Pith. sign in

REVIEW 4 major objections 7 minor 28 references

MSEMG: Surface Electromyography Denoising with a Mamba-based Efficient Network

T0 review · 4 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read MSEMG, a Mamba-based convolutional network, is claimed to remove ECG interference from surface electromyography better than prior neural and classical methods while using fewer parameters.

desk verdict A useful first Mamba-for-sEMG denoising result with a strong held-out evaluation, but the 'outperforms all others' claim is undercut by a missing transformer baseline and no error bars. read the letter →

arxiv 2411.18902 v2 pith:GAADWSYL submitted 2024-11-28 eess.SP cs.LG

Yu-Tung Liu , Kuan-Chen Wang , Rong Chao , Sabato Marco Siniscalchi , Ping-Cheng Yeh , Yu Tsao This is my paper · ORCID
classification eess.SPcs.LG
keywords surfaceelectromyographyECGartifactremovalMambastatespacemodelsignaldenoisingconvolutionalneuralnetworklightweightsEMGenhancement
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

This paper proposes MSEMG, a lightweight neural network that combines the Mamba state space model with convolutional layers to remove electrocardiogram (ECG) interference from surface electromyography (sEMG) recordings. The authors argue that, because sEMG and ECG overlap in frequency, conventional filters and template subtraction lose signal content, while prior neural methods are either distorted or computationally expensive. On simulated mixtures built from NINAPro DB2 sEMG and MIT-BIH ECG, MSEMG reports the highest SNR improvement (20.317 dB) and lowest reconstruction and feature errors among tested methods, using 279,937 parameters versus 1,233,857 for the diffusion baseline. If the simulation transfers to real trunk recordings, MSEMG would offer a practical, real-time-capable denoiser for prosthetics, rehabilitation, and wearable sensing.

What carries the argument

The central machinery is a selective state space model (Mamba) whose learnable time step Δ adaptively weights the hidden state against the current input, giving linear-time sequence modeling. MSEMG places a Mamba block between two Half Normalized Filter (HNF) convolutional blocks, so local features are extracted at multiple resolutions and long-range dependencies are modeled in the latent space.

What would settle it

If real trunk sEMG recordings with naturally occurring ECG contamination are processed with MSEMG and the residual ECG peaks, estimated by ECG-triggered averaging, are no smaller than the residuals left by high-pass filtering or template subtraction, then the claimed practical superiority over classical methods would fail.

Watch

Extended reading notes

Core claim

The central claim is that a selective state space Mamba block, inserted between two multi-resolution convolutional feature extraction blocks (HNF), can jointly model local and long-range temporal dependencies in sEMG, making it a stronger and cheaper ECG artifact remover than the fully convolutional FCN and the score-based diffusion SDEMG. The paper reports consistent superiority across all metrics and across a separate testing setup with different subjects, channels, movements, and SNR levels from the training setup. It also reports a simulated trunk-contamination scenario using biceps brachii data at -10 dB where MSEMG retains the best SNRimp and lowest RMSE values.

Load-bearing premise

The evaluation assumes that superimposing ECG segments from a public sinus-rhythm database onto clean sEMG at prescribed signal-to-noise ratios produces contaminated signals that match how ECG interference actually appears on trunk muscles.

Editorial extensions

If this is right

  • MSEMG can replace diffusion-based sEMG denoising in resource-constrained settings because it uses about one-fifth the parameters with no repeated sampling steps.
  • MSEMG produces lower errors in ARV and mean-frequency features, suggesting cleaner inputs to downstream myoelectric control and gesture recognition.
  • The per-SNR results indicate the gain holds from -14 to 0 dB input SNR, including very heavy ECG contamination.
  • The architecture can be applied to any single-channel sEMG enhancement task, not only trunk muscles, because it operates on raw waveform segments.

Reading between the lines

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

  • If the simulated mixtures are representative, MSEMG's long-range modeling could help in other bio-signal denoising tasks where artifact and signal overlap spectrally, such as ECG denoising of fetal or wearable recordings.
  • The reported parameter-efficiency gain does not yet measure inference latency or energy; a direct runtime comparison with SDEMG would show whether the smaller model translates into faster real-time operation.
  • Testing on real trunk sEMG with naturally occurring ECG artifact would settle whether the supervised simulation setup, rather than the architecture itself, drives the reported gains.
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.

Referee Report

4 major / 7 minor

Summary. The paper proposes MSEMG, a hybrid Mamba state-space-model and convolutional network for removing ECG contamination from surface electromyography (sEMG) signals. The method is evaluated on NINAPro DB2 sEMG data artificially contaminated with ECG segments from the MIT-BIH Normal Sinus Rhythm Database, and is compared against high-pass filtering, template subtraction, a fully convolutional network (FCN), and a score-based diffusion model (SDEMG). The authors report that MSEMG achieves the best SNR improvement (20.317 dB), lowest RMSE (8.603e-3), and lowest RMSE values for ARV and MF features, while using 279,937 parameters, about one-fifth the size of SDEMG. The paper claims that MSEMG consistently outperforms existing methods and is the first Mamba-based approach for sEMG processing.

Significance. If the reported results hold, the paper would make a useful contribution by showing that a selective state-space model can be combined with a lightweight convolutional front-end for efficient sEMG denoising. Strengths include the use of public datasets (NINAPro DB2 and MIT-BIH NSRD), a held-out test split across subjects, channels, movements, and SNR levels, and a comparison against both classical and neural baselines. The work also addresses an application-relevant problem, since ECG contamination is common in trunk and proximal-limb sEMG recordings. However, the central empirical claim is currently supported only by point estimates without error bars, and the most relevant transformer-based sEMG baseline is cited but not evaluated. The efficiency claim is also only established relative to one baseline. These issues mean the core contribution is plausible but not yet fully substantiated.

major comments (4)
  1. [Section IV-C, Table I and Table II] The comparison set omits TrustEMG-Net [14], a representation-masking transformer U-Net for sEMG enhancement published by overlapping authors and cited in the manuscript. Since the paper claims that MSEMG "consistently outperforms all other methods" and that SDEMG was the previous state of the art, the absence of the most relevant transformer baseline leaves the central superiority claim unsupported. Please add TrustEMG-Net results under the identical evaluation protocol, or revise the claims to be explicitly limited to the compared baselines.
  2. [Section IV-C, Table I and Fig. 3] All reported results are single point estimates with no error bars, confidence intervals, or significance tests, and no repeated-seed training is described. The 1.85 dB SNRimp advantage over SDEMG could fall within run-to-run variability, so the claim of consistent superiority is not statistically quantified. Please report means and standard deviations over multiple training runs, or at least provide significance tests on the held-out test metrics.
  3. [Table II and Abstract] The efficiency claim that MSEMG uses "fewer parameters than existing state-of-the-art sEMG denoising methods" is only true relative to SDEMG; Table II shows that FCN uses 137,801 parameters, considerably fewer than MSEMG's 279,937. Moreover, no inference time, FLOPs, or latency measurements are reported, so the practical efficiency advantage is not fully established. Please clarify the efficiency comparison and include computational-cost metrics beyond parameter count.
  4. [Section IV-A and Section IV-C, Fig. 4] The evaluation is conducted entirely on simulated contamination: NINAPro DB2 upper-limb sEMG signals are superimposed with ECG segments from MIT-BIH NSRD. The paper calls the scenario in Fig. 4 a simulation of trunk sEMG with ECG contamination, but the NINAPro DB2 recordings are from the upper limb (e.g., biceps brachii for Channel 11), and electrode placement, tissue coupling, and movement artifacts in real trunk recordings may differ substantially. Please add validation on a real ECG-contaminated sEMG dataset or clearly restrict the claims to the simulated setting.
minor comments (7)
  1. [Abstract and Introduction] "closed to the heart" should be "close to the heart" in the abstract and introduction.
  2. [Section III-B] The HNF block is referenced to DeepFilter [23] but its detailed configuration (kernel sizes, number of channels, normalization specifics) is not described, which makes the architecture difficult to reproduce.
  3. [Section IV-A] The preprocessing description does not specify the training segment length, test segment length, or any overlap between segments; please provide these details.
  4. [Section IV-A and Section IV-B] No information is given about the loss function, optimizer, learning rate, number of epochs, batch size, or hardware used for training; these are essential for reproducibility of an empirical methods paper.
  5. [Section III-A, Eqs. (5)-(8)] The notation "CA^kB" in Eq. (7) is formatted unclearly, and the sentence "we could only set k = 0 in Mamba" is confusing; please rewrite for clarity.
  6. [Table I and Fig. 4] The table header "RMSEM F(Hz)" is ambiguous; please define the units and abbreviation clearly in the caption or text. Also, axis labels in Fig. 4 such as "Noisy sEMGSNR (dB)" are missing a space and are hard to read.
  7. [Section III] The phrase "a key proportion of this novel integration" should be "a key component".

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claims are empirical test-set comparisons on public data, and self-citations serve only as baselines or related work.

full rationale

The paper's derivation chain is an empirical architecture evaluation. MSEMG is trained on clean NINAPro DB2 sEMG with ECG artifacts from the MIT-BIH NSRD superimposed at prescribed SNR levels, and its SNRimp, RMSE, ARV-RMSE, and MF-RMSE values are computed on held-out subjects, channels, movements, and SNR levels. These reported quantities are test-set metrics, not fitted parameters or outputs defined in terms of the model's own inputs. The self-citations to prior work by overlapping authors (FCN [12], SDEMG [13], TrustEMG-Net [14]) are used as comparison baselines or related-work context; none is invoked as a uniqueness theorem, a forced ansatz, or a definitional constraint on the result. The Mamba and HNF components are attributed to external sources ([15], [23]), and the evaluation follows the publicly available simulation protocol of previous work without building the target result into the loss or metric. Omission of TrustEMG-Net as a compared baseline and absence of repeated-seed statistics weaken the strength of the 'consistently outperforms' claim, but these are completeness/robustness concerns, not circularity. Therefore no circular step can be quoted or reduced to its own input.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on four classes of assumptions: the simulation adequately represents real ECG contamination, the public databases are representative, the Mamba implementation is correct, and the chosen metrics reflect useful denoising. The main free parameters are unreported architecture and training choices that cannot be audited from the paper. No new physical or theoretical entities are introduced.

free parameters (5)
  • Mamba hidden state dimension H
    Architecture capacity parameter controlling the selective state space. Its value is not reported, so its effect on the claimed performance cannot be audited.
  • HNF filter bank configuration
    The Half Normalized Filters block is taken from DeepFilter [23], but the exact kernel sizes, number of filters, and channel widths are not specified in this paper, and they determine the local feature extraction the results depend on.
  • Training SNR grid = -15 to -5 dB in 2 dB steps for training and validation
    The contamination levels and range are chosen by hand in Section IV-A; performance at other SNR levels could differ if this grid changed.
  • Loss function and training hyperparameters
    The optimizer, learning rate, loss, epochs, and batch size are not reported. These choices materially affect the performance of any learned denoiser.
  • Test segment length = 10 s
    Segmentation length is chosen by hand and can affect the temporal context available to the Mamba block and the resulting error metrics.
assumptions (5)
  • domain assumption Linear superposition of clean sEMG and ECG at controlled SNR produces realistic contaminated sEMG.
    Used throughout Section IV-A to generate every training, validation, and test example. If real ECG contamination is nonlinear or depends on electrode placement, the benchmark does not represent the target scenario.
  • domain assumption NINAPro DB2 sEMG and MIT-BIH NSRD ECG are representative of the target clinical population.
    All results come from these two public databases; no real trunk-sEMG recordings with ECG contamination are used (Section IV-A).
  • standard math The Mamba selective scan implementation follows Gu and Dao [15] and the zero-order hold discretization is valid.
    Equations (1)-(8) restate Mamba's continuous-time state space model. The correctness of the learned denoiser depends on the selective scan operating as described.
  • domain assumption SNRimp, RMSE, ARV, and MF capture clinically meaningful denoising quality.
    These metrics are standard in prior work [10], [12], [13], but no perceptual or downstream task validation is performed. A model that scores well on these metrics may not improve prosthesis control or clinical readings.
  • domain assumption Upper-limb NINAPro sEMG contaminated with ECG is an appropriate proxy for trunk sEMG contamination.
    The motivating scenario is trunk sEMG, but training data are upper-limb recordings. The simulated trunk scenario in Fig. 4 uses biceps brachii data, which is not a trunk muscle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MSEMG: Surface Electromyography Denoising with a Mamba-based Efficient Network." pith.science (2026). https://pith.science/paper/GAADWSYL

@misc{pith2026241118902,
  author       = {Pith},
  title        = {Pith review of: MSEMG: Surface Electromyography Denoising with a Mamba-based Efficient Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GAADWSYL}},
  note         = {Machine review of arXiv:2411.18902}
}
read the original abstract

Surface electromyography (sEMG) recordings can be contaminated by electrocardiogram (ECG) signals when the monitored muscle is closed to the heart. Traditional signal processing-based approaches, such as high-pass filtering and template subtraction, have been used to remove ECG interference but are often limited in their effectiveness. Recently, neural network-based methods have shown greater promise for sEMG denoising, but they still struggle to balance both efficiency and effectiveness. In this study, we introduce MSEMG, a novel system that integrates the Mamba state space model with a convolutional neural network to serve as a lightweight sEMG denoising model. We evaluated MSEMG using sEMG data from the Non-Invasive Adaptive Prosthetics database and ECG signals from the MIT-BIH Normal Sinus Rhythm Database. The results show that MSEMG outperforms existing methods, generating higher-quality sEMG signals using fewer parameters.

Figures

Figures reproduced from arXiv: 2411.18902 by the authors.

Figure 1
Figure 1. The Selective State Space Model. III. THE PROPOSED METHOD In this section, we introduce the Mamba model and the se￾lection mechanism as a key proportion of this novel integration of CNN and the state space model. The implementation detail of the NN in this work will be further elaborated. A. Selective state space model The Mamba SSM ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The denoising framework and the model architecture of MSEMG. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Performance under the scenario simulating trunk sEMG with ECG [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [14]

    Trustemg- net: Using representation-masking transformer with u-net for surface electromyography enhancement,

    K.-C. Wang, K.-C. Liu, P.-C. Yeh, S.-Y . Peng, and Y . Tsao, “Trustemg- net: Using representation-masking transformer with u-net for surface electromyography enhancement,” IEEE Journal of Biomedical and Health Informatics, 2024

    work page 2024

  2. [1]

    A novel interpretation of sample entropy in surface electromyographic examination of complex neuromuscular alternations in subacute and chronic stroke,

    X. Tang, X. Zhang, X. Gao, et al., “A novel interpretation of sample entropy in surface electromyographic examination of complex neuromuscular alternations in subacute and chronic stroke,” IEEE Transactions on Neural Systems and Rehabilitation Engineering , vol. 26, no. 9, pp. 1878–1888, 2018

    work page 2018

  3. [2]

    Surveying the interest of individuals with upper limb loss in novel prosthetic control techniques,

    S. M. Engdahl, B. P. Christie, et al., “Surveying the interest of individuals with upper limb loss in novel prosthetic control techniques,” Journal of Neuroengineering and Rehabilitation , vol. 12, no. 1, pp. 1– 11, 2015

    work page 2015

  4. [3]

    Wearable physiological sensors reflect mental stress state in office-like situations,

    J. Wijsman, B. Grundlehner, H. Liu, et al., “Wearable physiological sensors reflect mental stress state in office-like situations,” in Proc. ACII, 2013

    work page 2013

  5. [4]

    Clinical utility of measuring inspiratory neural drive during cardiopulmonary exercise testing (cpet),

    N. J. Domnik, E. S. Walsted, and D. Langer, “Clinical utility of measuring inspiratory neural drive during cardiopulmonary exercise testing (cpet),” Frontiers in Medicine, vol. 7, pp. 483, 2020

    work page 2020

  6. [5]

    Assessment of respiratory effort during sleep: esophageal pressure versus noninvasive monitoring techniques,

    N. L. Vandenbussche, S. Overeem, J. P. van Dijk, P. J. Simons, and D. A. Pevernagie, “Assessment of respiratory effort during sleep: esophageal pressure versus noninvasive monitoring techniques,” Sleep Medicine Reviews, vol. 24, pp. 28–36, 2015

    work page 2015

  7. [6]

    Hand and wrist movement control of myoelectric prosthesis based on synergy,

    J. Ma, N. V . Thakor, and F. Matsuno, “Hand and wrist movement control of myoelectric prosthesis based on synergy,” IEEE Transactions on Human-Machine Systems , vol. 45, no. 1, pp. 74–83, 2014

    work page 2014

  8. [7]

    A transferable adaptive domain adversarial neural network for virtual reality augmented emg- based gesture recognition,

    U. Cote-Allard, G. Gagnon-Turcotte, A. Phinyomark, K. Glette, E. Scheme, F. Laviolette, and B. Gosselin, “A transferable adaptive domain adversarial neural network for virtual reality augmented emg- based gesture recognition,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 29, pp. 546–555, 2021

    work page 2021

Show all 28 references
  1. [8]

    Interactive virtual ankle movement controlled by wrist semg improves motor imagery: An exploratory study,

    Y . Xiao, H. Bai, Y . Gao, B. Hu, J. Zheng, X. Cai, J. Rao, X. Li, and A. Hao, “Interactive virtual ankle movement controlled by wrist semg improves motor imagery: An exploratory study,” IEEE Transactions on Visualization and Computer Graphics , 2023

    2023

  2. [9]

    D. A. Winter, Biomechanics and motor control of human movement , John Wiley & Sons, 2009

    2009

  3. [10]

    Comparative review of the algorithms for removal of electrocardiographic interference from trunk electromyography,

    L. Xu, E. Peri, R. Vullings, et al., “Comparative review of the algorithms for removal of electrocardiographic interference from trunk electromyography,” Sensors, vol. 20, no. 17, pp. 4890, 2020

    2020

  4. [11]

    Elimination of electrocardiogram contamination from electromyogram signals: An evaluation of currently used removal techniques,

    J. D. Drake and J. P. Callaghan, “Elimination of electrocardiogram contamination from electromyogram signals: An evaluation of currently used removal techniques,” Journal of Electromyography and Kinesiol- ogy, vol. 16, no. 2, pp. 175–187, 2006

    2006

  5. [12]

    Ecg artifact removal from single-channel surface emg using fully convolutional networks,

    K.-C. Wang, K.-C. Liu, S.-Y . Peng, and Y . Tsao, “Ecg artifact removal from single-channel surface emg using fully convolutional networks,” in Proc. ICASSP, 2023

    2023

  6. [13]

    Sdemg: Score-based diffusion model for surface electromyographic signal de- noising,

    Y .-T. Liu, K.-C. Wang, K.-C. Liu, S.-Y . Peng, and Y . Tsao, “Sdemg: Score-based diffusion model for surface electromyographic signal de- noising,” in Proc. ICASSP, 2024

    2024

  7. [15]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  8. [16]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, vol. 30, pp. 5998–6008, 2017

    2017

  9. [17]

    Jamba: A hybrid transformer-mamba language model,

    O. Lieber, B. Lenz, H. Bata, G. Cohen, J. Osin, I. Dalmedigos, E. Safahi, S. Meirom, Y . Belinkov, S. Shalev-Shwartz, et al., “Jamba: A hybrid transformer-mamba language model,” arXiv preprint arXiv:2403.19887, 2024

    2024 arXiv

  10. [18]

    An investigation of incorporating mamba for speech enhancement,

    R. Chao, W.-H. Cheng, M. La Quatra, S. M. Siniscalchi, C.-H. H. Yang, S.-W. Fu, and Y . Tsao, “An investigation of incorporating mamba for speech enhancement,” arXiv preprint arXiv:2405.06573 , 2024

    2024

  11. [19]

    A template subtraction method for reducing electrocardiographic artifacts in emg signals of low intensity,

    J. D. C. Junior, J. M. de Seixas, et al., “A template subtraction method for reducing electrocardiographic artifacts in emg signals of low intensity,” Biomedical Signal Processing and Control , vol. 47, pp. 380–386, 2019

    2019

  12. [20]

    Ssamba: Self- supervised audio representation learning with mamba state space model,

    S. Shams, S. S. Dindar, X. Jiang, and N. Mesgarani, “Ssamba: Self- supervised audio representation learning with mamba state space model,” arXiv preprint arXiv:2405.11831 , 2024

    2024 arXiv

  13. [21]

    ChiMamba: Predicting Chromatin Interactions Based on Mamba,

    J. Zhang, C. Song, T. Cui, C. Li, and J. Ma, “ChiMamba: Predicting Chromatin Interactions Based on Mamba,” in Proc. ICIC, 2024

    2024

  14. [22]

    MSAMamba: Adapting Subquadratic Sequence Models to Long-Context DNA MSA Analysis,

    V . Thoutam and D. Ellsworth, “MSAMamba: Adapting Subquadratic Sequence Models to Long-Context DNA MSA Analysis,” in Proc. ICIST, 2024

    2024

  15. [23]

    Deepfilter: An ecg baseline wander removal filter using deep learning techniques,

    F. P. Romero, D. C. Pi ˜nol, and C. R. V ´azquez-Seisdedos, “Deepfilter: An ecg baseline wander removal filter using deep learning techniques,” Biomedical Signal Processing and Control , vol. 70, pp. 102992, 2021

    2021

  16. [24]

    Electromyography data for non-invasive naturally-controlled robotic hand prostheses,

    M. Atzori, A. Gijsberts, C. Castellini, et al., “Electromyography data for non-invasive naturally-controlled robotic hand prostheses,” Scientific data, vol. 1, no. 1, pp. 1–13, 2014

    2014

  17. [25]

    Deep learning for surface electromyography artifact contamination type detection,

    J. Machado, A. Machado, and A. Balbinot, “Deep learning for surface electromyography artifact contamination type detection,” Biomedical Signal Processing and Control , vol. 68, pp. 102752, 2021

    2021

  18. [26]

    Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals,

    A. L. Goldberger, L. A. Amaral, L. Glass, et al., “Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. e215–e220, 2000

    2000

  19. [27]

    Noise reduction in ecg signals using fully convolutional denoising autoencoders,

    H.-T. Chiang, Y .-Y . Hsieh, S.-W. Fu, et al., “Noise reduction in ecg signals using fully convolutional denoising autoencoders,” IEEE Access, vol. 7, pp. 60806–60813, 2019

    2019

  20. [28]

    Eliminating cardiac contamination from myoelectric control signals developed by targeted muscle reinnervation,

    P. Zhou and T. A. Kuiken, “Eliminating cardiac contamination from myoelectric control signals developed by targeted muscle reinnervation,” Physiological Measurement, vol. 27, no. 12, pp. 1311, 2006

    2006

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.