{"id":"da3fd5ce-75aa-4d1a-b103-e13aa4dd523d","arxiv_id":"2411.15533","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A low-cost myoelectric prosthetic hand with wrist rotation and tactile feedback reportedly classifies five gestures with 97.25% accuracy using time-domain EMG features and a shallow feedforward network.","lead":"This paper reports a low-cost, 3D printed prosthetic hand that reads muscle signals from three forearm muscles and uses a small neural network to recognize five gestures, with claimed accuracy near 97%. It is relevant because affordable myoelectric hands could improve prosthetic access in developing countries.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I's 'Average Accuracy' is not a held-out metric: 97.25% equals the mean of 99.7% training and 94.8% second-test accuracy, so the headline and Section V comparison are unsupported.","rationale":"The paper is a plausible prototype description, but its central claim is a specific accuracy number. The reported per-partition numbers in Section IV do not reproduce Table I's 'Average Accuracy' under any standard formula; the exact match for time-domain (99.7 + 94.8)/2 = 97.25 makes a training-inclusive average the most likely explanation. This is a correctness risk, not a disagreement with field consensus. The reader's weakest assumption was generalization across users and sessions; my primary concern is more fundamental and more checkable: even within the reported data, the claimed accuracy cannot be reconstructed. Thus I partially agree with the reader. A single arithmetic check on the confusion matrices can settle the issue. If the check confirms a training-inclusive average, the headline accuracy and the Section V superiority claim fail; if a defined held-out aggregate yields 97.25%, only reporting and protocol fixes would be needed. The mechanical and tactile-feedback work is not in question. I recommend keeping the reader's REJECT verdict because the quantitative basis for the paper's main claim is currently unsupported.","tokens_in":6614,"tokens_out":8860,"duration_ms":78108,"concrete_test":"Recompute the confusion matrices in Figures 9 and 10 with Eq. (7): count total correct and total classifications per class and overall for the 500-entry and 625-entry 'further testing' sets. Then determine whether Table I's 'Average Accuracy' equals (training accuracy + further-test accuracy)/2 for each domain; if it does, re-run the Section V comparison using only the held-out further-test accuracies (94.8% time, 93.2% frequency) against the same baselines. The headline should be accepted only if a defined held-out aggregate independently yields 97.25% and still beats matched baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 97.25% / 95.85% average accuracy in the abstract and Table I, together with the Section V statement that these results exceed the state of the art. Section IV reports four numbers per domain: training, validation, initial test, and a 'further testing' set. No formula or data partition is given for Table I's 'Average Accuracy.' For the time domain, (99.7 + 94.8)/2 = 97.25 exactly; for the frequency domain, (98.6 + 93.2)/2 = 95.9, within rounding of 95.85. The most plausible reconstruction is that the headline averages training accuracy with the second held-out test accuracy. That is not a valid aggregate for comparison with cited works, which report test-set accuracies. Using the held-out numbers alone gives 94.8% (time) and 93.2% (frequency), materially below the headline. Since Section V compares these averages to literature test-set accuracies such as 96.38% in [16], the 'better results with fewer resources' claim rests on mismatched metrics. The paper also gives no subject count, session count, or electrode-relocation protocol (Section III.A.2), so even the held-out numbers may not generalize; but the undefined accuracy aggregate is the more immediate obstacle to accepting the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6926,"tokens_out":3287,"duration_ms":30044,"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":[{"comment":"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.","section":"Section IV.A and Table I"},{"comment":"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.","section":"Section III.A.2"},{"comment":"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.","section":"Section III.A.2.a (Frequency Domain)"},{"comment":"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.","section":"Section IV.A"}],"minor_comments":[{"comment":"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.","section":"Abstract and Section III.B"},{"comment":"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.","section":"Section II.C"},{"comment":"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.","section":"Figures 9 and 10"},{"comment":"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.","section":"General"},{"comment":"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.","section":"Section V"}],"recommendation":"reject","confidential_remarks":"The paper reports an engineering prototype with a useful humanitarian motivation, but the evaluation is not yet of journal quality. The central accuracy claim is not reproducible from the reported numbers and appears to mix training accuracy with held-out test accuracy. In addition, the lack of subject/session information means the generalization claim is unsupported. I would encourage the authors to re-run their experiments with a clearly defined held-out test set, report per-subject and per-session results, and describe feature selection inside a cross-validation loop; a substantially revised manuscript could then be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline accuracy in this paper is not a held-out number. Table I's 'Average Accuracy' reconstructs to the mean of training accuracy and the second held-out test accuracy (99.7 and 94.8 give 97.25 exactly). That is not a valid metric for comparing against cited works that report test-set accuracies. The actual held-out numbers are 94.8% (time) and 93.2% (frequency), which are still reasonable for three channels but below the headline and likely below the comparison in [16]. So the Section V claim of 'better results with fewer resources' is unsupported.\n\nThat said, the paper has real engineering substance. It presents a complete low-cost prosthetic hand: 3D-printed, two finger servos plus a wrist rotation mechanism, three-channel dry electrodes, an Arduino-based classifier, and a tactile feedback loop with FSRs and vibration motors. The total cost is about 260 USD. The time-domain vs frequency-domain comparison under equal sampling and window size is useful, and the choice of the faster time-domain approach for the embedded classifier is sensible.\n\nThe softer spots beyond the metric: there is no subject count, no cross-session or electrode-relocation protocol, and no variance bars. The electrodes were placed 'arbitrarily' on the upper forearm, so the accuracy numbers may apply to one person's single session. Without multi-subject or cross-session testing, the generalization to amputee users is unknown. The feature selection (choosing the eight highest-power FFT bins) is done on the same data used for training and evaluation, which can inflate accuracy. That is a secondary issue but worth noting.\n\nThe paper is not a breakthrough, but it is a plausible system paper for low-cost prosthetics. The main quantitative claim needs to be redone with a defined held-out metric and at least a small multi-subject protocol. The hardware description and tactile feedback design are worth preserving. I would not cite it at present, but I would send it to peer review because a competent referee could extract usable engineering and quantify the true classification performance.\n\nFor your reading group, it's a maybe—useful as a case study in how averaging training and test accuracy can mislead.\n\nTake care.","headline":"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.","tokens_in":7434,"tokens_out":4056,"would_cite":false,"duration_ms":31634,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A $260 EMG prosthetic hand hits 97% gesture accuracy","keywords":["electromyography","myoelectric prosthetic hand","gesture classification","time-domain features","frequency-domain features","feedforward neural network","tactile feedback","low-cost prosthetics"],"falsifier":"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.","tokens_in":6407,"feed_emoji":"🦾","tokens_out":6128,"duration_ms":50813,"temperature":0.7,"pith_summary":"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.","feed_headline":"A $260 EMG prosthetic hand hits 97% gesture accuracy","feed_subtitle":"With three forearm sensors and a small neural network, five grips plus wrist rotation run on an Arduino, the authors report.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"It supplies the EMG amplitude and frequency ranges, the filtering rationale, and the electrode-placement guidelines that the authors say they followed.","marker":"[3]"},{"why":"It states the criteria for choosing EMG feature spaces for prosthetic-hand control, justifying the selected time- and frequency-domain features.","marker":"[6]"},{"why":"It lists the common time- and frequency-domain features and classifiers, including waveform length and the neural-network approach used here.","marker":"[8]"},{"why":"It provides the feedforward neural-network design background, including flexibility and low classification cost, that the classification stage relies on.","marker":"[10]"},{"why":"It gives the SVM/ANN baseline with PCA and LDA that the paper compares against to support its better-accuracy claim.","marker":"[11]"},{"why":"It reports 85.25% accuracy with three channels using dictionary learning, serving as the direct three-channel comparison the paper claims to beat.","marker":"[15]"},{"why":"It reports 96.38% accuracy for six gestures with a hand-modeled feature network, a recent baseline against which the 97.25% result is positioned.","marker":"[16]"}],"fun_headline_variants":["5 gestures, 3 sensors, $260: EMG hand hits 97%","EMG hand with 3 sensors nails 5 gestures at 97%","Low-cost EMG hand: 97% accuracy, 3DoF, $260","3-channel EMG to 5 gestures: 97% on a budget"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["5 gestures, 3 sensors, $260: EMG hand hits 97%","EMG hand with 3 sensors nails 5 gestures at 97%","Low-cost EMG hand: 97% accuracy, 3DoF, $260","3-channel EMG to 5 gestures: 97% on a budget"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000324,"raw_usage":{"total_tokens":1867,"prompt_tokens":1044,"completion_tokens":823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":735}},"tokens_in":660,"tokens_out":823,"duration_ms":6768,"temperature":1.0,"reasoning_tokens":735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:10:30.603221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"The abc of emg,","cited_arxiv_id":null,"evidence_quote":"It supplies the EMG amplitude and frequency ranges, the filtering rationale, and the electrode-placement guidelines that the authors say they followed."},{"cited_title":"Evaluation of the forearm emg signal features for the control of a prosthetic hand,","cited_arxiv_id":null,"evidence_quote":"It states the criteria for choosing EMG feature spaces for prosthetic-hand control, justifying the selected time- and frequency-domain features."},{"cited_title":"Comparison of different time and frequency domain feature extraction methods on elbow gesture’s emg,","cited_arxiv_id":null,"evidence_quote":"It lists the common time- and frequency-domain features and classifiers, including waveform length and the neural-network approach used here."},{"cited_title":"Hagan, H","cited_arxiv_id":null,"evidence_quote":"It provides the feedforward neural-network design background, including flexibility and low classification cost, that the classification stage relies on."},{"cited_title":"SVM and ANN based Classification of EMG signals by using PCA and LDA","cited_arxiv_id":"2110.15279","evidence_quote":"It gives the SVM/ANN baseline with PCA and LDA that the paper compares against to support its better-accuracy claim."},{"cited_title":"Myoelectric signal classification of targeted muscles using dictionary learning,","cited_arxiv_id":null,"evidence_quote":"It reports 85.25% accuracy with three channels using dictionary learning, serving as the direct three-channel comparison the paper claims to beat."},{"cited_title":"A hand-modeled feature extraction-based learning network to detect grasps using semg signal,","cited_arxiv_id":null,"evidence_quote":"It reports 96.38% accuracy for six gestures with a hand-modeled feature network, a recent baseline against which the 97.25% result is positioned."}],"review_version":1}