arxiv: 2604.02670 · v1 · submitted 2026-04-03 · 💻 cs.LG

Recognition: 1 theorem link

· Lean Theorem

Cross-subject Muscle Fatigue Detection via Adversarial and Supervised Contrastive Learning with Inception-Attention Network

Zitao Lin , Chang Zhu , Wei Meng

Authors on Pith no claims yet

Pith reviewed 2026-05-13 20:35 UTC · model grok-4.3

classification 💻 cs.LG

keywords muscle fatigue detectionsEMGcross-subjectadversarial learningcontrastive learningdomain adaptationInception-attentionrehabilitation

0 comments

The pith

A neural network with adversarial domain training and contrastive learning detects muscle fatigue from sEMG signals across different subjects.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a neural network that identifies muscle fatigue using surface electromyography signals in a way that works for new people without retraining. It processes the signals with an Inception-attention module to extract features, then uses a domain classifier with gradient reversal to strip out subject-specific differences. A supervised contrastive loss pulls together signals of the same fatigue level while pushing apart different levels. This targets the instability caused by signal changes during movement and across individuals. Successful cross-subject performance would allow consistent fatigue tracking in rehabilitation without per-person adjustments.

Core claim

The paper claims that an Inception-attention network as feature extractor, paired with a domain classifier using gradient reversal and a supervised contrastive loss, learns subject-invariant common fatigue features from sEMG signals and delivers 93.54 percent accuracy, 92.69 percent recall, and 92.69 percent F1-score in three-class cross-subject classification tasks.

What carries the argument

Inception-attention module for feature extraction combined with gradient-reversal adversarial domain classification and supervised contrastive loss to enforce subject-invariant fatigue representations.

If this is right

The approach supports real-time fatigue monitoring in rehabilitation programs without requiring subject-specific calibration.
Stable performance during dynamic contractions enables practical use in physical therapy sessions.
High three-class accuracy provides a consistent basis for guiding training and recovery assistance across varied patients.
The architecture offers a template for other biosignal tasks that must ignore individual differences.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same adversarial and contrastive components could transfer to EEG or other biosignals for cross-subject classification problems.
The invariant features extracted may map to shared physiological fatigue processes observable across populations.
Integration into wearable sensors could support ongoing monitoring in sports or daily activity settings.
Further tests on larger multi-activity datasets would clarify the limits of generalization.

Load-bearing premise

Gradient reversal on the domain classifier together with the contrastive loss will consistently separate fatigue-related signal patterns from subject-specific and dynamic-contraction variations.

What would settle it

Applying the trained model to sEMG recordings from an entirely new set of subjects during previously unseen dynamic exercises and measuring accuracy below 80 percent would show the extracted features are not invariant enough.

Figures

Figures reproduced from arXiv: 2604.02670 by Chang Zhu, Wei Meng, Zitao Lin.

**Figure 1.** Figure 1: The structure of the Inception module. B. Adversarial and Supervised Contrastive Learning Domain adversarial training based on GRL is introduced for cross-subject generalization. During forward propagation, GRL functions as an identity operator, denoting GRL(x) = x. Conversely, during backpropagation, gradients are multiplied by a negative scalar defined as −λ to reverse the gradient flow, denoting d dxGRL… view at source ↗

**Figure 2.** Figure 2: The overall structure of the proposed IADAN. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: The adversarial and contrastive learning process. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: The overall experimental setup. (a) Sensors attachment and data [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: The signal preprocessing protocol. (a) The segmented IMU data. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

**Figure 6.** Figure 6: The Model performance analysis result. (a) The loss curve. (b) The [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

**Figure 7.** Figure 7: The t-SNE visualization result. (a) Feature vectors colored according [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗

read the original abstract

Muscle fatigue detection plays an important role in physical rehabilitation. Previous researches have demonstrated that sEMG offers superior sensitivity in detecting muscle fatigue compared to other biological signals. However, features extracted from sEMG may vary during dynamic contractions and across different subjects, which causes unstability in fatigue detection. To address these challenges, this research proposes a novel neural network comprising an Inception-attention module as a feature extractor, a fatigue classifier and a domain classifier equipped with a gradient reversal layer. The integrated domain classifier encourages the network to learn subject-invariant common fatigue features while minimizing subject-specific features. Furthermore, a supervised contrastive loss function is also employed to enhance the generalization capability of the model. Experimental results demonstrate that the proposed model achieved outstanding performance in three-class classification tasks, reaching 93.54% accuracy, 92.69% recall and 92.69% F1-score, providing a robust solution for cross-subject muscle fatigue detection, offering significant guidance for rehabilitation training and assistance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper combines Inception-attention with adversarial domain adaptation and supervised contrastive loss for cross-subject sEMG fatigue detection and reports 93.54% three-class accuracy, but skips the basic diagnostic on whether the domain classifier actually lost subject information.

read the letter

The main takeaway is that this work applies a standard mix of Inception modules, gradient-reversal adversarial training, and supervised contrastive loss to the practical task of making sEMG fatigue classifiers work across different people. The reported 93.54% accuracy, 92.69% recall, and matching F1 on three-class fatigue detection is the central result they highlight for rehabilitation use cases.

Referee Report

1 major / 1 minor

Summary. The paper proposes an Inception-attention network for sEMG-based muscle fatigue detection that incorporates a domain classifier with gradient reversal layer to extract subject-invariant features and a supervised contrastive loss to improve generalization. It reports 93.54% accuracy, 92.69% recall, and 92.69% F1-score on a three-class cross-subject classification task, claiming a robust solution for rehabilitation applications.

Significance. If the invariance claim holds, the work could meaningfully advance practical cross-subject fatigue monitoring by mitigating inter-subject and dynamic-contraction variability in sEMG signals. The adversarial-plus-contrastive design is a reasonable technical choice, but its value depends on verifiable evidence that subject-specific cues have been suppressed.

major comments (1)

[Experimental Results] The accuracy of the domain classifier on held-out subjects after training is not reported. This diagnostic is load-bearing for the central claim: if domain-classifier accuracy remains substantially above chance (1/N_subjects), residual subject information persists and the reported fatigue-classification numbers cannot be attributed to subject-invariant features. (See description of the domain classifier with gradient reversal layer and the experimental results section.)

minor comments (1)

[Abstract] Dataset size, number of subjects, cross-validation scheme, and statistical significance tests are not summarized in the abstract and should be stated explicitly to allow immediate assessment of the reported metrics.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The single major comment is addressed point-by-point below. We agree that the requested diagnostic is important for supporting the central invariance claim and will incorporate it in the revision.

read point-by-point responses

Referee: [Experimental Results] The accuracy of the domain classifier on held-out subjects after training is not reported. This diagnostic is load-bearing for the central claim: if domain-classifier accuracy remains substantially above chance (1/N_subjects), residual subject information persists and the reported fatigue-classification numbers cannot be attributed to subject-invariant features. (See description of the domain classifier with gradient reversal layer and the experimental results section.)

Authors: We agree that the domain-classifier accuracy on held-out subjects is a necessary diagnostic to verify that subject-specific information has been suppressed. In the revised manuscript we will add this result to the experimental section, reporting the post-training domain-classifier accuracy on the held-out subjects together with the chance-level baseline (1/N_subjects). This addition will directly address the concern and allow readers to assess whether the reported fatigue-classification performance can be attributed to subject-invariant features. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML performance on held-out data

full rationale

The paper's central claim consists of measured classification accuracies (93.54% etc.) obtained by training an Inception-attention network with gradient-reversal domain classifier and supervised contrastive loss on sEMG data. No derivation chain, equations, or first-principles result is presented that reduces by construction to its own inputs; the reported metrics are external test-set outcomes rather than fitted quantities renamed as predictions. Standard architectural choices and loss terms are described without self-definitional loops or load-bearing self-citations that collapse the result.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The model relies on standard neural-network assumptions and data-driven feature learning rather than new physical postulates. Free parameters consist of the many tunable hyperparameters in the network and loss weights. No new entities are invented.

free parameters (1)

network hyperparameters and loss weights
Architecture sizes, learning rates, and the relative weighting of classification, adversarial, and contrastive losses are chosen or tuned on data.

axioms (1)

domain assumption sEMG signals contain extractable subject-invariant fatigue features separable from subject-specific variations
This premise underpins the design of the domain classifier with gradient reversal.

pith-pipeline@v0.9.0 · 5471 in / 1227 out tokens · 36731 ms · 2026-05-13T20:35:16.068795+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

[1]

Molecular mechanisms of muscle fatigue,

D. Constantin-Teodosiu and D. Constantin, “Molecular mechanisms of muscle fatigue,”Int. J. Mol. Sci., vol. 22, no. 21, p. 11587, 2021

work page 2021
[2]

Validation of the borg cr10 scale for the evaluation of shoulder perceived fatigue during work-related tasks,

A. Frasie, M. Bertrand-Charette, M. Compagnat, L. J. Bouyer, and J.-S. Roy, “Validation of the borg cr10 scale for the evaluation of shoulder perceived fatigue during work-related tasks,”Appl. Ergon., vol. 116, p. 104200, 2024

work page 2024
[3]

Multilevel assessment of exercise fatigue utilizing multiple attention and convolution network (MACNet) based on surface electromyography,

G. Zhang, B. Yang, P. Zan, and D. Zhang, “Multilevel assessment of exercise fatigue utilizing multiple attention and convolution network (MACNet) based on surface electromyography,”IEEE Trans. Neural Syst. Rehab. Eng., vol. 33, pp. 243–254, 2024

work page 2024
[4]

The behavior of muscle oxygen saturation, oxy and deoxy hemoglobin during a fatigue test in fibromyalgia,

S. Villafaina, P. Tomas-Carus, V . Silva, A. R. Costa, O. Fernandes, and J. A. Parraca, “The behavior of muscle oxygen saturation, oxy and deoxy hemoglobin during a fatigue test in fibromyalgia,”Biomedicines, vol. 11, no. 1, p. 132, 2023

work page 2023
[5]

Non-invasive techniques for muscle fatigue monitoring: A comprehensive survey,

N. Li, R. Zhou, B. Krishna, A. Pradhan, H. Lee, J. He, and N. Jiang, “Non-invasive techniques for muscle fatigue monitoring: A comprehensive survey,”ACM Comput. Surv., vol. 56, no. 9, p. 40, 2024

work page 2024
[6]

A hybrid model for muscle fatigue detection: Integrating time and frequency domain features of semg signals with rpe scale during isometric contraction,

W.-T. Wu, S.-J. Ruan, Y .-W. Tu, Y .-L. Huang, and H.-J. Guo, “A hybrid model for muscle fatigue detection: Integrating time and frequency domain features of semg signals with rpe scale during isometric contraction,”IEEE Trans. Instrum. Meas., vol. 74, pp. 1–11, 2025

work page 2025
[7]

Detecting muscle fatigue during lower limb isometric contractions tasks: a machine learning approach,

J. Sun, C. Zhang, G. Liu, W. Cui, Y . Sun, and C. Zhang, “Detecting muscle fatigue during lower limb isometric contractions tasks: a machine learning approach,”Front. Physiol., vol. 16, p. 1547257, 2025

work page 2025
[8]

Dynamic muscle fatigue state recognition based on deep learning fusion model,

J. Liu, Q. Tao, and B. Wu, “Dynamic muscle fatigue state recognition based on deep learning fusion model,”IEEE Access, vol. 11, pp. 95 079–95 091, 2023

work page 2023
[9]

Complexity analysis of surface electromyography for assessing the myoelectric manifestation of muscle fatigue: A review,

S. Rampichini, T. M. Vieira, P. Castiglioni, and G. Merati, “Complexity analysis of surface electromyography for assessing the myoelectric manifestation of muscle fatigue: A review,”Entropy, vol. 22, no. 5, p. 529, 2020

work page 2020
[10]

MFFNet: Multi-dimensional feature fusion network based on attention mechanism for sEMG analysis to detect muscle fatigue,

Y . Zhang, S. Chen, W. Cao, P. Guo, D. Gao, M. Wang, J. Zhou, and T. Wang, “MFFNet: Multi-dimensional feature fusion network based on attention mechanism for sEMG analysis to detect muscle fatigue,” Expert Syst. Appl., vol. 185, p. 115639, 2021

work page 2021
[11]

Multi-Scale Context Aggregation by Dilated Convolutions

F. Yu and V . Koltun, “Multi-scale context aggregation by dilated convolutions,”arXiv preprint arXiv:1511.07122, 2015

work page Pith review arXiv 2015
[12]

Cbam: Convolutional block attention module,

S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” inProc. Eur . Conf. Comput. Vis., 2018, pp. 3–19

work page 2018
[13]

Re- thinking the inception architecture for computer vision,

C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Re- thinking the inception architecture for computer vision,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 2818–2826

work page 2016
[14]

Domain-adversarial training of neural networks,

Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,”J. Mach. Learn. Res., vol. 17, no. 59, pp. 1–35, 2016

work page 2016
[15]

Supervised contrastive learning,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,”Adv. Neural Inf. Process. Syst., vol. 33, pp. 18 661–18 673, 2020

work page 2020
[16]

Continuous wavelet transform analysis of surface electromyography for muscle fatigue assessment on the elbow joint motion,

T. Triwiyanto, O. Wahyunggoro, H. A. Nugroho, and H. Herianto, “Continuous wavelet transform analysis of surface electromyography for muscle fatigue assessment on the elbow joint motion,”Recent Adv. Electr . Electron. Eng., vol. 15, no. 3, p. 424, 2017

work page 2017
[17]

Hybrid CNN-LSTM-Transformer model for robust muscle fatigue detection during rehabilitation using sEMG signals,

J. Yu, J. Yang, C. Zhu, W. Meng, Q. Liu, and Z. Zhou, “Hybrid CNN-LSTM-Transformer model for robust muscle fatigue detection during rehabilitation using sEMG signals,” inProc. 2024 Int. Conf. Adv. Robot. Mechatron. (ICARM), 2024, pp. 51–56

work page 2024