An Exploratory Study of Single Channel Surface Electromyography for Hand Gesture Classification

Daanish Hindustani

REVIEW 4 major objections 5 minor 12 references

An Exploratory Study of Single Channel Surface Electromyography for Hand Gesture Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A single surface-electromyography channel can drive a compact neural network that recognizes ten hand gestures with up to 90 percent accuracy.

desk verdict 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. read the letter →

arxiv 2607.15972 v1 pith:RQQNGSFI submitted 2026-07-17 cs.LG

Daanish Hindustani This is my paper · ORCID

classification cs.LG

keywords surfaceelectromyographyhandgesturerecognitionsingle-channelsEMGfeatureextractionPearsoncorrelationfilteringneuralnetworkclassificationdimensionalityreductionlow-powerembeddedsystems

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 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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

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

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.

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.

Desk Editor's Note

read the letter

Dear colleague,

The 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.

That 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.

The 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.

Who 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.

My 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.

Referee Report

4 major / 5 minor

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.

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 (4)

[Section 3.3 and Section 4.2] 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 3.2 and Section 4.1] 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 4.1-4.4] 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 4.2] 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.

minor comments (5)

[Section 3.1] Figures are numbered out of order: Figure 2 and Figure 3 are referenced before Figure 1.
[Section 1] The phrase 'comprising of' should be 'comprising' or 'consisting of'.
[Section 3.2] 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.
[General] 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.
[References] Some references (e.g., [5], [10]) are tangential to the method. Consider citing more directly relevant work on sEMG feature selection and evaluation protocols.

Circularity Check

1 steps flagged · score 5.0 of 10

The 90% accuracy claim is partially circular because Pearson feature selection uses the class labels of the full dataset before the train/test split.

fitted input called prediction [Section 3.3 (Data Pre-processing) and Section 4.2 (Experiment 2)]
"Pearson filtering: Features with an absolute Pearson correlation below 0.1 with respect to the class labels were removed, eliminating weakly informative features... In Experiment 2, the data was preprocessed using only the Pearson correlation filter, reducing the feature dimensionality to 26 features. The dataset was split into 80% training and 20% testing sets."

The Pearson filter is a target-dependent feature-selection step applied to all samples before the 80/20 split. The class labels of the test windows therefore determine which features survive. The reported 90% test accuracy is not an out-of-sample estimate of the full pipeline: the feature-selection stage has already seen the target values of the test data, so part of the 'prediction' reduces to information injected into the model inputs by construction.

full rationale

The paper's central claim is empirical: a compact NN on single-channel sEMG with Pearson filtering reaches 90% accuracy (Experiment 2). This is not a mathematical derivation, and the main circularity risks (load-bearing self-citation, uniqueness imported from the authors, ansatz smuggling, renaming) do not apply: the references are independent prior works and no self-citation bears the result. However, the evaluation pipeline contains a concrete circular step. Section 3.3 removes features with absolute Pearson correlation below 0.1 with respect to class labels; Section 4.2 then splits the already-filtered dataset into 80/20. Thus feature selection is performed on the entire dataset using test labels, so the accuracy reported in Experiment 2 partially reflects target information already present in the input representation. This is a fitted-input-called-prediction pattern: the feature subset is chosen using Y, and then Y is 'predicted' on the same windows. A second, non-circular but compounding validity threat is the random split of 300 ms windows with 50 ms overlap from the same 5-second contractions, which destroys train/test independence and further inflates accuracy. Because the classifier weights themselves are trained only on the training split, the result is not entirely forced by construction, but the headline 90% figure is not a clean out-of-sample result. Score 5 reflects partial circularity via label-informed feature selection.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's central claim rests on a standard feature-engineering pipeline with several hand-chosen thresholds (0.1, 0.95, 9 LDA components, 5 PCA components) and empirically selected architectures. No new physical entities are introduced. The most serious assumptions are statistical: that windows are independent and that label-based feature selection is not leakage. These assumptions are not stated explicitly and are likely violated, which directly affects the validity of the reported accuracy.

free parameters (6)

Pearson filter threshold = 0.1
Features with absolute correlation to class labels below 0.1 are removed. Chosen by hand in Section 3.3; no sensitivity analysis or nested selection is provided.
Correlation filter threshold = 0.95
Features with pairwise correlation above 95% are removed. Chosen by hand in Section 3.3; the paper notes high redundancy but provides no justification for the exact cutoff.
LDA components = 9
In Experiment 3, the 26 Pearson-filtered features are reduced to 9 LDA components. No rationale is given for selecting 9 components (Section 4.3).
PCA components = 5
In Experiment 4, PCA reduces to 5 components chosen by the elbow method (Section 4.4, Figure 18). The elbow is a subjective heuristic.
Neural network architectures = e.g., 32-64-32, 256-128-64-32
Layer sizes and depths are described as 'chosen empirically' (Section 4.1, 4.2). These choices are tuned on the data, likely on the test set, without nested validation.
KNN k values = k=7,1,7,9
k is tuned per experiment via hyperparameter search (Section 4). No cross-validation on a held-out set is described, so tuning may use test feedback.

assumptions (4)

domain assumption Single flexor channel (Channel 1) captures sufficiently discriminative gesture information.
Section 3.1 states only Channel 1 is used because flexor muscles are 'known to produce stronger and more discriminative sEMG signals'; this is based on a prior study [8] and not re-validated here.
domain assumption Windows from the same subject and repeated trials are independent and identically distributed.
The 80/20 random split (Section 4.1) assumes that overlapping 300ms windows from the same contraction are independent samples. This is violated by the data's temporal structure and is the load-bearing assumption for the reported accuracy.
domain assumption Feature selection can safely use class labels from the entire dataset without biasing the test estimate.
The Pearson filter (Section 3.3) removes features based on correlation with labels computed over the whole dataset, before the split. This assumes no leakage, which is false in a supervised evaluation.
standard math Standard machine learning mathematical tools (Pearson correlation, LDA, PCA, neural network training) operate as expected.
These are standard methods assumed without proof; the paper invokes them per standard textbook treatments [4].

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Exploratory Study of Single Channel Surface Electromyography for Hand Gesture Classification." pith.science (2026). https://pith.science/paper/RQQNGSFI

@misc{pith2026260715972,
  author       = {Pith},
  title        = {Pith review of: An Exploratory Study of Single Channel Surface Electromyography for Hand Gesture Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQQNGSFI}},
  note         = {Machine review of arXiv:2607.15972}
}

read the original abstract

Accurate hand gesture recognition using surface electromyography (sEMG) typically relies on multichannel sensor arrays and computationally intensive models. This limits practical deployment in low-power and embedded systems. This study investigates the feasibility of classifying ten hand gestures using a single sEMG channel combined with lightweight machine learning architectures. Raw sEMG signals were transformed into a comprehensive feature-based representation, including time-domain, frequency-domain, higher-order-crossing, and relative-intensity features. Feature redundancy was reduced using Pearson correlation filtering and the removal of highly correlated features, while dimensionality-reduction techniques (LDA and PCA) were applied selectively. Three classifiers, a feed-forward neural network (NN), k-nearest neighbors (KNN), and a support vector machine (SVM), were systematically evaluated across four experiments. Results demonstrate that combining time and frequency features with Pearson filtering and a compact NN can achieve up to 90 percent accuracy, even with limited temporal and spatial information. These findings highlight the potential of single-channel sEMG systems for cost-effective, low-power gesture-recognition applications.

Figures

Figures reproduced from arXiv: 2607.15972 by the authors.

**Figure 9.** Confusion matrix for the SVM classifier in Experiment 1. 4.2 Experiment 2 and Results In Experiment 2, the data was preprocessed using only the Pearson correlation filter, reducing the feature dimensionality to 26 features. The dataset was split into 80% training and 20% testing sets. Then the parameters were normalized and fit on the training set and applied to the test set to avoid data leakage [4]. 1) Neural … view at source ↗

**Figure 11.** Confusion matrix for the neural network classifier in Experiment 2. 2) k-Nearest Neighbors (KNN): The second sub experiment applied a KNN classifier. Hyperparameter tuning identified an optimal k = 1. The KNN achieved an overall accuracy of 31%. This poor performance is due to the presence of highly correlated features remaining after Pearson filtering, which adversely affected distance-based classification [PIT… view at source ↗

Figures from the paper (3 more)

**Figure 14.** Figure 14: , and the confusion matrix is shown in [PITH_FULL_IMAGE:figures/full_fig_p008_14.png]

**Figure 16.** Figure 16: Confusion matrix for the KNN classifier in Experiment 3 [PITH_FULL_IMAGE:figures/full_fig_p009_16.png]

**Figure 19.** Figure 19: Test accuracy versus training epoch for the neural network classifier in Experiment 4 [PITH_FULL_IMAGE:figures/full_fig_p009_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 1 canonical work pages

[1]

Machine learning-based sEMG signal classification for hand gesture recognition,

P. N. Aarotale and A. Rattani, “Machine learning-based sEMG signal classification for hand gesture recognition,” arXiv pre- print arXiv:2411.15655, 2024. Available: https://arxiv.org/abs/2411.15655

arXiv 2024
[2]

A systematic review on surface electromyography-based classification of hand and finger gestures,

M. A. Al-Timemy, A. M. Al-Faiz, and S. Sumari, “A systematic review on surface electromyography-based classification of hand and finger gestures,” Journal of King Saud University – Computer and Infor- mation Sciences, vol. 35, no. 1, pp. 101– 115, Jan. 2023, doi: 10.1016/j.jksuci.2022.11.004

work page doi:10.1016/j.jksuci.2022.11.004 2023
[3]

Comparison of five time series EMG features extractions using Myo arm- band,

Z. Arief, I. A. Sulistijono, and R. A. Ar- diansyah, “Comparison of five time series EMG features extractions using Myo arm- band,” in Proc. Int. Electron. Symp. (IES), Surabaya, Indonesia, 2015, pp. 11–14

2015
[4]

C. M. Bishop, Pattern Recognition and Machine Learning, New York, NY, USA: Springer, 2006

2006
[5]

Towards a theory of model distillation,

E. Boix-Adserà, “Towards a theory of model distillation,” arXiv preprint arXiv:2403.09053, 2024, doi: 10.48550/arXiv.2403.09053

arXiv doi:10.48550/arxiv.2403.09053 2024
[6]

An investigation of dimensionality reduc- tion techniques for EMG-based force esti- mation,

G. Hajian, A. Etemad, and E. Morin, “An investigation of dimensionality reduc- tion techniques for EMG-based force esti- mation,” in Proc. 41st Annual Int. Conf. IEEE Engineering in Medicine and Biology Society (EMBC), 2019, pp. 698–701, doi: 10.1109/EMBC.2019.8856293

arXiv 2019
[7]

Ant Colony Optimization-Based Feature Selec- tion for EMG Pattern Recognition,

Q. Huang, Z. Zhou, and X. Li, “Ant Colony Optimization-Based Feature Selec- tion for EMG Pattern Recognition,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 20, no. 6, pp. 759–767, 2012

2012
[8]

Toward im- proved control of prosthetic fingers using surface electromyogram (EMG) signals,

R. N. Khushaba, M. Takruri, S. Koda- goda, and G. Dissanayake, “Toward im- proved control of prosthetic fingers using surface electromyogram (EMG) signals,” Expert Systems with Applications, vol. 39, no. 12, pp. 10731–10738, 2012

2012

Show all 12 references

[9]

Hand gestures recognition based on one- channel surface EMG signals,

Y. Li, X. Yu, X. Li, and Y. Zhang, “Hand gestures recognition based on one- channel surface EMG signals,” Journal of Signal and Information Processing, vol. 10, no. 4, pp. 253–266, 2019, doi: 10.4236/jsip.2019.104015

2019
[10]

Gesture recognition using surface electromyography and deep learning for prostheses hand: State-of-the-art, challenges and future,

W. Li, P. Shi, and H. Yu, “Gesture recognition using surface electromyography and deep learning for prostheses hand: State-of-the-art, challenges and future,” Frontiers in Neuroscience, vol. 15, Art. no. 621885, 2021, doi: 10.3389/fnins.2021.621885

2021
[11]

Single- channel sEMG hand gesture classification using an artificial neural network imple- mented on an ESP32 microcontroller,

J. A. Marbán Salgado, E. R. Salgado, J. L. Gordillo, and M. A. Martínez, “Single- channel sEMG hand gesture classification using an artificial neural network imple- mented on an ESP32 microcontroller,” IEEE Access, vol. 13, pp. 142457–142466, Daanish Hindustani 2025, doi: 10...

2025
[12]

Support vector machine-based classification scheme for myoelectric control applied to upper limb,

M. A. Oskoei and H. Hu, “Support vector machine-based classification scheme for myoelectric control applied to upper limb,” IEEE Trans. Biomed. Eng., vol. 55, no. 8, pp. 1956–1965, Aug. 2008, doi: 10.1109/TBME.2008.919734

1956

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Machine learning-based sEMG signal classification for hand gesture recognition,

P. N. Aarotale and A. Rattani, “Machine learning-based sEMG signal classification for hand gesture recognition,” arXiv pre- print arXiv:2411.15655, 2024. Available: https://arxiv.org/abs/2411.15655

arXiv 2024

[2] [2]

A systematic review on surface electromyography-based classification of hand and finger gestures,

M. A. Al-Timemy, A. M. Al-Faiz, and S. Sumari, “A systematic review on surface electromyography-based classification of hand and finger gestures,” Journal of King Saud University – Computer and Infor- mation Sciences, vol. 35, no. 1, pp. 101– 115, Jan. 2023, doi: 10.1016/j.jksuci.2022.11.004

work page doi:10.1016/j.jksuci.2022.11.004 2023

[3] [3]

Comparison of five time series EMG features extractions using Myo arm- band,

Z. Arief, I. A. Sulistijono, and R. A. Ar- diansyah, “Comparison of five time series EMG features extractions using Myo arm- band,” in Proc. Int. Electron. Symp. (IES), Surabaya, Indonesia, 2015, pp. 11–14

2015

[4] [4]

C. M. Bishop, Pattern Recognition and Machine Learning, New York, NY, USA: Springer, 2006

2006

[5] [5]

Towards a theory of model distillation,

E. Boix-Adserà, “Towards a theory of model distillation,” arXiv preprint arXiv:2403.09053, 2024, doi: 10.48550/arXiv.2403.09053

arXiv doi:10.48550/arxiv.2403.09053 2024

[6] [6]

An investigation of dimensionality reduc- tion techniques for EMG-based force esti- mation,

G. Hajian, A. Etemad, and E. Morin, “An investigation of dimensionality reduc- tion techniques for EMG-based force esti- mation,” in Proc. 41st Annual Int. Conf. IEEE Engineering in Medicine and Biology Society (EMBC), 2019, pp. 698–701, doi: 10.1109/EMBC.2019.8856293

arXiv 2019

[7] [7]

Ant Colony Optimization-Based Feature Selec- tion for EMG Pattern Recognition,

Q. Huang, Z. Zhou, and X. Li, “Ant Colony Optimization-Based Feature Selec- tion for EMG Pattern Recognition,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 20, no. 6, pp. 759–767, 2012

2012

[8] [8]

Toward im- proved control of prosthetic fingers using surface electromyogram (EMG) signals,

R. N. Khushaba, M. Takruri, S. Koda- goda, and G. Dissanayake, “Toward im- proved control of prosthetic fingers using surface electromyogram (EMG) signals,” Expert Systems with Applications, vol. 39, no. 12, pp. 10731–10738, 2012

2012