Pith. sign in

REVIEW 4 major objections 5 minor 16 references

Development of a Low-Cost Prosthetic Hand Using Electromyography and Machine Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A $260 EMG prosthetic hand hits 97% gesture accuracy

desk verdict The hardware is real, but the headline accuracy is an apples-to-oranges average that undersells the actual held-out numbers and overstates the state-of-the-art comparison. read the letter →

arxiv 2411.15533 v1 pith:ZY3UAB7B submitted 2024-11-23 cs.RO cs.HC

Mosab Diab , Ashraf Mohammed , Yinlai Jiang This is my paper · ORCID
classification cs.ROcs.HC
keywords electromyographymyoelectricprosthetichandgestureclassificationtime-domainfeaturesfrequency-domainfeedforwardneuralnetworktactilefeedbacklow-costprosthetics
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 reports a complete myoelectric prosthetic hand built from low-cost, off-the-shelf electronics and a 3D-printed frame, aimed at amputees in developing countries. The authors' central claim is that with only three electrodes placed on the forearm, five hand gestures can be classified with an average accuracy of 97.25% using time-domain features and a shallow feedforward neural network, and 95.85% using frequency-domain features. They further claim that the time-domain route is faster (26 ms per window versus 198 ms) and was therefore adopted, and that the complete hand, including a wrist-rotation mechanism and tactile force feedback, costs about 260 USD. If these claims hold, the work would show that high-accuracy myoelectric control is achievable at a fraction of the cost of commercial prosthetics, using a small number of channels and modest computation.

What carries the argument

The load-bearing mechanism is the pairing of a deliberately minimal signal chain with a shallow feedforward neural network. Three dry electrodes on the upper forearm feed an AD620 instrumentation amplifier (gain 100) and an AD8618-based notch filter that removes 50 Hz power-line noise; the Arduino Mega 2560 samples at 2000 Hz and forms 128-sample windows. In the time domain the features are variance, mean absolute deviation, and waveform length per channel (nine inputs); in the frequency domain they are the eight highest-power FFT bins per channel (24 inputs). A feedforward neural network with one 10-neuron hidden layer (tansig) and a 5-neuron softmax output layer classifies the five gestures, and the classified gesture drives three servo motors, one pair for the fingers and thumb and one for the 180-degree wrist rotation, while force-sensing resistors on the fingertips modulate three coin vibration motors for tactile feedback.

What would settle it

Record the same five gestures from several users, or from the same user on different days with the electrodes removed and replaced, using the paper's exact feature set and network; if the average classification accuracy falls to roughly 85% or below—or near chance for users not seen in training—the 97% claim is specific to a single session and does not generalize.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a three-channel surface-EMG system, using dry carbon-silicon electrodes, an instrumentation amplifier with a 50 Hz notch filter, and an Arduino sampling at 2 kHz, can drive a five-gesture classifier with near-97% accuracy when features are computed from 128-sample moving windows in the time domain (variance, mean absolute deviation, waveform length) and fed to a 9-10-5 feedforward network with tansig and softmax layers. The same network with 24 inputs from the eight highest-power FFT bins per channel reached 95.85% average accuracy but needed 198 ms per window, against 26 ms for the time features. The authors compare these numbers with published results using SVM, ANN, PCA, LDA, and dictionary-learning methods, and conclude that their approach produces better accuracy with fewer electrodes and lower computational cost, with a total component cost of 259.86 USD and three degrees of freedom including wrist rotation.

Load-bearing premise

The load-bearing assumption is that the three forearm sensors, placed on one person's arm without a tested positioning rule, pick up muscle signals that separate the five gestures just as reliably for other users and on later days.

Editorial extensions

If this is right

  • A five-gesture myoelectric control system can be reproduced with roughly 260 USD of components and a single 8-bit microcontroller, which would put functional prosthetics within reach of users in low-resource settings.
  • Time-domain features alone (variance, mean absolute deviation, waveform length) are sufficient to match or beat frequency-domain classification while cutting per-window computation by nearly a factor of eight, so FFT hardware or DSP coprocessors are unnecessary.
  • Three forearm channels, not the six to ten used in many studies, can support five separable gestures, suggesting that electrode count and cost can be reduced without sacrificing accuracy.
  • Adding a user-controlled wrist-rotation gesture pair and fingertip force feedback gives the low-cost hand two of the functions that commercial prosthetics use to improve embodiment and acceptance.

Reading between the lines

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

  • The results were collected under conditions the paper does not fully specify; a multi-user, cross-session study with electrodes removed and reapplied would determine whether the 97% figure transfers beyond the original recording setup, and that test is a natural extension of the authors' claim.
  • If the accuracy does transfer, the same three-electrode, time-domain recipe could be ported to other low-cost platforms and to other gesture sets, since the features and network are generic rather than gesture-specific.
  • The 'arbitrary' electrode placement suggests an untested optimization opportunity: even a modest systematic search over forearm electrode positions and orientations might improve robustness, while the reported result implies that the method is forgiving of placement.
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. This paper reports the development of a low-cost prosthetic hand that uses surface EMG signals from three forearm electrodes to classify five hand gestures. The authors compare time-domain and frequency-domain feature sets, each feeding a shallow feedforward neural network, and report an average classification accuracy of 97.25% (time) and 95.85% (frequency). They also describe a 3D-printed hand with two servo motors plus a wrist rotation mechanism, and a tactile feedback system using force-sensing resistors and vibration motors. The paper claims that this system achieves better accuracy than current state-of-the-art methods with fewer resources and at a total cost of 259.86 USD.

Significance. If the accuracy claims were valid and reproducible, the paper would be a valuable engineering contribution: a complete, low-cost myoelectric prosthetic system integrating acquisition hardware, machine-learning classification, mechanical actuation, and sensory feedback. The authors are transparent about the system's low cost and make a public-health-motivated case for access to prosthetics in developing countries. However, the central quantitative result is not currently supportable. The reported 'Average Accuracy' in Table I cannot be reconstructed as a held-out metric from the numbers given in Section IV, and the paper omits basic experimental details such as subject count, session count, and electrode-relocation protocol. As a result, the headline accuracy and the state-of-the-art comparison in Section V rest on an undefined and likely invalid performance aggregate. The system-building aspects are described with useful detail, but the evaluation does not meet the bar for a journal publication.

major comments (4)
  1. [Section IV.A and Table I] The 'Average Accuracy' in Table I is not defined in the text, and the only way to reproduce the time-domain value of 97.25% from the Section IV numbers is to average the training accuracy (99.7%) with the 'further testing' accuracy (94.8%), i.e., (99.7 + 94.8)/2 = 97.25. For the frequency domain, (98.6 + 93.2)/2 = 95.9, which matches 95.85% within rounding. Training accuracy is not a held-out metric, so this average is not comparable to the test-set accuracies quoted for the literature in Section V. Using only the held-out 'further testing' numbers gives 94.8% (time) and 93.2% (frequency), materially lower than the headline. The Section V claim that the methods 'yield better results with fewer resources' therefore rests on a mismatched and unjustified metric.
  2. [Section III.A.2] The manuscript states that the three electrode channels were 'chosen arbitrarily from the upper forearm,' but it reports no subject count, no number of recording sessions, no electrode-relocation procedure, and no cross-session or cross-user validation. For a myoelectric prosthetic intended for amputee users, this is a load-bearing omission: the reported accuracies could be specific to one person and one electrode placement, and there is no evidence they transfer to the target population. The absence of this information directly undermines the generalization claim implicit in the abstract and Section V.
  3. [Section III.A.2.a (Frequency Domain)] The frequency-domain feature selection chooses the eight frequency bins with the 'highest power values' per channel, but the manuscript does not state whether this selection is performed on the same data used to train and evaluate the neural network, or within a cross-validation loop. If the selection is based on the full dataset, the reported frequency-domain accuracy is optimistically biased. The feature-selection procedure and its relationship to the training/test split need to be described precisely.
  4. [Section IV.A] The dataset partitions are not described: the reader is told that training used 725 entries (frequency) or 1000 entries (time) and that 'further testing' used 500 or 625 entries, but the sizes and provenance of the validation and initial test sets are not given, and the overlap between sets is unknown. This makes it impossible to assess whether the reported training, validation, and test accuracies come from independent samples, which is another reason the Table I average cannot be interpreted.
minor comments (5)
  1. [Abstract and Section III.B] The abstract says the hand has 'two degrees of freedom,' while Section III.B describes the wrist rotation mechanism as adding a 'third degree of freedom.' Please clarify whether the two finger servos count as one degree of freedom or two, and state the total number of active DOFs consistently.
  2. [Section II.C] The sentence 'Sampling with lower frequencies will result in signal aliasing' should read 'Sampling at a lower rate' or 'using a lower sampling frequency'; the current phrasing is technically incorrect.
  3. [Figures 9 and 10] The confusion matrices are not discussed in the text. A brief description of which gestures are most often confused would help the reader judge the practical utility of the classifier.
  4. [General] The paper reports human EMG data but includes no statement on informed consent or ethical approval for human-subject recording; such a statement is expected in a biomedical engineering journal.
  5. [Section V] The phrase 'It is clearly visible that the methods provided in this paper yield better results' is an overstatement given the metric issues noted above; please temper the conclusion to match the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper's accuracy claims are empirical measurements, not consequences of fitted definitions or self-citations.

full rationale

This is an empirical systems paper rather than a formal derivation. The load-bearing claims are the reported classification accuracies (97.25% time-domain, 95.85% frequency-domain), the faster time-domain computation, and the functional prosthetic hand. These are supported by measurements on recorded EMG data, a trained neural network, and hardware tests; there is no equation in which an output is defined in terms of the target result. The neural network was trained on one dataset and then evaluated on separate further-testing datasets (94.8% time-domain, 93.2% frequency-domain), so the held-out numbers are independent empirical results rather than fitted quantities relabeled as predictions. There is no load-bearing self-citation: the reference list contains no works by the present authors. The main numerical concern is that Table I's 'Average Accuracy' is not defined in the paper and may mix training accuracy with held-out test accuracy; that is a reporting or metrics-construction flaw, not a circular reduction, because the underlying held-out results do not reduce by construction to the training inputs. Similarly, choosing the eight highest-power frequency bins from the same dataset used for training and evaluation is a data-leakage concern that can inflate accuracy, but it is not an input-output equivalence of the kind required for a circularity finding. Section V's comparison with prior work rests on benchmark assumptions and metric comparability, which are correctness risks rather than circular reasoning. Therefore no specific circular step can be quoted and exhibited, and the paper should not receive a circularity score above 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central accuracy claim depends on several hand-chosen parameters (window size, sampling rate, hidden layer size, number of selected frequency bins) and on domain assumptions about EMG stability and electrode placement. No dataset or code is provided, so these choices cannot be independently validated.

free parameters (5)
  • Number of frequency bins selected per channel = 8
    The eight bins with the highest power are chosen from the same data used to train and evaluate the network, making this a data-dependent feature selection that can inflate accuracy (Section III.A.2.a).
  • Neural network hidden layer size = 10 neurons
    Network architecture chosen without reported search or justification (Section III.A.2).
  • Moving window size = 128 samples
    Chosen to meet Arduino memory limits; directly affects feature values and accuracy (Section III.A.2).
  • Sampling rate = 2000 Hz
    Set above Nyquist for 500 Hz EMG; a design choice that affects frequency resolution and computational load (Section III.A.2).
  • Number of EMG channels / electrode positions = 3 channels, arbitrary forearm positions
    The choice of muscles is described as arbitrary (Section III.A.2), and the classification accuracy depends on placement.
assumptions (4)
  • domain assumption EMG signals from three forearm muscles contain enough information to distinguish the five gestures.
    The entire classification system rests on this; the paper provides no ablation or feature-selection evidence beyond the reported accuracy.
  • domain assumption The recorded dataset is representative of the target user population and stable across sessions.
    No subject demographics or session protocol are given, so generalization is assumed rather than shown (Section IV).
  • standard math Standard FFT and neural network mathematics behave as expected on the Arduino and MATLAB implementations.
    The paper relies on textbook signal processing and network training (Sections II.D, III.A).
  • ad hoc to paper The 'highest power' frequency bins are the most discriminative features.
    This criterion is selected ad hoc for the five gestures and is not justified by analysis (Section III.A.2.a).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Development of a Low-Cost Prosthetic Hand Using Electromyography and Machine Learning." pith.science (2026). https://pith.science/paper/ZY3UAB7B

@misc{pith2026241115533,
  author       = {Pith},
  title        = {Pith review of: Development of a Low-Cost Prosthetic Hand Using Electromyography and Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZY3UAB7B}},
  note         = {Machine review of arXiv:2411.15533}
}
read the original abstract

Electromyography (EMG) is a measure of muscular electrical activity and is used in many clinical/biomedical disciplines and modern human computer interaction. Myo-electric prosthetics analyze and classify the electrical signals recorded from the residual limb. The classified output is then used to control the position of motors in a robotic hand and a movement is produced. The aim of this project is to develop a low-cost and effective myo-electric prosthetic hand that would meet the needs of amputees in developing countries. The proposed prosthetic hand should be able to accurately classify five different patterns (gestures) using EMG recordings from three muscles and control a robotic hand accordingly. The robotic hand is composed of two servo motors allowing for two degrees of freedom. After establishing an efficient signal acquisition and amplification system, EMG signals were thoroughly analyzed in the frequency and time domain. Features were extracted from both domains and a shallow neural network was trained on the two sets of data. Results yielded an average classification accuracy of 97.25% and 95.85% for the time and frequency domains respectively. Furthermore, results showed a faster computation and response for the time domain analysis; hence, it was adopted for the classification system. A wrist rotation mechanism was designed and tested to add significant functionality to the prosthetic. The mechanism is controlled by two of the five gestures, one for each direction. Which added a third degree of freedom to the overall design. Finally, a tactile sensory feedback system which uses force sensors and vibration motors was developed to enable sensation of the force inflicted on the hand for the user.

Figures

Figures reproduced from arXiv: 2411.15533 by the authors.

Figure 1
Figure 1. EMG signal of relaxed versus contracted muscle [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Frequency domain neural network structure [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. EMG electrodes F. Feedforward Artificial Neural Networks (FFANN) A computational architecture consisting of an input layer, hidden layer/s, and an output layer. Each layer is defined by weights and bias matrices connected to it, and a transfer function that determines the output of that layer. The objec￾tive is to train this architecture to be able to successfully classify input data. This is done by updating the we… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Time domain features [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Mechanical design of the prosthetic hand [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 9
Figure 9. Figure 9: Frequency domain testing data confusion matrix [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [16]

    A hand-modeled feature extraction-based learning network to detect grasps using semg signal,

    M. Baygin, P. D. Barua, S. Dogan, T. Tuncer, S. Key, U. R. Acharya, and K. H. Cheong, “A hand-modeled feature extraction-based learning network to detect grasps using semg signal,” Sensors, vol. 22, no. 5, 2022. [Online]. Available: https://www.mdpi.com/1424-8220/ 22/5/2007 6

    work page 2022

  2. [1]

    (2018, 09) Why limbs

    Limbs International. (2018, 09) Why limbs. [Online]. Available: https://www.limbsinternational.org/why-limbs.html

    work page 2018

  3. [2]

    Access to prosthetic devices in developing countries: Pathways and challenges,

    M. Marino, S. Pattni, M. Greenberg, A. Miller, E. Hocker, S. Ritter, and K. Mehta, “Access to prosthetic devices in developing countries: Pathways and challenges,” in 2015 IEEE Global Humanitarian Tech- nology Conference (GHTC) , 2015, pp. 45–51

    work page 2015

  4. [3]

    The abc of emg,

    P. Konrad, “The abc of emg,” A Practical Introduction to Kinesiolog- ical Electromyography, vol. 1, 01 2005

    work page 2005

  5. [4]

    Bronzino and Donald R

    The Biomedical Engineering Handbook / edited by Joseph D. Bronzino and Donald R. Peterson., 3rd ed., ser. Electrical Engineering Handbook Series. CRC/Taylor & Francis, 2006

    work page 2006

  6. [5]

    Dry-contact and non- contact biopotential electrodes: Methodological review,

    Y . M. Chi, T.-P. Jung, and G. Cauwenberghs, “Dry-contact and non- contact biopotential electrodes: Methodological review,” IEEE Reviews in Biomedical Engineering , vol. 3, pp. 106–119, 2010

    work page 2010

  7. [6]

    Evaluation of the forearm emg signal features for the control of a prosthetic hand,

    R. Boostani and M. Moradi, “Evaluation of the forearm emg signal features for the control of a prosthetic hand,” Physiological measure- ment, vol. 24, pp. 309–19, 06 2003

    work page 2003

  8. [7]

    Emg-controlled prosthetic hand with fuzzy logic classification algorithm,

    B. Tas ¸ar and A. G ¨ulten, “Emg-controlled prosthetic hand with fuzzy logic classification algorithm,” in Modern Fuzzy Control Systems and Its Applications , S. Ramakrishnan, Ed. Rijeka: IntechOpen, 2017, ch. 16. [Online]. Available: https://doi.org/10.5772/intechopen.68242

    work page doi:10.5772/intechopen.68242 2017

Show all 16 references
  1. [8]

    Comparison of different time and frequency domain feature extraction methods on elbow gesture’s emg,

    C. Altin and O. Er, “Comparison of different time and frequency domain feature extraction methods on elbow gesture’s emg,” European Journal of Interdisciplinary Studies , vol. 5, p. 35, 08 2016

    2016

  2. [9]

    Feature extraction of emg signals in time and frequency domain for myopathy, neuropathy and healthy muscle,

    A. Kanwade and V . Bairagi, “Feature extraction of emg signals in time and frequency domain for myopathy, neuropathy and healthy muscle,” 10 2016

    2016

  3. [10]

    Hagan, H

    M. Hagan, H. Demuth, M. Beale, and O. De Jes ´us, Neural Network Design . Martin Hagan, 2014. [Online]. Available: https: //books.google.nl/books?id=4EW9oQEACAAJ

    2014

  4. [11]

    Svm and ann based classification of emg signals by using pca and lda,

    H. Basak, A. Roy, J. B. Lahiri, S. Bose, and S. Patra, “Svm and ann based classification of emg signals by using pca and lda,” ArXiv, vol. abs/2110.15279, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:240070619

    2021 arXiv

  5. [12]

    Classi- fication of myoelectric surface signals of hand movements using su- pervised learning techniques,

    M. C. G. Flores., J. F. M. Medina., and C. L. del ´Alamo., “Classi- fication of myoelectric surface signals of hand movements using su- pervised learning techniques,” in Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (...

    2021

  6. [13]

    Feature extraction and classifi- cation for emg signals using linear discriminant analysis,

    S. Negi, Y . Kumar, and V . M. Mishra, “Feature extraction and classifi- cation for emg signals using linear discriminant analysis,” in 2016 2nd International Conference on Advances in Computing, Communication, & Automation (ICACCA) (Fall) , 2016, pp. 1–6

    2016

  7. [14]

    Multiclass motion identification using myoelectric signals and support vector machines,

    M. Le ´on, J. Guti´errez, L. Leija, R. Mu˜noz, J. de la Cruz, and M. Santos, “Multiclass motion identification using myoelectric signals and support vector machines,” in 2011 Third World Congress on Nature and Biologically Inspired Computing , 2011, pp. 189–194

    2011

  8. [15]

    Myoelectric signal classification of targeted muscles using dictionary learning,

    H.-J. Yoo, H.-j. Park, and B. Lee, “Myoelectric signal classification of targeted muscles using dictionary learning,” Sensors, vol. 19, no. 10, 2019. [Online]. Available: https://www.mdpi.com/1424-8220/ 19/10/2370

    2019

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.