{"id":"aed52a34-21f9-4c8e-a11b-5ebe41f45ed8","arxiv_id":"2411.14131","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new open dataset and benchmark show that wrist sEMG can classify finger force intentions during isometric (no-movement) contractions, with about 93 percent single-day accuracy and lower cross-day and cross-subject accuracy.","lead":"Researchers built a wristband that reads muscle signals while a person grips an object and tenses fingers without moving, aiming to send messages covertly during meetings or sports. They recorded a 10-person dataset, released a toolbox, and benchmarked ten algorithms, finding single-session accuracy near 93 percent but much lower performance across days and people.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Covertness premise is untested: the 'gesture-free' interaction may be visually detectable, and the Section I overclaim that sEMG can 'always accurately' determine intention is contradicted by the paper's own cross-day and cross-subject results in Table I.","rationale":"The reader's weakest_assumption identifies the covertness premise, and I agree this is the most load-bearing soft spot. The paper's scientific contribution is a dataset and benchmark for decoding isometric finger-force patterns from wrist sEMG; that part is coherent and reasonably well documented. The failure mode is not that the classification accuracy is fake, but that the framing as 'gesture-free' and 'covert' is unsupported. The protocol itself—gripping a cylinder and varying finger force—has obvious potential for observable cues, and no experiment addresses detectability. I also share the reader's concern about the 'always' claim, which the authors' own cross-day/cross-subject numbers contradict; however, that is a wording fix, whereas the untested covertness is central to the paper's motivation. The proposed video-based perceptual study would settle the issue concretely. Meanwhile, the verdict CONDITIONAL remains appropriate: the technical benchmark should be accepted only with the condition that the authors either provide evidence for non-detectability or materially soften the covertness claim. The dataset/toolbox inaccessibility is a separate reproducibility issue that also supports keeping the verdict conditional.","tokens_in":19109,"tokens_out":3463,"duration_ms":37554,"concrete_test":"Run a controlled perceptual study: video-record subjects performing the Section III.B protocol with the wristband and cylindrical object visible, including rest periods and all 12 force modes; show randomly ordered clips to naive observers and ask them to (1) indicate whether the subject is sending a message and (2) identify which force mode was executed. If message-detection or mode-classification accuracy significantly exceeds chance (e.g., via a McNemar test against a chance baseline), the covertness premise is falsified and the 'gesture-free' central claim must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central motivation is covert, gesture-free communication, asserted in Section I and the Abstract. For that claim to hold, the act of sending a message must be observationally undetectable. The paper never tests this. Section III.B instructs subjects to hold a cylindrical object and 'maintain a tight grip' while applying different finger-force patterns; this is a physical activity that may be visible through hand tension, grip changes, or the conspicuous wristband and object. No perceptual study, video analysis, or observational experiment is reported, so the claim that messages are sent 'without any gesture' and 'difficult for non-collaborators to detect' rests on assertion rather than evidence. A second, related load-bearing issue is the overgeneralized claim that 'regardless of whether the limbs produce movements or not, sEMG signals can always accurately determine whether a person has the intention to control limb movements.' The authors' own benchmarks show this is false in general: the best cross-day accuracy for 6 classes is about 0.71 (L-EMGNet, 500 ms) and the best cross-subject accuracy is about 0.62 (2DCNN, 500 ms), which are far below the single-day 0.9367 headline. While single-day feasibility may be real, the advertised capability of deciding intention 'always accurately' and covertly is not supported by the measurements. The covertness failure would not invalidate the dataset as a technical artifact, but it would collapse the paper's stated purpose of enabling hidden message exchange in sensitive scenarios.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a gesture-free hand intention recognition scheme based on wrist sEMG and isometric contraction, with the stated purpose of enabling covert message transmission in settings where collaborators and non-collaborators coexist. The authors built a custom eight-channel myoelectric wristband and host software, collected a dataset from ten subjects in two sessions spaced 5–10 days apart, and developed an analysis toolbox. They benchmarked ten sEMG decoding methods under single-day, cross-day, and cross-subject protocols for 6-class and 12-class tasks at 250 ms, 500 ms, and 750 ms windows, and report online accuracies of 94%, 90%, and 88% on three subjects. The central claim is that the system recognizes hand intentions without any detectable gesture, enabling covert communication.","tokens_in":19345,"tokens_out":2726,"duration_ms":26933,"significance":"If the results are substantiated, the dataset and toolbox could be a useful community resource, and the benchmark itself is honest in reporting that cross-day and cross-subject performance drops substantially and that the authors' own method (L-EMGNet) does not always win, with 2DCNN outperforming it in cross-subject settings. The single-day accuracy of 0.9367 ± 0.0483 (6-class, 750 ms) suggests practical feasibility under the same-session condition. However, the paper's central motivation, covertness, is asserted rather than tested, the abstract and Section I overclaim generality, and the claimed open-source release is not actually verifiable because the manuscript gives only a placeholder link. These issues bear directly on whether the paper's headline contribution is supported.","major_comments":[{"comment":"The covertness premise is untested. The Abstract and Section I claim that messages can be conveyed \"without detection by non-collaborators\" and that the scheme \"hides the action of sending messages,\" but no perceptual, observational, or video-analysis study is reported. Section III.B instructs subjects to hold a cylindrical object and \"maintain a tight grip on the object with their fingers during the force application process\" to keep signals gesture-free; visible finger tension, grip changes, or the wristband itself may be observable to others. The claim of covert, undetectable interaction is load-bearing for the paper's motivation and currently rests on assertion, not evidence. Either add an explicit observer study or substantially temper the claim to \"isometric, low-amplitude force patterns\" without stating they are hidden from non-collaborators.","section":"Abstract; Section I; Section III.B"},{"comment":"The statement in Section I that \"regardless of whether the limbs produce movements or not, sEMG signals can always accurately determine whether a person has the intention to control limb movements\" is contradicted by the paper's own results. The best cross-day 6-class accuracy is 0.7136 (L-EMGNet, 500 ms) and the best cross-subject 6-class accuracy is 0.6216 (2DCNN, 500 ms), both far below the single-day 0.9367. The word \"always\" and the implication of near-perfect intention decoding are not supported by the measurements. This sentence should be rewritten as a conditional, evidence-based claim, and the related claims in the Abstract and Conclusion should be aligned with the reported cross-day and cross-subject gaps.","section":"Section I (Introduction); Table I"},{"comment":"The benchmark ranking claims are made without any statistical hypothesis testing. For example, in the 6-class 250 ms single-day condition, L-EMGNet (0.9107 ± 0.0534) and 2DCNN (0.9047 ± 0.0595) differ by 0.6 percentage points with large per-subject standard deviations; similar small gaps occur in several other conditions. The text states that one method \"achieves the best performance\" and ranks first/second/third, but with ten subjects and overlapping standard deviations, these differences may not exceed chance. Add paired statistical tests (e.g., paired t-test or Wilcoxon signed-rank test with multiple-comparison correction), or at minimum report per-condition effect sizes and explicitly qualify the rankings as descriptive.","section":"Section V.A and V.B; Table I"},{"comment":"The paper repeatedly states that \"all data, hardware, software, and methods are open-sourced\" and refers to a project website, but the only URL in the text is the literal placeholder \"click here\" (in the Abstract and at the end of Section IV.B). A dataset-oriented submission cannot be evaluated or reproduced without a working link or a stable repository identifier. Provide the actual URL, a DOI, or an institutional repository link, and describe access conditions.","section":"Abstract; Section IV; Conclusion; project website"}],"minor_comments":[{"comment":"The sentence \"the system selects MX-06 and uses Bluetooth SPP protocol\" is unclear about what MX-06 is; please specify the Bluetooth module or clarify the component name.","section":"Section II.A"},{"comment":"The phrase \"trail number\" should be \"trial number,\" and similar typos occur elsewhere (e.g., \"euqated\" in Section VI.E).","section":"Section III.B"},{"comment":"The function name \"data segmentation\" mixes English and spacing; consider standardizing code identifiers, but this is purely stylistic.","section":"Section IV.A.1"},{"comment":"The Shapiro-Wilk test is misspelled as \"Shapro-Wilk\" in Section V.B; also, the claim that all benchmark results follow a normal distribution is based on a small sample (ten subjects per condition) and should be stated with that caveat.","section":"Section V.A"},{"comment":"The confusion matrix subfigures contain residual font-encoding artifacts (e.g., repeated \"/uni00000030...\" strings); please regenerate or replace the figure files with clean vector graphics.","section":"Section VI.B; Fig. 5"},{"comment":"The caption says \"results on the left are from [8], results on the right from this study,\" but the table uses slash-separated pairs inside each cell; the intended column layout should be clarified or restructured.","section":"Table II"},{"comment":"The force-intensity analysis covers only three subjects, and the ranges include 0–100 N as a single bin; please state explicitly that these results are preliminary and not statistically powered.","section":"Section VI.D"}],"recommendation":"major_revision","confidential_remarks":"The dataset and benchmark are potentially valuable, and the authors are honest about cases where their own method loses. However, the manuscript as submitted cannot be accepted because the central covertness claim is untested, the abstract overstates the accuracy, and the open-source availability is unverifiable. The first two issues are fixable by tempering language and adding an observational experiment or an explicit limitation; the third is fixable with a working link. I recommend major revision rather than rejection because the technical artifact, if made available and accurately described, could still make a useful contribution to the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine dataset-plus-benchmark paper. The core technical claim—isometric single-finger and combined-finger force patterns are decodable from wrist sEMG without an intended gesture—holds up for single-day use, with best 6-class accuracy near 0.94 and online results at 94/90/88%. What is new is the task definition and the dataset: ten subjects, two sessions 5–10 days apart, four locomotion speeds, plus a toolbox reproducing ten baselines. The benchmark is honestly reported: chronological single-day split, cross-day and cross-subject splits, variance, and cases where their own L-EMGNet loses to 2DCNN. Signal-quality metrics and the force-intensity analysis add value.\n\nThe soft spots are real but mostly fixable. The most serious is the covert-communication framing. The paper claims concealment and protection from non-collaborator detection, but nothing tests whether the grip forces are visually detectable. The protocol tells subjects to maintain a tight grip on a cylinder; a wristband plus a gripped object is not obviously invisible. 'Gesture-free' is fine as a protocol description; 'covert' and 'difficult to detect' are assertions. Either test detectability or drop the framing.\n\nSecond, Section I states sEMG 'can always accurately determine' intention. Their own Table I shows best cross-day 6-class accuracy around 0.71 and cross-subject around 0.62. The word 'always' is not supported, and the conclusion itself acknowledges the cross-day and cross-subject weakness. That should be reworded.\n\nThird, the dataset and code are promised but the URL is a placeholder ('click here'). For a dataset paper that blocks the contribution. Fourth, the benchmark lacks statistical comparison between methods; top gaps are often under 1–2% on ten subjects, so the method rankings may be noise. Mean±std without paired tests is not enough. Minor: n=10 is small and online n=3 is small, but acceptable for a feasibility dataset.\n\nWho this is for: sEMG/myoelectric-control and HCI researchers looking for a new benchmark with cross-day and cross-subject splits on an isometric force task. It deserves a serious referee: the hardware, paradigm, and benchmark breadth are substantial. I would not accept it as-is; the repository must be live, the covertness claim tested or removed, and the 'always' language cut. The central feasibility result is probably right; the framing oversells it.","headline":"Real dataset and benchmark for isometric wrist-sEMG intention decoding; single-day feasibility is solid, but the covert-communication claim is untested and the 'always' overclaim should be cut.","tokens_in":19950,"tokens_out":3355,"would_cite":false,"duration_ms":31224,"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":"Surface electromyography recorded at the wrist can decode which finger a person intends to press even when the hand produces no visible movement, because isometric muscle contractions generate sEMG signals without limb displacement.","keywords":["sEMG","gesture-free hand intention recognition","isometric contraction","covert communication","myoelectric wristband","benchmark dataset","human-computer interaction"],"falsifier":"Videotape subjects performing the 6 or 12 finger-force patterns while holding a cylindrical object, and ask naive observers to guess which (if any) finger is being pressed; if detection accuracy is significantly above chance, the 'gesture-free' and covert-transmission claim fails even though the sEMG classification may be accurate. Alternatively, an experiment where subjects press with the same force patterns but with visible finger movement should yield similar sEMG classification; if it does not, isometric contraction is not the actual carrier.","tokens_in":18865,"feed_emoji":"💪","tokens_out":4065,"duration_ms":36022,"temperature":0.7,"pith_summary":"This paper claims that wrist-worn surface electromyography (sEMG) can reliably identify which hand movement a person intends even when the hand produces no visible motion, because muscles can contract isometrically—generating electrical activity and force without changing length. To support this, the authors built a custom eight-channel myoelectric wristband, recorded a ten-subject dataset at four movement speeds, released a preprocessing and decoding toolbox, and benchmarked ten classifiers. The headline result is a 6-class single-day accuracy of 93.67% with a 750 ms window, and online tests on three subjects reaching 94%, 90%, and 88% accuracy. If the claim holds, this enables a covert communication channel: discrete finger-force patterns, not gestures, carry messages that non-collaborators cannot see.","feed_headline":"Wrist sEMG decodes hidden hand intentions with no visible gesture","feed_subtitle":"Finger forces pressed against an object carry 6–12 messages; best single-day accuracy is 93.7%.","key_machinery":"The load-bearing mechanism is isometric muscle contraction: muscles generate force and sEMG activity while maintaining constant length, so finger forces can be applied against an object without visible limb movement. The experimental apparatus is a self-built eight-channel myoelectric wristband (with a three-axis accelerometer) plus a paradigm in which subjects hold a cylindrical object and apply finger forces in twelve prescribed patterns while standing, walking, or jogging. The decoding side uses a benchmark of ten classifiers, with the best performer being L-EMGNet, a lightweight convolutional network integrated into the toolbox.","core_discovery":"The paper's central claim is that 'regardless of whether the limbs produce movements or not, sEMG signals can always accurately determine whether a person has the intention to control limb movements.' Concretely, it asserts that eight-channel sEMG from a wristband, combined with isometric contraction of the fingers around a cylindrical object, encodes enough information to distinguish up to twelve hand-force intentions (single-finger and multi-finger combinations) with no externally visible gesture. The benchmark supports this within a single recording day, with L-EMGNet achieving 0.9367 ± 0.0483 accuracy on 6 classes and 0.7917 ± 0.0740 on 12 classes at 750 ms windows, while cross-day and cross-subject performance drops substantially (best 6-class cross-day 0.7136; best cross-subject 0.6216). The paper attributes the drop to electrode displacement and physiological variability across sessions.","pith_inferences":["Beyond the paper: the 'gesture-free' claim is never tested against human perception; a study where observers try to detect which finger is being pressed while subjects grip an object could falsify the covertness motivation even if classification accuracy holds.","Beyond the paper: the same wrist-sEMG signal might support a continuous force-regression channel rather than discrete classes, potentially increasing information rate beyond 12 states.","Beyond the paper: the strong single-day but weak cross-subject results suggest the signal depends heavily on individual muscle-recruitment patterns, so calibration-free use across people may require domain adaptation or much larger training corpora.","Beyond the paper: combining the already-recorded IMU channels with sEMG might improve cross-day robustness, since the accelerometer could help align wrist position across sessions."],"forward_implications":["If the scheme works as claimed, covert messaging can be performed by applying finger forces inside a hand holding an object, with no visible gesture, and decoded by a wrist-worn device in about 0.3–0.35 seconds.","The benchmark provides a public dataset and toolbox so other groups can reproduce and extend gesture-free sEMG recognition.","Longer analysis windows (750 ms vs 250 ms) improve accuracy at the cost of latency, giving a tunable accuracy/speed trade-off for real-time use.","The large cross-day and cross-subject accuracy drops imply that practical deployment needs adaptation or per-session calibration, a direction the paper states for future work."],"supporting_citations":[{"why":"Supplies the isometric contraction theory that underpins the claim that muscle can contract without limb displacement, the theoretical basis for gesture-free intent detection.","marker":"[7]"},{"why":"Shows wrist EMG outperforms forearm EMG for single-finger and multi-finger gestures, supporting the choice of a wristband electrode layout.","marker":"[8]"},{"why":"Demonstrates feasibility of wrist-worn real-time hand gesture recognition via sEMG and IMU sensing, a direct precedent for the wearable system.","marker":"[9]"},{"why":"Provides the L-EMGNet model that achieves the best single-day and cross-day benchmark results, the key classifier behind the headline accuracy.","marker":"[19]"},{"why":"Informs the choice of 250/500/750 ms analysis windows and the associated latency-accuracy trade-off for myoelectric control.","marker":"[22]"}],"fun_headline_variants":["Wrist sEMG reads hidden hand intentions without any gesture","Hidden hand intentions decoded by wrist sEMG, no moves needed","sEMG wristband reveals 12 hand intentions, no gesture visible","Gesture-free hand intention: sEMG hits 93.7% same-day accuracy","Wrist sensors decode silent hand signals, up to 12 classes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that holding an object and pressing with individual fingers produces no visible cue—finger tension, grip change, or arm stiffening—that an observer could notice, since the paper never tests this.","fun_headline_variants_meta":{"raw":{"variants":["Wrist sEMG reads hidden hand intentions without any gesture","Hidden hand intentions decoded by wrist sEMG, no moves needed","sEMG wristband reveals 12 hand intentions, no gesture visible","Gesture-free hand intention: sEMG hits 93.7% same-day accuracy","Wrist sensors decode silent hand signals, up to 12 classes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1356,"prompt_tokens":1016,"completion_tokens":340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":245}},"tokens_in":632,"tokens_out":340,"duration_ms":3412,"temperature":1.0,"reasoning_tokens":245,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:30:29.371910+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Videotape subjects performing the 6 or 12 finger-force patterns while holding a cylindrical object, and ask naive observers to guess which (if any) finger is being pressed; if detection accuracy is significantly above chance, the 'gesture-free' and covert-transmission claim fails even though the sEMG classification may be accurate. Alternatively, an experiment where subjects press with the same force patterns but with visible finger movement should yield similar sEMG classification; if it does not, isometric contraction is not the actual carrier.","supporting_citations":[{"cited_title":"Regulation of isometric contraction in skeletal muscle,","cited_arxiv_id":null,"evidence_quote":"Supplies the isometric contraction theory that underpins the claim that muscle can contract without limb displacement, the theoretical basis for gesture-free intent detection."},{"cited_title":"Electromyography- based gesture recognition: Is it time to change focus from the forearm to the wrist?","cited_arxiv_id":null,"evidence_quote":"Shows wrist EMG outperforms forearm EMG for single-finger and multi-finger gestures, supporting the choice of a wristband electrode layout."},{"cited_title":"Feasibility of wrist-worn, real-time hand, and surface gesture recognition via sEMG and IMU sensing,","cited_arxiv_id":null,"evidence_quote":"Demonstrates feasibility of wrist-worn real-time hand gesture recognition via sEMG and IMU sensing, a direct precedent for the wearable system."},{"cited_title":"Command and control method of grouping based on covert gesture interaction technology,","cited_arxiv_id":null,"evidence_quote":"Provides the L-EMGNet model that achieves the best single-day and cross-day benchmark results, the key classifier behind the headline accuracy."},{"cited_title":"Decoding HD-EMG signals for myoelectric control-how small can the analysis window size be?","cited_arxiv_id":null,"evidence_quote":"Informs the choice of 250/500/750 ms analysis windows and the associated latency-accuracy trade-off for myoelectric control."}],"review_version":1}