Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Multimodal Surface EMG Hand Gesture Recognition Using Query-Based Transformers for Prosthetic Control

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 ↗

Pith's one-line read A query-based transformer that fuses muscle and accelerometer signals raises balanced hand-gesture accuracy from about 73% to past 92% on standard benchmarks, and sharply narrows the gap between intact users and amputees.

desk verdict A serious and reproducible multimodal sEMG benchmark whose large fusion gains are credible, but whose architecture-specific claims are confounded by unequal training recipes. read the letter →

arxiv 2607.22779 v1 pith:KZU4RKD2 submitted 2026-07-24 cs.LG cs.AI

Federico Del Pup , Elisa Tentori , Manfredo Atzori This is my paper
classification cs.LGcs.AI
keywords surfaceelectromyographyhandgesturerecognitionprostheticcontrolcross-attentionfusionquery-basedtransformermultimodallearningaccelerometerdecoding
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

The paper is trying to establish that the usual reasons gesture-decoding accuracy stalls—convolutional models seeing only local signal structure, and models using only one signal—are addressable in one architecture. It introduces EMG-CrossFormer, which sends each modality through its own encoder, merges the resulting token streams through cascaded cross-attention fusion layers, and decodes the fused representation using learnable gesture queries. On four public benchmarks with up to 40 hand movements, the model reports the best sEMG-only balanced accuracy among the compared models, and adding accelerometer signals lifts accuracy by up to roughly 20 percentage points, with the healthy-versus-amputee gap shrinking from about 15 points to under 5. If the attribution holds, the design principle—capture local and global features and fuse modalities before decoding—is a scalable recipe for prosthetic hand control.

What carries the argument

The load-bearing mechanism is cascaded cross-attention fusion: a stack of transformer decoder layers in which the current fused representation first self-attends, then cross-attends to the token sequence of the next auxiliary modality, so each signal can condition its features on the others before classification. A query-based decoder, using a set of learned hand-gesture query vectors that cross-attend to the fused tokens, lets the model compare candidate gestures jointly. A learned mode token marks each auxiliary modality, and a learnable linear projection aligns all encoders to a shared embedding dimension, which is what makes the number of modalities arbitrary.

What would settle it

Run the strongest baseline CNN under the exact EMG-CrossFormer training protocol—same optimizer, same composite loss, same augmentation, same epochs—and compare; if its accuracy rises to parity, the training recipe rather than the architecture explains the results. Conversely, ablate EMG-CrossFormer by removing the auxiliary contrastive and L1 losses and switching to plain cross-entropy training; if performance barely changes, the losses were not the driver.

Watch

Extended reading notes

Core claim

EMG-CrossFormer is an end-to-end hybrid convolutional-transformer. Each input modality (sEMG, accelerometer, optionally gaze) is turned by a unimodal backbone into a compact token sequence; a cascade of non-causal transformer decoder layers then lets each auxiliary modality attend to, and be attended by, the current fused representation via multi-head cross-attention, ending with a fused token sequence at the native sEMG length. A transformer decoder fed by learnable hand-gesture queries reasons jointly over all candidate gestures, and a shared feed-forward network outputs class probabilities. Evaluated on four standard public benchmarks, the two-backbone variant attains the highest sEMG-onl

Load-bearing premise

The paper credits the cross-attention fusion and query-decoder architecture for the gains, but the proposed model was also trained with a different optimizer, a composite loss, and auxiliary losses that the compared models did not receive, so the architectural attribution is not yet proven.

Editorial extensions

If this is right

  • Multimodal fusion by cross-attention, rather than simple concatenation, yields large accuracy gains: adding accelerometer data improves balanced accuracy by up to +20.65 percentage points in the best configuration.
  • The decoding gap between intact and amputee users shrinks: for the 2D variant, the healthy-amputee difference drops from roughly 15 points on sEMG alone to 3.8–4.8 points once accelerometer data are added.
  • Even in the sEMG-only setting, the model beats classical and convolutional baselines on all four benchmarks, though the advantage is modest (roughly 0.8–2.1 points).
  • A lightweight depthwise variant reaches sub-millisecond inference latency (about 0.77 ms on a GPU after compilation), so the design is not only accurate but deployable in real-time control loops.
  • Incorporating eye-tracking as a third modality does not consistently improve over sEMG+accelerometer, suggesting gaze carries more information about imminent movement than about the executed gesture.

Reading between the lines

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

  • Because no ablation removes the auxiliary supervised-contrastive and handcrafted-feature L1 losses, or changes the optimizer, the architecture-specific claim is not yet isolated; a fair test would train the strongest baseline under the proposed model's full training recipe.
  • The same fusion blocks should transfer to other paired modalities, such as EEG or video, without architectural changes, since the only modality-specific parts are the linear projection and mode token.
  • The convergence between amputee and intact performance once accelerometer data are added suggests kinematic signals partly substitute for missing proprioception; a targeted analysis could test whether the gains concentrate in movements with large limb displacement rather than fine finger gestures.
  • The lack of gaze benefit for executed gestures points to a natural extension: use gaze for predicting gesture onset or imminent intent rather than classifying an already-executed movement.
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.

Referee Report

4 major / 5 minor

Summary. The paper introduces EMG-CrossFormer, a hybrid convolutional-transformer architecture for multimodal hand gesture recognition. Modality-specific backbones produce token sequences, which are fused through cascaded cross-attention decoder layers and decoded with learnable gesture queries. The model is evaluated on NinaPro DB2, DB3, DB7, and DB10, using sEMG only, sEMG+ACC, and sEMG+ACC+gaze, against six baselines. The main claims are that joint local-global feature modeling improves sEMG-only decoding (best balanced accuracies 72.33/52.48/79.16/73.49%) and that multimodal fusion substantially amplifies this benefit (90.66/80.40/92.79/92.06% with sEMG+ACC). Results are reported with subject-level paired statistics, FDR-corrected Wilcoxon tests, and bootstrap confidence intervals.

Significance. If the architectural attribution were supported, this would be a useful contribution: the code is openly available, the evaluation spans four public datasets, the statistical treatment is careful, and the paper reports honest non-significant comparisons. However, the experiments as designed do not isolate the proposed architecture from training-protocol differences or from an extra split-sEMG fusion condition in the 'unimodal' setting. The benchmark numbers may stand, but the central causal claim that cross-attention fusion and query-based decoding drive the gains is not currently established. The missing ablations and missing DB10 baselines are fixable within the paper's scope, so the result is potentially salvageable.

major comments (4)
  1. [§2.4.4–2.4.5 vs. §3.1–3.2] The central attribution claim is confounded by an unequal training protocol. EMG-CrossFormer is trained with the composite loss in Eq. (4) (CE + supervised contrastive on z_dec and all z_enc + auxiliary CE + L1 on handcrafted features) and with separate learning rates for backbones (5e-4) vs. fusion/decoder (5e-5), whereas the deep baselines use a uniform 5e-4 learning rate and plain categorical cross-entropy. No ablation removes the auxiliary losses from EMG-CrossFormer, and no baseline is retrained with the composite loss. In the sEMG-only setting the reported gains over MKCNN are only 1.28–2.09 pp (Table 2), a range in which additional supervision and per-module optimization differences could plausibly account for the effect. The claim that cross-attention fusion and query decoding are responsible for the improvements therefore needs an ablation that matches training recipes across mo
  2. [§2.3.5] The 'unimodal' (sEMG-only) setting is not unimodal for EMG-CrossFormer. The sEMG signal is split into forearm (8 channels) and upper-arm (4 channels) streams, which are provided as two separate input modalities through the fusion cascade. Baselines receive all 12 channels as a single input. This gives EMG-CrossFormer an additional fusion condition and a different input representation, making the sEMG-only comparisons in Table 2 not apples-to-apples. A configuration with all 12 sEMG channels as a single modality should be reported, or the baselines should be adapted to the same split setting.
  3. [Table 4, §3.3] The abstract claims EMG-CrossFormer 'consistently achieve[s] the highest performance' across four datasets, but DB10 is evaluated only with the two EMG-CrossFormer variants; no competing model is run on DB10. Thus the 'consistently highest' claim is not directly evidenced on DB10, and no paired statistical comparison to baselines is available there. The authors should either run the baselines on DB10 or restrict the claim to DB2/DB3/DB7.
  4. [§2.4.2, §3.2] No transformer-based sEMG baseline is included even though TraHGR [10] is cited and used to motivate the approach, and no ablation isolates the two named architectural components. In particular, the paper never replaces the cascaded cross-attention fusion with concatenation, nor the query-based decoder with a global-pooling classifier head. Without such ablations, the contribution of the cross-attention fusion and query-decoder modules is not separated from the overall pipeline or from the training recipe.
minor comments (5)
  1. [Header / title] The running header reads 'Quey-Based Transformers'; the intended wording is 'Query-Based'.
  2. [Table 1] The DB3 row lists 11(A) subjects, but the footnote says subjects 6 and 7 were discarded; the same issue appears for DB10, where 30+15(A) is listed while the text says six subjects were discarded. Please clarify whether the table reports original or post-exclusion counts.
  3. [§2.4.5] The phrase 'trained using the same set of training hyperparameters' is ambiguous: it is not clear whether data augmentation and the LAMB optimizer were applied to the baselines or only to EMG-CrossFormer. Please state explicitly which hyperparameters are shared and which differ.
  4. [Eq. (4)] The ground-truth handcrafted features h are described as 'sEMG features', but in the multimodal setting the model receives ACC and gaze inputs as well. Please clarify whether h is defined only from sEMG or from all modalities, and specify how it is computed for the auxiliary modalities.
  5. [§3.1] In the text, the comparison 'EMG-CF2D vs MKCNN on DB3 at 100 ms' reports CI = [−0.16, 2.42], which is identical to the CI reported for DB2 at 100 ms in the same paragraph. Please check whether this is a transcription error.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: EMG-CrossFormer results are empirical benchmarks on held-out NinaPro repetitions; no prediction reduces to a fitted input or self-cited theorem.

full rationale

The paper's central claim is an empirical comparison on public NinaPro data, not a derivation from fitted constants. No equation defines a target quantity in terms of the claimed result; no fitted parameter is renamed as a prediction; and no uniqueness theorem from prior work is used to force the architecture. The self-cited TransformEEG [26] and the co-authored MKCNN [7] are used as implemented backbone components whose designs are described and benchmarked in this paper; they are not invoked as external proof of the fusion/decoder benefit. Comparing EMG-CF2D against MKCNN is a legitimate head-to-head that isolates the added fusion/decoder modules. The composite loss (Eq. 4) and different optimizers/losses for baselines create a training-recipe confound, which may weaken causal attribution of the gains, but this is a validity/confounding issue, not circularity. The paper is therefore self-contained with respect to its empirical benchmarks, and the correct circularity finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 7 assumptions · 2 invented entities

The central claim rests on standard deep-learning assumptions: the public datasets are correctly labeled, the intra-subject split is a meaningful performance measure, and the transformer/attention mechanisms are transferable to sEMG. The main free parameters are training hyperparameters chosen by hand or by empirical tuning. No new physical entities are introduced; the mode tokens and gesture queries are learned model parameters, not independently testable inventions.

free parameters (6)
  • Composite loss weights λ_ce, λ_supcon, λ_aux, λ_hand = 3.0, 1.5, 1.0, 2.0
    Set 'based on empirical tuning' (Section 2.4.4); these directly shape the objective that produces the reported accuracies.
  • Initial learning rates = 5.0e-4 (backbones), 5.0e-5 (fusion/decoder)
    Chosen by hand (Section 2.4.5); no sensitivity analysis is reported.
  • Gradient clipping max norm = 0.1
    Selected because 1.0 'yielded worse results' (Section 2.4.5).
  • Transformer embedding dimension, decoder layers, heads = d=128, 4 decoder layers, 8 heads, SwiGLU
    Fixed architecture hyperparameters (Section 2.3.3 and supplementary A.1); not swept in the paper.
  • Window lengths and overlap = 100/150/200 ms, 10% shift
    Three window sizes are evaluated; results vary by window length (Section 2.2).
  • Augmentation hyperparameters = segments 4/6/8; compression 1.25/1.5/2.0; stretch 1.0/1.5/2.0; p=0.7
    Randomly sampled from a predefined grid (Section 2.4.3); no ablation of augmentation strength.
assumptions (7)
  • domain assumption NinaPro ground-truth labels and per-subject repetition data are reliable.
    The whole benchmark depends on the correctness and consistency of publicly released NinaPro labels and preprocessing (Section 2.1).
  • domain assumption Intra-subject, repetition-based train/test split is a valid proxy for prosthetic decoding performance.
    Section 2.4.1 uses train reps 1,3,4,6 and test reps 2,5; if cross-repetition generalization does not reflect real user adaptation, the reported accuracies overstate practical performance.
  • domain assumption DETR-style learnable queries and transformer decoder layers transfer to sEMG gesture decoding.
    Section 2.3.3, inspired by DETR [20]; no proof that this inductive bias suits physiological signals.
  • domain assumption Splitting sEMG channels into forearm/upper-arm groups yields complementary muscle-synergy information.
    Used in Section 2.3.5 for the unimodal EMG-CrossFormer design, referencing muscle synergies [29,30].
  • standard math Standard transformer multi-head attention, layer norm, and FFN definitions are correct.
    Equations 2–3 rely on published MHA/LN/FFN definitions from Vaswani et al. [27].
  • standard math Wilcoxon signed-rank test and bootstrap confidence intervals are valid on subject-level balanced accuracy differences.
    Section 2.4.6; assumes exchangeable subjects and paired differences.
  • domain assumption Upsampling and imputation of accelerometer/gaze signals to the sEMG sampling rate preserved temporal alignment.
    Sections 2.1–2.2 rely on the original dataset authors' upsampling; if alignment is imperfect, multimodal gains could be artifacts.
invented entities (2)
  • Mode token per auxiliary modality
    purpose: Learnable additive bias that marks non-sEMG modalities before cross-attention fusion (Eq. 1).
    A learned architectural component; it has no external falsifiable handle.
  • Learnable hand-gesture query embeddings
    purpose: Decoder inputs that attend to the fused representation and produce class logits per query (Section 2.3.3).
    Standard DETR-style learned parameters; not a physical entity, no independent evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Surface EMG Hand Gesture Recognition Using Query-Based Transformers for Prosthetic Control." pith.science (2026). https://pith.science/paper/KZU4RKD2

@misc{pith2026260722779,
  author       = {Pith},
  title        = {Pith review of: Multimodal Surface EMG Hand Gesture Recognition Using Query-Based Transformers for Prosthetic Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KZU4RKD2}},
  note         = {Machine review of arXiv:2607.22779}
}
read the original abstract

Hand gesture recognition via surface electromyography (sEMG) is fundamental to prosthetic control. In this field, deep learning approaches have become the gold standard. However, current architectures struggle to scale; model performance typically decreases as the number of hand movements increases. Performance degradation is tied to the increased statistical complexity of decoding expanded gesture sets and compounded by the limitations of state-of-the-art methods, which primarily rely on low-latency unimodal convolutional architectures. Convolutions operate locally, limiting model's ability to capture long-range sequential patterns. Unimodal setups cannot leverage complementary information from coordinated signals characterizing movement execution, such as inertial and eye-tracking data. These limitations motivate architectures that integrate local and global features across multimodal physiological sequences. To bridge this gap, this study introduces EMG-CrossFormer, an end-to-end hybrid convolutional-transformer for seamless multimodal integration. EMG-CrossFormer combines representations from an arbitrary number of unimodal encoders through cascaded cross-attention fusion layers, and decodes the fused representations using learnable gesture queries. EMG-CrossFormer was evaluated on four NinaPro datasets (DB2, DB3, DB7, and DB10) and benchmarked against six state-of-the-art models using an increasing number of modalities. Using only sEMG, EMG-CrossFormer achieved mean accuracies of 72.33%, 52.48%, 79.16%, and 73.49% on DB2, DB3, DB7, and DB10, respectively. Incorporating inertial signals improved performance to 90.66%, 80.40%, 92.79%, and 92.06%. These results show that joint local-global feature modeling improves sEMG-only decoding and that multimodal fusion substantially amplifies this benefit, underscoring the value of both design principles for complex hand gesture recognition.

Figures

Figures reproduced from arXiv: 2607.22779 by the authors.

Figure 1
Figure 1. Schematic representation of EMG-CrossFormer. The model combines representations (𝐳enc𝑚 ) from unimodal backbones through cascaded fusion layers (transformer decoder layers). The transformer decoder block decodes the fused representation (𝐳fuse) using learnable hand movement queries. A final shared FFN outputs the hand movement predictions from the decoded representation (𝐳dec). EMG-CrossFormer can integrate an arbit… view at source ↗
Figure 2
Figure 2. Hand-movement decoding performance of EMG-CrossFormer across datasets, window lengths, model variants, and input modalities. Balanced accuracy is shown for EMG-CF1D, based on a 1D depthwise convolutional backbone, and EMG-CF2D, based on a 2D multi-scale convolutional backbone. Results are stratified by window length (100 ms, 150 ms, and 200 ms) and input configuration: sEMG-only decoding is shown in blue, sEMG + ACC… view at source ↗
Figure 1
Figure 1. expands on the EMG-CrossFormer architecture, reporting the structure of the network modules used to train the model, the unimodal backbones, and the initialization hyperparameters. Linear GAP 𝐡! Linear GAP 𝐬𝐟𝐮𝐬𝐞 Linear GAP 𝐬𝐞𝐧𝐜𝟏 one for each 𝐳𝐞𝐧𝐜! Training Only Depthwise Conv 1D Batch Norm Batch Norm ELU Avg Pooling Dropout Depthwise Conv 1D ELU + x 2 Batch Norm ELU Max Pooling Dropout Circular Conv 2D Kernel: (H,10… view at source ↗
Figures from the paper (4 more)
Figure 2
Figure 2. Figure 2: sEMG-only paired comparisons between EMG-CF2D and the competing models for NinaPro DB2, DB3, and DB7 across 100, 150, and 200 ms windows. Boxplots summarize the subject-level balanced accuracies; points represent individual subjects, and gray lines connect paired obser…
Figure 3
Figure 3. Figure 3: sEMG-only paired comparisons between EMG-CF1D and the competing models for NinaPro DB2, DB3, and DB7 across 100, 150, and 200 ms windows. Boxplots summarize the subject-level balanced accuracies; points represent individual subjects, and gray lines connect paired obser…
Figure 4
Figure 4. Figure 4: sEMG + ACC paired comparisons between EMG-CF2D and the competing multimodal models for NinaPro DB2, DB3, and DB7 across 100, 150, and 200 ms windows. Boxplots summarize the subject-level balanced accuracies; points represent individual subjects, and gray lines connect …
Figure 5
Figure 5. Figure 5: sEMG + ACC paired comparisons between EMG-CF1D and the competing multimodal models for NinaPro DB2, DB3, and DB7 across 100, 150, and 200 ms windows. Boxplots summarize the subject-level balanced accuracies; points represent individual subjects, and gray lines connect …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 5 canonical work pages

  1. [10]

    Zabihi, E

    S. Zabihi, E. Rahimian, A. Asif, A. Mohammadi, TraHGR: Trans- former for hand gesture recognition via electromyography, IEEE Trans. Neural Syst. Rehabil. Eng. 31 (2023) 4211–4224. doi: 10.1109/TNSRE.2023.3324252

    arXiv 2023

  2. [1]

    N.Jiang,C.Chen,J.He,J.Meng,L.Pan,S.Su,X.Zhu,Bio-robotics researchfornon-invasivemyoelectricneuralinterfacesforupper-limb prostheticcontrol:a10-yearperspectivereview,Natl.Sci.Rev.10(5) (2023) nwad048.doi:10.1093/nsr/nwad048

    work page doi:10.1093/nsr/nwad048 2023

  3. [2]

    X. Lv, C. Dai, H. Liu, Y. Tian, L. Chen, Y. Lang, R. Tang, J. He, GesturerecognitionbasedonsEMGusingmulti-attentionmechanism for remote control, Neural Comput. Appl. 35 (19) (2023) 13839– 13849. doi:10.1007/s00521-021-06729-6

    work page doi:10.1007/s00521-021-06729-6 2023

  4. [3]

    Marinelli, N

    A. Marinelli, N. Boccardo, F. Tessari, D. Di Domenico, G. Caserta, M. Canepa, G. Gini, G. Barresi, M. Laffranchi, L. De Michieli, M. Semprini, Active upper limb prostheses: a review on current state and upcoming breakthroughs, Prog. Biomed. Eng. 5 (1) (2023) 012001. doi:10.1088/2516-1091/acac57

    work page doi:10.1088/2516-1091/acac57 2023

  5. [4]

    Jiang, S

    N. Jiang, S. Dosen, K.-R. Muller, D. Farina, Myoelectric control of artificial limbs—is there a need to change focus? [in the spotlight], IEEESignalProcess.Mag.29(5)(2012)152–150. doi:10.1109/ MSP.2012.2203480

    arXiv 2012

  6. [5]

    Ozhikenova, N

    B.Abdikenov,D.Zholtayev,K.Suleimenov,N.Assan,K.Ozhikenov, A. Ozhikenova, N. Nadirov, A. Kapsalyamov, Emerging frontiers in roboticupper-limbprostheses:Mechanisms,materials,tactilesensors and machine learning-based EMG control: A comprehensive review, Sensors 25 (13) (2025).doi:10.3390/s25133892

    work page doi:10.3390/s25133892 2025

  7. [6]

    LeCun, Y

    Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444

    2015

  8. [7]

    Fratti, N

    R. Fratti, N. Marini, M. Atzori, H. Müller, C. Tiengo, F. Bassetto, A multi-scale CNN for transfer learning in sEMG-based hand gesture recognition for prosthetic devices, Sensors 24 (22) (2024).doi: 10.3390/s24227147

    doi:10.3390/s24227147 2024

Show all 47 references
  1. [8]

    N.K.Karnam,S.R.Dubey,A.C.Turlapaty,B.Gokaraju,EMGHand- Net: A hybrid CNN and Bi-LSTM architecture for hand activity classification using surface EMG signals, Biocybern. Biomed. Eng. 42 (1) (2022) 325–340.doi:10.1016/j.bbe.2022.02.005

    2022 doi

  2. [9]

    3302–3305

    M.Jabbari,R.N.Khushaba,K.Nazarpour,EMG-basedhandgesture classification with long short-term memory deep recurrent neural networks,in:202042ndAnnualInternationalConferenceoftheIEEE Engineering in Medicine & Biology Society (EMBC), 2020, pp. 3302–3305. doi:10.1109/EMBC44109.2020.9175279

    2020

  3. [11]

    Stival, S

    F. Stival, S. Michieletto, M. Cognolato, E. Pagello, H. Müller, M. Atzori, A quantitative taxonomy of human hand grasps, J. NeuroEng. Rehabil. 16 (1) (2019) 28. doi:10.1186/ s12984-019-0488-x

    2019

  4. [12]

    Atzori, A

    M. Atzori, A. Gijsberts, C. Castellini, B. Caputo, A.-G. M. Hager, S. Elsig, G. Giatsidis, F. Bassetto, H. Müller, Electromyography data for non-invasive naturally-controlled robotic hand prostheses, Sci. Data 1 (1) (2014) 140053.doi:10.1038/sdata.2014.53

    2014 doi

  5. [13]

    Atzori, M

    M. Atzori, M. Cognolato, H. Müller, Deep learning with convolu- tional neural networks applied to electromyography data: A resource F. Del Pup et al.:Preprint Page 11 of 21 Multimodal Hand Gesture Recognition Using Quey-Based Transformers for the classification of movements fo...

    2016

  6. [14]

    Atzori, A

    M. Atzori, A. Gijsberts, S. Heynen, A.-G. M. Hager, O. Deriaz, P. van der Smagt, C. Castellini, B. Caputo, H. Müller, Building the Ninapro database: A resource for the biorobotics community, in: 2012 4th IEEE RAS & EMBS International Conference on Biomed- ical Robotics and Bio...

    2012

  7. [15]

    T.R.Farrell,R.F.Weir,Theoptimalcontrollerdelayformyoelectric prostheses,IEEETrans.NeuralSyst.Rehabil.Eng.15(1)(2007)111–

    2007

  8. [16]

    Rasool, Transformers in time-series analysis: A tutorial, Circuits Syst

    S.Ahmed,I.E.Nielsen,A.Tripathi,S.Siddiqui,R.P.Ramachandran, G. Rasool, Transformers in time-series analysis: A tutorial, Circuits Syst. Signal Process. 42 (12) (2023) 7433–7466.doi:10.1007/ s00034-023-02454-8

    2023

  9. [17]

    NeuroEng

    A.Krasoulis,I.Kyranou,M.S.Erden,K.Nazarpour,S.Vijayakumar, Improved prosthetic hand control with concurrent use of myoelectric and inertial measurements, J. NeuroEng. Rehabil. 14 (1) (2017) 71. doi:10.1186/s12984-017-0284-4

    2017 doi

  10. [18]

    Atzori, A

    M. Atzori, A. Gijsberts, H. Müller, B. Caputo, Classification of hand movements in amputated subjects by sEMG and accelerometers, in: 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2014, pp. 3545–3549.doi:10. 1109/EMBC.2014.6944388

    2014

  11. [19]

    Jiang, H

    B. Jiang, H. Wu, Q. Xia, G. Li, H. Xiao, Y. Zhao, NKDFF-CNN: A convolutional neural network with narrow kernel and dual-view feature fusion for multitype gesture recognition based on sEMG, Digit.SignalProcess.156(2025)104772. doi:10.1016/j.dsp. 2024.104772

    2025

  12. [20]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, S. Zagoruyko, End-to-end object detection with transformers, in: Computer Vision – ECCV 2020, Springer International Publishing, Cham, 2020, pp. 213–229. doi:10.1007/ 978-3-030-58452-8\_13

    2020

  13. [21]

    M.Cognolato,A.Gijsberts,V.Gregori,G.Saetta,K.Giacomino,A.- G. M. Hager, A. Gigli, D. Faccio, C. Tiengo, F. Bassetto, B. Caputo, P. Brugger, M. Atzori, H. Müller, Gaze, visual, myoelectric, and inertialdataofgraspsforintelligentprosthetics,Sci.Data7(1)(2020)

    2020

  14. [22]

    doi:10.1016/ j.jbiomech.2010.01.027

    C.J.DeLuca,L.DonaldGilmore,M.Kuznetsov,S.H.Roy,Filtering the surface EMG signal: Movement artifact and baseline noise con- tamination,J.Biomech.43(8)(2010)1573–1579. doi:10.1016/ j.jbiomech.2010.01.027

    2010

  15. [23]

    F. R. Hampel, The influence curve and its role in robust estimation, JournaloftheAmericanStatisticalAssociation69(346)(1974)383–

    1974

  16. [24]

    J.W.Grootjen,H.Weingärtner,S.Mayer,Uncoveringandaddressing blink-related challenges in using eye tracking for interactive systems, in: Proceedings of the 2024 CHI Conference on Human Factors in ComputingSystems,CHI’24,AssociationforComputingMachinery, New York, NY, USA, 2024, ...

    2024 doi

  17. [25]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperative style, high-performance deep learning library, Proc. Adv. Neural Inf. Process. Syst. 32 (2019). doi:10.48550/arXiv. 1912.01703

    2019 doi arXiv

  18. [26]

    Zanola, L

    F.DelPup,R.Brun,F.Iotti,E.Paccagnella,M.Pezzato,S.Bertozzo, A. Zanola, L. F. Tshimanga, H. Müller, M. Atzori, TransformEEG: Towards improving model generalizability in deep learning-based EEG Parkinson’s disease detection, Neurocomputing 664 (2026) 132075. doi:10.1016/j.neucom...

    2026

  19. [27]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, I. Polosukhin, Attention is all you need, in: AdvancesinNeuralInformationProcessingSystems,Vol.30,Curran Associates, Inc., 2017, pp. 5998–6008.doi:10.48550/arXiv. 1706.03762

    2017 doi arXiv

  20. [28]

    Shazeer, GLU variants improve transformer, arXiv Preprint (Feb 2020)

    N. Shazeer, GLU variants improve transformer, arXiv Preprint (Feb 2020). doi:10.48550/arXiv.2002.05202

    2020 doi arXiv

  21. [29]

    K. Zhao, Z. Zhang, H. Wen, B. Liu, J. Li, A. d’Avella, A. Scano, Muscle synergies for evaluating upper limb in clinical applications: A systematic review, Heliyon 9 (5) (May 2023).doi:10.1016/ j.heliyon.2023.e16202

    2023

  22. [30]

    Brambilla, M

    C. Brambilla, M. Atzori, H. Müller, A. d’Avella, A. Scano, Spa- tial and temporal muscle synergies provide a dual characterization of low-dimensional and intermittent control of upper-limb move- ments, Neuroscience 514 (2023) 100–122. doi:10.1016/j. neuroscience.2023.01.017

    2023 doi

  23. [31]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: Machine learning in Python, Journal of Machine Learning Res...

    2011

  24. [32]

    Virtanen, R

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, et al., SciPy 1.0: fundamental algorithms for scientific computing in python, Nat. Methods 17 (3) (2020) 261–272.doi:10.1038/ s41592-019-0686-2

    2020

  25. [33]

    S.Seabold,J.Perktold,statsmodels:Econometricandstatisticalmod- eling with python, in: 9th Python in Science Conference, 2010, pp. 92–96. doi:10.25080/Majora-92bf1922-011

    2010 doi

  26. [34]

    6 (60) (2021) 3021.doi:10.21105/joss.03021

    M.L.Waskom,Seaborn:statisticaldatavisualization,J.OpenSource Softw. 6 (60) (2021) 3021.doi:10.21105/joss.03021

    2021 doi

  27. [35]

    R.Ganiga,M.S.N.,W.Choi,S.Pan,ResNet1D-basedpersonaliden- tification with multi-session surface electromyography for electronic health record integration, Sensors 24 (10) (2024).doi:10.3390/ s24103140

    2024

  28. [36]

    doi: 10.1016/j.mlwa.2021.100152

    A.Saeed,V.Ungureanu,B.Gfeller,Senseandlearn:Self-supervision foromnipresentsensors,Mach.Learn.Appl.6(2021)100152. doi: 10.1016/j.mlwa.2021.100152

    2021

  29. [37]

    F. N. Fritsch, J. Butland, A method for constructing local monotone piecewisecubicinterpolants,SIAMJ.Sci.Stat.Comput.5(2)(1984) 300–304. doi:10.1137/0905021

    1984 doi

  30. [38]

    Khosla, P

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y. Tian, P. Isola, A. Maschinot, C. Liu, D. Krishnan, Supervised contrastive learning, in: Advances in neural information processing systems, Vol. 33, CurranAssociates,Inc.,2020,pp.18661–18673. doi:10.48550/ arXiv.2004.11362

    2020 doi arXiv

  31. [39]

    Y. You, J. Li, S. Reddi, J. Hseu, S. Kumar, S. Bhojanapalli, X. Song, J. Demmel, K. Keutzer, C.-J. Hsieh, Large batch optimization for deeplearning:TrainingBERTin76minutes,in:InternationalConfer- ence on Learning Representations, 2020, p. 17.doi:10.48550/ arXiv.1904.00962

    2020 doi arXiv

  32. [40]

    Kingma, J

    D. Kingma, J. Ba, Adam: A method for stochastic optimization, in: Proc. Int. Conf. Learn. Represent. (ICLR), San Diego, CA, USA, 2015, p. 13.doi:10.48550/arXiv.1412.6980

    2015 doi arXiv

  33. [41]

    1 (6) (1945) 80–83.doi:10.2307/3001968

    F.Wilcoxon,Individualcomparisonsbyrankingmethods,Biometrics Bull. 1 (6) (1945) 80–83.doi:10.2307/3001968

    1945 doi

  34. [42]

    Benjamini, Y

    Y. Benjamini, Y. Hochberg, Controlling the false discovery rate: A practical and powerful approach to multiple testing, J. R. Stat. Soc., B (Methodol.) 57 (1) (1995) 289–300. URL http://www.jstor.org/stable/2346101

    1995

  35. [43]

    doi:10.1038/s41597-020-0380-3

    doi

  36. [44]

    Abnar, W

    S. Abnar, W. Zuidema, Quantifying attention flow in transformers, in: Proceedings of the 58th annual meeting of the association for computational linguistics, 2020, pp. 4190–4197.doi:10.18653/ v1/2020.acl-main.385. F. Del Pup et al.:Preprint Page 12 of 21 Multimodal Hand Gestu...

    2020

  37. [46]

    L. T. Taverne, M. Cognolato, T. Bützer, R. Gassert, O. Hilliges, Video-based prediction of hand-grasp preshaping with application to prosthesiscontrol,in:2019InternationalConferenceonRoboticsand Automation(ICRA),2019,pp.4975–4982. doi:10.1109/ICRA. 2019.8794175

    2019

  38. [118]

    doi:10.1109/TNSRE.2007.891391

    2007

  39. [393]

    doi:10.1080/01621459.1974.10482962

    1974

Pith tools

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