Attention-Based Convolutional Neural Network Model for Human Lower Limb Activity Recognition using sEMG

Farshad Haghgoo Daryakenari; Gholamreza Vossoughi; Mohammad Hossein Rohban; Mojtaba Mollahossein

REVIEW 4 major objections 6 minor 14 references

Attention-Based Convolutional Neural Network Model for Human Lower Limb Activity Recognition using sEMG

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 62,876-parameter attention CNN classifies raw sEMG lower-limb activity at 85.38% test accuracy with no preprocessing.

desk verdict A plausible lightweight attention-CNN for sEMG lower-limb classification, but the 'leave-one-out' claim is overstated: the evaluation is a single split with two test subjects, so the headline accuracy is a point estimate, not a robust generalization result. read the letter →

arxiv 2506.06624 v1 pith:NL5V4ANA submitted 2025-06-07 cs.RO

Mojtaba Mollahossein , Farshad Haghgoo Daryakenari , Mohammad Hossein Rohban , Gholamreza Vossoughi This is my paper

classification cs.RO

keywords sEMGclassificationlowerlimbactivityrecognitionattentionmechanismconvolutionalneuralnetworkleave-one-outvalidationreal-timeinferenceBASANdatasetrehabilitationrobotics

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 a small attention-based convolutional network can classify three lower-limb activities (walking, standing with knee flexion, and sitting with knee extension) directly from raw multi-channel surface EMG, with no denoising or handcrafted features, while staying light enough for real-time use. On the public BASAN dataset, with one healthy and one abnormal subject held out, the model reports 86.74% validation and 85.38% test accuracy from 62,876 parameters and sub-millisecond inference. The authors argue that their leave-one-out evaluation gives a more realistic measure of generalization to new people than splits that let the network see all subjects during training, and that comparable or better accuracy is reached at far lower parameter counts than prior baselines. A sympathetic reader would care because real-time assistive robotics and prostheses need classifiers that are both accurate on unseen users and cheap enough to run on embedded hardware.

What carries the argument

The load-bearing mechanism is a per-channel Bahdanau attention layer inserted after convolutional feature extraction; Bahdanau attention is a learned weighting scheme that lets the network focus on the most informative temporal segments within each 256-sample EMG window. The concatenation layer acts as a skip pathway so the final dense layers see both early low-level patterns and later high-level features. This combination, in a network of 62,876 parameters, is what the paper credits for keeping accuracy high without denoising or feature engineering.

What would settle it

A concrete check: repeat the leave-one-out evaluation over many different pairings of one healthy and one abnormal test subject on the BASAN dataset and report the distribution of test accuracy. If most pairings fall well below the reported 85.38%, the claim that the model generalizes to unseen subjects would be weakened.

Watch

Extended reading notes

Core claim

The paper's central claim is that an attention-augmented 1D CNN is enough to separate three lower-limb movements from four raw sEMG channels without preprocessing. The architecture uses three convolution-pooling stages per channel, a Bahdanau attention layer, and a concatenation shortcut that mixes low- and high-level features before the dense classifier. With a 256 ms window and 64 ms overlap, the model reaches 86.74% accuracy on a held-out validation subject and 85.38% on a held-out test subject, with class-wise F1 scores between 82.83% and 87.66%. The paper also reports that adding wavelet denoising raises test accuracy to 86.39%, and that the model beats the majority-voting CNN baseline on the same leave-one-out protocol while using about a fifteenth of the parameters. The intended contribution is a lightweight, generalizable, real-time-capable classifier for upper-level robot and prosthetic controllers.

Load-bearing premise

The central claim assumes that two held-out subjects—one healthy and one abnormal—represent how well the model will do on all other new users, and with a sample that small the 85% test accuracy could change substantially with different subjects.

Editorial extensions

If this is right

If the reported accuracy holds, raw sEMG can serve as a direct control signal for lower-limb assistive devices without a denoising stage, removing preprocessing latency.
With inference below 1 ms and 62,876 parameters, the model fits on modest embedded hardware, making subject-independent movement classification feasible in real time.
The leave-one-out protocol sets a stricter evaluation standard: models trained and tested with this split are directly comparable, and earlier accuracy numbers obtained without subject separation likely overstate real-world performance.
Adding wavelet denoising improves the model to 86.39% test accuracy, suggesting further gains are available from filters that can be implemented without blocking real-time use.

Reading between the lines

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

A direct testable extension would be to run the same architecture on other sEMG activity datasets that include transition phases (sit-to-stand), because the BASAN data lacks transitions and the paper flags that as a limitation.
The two-subject test set means the 85.38% figure is an anecdotal estimate of subject generalization; a confidence interval over many leave-one-subject-out folds would separate architecture quality from lucky subject choice.
The attention maps themselves could be inspected to see which muscles and time segments drive each class decision, which may reveal physiologically interpretable features without extra sensors.
The comparison with the 141,000-parameter ensemble suggests that parameter count, inference time, and generalization accuracy should be reported together in future sEMG classifiers.

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

This one is a modest empirical contribution with a real evaluation flaw. The model is a small attention-CNN (62,876 parameters) that classifies three lower-limb movements from raw multi-channel sEMG on the BASAN dataset. What is actually new: a specific lightweight architecture with subject-independent evaluation on that dataset, plus a direct comparison to the voting-based 1D CNN in [4]. The authors show their model beats that baseline at lower parameter count, and they are honest about the lack of transition-phase data and the need for real-time filtering studies. The paper does several things well: the architecture is simple and clearly described, the confusion matrix is internally consistent with the reported accuracy, balanced accuracy matches overall accuracy, and the inference time claim (<1 ms) is relevant for deployment. Credit is earned for the comparison with [4] and for avoiding heavy preprocessing.

The soft spot is load-bearing. The abstract and conclusion say 'leave-one-out validation,' but Section II-C describes a single split into 18 train, 2 validation, and 2 test subjects, with one healthy and one abnormal per held-out group. That is not leave-one-out; it is one draw from the set of possible two-subject test pairs. The 85.38% test accuracy rests entirely on two individuals. There are no per-subject accuracies, no error bars, no repeated random splits, and no random-seed details. The thousands of test windows come from the same two subjects, so they are not independent samples. The claim that the model 'generalizes to unseen subjects' is therefore not supported by the evidence as presented. This is a serious issue, though not a fatal one: the architecture and approach are plausible, and a properly cross-validated evaluation could well produce similar numbers. But as reported, the headline is a point estimate for two people.

Minor concerns: training hyperparameters (optimizer, learning rate, batch size) are missing, and the attention mechanism description is vague—'Bahdanau attention' is named but the exact placement and implementation are not fully specified. The authors' comparison to [4] appears to be their own reimplementation under their split, which is fine, but it should be stated explicitly. Citation practice is acceptable; the self-citation [6] is relevant prior work, not padding.

Who this is for: researchers working on lightweight sEMG classifiers for real-time assistive devices, who will read it as an incremental data point, not a methodological advance. It deserves a serious referee because the underlying question—how well can a small attention-CNN generalize subject-independently on BASAN—is legitimate and the authors supply a concrete architecture and comparisons. But I would send it back for major revision: fix the 'leave-one-out' language, report per-subject results and variance via true leave-one-subject-out or repeated k-fold, and provide full training details. Without those, the central generalization claim does not hold as stated.

Referee Report

4 major / 6 minor

Summary. This paper proposes a lightweight attention-based convolutional neural network for classifying three lower-limb activities (walking, standing with knee flexion, sitting with knee extension) from four-channel sEMG signals in the BASAN/UCI lower-limb dataset. The model has 62,876 parameters, uses a sliding window of 256 ms with 64 ms overlap, and is trained for 50 epochs. The authors report an 86.74% validation accuracy and 85.38% test accuracy on held-out subjects, along with AUC values and a confusion matrix, and claim the model is suitable for real-time deployment. The paper also compares the model with existing 1D-CNN and CNN-LSTM baselines.

Significance. If the reported generalization performance holds, the paper would provide a computationally efficient sEMG classifier with a public dataset, which is useful for assistive robotics. The reported metrics are internally consistent: the confusion matrix counts sum to 3,214 windows and the resulting accuracy (85.38%), precision, recall, and F1-scores in Table I all match. The paper also openly reports the parameter count and the lack of handcrafted features. However, the significance is limited by the evaluation protocol, which currently does not substantiate the subject-independent generalization claim, and by the lack of a comparison under a common protocol.

major comments (4)

[Section II-C] The paper claims to employ a leave-one-out strategy (Abstract, Sections II-A and II-C), but Section II-C describes a single split of 22 subjects into 18 training, 2 validation, and 2 test subjects, not a leave-one-subject-out procedure. The reported test accuracy of 85.38% is therefore a point estimate from exactly two test subjects (one healthy, one abnormal). This cannot support the abstract's claim that the model 'ensures generalizability across subjects.' Please provide per-subject results (e.g., a confusion matrix for each test subject) and either a true leave-one-subject-out cross-validation over all 22 subjects or repeated random splits with mean and standard deviation.
[Section II-D and III] The architecture description is not reproducible. The text says 'these four one-dimensional signals are then passed through two layers, which transform them into two separate one-dimensional signals,' but the figure appears to show a different flow, and neither the text nor the figure specifies the layer types (dense? conv?) or their sizes. In addition, the training procedure is incompletely specified: the optimizer, learning rate, batch size, weight initialization, and random seed are not reported. Without these, the reader cannot reproduce the 85.38% test result, which is a central claim of the paper.
[Section III (Tables II and III)] The comparison with [4] and [14] is not performed under a common protocol. [4] uses a majority-voting ensemble of four 1D-CNNs, and [14] uses 3-fold cross-validation with ADASYN oversampling. The proposed model is evaluated on a single 18/2/2 split. The statement 'our proposed network outperforms the Majority-Voting-based network' is only valid for this particular test split and does not account for the variance of the evaluation. To support the claim of superiority, the authors should report results for all methods on the same test subjects and the same number of trials.
[Section II-D and III] The paper does not provide an ablation study for the attention mechanism. Since the title and abstract emphasize the attention-based model, the reader cannot determine whether the attention mechanism contributes to the reported accuracy or whether the result is dominated by the convolutional layers and the concatenation pathway. An ablation with and without attention, evaluated on the same protocol, is needed to justify the proposed architecture.

minor comments (6)

[Section II-C] The term 'leave-one-out' is used inconsistently; the described split is a single hold-out split. Please rename it throughout.
[Section II-B] The BASAN dataset is referenced as [12] with a UCI repository URL, but the UCI dataset is titled 'EMG dataset in Lower Limb'; clarify the naming (BASAN vs. UCI).
[Section III] The ROC curves in Fig. 3(c) are not described; indicate whether they are per-class one-vs-rest curves and how AUC was computed.
[Tables II and III] The header 'Numer of Parameters (%)' has a typo and the unit '%' is inappropriate; use 'Number of Parameters'.
[Fig. 2] The kernel '5×5' and '3×3' labels appear inconsistent with 1D convolutions; use 'kernel size' notation such as 5×1 or 'kernel length 5'.
[Abstract and II-C] The phrase 'no preprocessing' is overstated; the dataset has already been band-pass filtered (20-460 Hz) as described in Section II-B. Rephrase to 'no additional handcrafted feature extraction.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported accuracies are empirical measurements on held-out subjects, and no load-bearing step reduces to its own inputs.

full rationale

The paper's central quantitative claims are the validation accuracy (86.74%) and test accuracy (85.38%) of the proposed attention-CNN. These are presented as measured outcomes of training on 18 subjects (9 healthy, 9 abnormal) and evaluating on two held-out subjects (1 healthy, 1 abnormal), as described in Section II-C. The test subjects are disjoint from the training set, so the test accuracy is not a fitted value or a re-expression of training data. No equation in the paper defines a model output in terms of the reported accuracy, and no parameter is tuned to the test set. The only self-citation (reference [6], the authors' prior work) is used to support the general statement that CNNs are well-suited for sEMG feature extraction; it is not load-bearing for the specific architecture or the reported results. The paper's use of the term 'leave-one-out' is inaccurate for the described single split (18/2/2), but this is a methodological/statistical issue about generalization confidence, not a circular derivation. The measured inference time (<1 ms) is also an empirical result. Consequently, the derivation chain is self-contained and non-circular.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical deep learning study; it introduces no new theoretical entities. Its central claims rest on the dataset, the hand-chosen architecture, and the assumption that two test subjects represent the population.

free parameters (6)

Window length and overlap = 256 samples with 64 sample overlap
Chosen by hand, affects input size and number of training windows.
Convolutional filter counts and kernel sizes = 16x5, 8x3, 4x3
Chosen by hand, defines the feature extraction layers.
Dense layer size = 100 neurons
Chosen by hand, controls the classification head capacity.
Dropout rate = 0.5
Chosen by hand to prevent overfitting.
Training epochs = 50
Chosen by hand, stopping point for training.
Optimizer, learning rate, batch size = Not reported
Unspecified training hyperparameters that directly affect results.

assumptions (3)

domain assumption The BASAN dataset labels are correct and the sEMG signals are representative of the three activities.
The whole evaluation relies on the dataset's fidelity and proper labeling.
ad hoc to paper The 256 ms windows with 64 ms overlap contain sufficient information to classify the activities.
The window choice is made without justification and affects all results.
domain assumption Classification accuracy on two held-out subjects estimates generalization to unseen subjects.
The paper treats n=2 as a generalization estimate, which is statistically fragile.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Attention-Based Convolutional Neural Network Model for Human Lower Limb Activity Recognition using sEMG." pith.science (2026). https://pith.science/paper/NL5V4ANA

@misc{pith2026250606624,
  author       = {Pith},
  title        = {Pith review of: Attention-Based Convolutional Neural Network Model for Human Lower Limb Activity Recognition using sEMG},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NL5V4ANA}},
  note         = {Machine review of arXiv:2506.06624}
}

read the original abstract

Accurate classification of lower limb movements using surface electromyography (sEMG) signals plays a crucial role in assistive robotics and rehabilitation systems. In this study, we present a lightweight attention-based deep neural network (DNN) for real-time movement classification using multi-channel sEMG data from the publicly available BASAN dataset. The proposed model consists of only 62,876 parameters and is designed without the need for computationally expensive preprocessing, making it suitable for real-time deployment. We employed a leave-oneout validation strategy to ensure generalizability across subjects, and evaluated the model on three movement classes: walking, standing with knee flexion, and sitting with knee extension. The network achieved 86.74% accuracy on the validation set and 85.38% on the test set, demonstrating strong classification performance under realistic conditions. Comparative analysis with existing models in the literature highlights the efficiency and effectiveness of our approach, especially in scenarios where computational cost and real-time response are critical. The results indicate that the proposed model is a promising candidate for integration into upper-level controllers in human-robot interaction systems.

Figures

Figures reproduced from arXiv: 2506.06624 by the authors.

**Figure 2.** Network architecture. TABLE I: Prediction results of test data on each class. Metrices Class Precision(%) Recall(%) F1-score(%) Movement 1 90.56 84.94 87.66 Movement 2 86.79 83.74 85.23 Movement 3 78.72 87.40 82.83 A similar result can be concluded from the F1-score, which provides a balanced measure of precision and recall. F1- score was computed for each class to evaluate the overall classification performance. As… view at source ↗

**Figure 3.** Results of the proposed classification network: a) Network training loss (blue line) and validation loss (red line) for [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

[4]

V oting-based 1d cnn model for human lower limb activity recognition using semg signal,

A. Vijayvargiya, Khimraj, R. Kumar, and N. Dey, “V oting-based 1d cnn model for human lower limb activity recognition using semg signal,” Physical and Engineering Sciences in Medicine, vol. 44, pp. 1297–1309, 12 2021

work page 2021
[6]

Classification of lower limb electromyographical signals based on autoencoder deep neural network transfer learning,

F. H. Daryakenari, M. Mollahossein, A. Taheri, and G. R. V ossoughi, “Classification of lower limb electromyographical signals based on autoencoder deep neural network transfer learning,” in10th RSI In- ternational Conference on Robotics and Mechatronics, ICRoM 2022, pp. 323–328, Institute of Electrical and Electronics Engineers Inc., 2022

work page 2022
[14]

Deep ensemble learning approach for lower limb movement recognition from multichannel semg signals,

P. Tokas, V . B. Semwal, and S. Jain, “Deep ensemble learning approach for lower limb movement recognition from multichannel semg signals,” Neural Computing and Applications, vol. 36, pp. 7373–7388, 5 2024

work page 2024
[1]

A review on appli- cations of activity recognition systems with regard to performance and evaluation,

S. Ranasinghe, F. A. MacHot, and H. C. Mayr, “A review on appli- cations of activity recognition systems with regard to performance and evaluation,” 8 2016

work page 2016
[2]

Human daily activity recognition for healthcare using wearable and visual sensing data,

X. Liu, L. Liu, S. J. Simske, and J. Liu, “Human daily activity recognition for healthcare using wearable and visual sensing data,” inProceedings - 2016 IEEE International Conference on Healthcare Informatics, ICHI 2016, pp. 24–31, Institute of Electrical and Electronics Engineers Inc., 12 2016

work page 2016
[3]

A survey on human activity recognition using wearable sensors,

´Oscar D. Lara and M. A. Labrador, “A survey on human activity recognition using wearable sensors,”IEEE Communications Surveys and Tutorials, vol. 15, pp. 1192–1209, 2013

work page 2013
[5]

Human lower limb activity recognition techniques, databases, challenges and its applications using semg signal: an overview,

A. Vijayvargiya, B. Singh, R. Kumar, and J. M. R. Tavares, “Human lower limb activity recognition techniques, databases, challenges and its applications using semg signal: an overview,” 11 2022

work page 2022
[7]

Biological cybernetics neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position,

K. Fukushima, “Biological cybernetics neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position,” tech. rep., 1980

work page 1980

Show all 14 references

[8]

Myonet: A transfer-learning-based lrcn for lower limb movement recognition and knee joint angle prediction for remote monitoring of rehabilitation progress from semg,

A. Gautam, M. Panwar, D. Biswas, and A. Acharyya, “Myonet: A transfer-learning-based lrcn for lower limb movement recognition and knee joint angle prediction for remote monitoring of rehabilitation progress from semg,”IEEE Journal of Translational Engineering in Health and Med...

2020
[9]

Long short-term memory,

S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Computation, vol. 9, pp. 1735–1780, Nov 1997

1997
[10]

Deep belief networks for phone recognition,

A.-R. Mohamed, G. Dahl, and G. Hinton, “Deep belief networks for phone recognition,”Nips workshop on deep learning for speech recognition and related applications, vol. 1, p. 39, 2009

2009
[11]

Attention is all you need,

A. Vaswani, G. Brain, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Łukasz Kaiser, and I. Polosukhin, “Attention is all you need,”ArXiv, p. 15, 2017

2017
[12]

EMG dataset in Lower Limb

O. Sanchez and J. Sotelo, “EMG dataset in Lower Limb.” UCI Machine Learning Repository, 2014. DOI: https://doi.org/10.24432/C5ZW3P

2014 doi
[13]

A deep learn- ing strategy for emg-based joint position prediction in hip exoskeleton assistive robots,

A. Foroutannia, M. R. Akbarzadeh-T, and A. Akbarzadeh, “A deep learn- ing strategy for emg-based joint position prediction in hip exoskeleton assistive robots,”Biomedical Signal Processing and Control, vol. 75, 5 2022

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [4]

V oting-based 1d cnn model for human lower limb activity recognition using semg signal,

A. Vijayvargiya, Khimraj, R. Kumar, and N. Dey, “V oting-based 1d cnn model for human lower limb activity recognition using semg signal,” Physical and Engineering Sciences in Medicine, vol. 44, pp. 1297–1309, 12 2021

work page 2021

[2] [6]

Classification of lower limb electromyographical signals based on autoencoder deep neural network transfer learning,

F. H. Daryakenari, M. Mollahossein, A. Taheri, and G. R. V ossoughi, “Classification of lower limb electromyographical signals based on autoencoder deep neural network transfer learning,” in10th RSI In- ternational Conference on Robotics and Mechatronics, ICRoM 2022, pp. 323–328, Institute of Electrical and Electronics Engineers Inc., 2022

work page 2022

[3] [14]

Deep ensemble learning approach for lower limb movement recognition from multichannel semg signals,

P. Tokas, V . B. Semwal, and S. Jain, “Deep ensemble learning approach for lower limb movement recognition from multichannel semg signals,” Neural Computing and Applications, vol. 36, pp. 7373–7388, 5 2024

work page 2024

[4] [1]

A review on appli- cations of activity recognition systems with regard to performance and evaluation,

S. Ranasinghe, F. A. MacHot, and H. C. Mayr, “A review on appli- cations of activity recognition systems with regard to performance and evaluation,” 8 2016

work page 2016

[5] [2]

Human daily activity recognition for healthcare using wearable and visual sensing data,

X. Liu, L. Liu, S. J. Simske, and J. Liu, “Human daily activity recognition for healthcare using wearable and visual sensing data,” inProceedings - 2016 IEEE International Conference on Healthcare Informatics, ICHI 2016, pp. 24–31, Institute of Electrical and Electronics Engineers Inc., 12 2016

work page 2016

[6] [3]

A survey on human activity recognition using wearable sensors,

´Oscar D. Lara and M. A. Labrador, “A survey on human activity recognition using wearable sensors,”IEEE Communications Surveys and Tutorials, vol. 15, pp. 1192–1209, 2013

work page 2013

[7] [5]

Human lower limb activity recognition techniques, databases, challenges and its applications using semg signal: an overview,

A. Vijayvargiya, B. Singh, R. Kumar, and J. M. R. Tavares, “Human lower limb activity recognition techniques, databases, challenges and its applications using semg signal: an overview,” 11 2022

work page 2022

[8] [7]

Biological cybernetics neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position,

K. Fukushima, “Biological cybernetics neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position,” tech. rep., 1980

work page 1980