{"id":"efd8b5a1-f9ac-469e-a36e-7df04c9ed450","arxiv_id":"2607.15972","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A single sEMG channel plus Pearson-filtered time/frequency features and a compact neural network reaches 90% accuracy on 10 hand gestures, though the evaluation likely leaks label information.","lead":"This paper tests whether one muscle sensor can tell apart ten hand gestures using simple computer models. It reports up to 90% accuracy, but the way it was measured may overstate how well it would work in practice.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 90% accuracy likely inflated by window-level split leakage and pre-split Pearson filtering; subject-independent evaluation is required.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: the random window-level split and pre-split Pearson filtering inflate the accuracy estimate. I agree that subject-independent evaluation is necessary to support the central claim. The concern is not about the novelty or the exploratory nature of the study, but about whether the reported 90% accuracy reflects real generalization. Given that the abstract and conclusion directly assert that single-channel classification 'can achieve up to 90 percent accuracy,' the methodological leakage undermines the main contribution. A concrete subject-independent re-evaluation would settle the issue; if accuracy holds above roughly 70–80%, the claim would be credible, but if it collapses, the paper's conclusion is not supported. Therefore the reader's REJECT verdict is appropriate, and no change is needed.","tokens_in":7661,"tokens_out":2536,"duration_ms":27109,"concrete_test":"Reproduce Experiment 2's best NN pipeline (Pearson filter + 4-hidden-layer NN) on Khushaba's dataset using leave-one-subject-out cross-validation, re-fitting the Pearson filter inside each training fold only (using only training labels), and using non-overlapping windows (e.g., one 300 ms window per contraction). If the mean accuracy is substantially lower than 90%—for example below 70%—the reported result is an artifact of leakage and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a single sEMG channel plus Pearson filtering and a compact NN achieves up to 90% accuracy—rests entirely on the Experiment 2 evaluation. That evaluation uses a random 80/20 split of 300 ms windows with 50 ms overlap (Section 3.2). Windows from the same 5-second contraction are separated by only 50 ms, so adjacent windows share 250 ms of signal and are highly autocorrelated. A random split places many such overlapping windows from the same contraction into both training and test sets, allowing the model to memorize contraction-specific artifacts rather than learn gesture-general patterns. Additionally, the Pearson filter (Section 3.3) is applied before the split using class labels from the entire dataset, so test-set labels influence which features are selected—leakage in feature selection. The paper reports 90% accuracy but provides no subject-independent cross-validation or non-overlapping window evaluation. Since the conclusion is specifically about generalizing with limited temporal/spatial information, the reported number does not establish that claim. The prior independent result by Marbán et al. (2025) of 92.3% on 8 gestures suggests single-channel feasibility is plausible, but this paper's 90% on 10 gestures is not supported by its evaluation protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an exploratory study of single-channel sEMG hand gesture classification on the Khushaba 10-gesture dataset. It extracts 32 time- and frequency-domain features, applies Pearson and correlation filtering, optional LDA/PCA, and evaluates NN/KNN/SVM in four experiments. The central claim is that a compact feed-forward NN with Pearson filtering reaches 90% accuracy. The manuscript provides a detailed experimental setup, but the evaluation protocol has critical leakage issues that undermine the reported result.","tokens_in":7895,"tokens_out":6818,"duration_ms":72604,"significance":"If the 90% result were valid, it would meaningfully support the feasibility of single-channel, low-power sEMG gesture recognition, complementing prior work such as Marbán et al.'s 92.3% accuracy on 8 gestures. The paper's strengths are its systematic comparison of feature-filtering and dimensionality-reduction choices, the use of a public benchmark dataset, and the explicit reporting of hyperparameters. However, because the main result is compromised by feature-selection leakage and overlapping-window leakage, the quantitative contribution is not currently established. The paper is potentially salvageable with a corrected evaluation protocol, but the present version does not support its central claim.","major_comments":[{"comment":"The Pearson filter ('Features with an absolute Pearson correlation below 0.1 with respect to the class labels were removed') is applied to the entire dataset before the 80/20 split. This uses test-set labels to select features, which is a form of selection leakage. Consequently, the 90% accuracy reported in Experiment 2 is not an unbiased estimate of generalization. The filter must be applied only to the training portion, or embedded in a cross-validation loop, before evaluating on a held-out test set.","section":"Section 3.3 and Section 4.2"},{"comment":"The 300ms windows with 50ms overlap mean consecutive windows from the same 5-second contraction share 250ms of signal. A random 80/20 window-level split places such strongly autocorrelated windows in both training and test sets, allowing the model to memorize contraction-specific artifacts. This inflates accuracy. The evaluation should use non-overlapping windows or, better, split by contraction/session, and should include subject-independent cross-validation (e.g., leave-one-subject-out).","section":"Section 3.2 and Section 4.1"},{"comment":"All experiments pool data from 8 subjects and use a random split, so test windows can come from the same subjects, and even the same contractions, as training windows. For a gesture-recognition claim, subject-independent evaluation is required. Without it, the results do not demonstrate generalization to new users, which is central to the stated application of low-power embedded systems.","section":"Section 4.1-4.4"},{"comment":"The stark contrast between the NN accuracy (90%) and the KNN/SVM accuracies (31%/19%) on the same Pearson-filtered features is not explained by the manuscript's brief comment about correlated features. Such a large discrepancy is consistent with the NN exploiting leakage or variance from the flawed split; it also suggests the reported single-number result is unstable. Please report repeated runs with standard deviations, cross-validation, and confusion matrices for all conditions.","section":"Section 4.2"}],"minor_comments":[{"comment":"Figures are numbered out of order: Figure 2 and Figure 3 are referenced before Figure 1.","section":"Section 3.1"},{"comment":"The phrase 'comprising of' should be 'comprising' or 'consisting of'.","section":"Section 1"},{"comment":"The manuscript states 'approximately 57,000 samples' but does not report the distribution across subjects, classes, or contractions. This should be added for reproducibility and to assess class balance.","section":"Section 3.2"},{"comment":"No code, split seeds, or feature-extraction scripts are provided. For an empirical machine-learning paper, these are needed to reproduce the reported results and to rerun the corrected evaluation.","section":"General"},{"comment":"Some references (e.g., [5], [10]) are tangential to the method. Consider citing more directly relevant work on sEMG feature selection and evaluation protocols.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's stress-test concern: the leakage issues are real and load-bearing. I recommend major revision rather than outright rejection because the described fixes—moving feature selection inside training, non-overlapping or subject-independent splits, and repeated evaluation—are within the scope of a re-analysis of the same public dataset. If the authors cannot rerun the full evaluation, the paper should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper's central claim—90% accuracy on ten gestures from a single sEMG channel—does not survive a close look at the evaluation protocol. The reader's concern is right, and it is not a minor quibble. The 80/20 split is at the window level, with 300 ms windows overlapping by 50 ms, so windows from the same contraction appear in both training and test sets. On top of that, the Pearson filter selects features by correlating with class labels on the full dataset before the split, so test information leaks into feature selection. The paper even notes that normalization was applied after the split \"to avoid data leakage,\" but misses that the filter itself is a leak. Without subject-independent cross-validation or at least non-overlapping windows, the 90% figure does not establish that single-channel sEMG generalizes.\n\nThat said, the paper has real strengths. It is clearly written, systematically compares Pearson filtering, correlation filtering, LDA, PCA, and three classifiers, and uses a public dataset (Khushaba's). The specific combination results—e.g., Pearson-only filtering with NN versus with KNN/SVM—are new, and the negative results (KNN at 31%, SVM at 19% with Pearson-only) are actually informative. The comparison to Marbán et al. (2025) is honest, and the paper correctly identifies itself as exploratory.\n\nThe soft spots are proportionate to the central claim. The 90% result is the best of many configurations, with no significance testing or correction for multiple comparisons. The poor KNN/SVM results in Experiment 2 suggest that the NN is not learning a robust feature space but rather exploiting window-level redundancy. The paper's own conclusion—that a feature-based pipeline with Pearson filtering and a compact NN can enable accurate single-channel classification—is exactly what the evaluation does not support in its current form.\n\nWho is this for? Readers working on low-power sEMG might find the comparison useful as a starting point, but they should not cite the 90% number. The paper would be a good teaching example of data leakage in myoelectric control.\n\nMy recommendation: engage with it, but only with a clear request for major revision. The evaluation needs to be redone with subject-independent cross-validation and feature selection nested inside the training folds. If that is done, the paper could become a legitimate empirical contribution. As is, the headline result is not trustworthy, but the underlying question is relevant and the writing is clear enough that referee time is justified.","headline":"The 90% single-channel accuracy claim is not supported by the evaluation: window-level splitting and pre-split Pearson filtering leak information, so the headline number is likely inflated.","tokens_in":780,"tokens_out":1786,"would_cite":false,"duration_ms":37210,"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":"A single surface-electromyography channel can drive a compact neural network that recognizes ten hand gestures with up to 90 percent accuracy.","keywords":["surface electromyography","hand gesture recognition","single-channel sEMG","feature extraction","Pearson correlation filtering","neural network classification","dimensionality reduction","low-power embedded systems"],"falsifier":"Compute a leave-one-subject-out accuracy on the same 32-feature, Pearson-filtered, neural-network pipeline and compare it with the random 80/20 split; also compute a trial-wise split in which no windows from the same five-second contraction appear in both training and test. If either evaluation drops accuracy well below 90 percent, the claim that a single channel plus compact NN generalizes is undermined.","tokens_in":7491,"feed_emoji":"✋","tokens_out":5461,"duration_ms":53459,"temperature":0.7,"pith_summary":"This paper tries to establish that one sEMG channel, placed on the flexor side of the forearm, can support ten-class hand gesture recognition when raw signals are converted into 32 time- and frequency-domain features, filtered by correlation with class labels, and fed to a compact neural network. The headline result is 90 percent accuracy, which the paper presents as evidence that single-channel sEMG is viable for low-power, embedded human-machine interfaces. A sympathetic reader would care because most high-accuracy sEMG systems use dense multichannel arrays and heavier models, which are costly and power-hungry. The paper also argues that neural networks benefit from Pearson filtering, while distance-based classifiers like KNN and SVM need LDA or PCA to remove redundant features.","feed_headline":"One forearm muscle sensor reads 10 hand gestures at 90% accuracy","feed_subtitle":"Filtered time-frequency features plus a compact neural net could make gesture control practical on low-power hardware.","key_machinery":"The load-bearing object is the feature-engineering pipeline: 32 features per 300 ms window spanning time-domain statistics (IEMG, MAV, RMS, WL, ZC, etc.), autoregressive coefficients, higher-order crossings, and relative intensity measures, plus a Yule–Burg spectral feature. Pearson filtering keeps only features whose absolute correlation with the class label exceeds 0.1, trimming the set to 26; the paper then optionally applies correlation filtering (removing pairwise-correlated features above 95 percent), LDA to 9 components, or PCA to 5 components. These filters and reductions are what make the compact neural network work, and the paper attributes the NN advantage to the label-correlation","core_discovery":"The paper's central claim is that on a ten-gesture, eight-subject dataset, a single flexor-channel sEMG signal, segmented into 300 ms windows with 50 ms overlap and described by 32 features, reaches up to 90 percent test accuracy when the features are pre-filtered by absolute Pearson correlation with the class labels and classified by a feed-forward neural network with four hidden layers. The same Pearson-filtered features make KNN and SVM perform poorly (31 percent and 19 percent), while applying LDA or PCA restores them to roughly 75–80 percent, suggesting that the redundancy structure of the features interacts strongly with classifier choice. The paper concludes that a carefully engineere","pith_inferences":["Inference: The 90 percent figure is likely optimistic, because the evaluation uses overlapping windows from the same contractions on both sides of an 80/20 split and selects the Pearson filter using the full dataset's class labels before splitting.","Inference: If a subject-independent evaluation still showed high accuracy, single-channel sEMG could make prosthetic and wearable gesture interfaces markedly cheaper and more power-efficient than today's multichannel systems.","Inference: A natural next test is online streaming classification, using the same 50 ms window hop for real-time inference and a short per-user calibration to re-fit the Pearson filter."],"forward_implications":["If the 90 percent result holds under realistic evaluation, a single sEMG channel is enough for ten-gesture control, eliminating the need for dense electrode arrays.","A compact NN on filtered features could run in real time on microcontroller-class hardware, enabling low-power prosthetics, wearables, and smart-home controls.","Feature selection via label correlation works for neural networks but hurts distance-based classifiers unless followed by LDA or PCA, so pipeline design should be classifier-aware.","The feature set's success suggests that rich time-frequency descriptors can substitute for spatial diversity in myoelectric control."],"fun_headline_variants":["Single sEMG channel hits 90% on 10 gestures with a compact NN","One EMG channel, ten gestures, 90% accuracy via a small neural net","Single sEMG sensor: 10 gestures, up to 90% accuracy on a tiny net","Single-channel EMG: 10 gestures at 90% via a lightweight net"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported accuracies assume that randomly splitting 300 ms windows into 80 percent training and 20 percent test sets creates independent examples, but overlapping windows from the same sustained contraction appear on both sides, and the Pearson filter is selected using the full dataset's class labels before the split, so the numbers may not reflect performance on new users.","fun_headline_variants_meta":{"raw":{"variants":["Single sEMG channel hits 90% on 10 gestures with a compact NN","One EMG channel, ten gestures, 90% accuracy via a small neural net","Single sEMG sensor: 10 gestures, up to 90% accuracy on a tiny net","Single-channel EMG: 10 gestures at 90% via a lightweight net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000681,"raw_usage":{"total_tokens":2916,"prompt_tokens":720,"completion_tokens":2196,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":2103}},"tokens_in":464,"tokens_out":2196,"duration_ms":16093,"temperature":1.0,"reasoning_tokens":2103,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:42:40.654418+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute a leave-one-subject-out accuracy on the same 32-feature, Pearson-filtered, neural-network pipeline and compare it with the random 80/20 split; also compute a trial-wise split in which no windows from the same five-second contraction appear in both training and test. If either evaluation drops accuracy well below 90 percent, the claim that a single channel plus compact NN generalizes is undermined.","supporting_citations":[],"review_version":1}