95% accuracy. The central claim is not verifiable from the manuscript.
- [Abstract] The abstract states that dominant frequencies around 50 Hz and 76 Hz are 'isolated' and that classification accuracy is ~94–95%. No information is given about whether the frequency selection and classifier evaluation were performed on the same dataset or on independent training/test partitions. If the same recordings were used to identify the peaks and to train/evaluate the classifier without a held-out set, the reported accuracy could be inflated by circular feature selection. This concern cannot be checked because the methods are missing.
- [Abstract] The abstract invokes 'bi-stability and bi-furcations' in phase-space images but does not define these operationally: what observable is bistable, over what time scale, how bistability is quantified, and how it relates to the gamma-band frequencies. Without operational definitions, the claimed structural differences between normal and ictal conditions are not testable.
- [Abstract] The classification claim is based on a publicly available EEG dataset from the University of Bonn. It is well documented that the healthy and ictal subsets of that dataset differ not only in pathology but also in recording conditions (e.g., electrode type, reference, acquisition hardware) and signal amplitude. Since no normalization, artifact handling, or control for these confounds is described, the reported 94–95% accuracy could reflect trivial amplitude or recording differences rather than the proposed gamma-band mechanism. This is a correctness risk that requires a concrete response (e.g., subject-level cross-validation, amplitude normalization, or feature ablation).
minor comments (4)
- [Abstract] Typographical issue: 'bi-furcations' should be 'bifurcations'.
- [Abstract] The term 'Fourier reconstruction' is used without specifying whether the inverse transform is applied to a band-limited subset of the Welch periodogram; this should be clarified if the intended method is band-pass filtering in the frequency domain.
- [Abstract] The manuscript should cite the specific Bonn dataset paper (Andrzejak et al., 2001) and describe which sets (A–E) were used for healthy versus ictal comparisons, including the number of subjects and segments.
- [Abstract] The abstract mentions 'linear regression and SVM-based machine learning models' but provides no details on the feature vector, kernel type, hyperparameter selection, or cross-validation scheme. Even if the full text were present, these are necessary for reproducibility.
Circularity Check
No circularity established; supplied full text is an unrelated federated-learning paper, so the claimed EEG derivation cannot be checked
full rationale
The abstract claims a derivation chain: Welch-transform identification of ~50/76 Hz gamma peaks, phase-space reconstruction, and 94–95% healthy-vs-ictal classification via linear regression/SVM on the Bonn EEG dataset. The supplied Full Text, however, is an entirely different paper ('Deploying Models to Non-participating Clients in Federated Learning without Fine-tuning: A Hypernetwork-based Approach'), containing no EEG data, no Welch transforms, no phase-space analysis, and no SVM/linear-regression experiments. Per the reviewing rule, I take that text at face value as the manuscript rather than dismissing it as an artifact. Because the claimed analysis is wholly absent, there is no derivation chain to walk and no equation or fitted parameter that can be shown to reduce to an input. Every circularity pattern in the priority list requires quoting a specific reduction (e.g., Eq. X = Eq. Y by construction, or a fitted parameter renamed as a prediction); none can be exhibited here. The absence of methods is a severe verification/integrity problem—it leaves the classification claim uncheckable and open to potential selection bias or confounding—but it is not itself a demonstrated circularity. Therefore the honest finding is 'no significant circularity established,' score 0.
Assumptions & free parameters
free parameters (1)
- Dominant frequencies f1, f2 =
approximately 50 Hz and 76 Hz
assumptions (3)
- domain assumption Fourier-based spectral analysis (Welch's transform) is valid for EEG signals
- domain assumption Phase-space reconstruction from a few frequency components preserves the relevant dynamics
- domain assumption The Bonn EEG dataset is representative of normal and epileptic populations
Cite this review
Pith. "Pith review of Discerning and quantifying high frequency activities in EEG under normal and epileptic conditions." pith.science (2026). https://pith.science/paper/7GOAL7MD
@misc{pith2026250812670,
author = {Pith},
title = {Pith review of: Discerning and quantifying high frequency activities in EEG under normal and epileptic conditions},
year = {2026},
howpublished = {\url{https://pith.science/paper/7GOAL7MD}},
note = {Machine review of arXiv:2508.12670}
}
read the original abstract
We investigate the nature of the modifications in the temporal dynamics manifested in the high-frequency EEG spectra of the normal human brain in comparison to the diseased brain undergoing epilepsy. For this purpose, the Fourier reconstruction is efficaciously made use of after Welch's transform, which helped identify the relevant frequency components undergoing significant changes in the case of epilepsy. The temporal dynamics involved in the EEG signals and their associated variations showed a well-structured periodic pattern characterized by bi-stability and significant quantifiable structural changes during epileptic episodes. In particular, we demonstrate and quantify the precise differences in the high-frequency gamma band (40-100 Hz) present in EEG recordings from neurologically normal participants compared to those with epilepsy. The periodic modulations at two dominant frequencies around 50 Hz and 76 Hz in power spectral density are isolated from high frequency noise through the use of Welch's transform, pinpointing their collective behaviors through a phase-space approach. The reconstructed signals from these restricted frequency domains revealed oscillatory motions showing a bi-stability and bi-furcations with distinct differences between normal and seizure conditions. These differences in the phase space images, when analyzed through linear regression and SVM-based machine learning models, support a classification accuracy of around 94-95% between healthy and ictal states using a publicly available EEG dataset from the University of Bonn (Germany). The partial reconstruction of the dynamics as compared to the earlier studies of the full phase space accurately pinpointed the destabilization of the collective high-frequency synchronous behavior and their precise differences in the normal and diseased conditions, avoiding the other chaotic components of the EEG signals.
Reference graph
Works this paper leans on
-
[4]
Granada. Nguyen, A. T.; Torr, P.; and Lim, S. N. 2022. Fedsr: A sim- ple and effective domain generalization method for feder- ated learning. Advances in Neural Information Processing Systems, 35: 38831–38843. Nguyen, D. C.; Pham, Q.-V .; Pathirana, P. N.; Ding, M.; Seneviratne, A.; Lin, Z.; Dobre, O.; and Hwang, W.-J. 2022. Federated learning for smart h...
work page 2022
-
[6]
Sch¨urholt, K.; Knyazev, B.; Gir´o-i Nieto, X.; and Borth, D
Robust and communication-efficient federated learn- ing from non-iid data.IEEE transactions on neural networks and learning systems, 31(9): 3400–3413. Sch¨urholt, K.; Knyazev, B.; Gir´o-i Nieto, X.; and Borth, D
-
[8]
arXiv preprint arXiv:2107.10663
Fed-ensemble: Improving generalization through model ensembling in federated learning. arXiv preprint arXiv:2107.10663. Shin, Y .; Lee, K.; Lee, S.; Choi, Y . R.; Kim, H.-S.; and Ko, J
-
[10]
Example-based Hypernetworks for Out-of-Distribution Generalization
Imbalance trouble: Revisiting neural-collapse geom- etry. Advances in Neural Information Processing Systems , 35: 27225–27238. Vamanan, B.; V oskuilen, G.; and Vijaykumar, T. 2010. Ef- fiCuts: Optimizing packet classification for memory and throughput. ACM SIGCOMM Computer Communication Review, 40(4): 207–218. Van der Maaten, L.; and Hinton, G. 2008. Visu...
work page Pith review arXiv 2010
-
[11]
Hypernetwork-based Personalized Federated Learning for Multi-Institutional CT Imaging
Hypernetwork-based personalized federated learn- ing for multi-institutional CT imaging. arXiv preprint arXiv:2206.03709. Ye, M.; Fang, X.; Du, B.; Yuen, P. C.; and Tao, D. 2023. Het- erogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3): 1–44. Zenke, F.; Poole, B.; and Ganguli, S. 2017. Continual learn- ing...
work page Pith review arXiv 2023
-
[13]
In Proceedings of the IEEE/CVF International Conference on Computer Vision, 5031–5040
Communication-efficient Federated Learning with Single-Step Synthetic Features Compressor for Faster Con- vergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 5031–5040. Zhou, Y .; Shi, M.; Tian, Y .; Li, Y .; Ye, Q.; and Lv, J. 2024a. Federated CINN Clustering for Accurate Clustered Feder- ated Learning. In ICASSP 2024-20...
work page 2024
-
[15]
Recently, FedIG (Seunghan et al
calibrates the aggregation weights in FL to achieve a tighter generalization bound. Recently, FedIG (Seunghan et al. 2024) introduced client-agnostic learning for zero-shot adaptation, but it relies on multi-domain training data, which is often unavailable or unlabeled in real-world FL scenarios. Despite the promising performance, existing literature rare...
work page 2024
-
[456]
PMLR. Karimireddy, S. P.; Kale, S.; Mohri, M.; Reddi, S.; Stich, S.; and Suresh, A. T. 2020. Scaffold: Stochastic controlled av- eraging for federated learning. In International conference on machine learning, 5132–5143. PMLR. Krizhevsky, A.; Hinton, G.; et al. 2009. Learning multiple layers of features from tiny images. Le, Y .; and Yang, X. S. 2015. Tin...
work page 2020
Show all 15 references
-
[2012]
Advances in neural information processing systems, 25
Large scale distributed deep networks. Advances in neural information processing systems, 25. Deng, L. 2012. The mnist database of handwritten digit im- ages for machine learning research. IEEE Signal Processing Magazine, 29(6): 141–142. Do, M. T.; Yoon, S.-e.; Hooi, B.; and S...
2012 arXiv
-
[2019]
arXiv preprint arXiv:1907.02189
On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189. Liu, Q.; Chen, C.; Qin, J.; Dou, Q.; and Heng, P.-A. 2021. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of...
1907 arXiv
-
[2021]
In International conference on machine learning, 2089–2099
Exploiting shared representations for personalized federated learning. In International conference on machine learning, 2089–2099. PMLR. Dean, J.; Corrado, G.; Monga, R.; Chen, K.; Devin, M.; Mao, M.; Ranzato, M.; Senior, A.; Tucker, P.; Yang, K.; et al
-
[2022]
Advances in Neural Information Processing Systems, 35: 27906–27920
Hyper-representations as generative models: Sam- pling unseen neural network weights. Advances in Neural Information Processing Systems, 35: 27906–27920. Scott, J.; Zakerinia, H.; and Lampert, C. H. 2023. Pefll: Personalized federated learning by learning to learn. arXiv prepr...
2023 arXiv
-
[2023]
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3954–3963
Federated domain generalization with generalization adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3954–3963. Zhao, D.; Kobayashi, S.; Sacramento, J.; and von Oswald, J. 2020. Meta-learning via hypernetworks. In 4th Work- shop...
2020
-
[2024]
In Proceed- ings of the 22nd ACM Conference on Embedded Networked Sensor Systems, 112–125
Effective Heterogeneous Federated Learning via Effi- cient Hypernetwork-based Weight Generation. In Proceed- ings of the 22nd ACM Conference on Embedded Networked Sensor Systems, 112–125. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; and Salakhutdinov, R. 2014. Dr...
2014
-
[2025]
arXiv preprint arXiv:2502.03092
E-3SFC: Communication-Efficient Federated Learn- ing with Double-way Features Synthesizing. arXiv preprint arXiv:2502.03092. Zhou, Y .; Ye, Q.; and Lv, J. 2021. Communication-efficient federated learning with compensated overlap-fedavg. IEEE Transactions on Parallel and Distri...
2021 arXiv
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.