{"id":"7597661c-b5c4-4dc6-b187-2d0a08ea2da8","arxiv_id":"2411.18902","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MSEMG, a Mamba plus CNN denoising network, outperforms prior sEMG denoising methods on simulated ECG-contaminated signals while using roughly one-fifth of SDEMG's parameters.","lead":"This paper introduces MSEMG, a small neural network that combines a Mamba state-space model with convolutional layers to remove electrocardiogram interference from surface electromyography recordings. On simulated noisy signals, it reports better denoising quality and fewer parameters than the previous best diffusion-based model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing TrustEMG-Net baseline and no repeated-seed statistics leave the 'consistently outperforms all other methods' claim unsubstantiated.","rationale":"The reader's stated weakest assumption is that the simulated ECG contamination transfers to real trunk sEMG recordings; that is a legitimate external-validity concern. My stress-test pass identifies a more immediate, checkable gap: the paper's own reference list contains TrustEMG-Net, a same-group transformer U-Net for sEMG enhancement, yet no experimental comparison against it is reported. Because the central claim is explicitly about outperforming existing methods, an omitted strong baseline directly threatens the claim, regardless of whether the simulation is realistic. The absence of error bars worsens this: the reported improvements over SDEMG could be within run-to-run variation, and no code is provided to check the exact training setup. I do not see an internal mathematical error in the Mamba formulation, and the train/test splits for both sEMG and ECG appear to be disjoint as described. Thus the correct response is to keep the reader's CONDITIONAL verdict: the paper should not be rejected outright, but it should not be accepted as 'state of the art' until the comparison is completed and uncertainty is quantified. I therefore leave the verdict unchanged while strengthening the condition.","tokens_in":6820,"tokens_out":5235,"duration_ms":49926,"concrete_test":"Re-run the identical train/test protocol (NINAPro DB2 channel 2, Exercises 1 and 3 for training; channels 9-12, Exercise 2 for test; MIT-BIH NSRD ECG subject split; SNR levels -14 to 0 dB) with TrustEMG-Net and with at least five random seeds for MSEMG, FCN, and SDEMG. Report mean plus/minus standard deviation for SNRimp and RMSE, along with parameter counts. If TrustEMG-Net's SNRimp falls within MSEMG's error bars, or if seed variance exceeds the MSEMG-SDEMG gap, the claim that MSEMG consistently outperforms existing methods is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is an empirical superiority claim: MSEMG 'consistently outperforms all other methods' and is more efficient than the previous state of the art. The comparison set, however, omits TrustEMG-Net [14], a representation-masking transformer U-Net for sEMG enhancement published by overlapping authors and cited in the manuscript. Because no result is reported for this baseline, the headline is established only against FCN and SDEMG, not against the current transformer-based approach. The reported 1.85 dB SNRimp advantage over SDEMG and the parameter-count advantage (279,937 vs 1,233,857) are also given without error bars or multiple-seed statistics, so it is unknown whether the gap is statistically reliable or reflects a single favorable training run. In addition, Table II shows that FCN uses fewer parameters (137,801) than MSEMG, so the efficiency claim is only meaningful relative to SDEMG unless additional baselines are considered. The central claim therefore rests on an incomplete and statistically unquantified comparison. This is a conditional-evidence concern, not an internal inconsistency; the table values may be correct, but they do not yet support the claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":7029,"tokens_out":3357,"duration_ms":31160,"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":[{"comment":"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.","section":"Section IV-C, Table I and Table II"},{"comment":"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.","section":"Section IV-C, Table I and Fig. 3"},{"comment":"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.","section":"Table II and Abstract"},{"comment":"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.","section":"Section IV-A and Section IV-C, Fig. 4"}],"minor_comments":[{"comment":"\"closed to the heart\" should be \"close to the heart\" in the abstract and introduction.","section":"Abstract and Introduction"},{"comment":"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.","section":"Section III-B"},{"comment":"The preprocessing description does not specify the training segment length, test segment length, or any overlap between segments; please provide these details.","section":"Section IV-A"},{"comment":"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.","section":"Section IV-A and Section IV-B"},{"comment":"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.","section":"Section III-A, Eqs. (5)-(8)"},{"comment":"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.","section":"Table I and Fig. 4"},{"comment":"The phrase \"a key proportion of this novel integration\" should be \"a key component\".","section":"Section III"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and addresses a practical problem, but the empirical case needs strengthening before publication. The most significant concern is the omission of TrustEMG-Net [14], which is cited in the paper and comes from the same research group; reviewers and readers will likely view this as a substantial gap in a superiority claim. In addition, the lack of any repeated-seed statistics means the reported gains cannot be distinguished from training noise. I would ask the authors to add the missing baseline and error bars, and to temper the efficiency claim in light of FCN's lower parameter count."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MSEMG is the first Mamba-based sEMG denoiser I know of, and the paper shows it works on a clean held-out protocol: different subjects, channels, movements, and SNRs, with consistent gains over FCN and SDEMG on SNRimp, RMSE, and feature metrics, while being about 4x smaller than SDEMG. That is a real, useful result.\n\nWhere it's soft: the headline claim 'consistently outperforms all other methods' is overreach. The comparison omits TrustEMG-Net [14], a transformer U-Net from the same group for the same task, and the paper doesn't report results for it. The stress-test is right about that; it matters because without that baseline the claim is only against two NN methods. Also, there are no error bars or repeated-seed runs, so the 1.85 dB advantage over SDEMG might be a single lucky run. The efficiency story is also partial: FCN has 137k parameters vs MSEMG's 280k, so 'fewer parameters' only holds relative to SDEMG unless you report FLOPs or runtime.\n\nThe simulation protocol is standard and inherited from prior work, so the external-validity concern is minor; that's how this literature tests ECG contamination.\n\nBottom line: this is a solid engineering contribution, not a breakthrough. The math is standard Mamba; the novelty is the application and the architecture integration. The evaluation protocol is better than many papers in this space. The missing baseline and lack of uncertainty estimates are fixable in revision. I'd send it to peer review, but require the TrustEMG-Net comparison (or a clear justification for omitting it) and some repeated-seed statistics before accepting.","headline":"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.","tokens_in":7595,"tokens_out":2078,"would_cite":true,"duration_ms":18595,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["surface electromyography","ECG artifact removal","Mamba state space model","signal denoising","convolutional neural network","lightweight network","sEMG enhancement"],"falsifier":"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.","tokens_in":6619,"feed_emoji":"💪","tokens_out":5297,"duration_ms":43957,"temperature":0.7,"pith_summary":"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.","feed_headline":"Mamba-based denoiser beats diffusion model on muscle signals","feed_subtitle":"MSEMG lifts SNR by 20.3 dB on ECG-contaminated sEMG with 279,937 parameters, about a fifth of the diffusion baseline.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the fully convolutional network baseline that MSEMG outperforms on ECG artifact removal from single-channel sEMG.","marker":"[12]"},{"why":"Supplies the score-based diffusion baseline (SDEMG) whose evaluation protocol and model size MSEMG is compared against.","marker":"[13]"},{"why":"Supplies the selective state space model (Mamba) that forms the core of the proposed architecture.","marker":"[15]"},{"why":"Supplies the Half Normalized Filter block used to extract multi-resolution features at the input and output of MSEMG.","marker":"[23]"},{"why":"Supplies the NINAPro DB2 database used as the clean sEMG source for training, validation, and testing.","marker":"[24]"},{"why":"Supplies the MIT-BIH Normal Sinus Rhythm Database used as ECG interference in the simulated mixtures.","marker":"[26]"},{"why":"Defines high-pass filtering and template subtraction baselines and the evaluation setup for ECG interference removal.","marker":"[10]"}],"fun_headline_variants":["Mamba-based net beats diffusion for ECG-corrupted sEMG","Lightweight Mamba model boosts sEMG SNR by 20.3 dB","MSEMG: efficient Mamba network surpasses diffusion denoiser","279K parameters: Mamba sEMG denoiser outperforms diffusion","Mamba model cleans muscle signals with fewer parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-based net beats diffusion for ECG-corrupted sEMG","Lightweight Mamba model boosts sEMG SNR by 20.3 dB","MSEMG: efficient Mamba network surpasses diffusion denoiser","279K parameters: Mamba sEMG denoiser outperforms diffusion","Mamba model cleans muscle signals with fewer parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001027,"raw_usage":{"total_tokens":4278,"prompt_tokens":844,"completion_tokens":3434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":3351}},"tokens_in":460,"tokens_out":3434,"duration_ms":24408,"temperature":1.0,"reasoning_tokens":3351,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:45:23.687931+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Ecg artifact removal from single-channel surface emg using fully convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the fully convolutional network baseline that MSEMG outperforms on ECG artifact removal from single-channel sEMG."},{"cited_title":"Sdemg: Score-based diffusion model for surface electromyographic signal de- noising,","cited_arxiv_id":null,"evidence_quote":"Supplies the score-based diffusion baseline (SDEMG) whose evaluation protocol and model size MSEMG is compared against."},{"cited_title":"Deepfilter: An ecg baseline wander removal filter using deep learning techniques,","cited_arxiv_id":null,"evidence_quote":"Supplies the Half Normalized Filter block used to extract multi-resolution features at the input and output of MSEMG."},{"cited_title":"Electromyography data for non-invasive naturally-controlled robotic hand prostheses,","cited_arxiv_id":null,"evidence_quote":"Supplies the NINAPro DB2 database used as the clean sEMG source for training, validation, and testing."},{"cited_title":"Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals,","cited_arxiv_id":null,"evidence_quote":"Supplies the MIT-BIH Normal Sinus Rhythm Database used as ECG interference in the simulated mixtures."},{"cited_title":"Comparative review of the algorithms for removal of electrocardiographic interference from trunk electromyography,","cited_arxiv_id":null,"evidence_quote":"Defines high-pass filtering and template subtraction baselines and the evaluation setup for ECG interference removal."}],"review_version":1}