{"id":"0a59b362-9cb5-4f53-9816-2daff166470d","arxiv_id":"2607.26401","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A three-pose-calibrated optical-flow sonomyography algorithm gives continuous 1-DOF cursor control from arbitrary sensor locations, plus a first pilot of continuous 2-DOF control.","lead":"Sonomyography uses ultrasound to watch muscle tissue move under the skin; this paper turns those movements into continuous cursor control with only a three-pose calibration, from sensors placed anywhere on the arm, neck, or torso. In a pilot with nine participants, including three with tetraplegia, the 1-DOF system worked across placements, and a preliminary 2-DOF version was demonstrated for the first time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sensor locations were pre-adjusted to satisfy feature-quality criteria; 'all tested sensor locations' is therefore unrepresentative and placement-agnosticism is not fully supported.","rationale":"The reader identified the key preconditions (sufficient features, motion magnitude, and orthogonality) and noted the asterisked incomplete trials as conflicting with the abstract's 'all tested sensor locations' claim. I agree, and I add a sharper framing: the evaluation protocol actively selected placements to satisfy those preconditions, so even the successful trials do not establish that the system is agnostic to sensor placement in the sense of working at arbitrary locations. This is a load-bearing concern because the title and abstract center on 'sensor-placement-agnostic' control, and the current evidence only shows the algorithm works when investigators can find a suitable spot. The proposed concrete test would either substantiate or falsify the agnosticism claim. A CONDITIONAL verdict is appropriate: the paper should be revised to acknowledge this selection bias and to report placement success rates, or the claim should be downgraded. The reader's conditional verdict already captures the need for re-reporting, but our refinement points to a specific methodological gap that a fixed-grid or acceptance-rate analysis would resolve.","tokens_in":12739,"tokens_out":6931,"duration_ms":70566,"concrete_test":"Run a fixed grid of probe positions (e.g., 9 positions centered over the biceps with 1 cm spacing) on a cohort of participants, without any real-time image-based placement adjustment. Apply the same 3-pose calibration and 1-DOF trajectory task at each position, and pre-register a success criterion (e.g., RMSE below a threshold derived from a chance baseline). Report the success rate and per-position RMSE. If most random positions succeed, the placement-agnostic claim is supported; if a substantial fraction fail, the claim should be qualified to 'locations with sufficient tissue deformation features.' Alternatively, re-analyze recorded ultrasound data to determine how many initial probe attachments were rejected or adjusted before a placement was 'tested'; if this acceptance rate is low, the current 'all tested' claim is biased.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim of sensor-placement-agnostic control is governed by preconditions stated in §III-A: the image must contain enough trackable features, features must move sufficiently between calibration poses, and for 2-DOF the up/down and right motions must be sufficiently orthogonal. However, the evaluation protocol in §IV-C explicitly allowed investigators to adjust probe placement 'based on real-time inspection of the ultrasound image to ensure sufficient adherence to §III-A-defined signal qualities' before testing. The six sensor locations were therefore not sampled independently or randomly; they were actively selected to satisfy the algorithm's requirements. The abstract's statement that 'all participants achieved continuous 1-DOF control at all tested sensor locations' is consequently an artifact of this pre-screening, not evidence that arbitrary locations work. The asterisked nonperformant configurations in Fig. 3 (participants 2, 5, and 7, at specific placement–motion pairs) further show that even after adjustment some configurations failed to complete the full protocol. A truly placement-agnostic system must be tested on locations chosen without real-time image-based correction, or at minimum the initial-attachment failure rate and the number of adjustments per placement should be reported. Without this, the headline claim overstates what has been demonstrated and the practical deployment value is unclear.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a real-time sonomyography (SMG) control system based on sparse Lucas–Kanade optical flow tracking of B-mode ultrasound features. A 1-DOF controller is calibrated from three user-defined poses (neutral, up, down) and maps the average normalized displacement of tracked points to cursor position (Eq. 1). A preliminary 2-DOF extension adds a computer-aided calibration that asks users to find a second motion approximately orthogonal to the first. The system is evaluated on 3 cervical SCI survivors and 6 uninjured participants across 6 sensor placements, with an abstract claim that all participants achieved continuous 1-DOF control at all tested sensor locations and <5.5% tracking error at least one placement, and that the 2-DOF results constitute the first location-agnostic multi-DOF continuous SMG control demonstration. The authors also release code and data via OpenMyoControl on SimTK.","tokens_in":12996,"tokens_out":3493,"duration_ms":38934,"significance":"If the claims hold, the work is significant: it offers a calibration-lean, placement-agnostic alternative to data-hungry SMG/sEMG decoders, with a plausible mechanism for exploiting passive tissue motion for users with tetraplegia. Strengths include the open-source release, real-time operation, inclusion of SCI participants, and exploration of anatomically diverse sensor locations. The central 1-DOF idea is simple and intuitive, and the paper is honest about several limitations in Section VI. However, the headline claims outrun the evidence: the evaluation protocol pre-screened sensor placements, incomplete trials are folded into averages, and the 2-DOF 'location-agnostic' claim rests on a single self-selected placement per participant. These issues are fixable with more transparent reporting and qualified claims, so the result is not fatally compromised.","major_comments":[{"comment":"The placement-agnostic claim is undermined by the evaluation protocol. §IV-C states that probe placement was 'adjusted as needed by investigators based on real-time inspection of the ultrasound image to ensure sufficient adherence to section III-A-defined signal qualities.' Because the tested locations were thus filtered to satisfy the algorithm's preconditions (trackable features, sufficient displacement, and, for 2-DOF, orthogonality), the abstract's 'all tested sensor locations' cannot be read as evidence for arbitrary placement. To support the central claim, the paper should report the number of initial attachment failures, the number/type of adjustments per placement, and ideally include a condition in which placements are chosen without real-time image-based correction.","section":"§IV-C and §III-A"},{"comment":"The reporting of incomplete trials is misleading. The Fig. 3 caption says values are averaged across 'all successful trials' while asterisks indicate participants who 'were unable to complete all trials' for given placements/motions (participants 2, 5, 7; BIC, TRA, DEL, FLE, SCM). These non-completions are silently absorbed into the per-placement means, so the statement in §V-A.2 that 'every participant demonstrated some level of control with every sensor placement' and the abstract's 'all participants achieved continuous 1-DOF control at all tested sensor locations' are not supported. The paper should report trial completion rates separately, treat incomplete configurations as failures in any aggregate claim, and provide a defined criterion for 'control' (e.g., RMSE relative to a no-control or chance baseline).","section":"Fig. 3 and §V-A.2"},{"comment":"No error bars, confidence intervals, or chance baseline are provided for the primary 1-DOF results. RMSE is reported as a unitless fraction of screen height, but without a comparison to chance-level tracking (e.g., cursor held at center, uncontrolled drift, or a random mapping) the numerical RMSE values do not themselves establish that a participant achieved voluntary 'control.' This is particularly important for the strongest claim, '<5.5% tracking error at at least one placement,' which needs a statistical or pre-specified threshold. Adding per-participant/per-placement variability and a null-condition comparison would make the claim falsifiable and reproducible.","section":"§V-A.1 and §V-A.2"},{"comment":"The 2-DOF 'location-agnostic' claim is not supported by the experimental design. Each participant was tested at a single self-selected sensor location, with unlimited recalibration and repositioning (§IV-D), and the results show substantial coupling, 'binding,' and inaccessible workspace regions for several participants (Fig. 6). The abstract's 'location-agnostic multi-DOF continuous SMG-based control' should be qualified to 'single arbitrary location per participant' or 'proof-of-concept at a user-selected location.' The quantitative success criteria for the drawing task (Fig. 7) should also be stated.","section":"§V-B, Fig. 6, Fig. 7"}],"minor_comments":[{"comment":"The regression in Fig. 5 has R²=0.0466; the phrase 'significantly anticorrelated' is technically true (p=0.002) but the effect size is very small. Please temper the wording and discuss practical relevance.","section":"§V-A.4 / Fig. 5"},{"comment":"The 2-DOF orthogonality threshold ('at least 50% of available points within 30°') is presented as an empirical choice. A sensitivity analysis or at least a note on how this threshold affects calibration success would strengthen the paper.","section":"§III-D"},{"comment":"The statement that functional attachment locations were 'largely easy to find' (footnote 1) is anecdotal. Please report the actual number of placement attempts and adjustments per participant.","section":"§IV-C"},{"comment":"Some formatting issues: '12×12cm' should be '12 cm × 12 cm'; '29.2±14.2y' should be formatted consistently; the OpenMyoControl release should include a version/access date.","section":"General"},{"comment":"The exemplar trajectory is illustrative but not quantitatively representative. Consider reporting median/quartile trajectories or a summary statistic for the qualitative claims about overshoot and saturation.","section":"§V-A.3 / Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a human-computer interaction / assistive technology venue and the open-source release is a genuine asset. The core 1-DOF algorithm is simple and plausible, and the inclusion of SCI participants is commendable. However, the abstract and contribution list currently overstate the strength of the evidence. The placement pre-screening and incomplete-trial handling are not minor presentation issues; they directly affect the central 'placement-agnostic' and 'all participants' claims. I believe the authors can address these with additional reporting (trial counts, adjustment counts, chance baseline, statistical precision) and by softening the claims for the preliminary 2-DOF result. That is why I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before the abstract sells you on it. The core idea is real: use sparse optical flow on B-mode ultrasound to get a continuous 1-DOF control signal from arbitrary body locations with only three calibration poses, plus a deliberately naive 2-DOF extension. That is a step beyond prior SMG work that needed location-specific training data. The evaluation includes 3 tetraplegic participants and 6 uninjured controls across six placements including neck and torso, which is a meaningful pilot. And they open-sourced the code and data, which makes the work reproducible and worth engaging with on that ground alone. The 1-DOF algorithm is simple, clearly explained in Eq. (1), and the heterogeneity results (best placement varies by user; passive tissue motion at non-innervated muscles can work) are genuinely informative for the field.\n\nThe soft spots are in the evaluation rather than the concept. The biggest is that sensor placements were adjusted in real time based on ultrasound image quality to satisfy the algorithm's feature requirements. That means the 'all tested sensor locations' in the abstract describes a pre-screened set, not arbitrary placement. The paper's methods and footnote do mention this, but the abstract doesn't carry the caveat. Asterisked incomplete trials in Fig. 3 are also silently absorbed into the reported averages, and there are no error bars or chance baseline, so a 5.5% RMSE is hard to interpret as 'control' rather than just smooth tracking. The 2-DOF results are explicitly preliminary, with coupling and binding acknowledged; that is honest but means the 'first location-agnostic multi-DOF' headline is more aspirational than established. None of this is a load-bearing mathematical error. The circularity worry you raised doesn't land: calibration-to-mapping and then testing tracking is a standard user-in-the-loop loop, not a fitted prediction.\n\nThis paper deserves a serious referee, not a desk reject. The right outcome is major revision: re-report with per-condition N, variance, and a chance or naive baseline, and soften the placement-agnostic wording to match what was actually tested. I'd bring it to a reading group if your group cares about assistive interfaces or biosignal control, and I'd cite the open-source system and the 1-DOF method with caveats. My recommendation: engage with it, but push the authors to align claims with evidence.","headline":"A resourceful pilot with a genuinely useful placement-agnostic SMG idea; the abstract sells the claim harder than the data support.","tokens_in":13538,"tokens_out":1808,"would_cite":true,"duration_ms":21632,"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":"The paper's central claim is that a single ultrasound probe placed almost anywhere on the body yields continuous cursor control from arbitrary tissue motion after a three-pose calibration, with a 2-DOF extension achieving multi-axis control","keywords":["sonomyography","ultrasound control","optical flow","tetraplegia","assistive technology","continuous control","sensor-placement-agnostic","motor rehabilitation"],"falsifier":"Take a sensor placement that passes the 1-DOF calibration but whose 'up' and 'right' feature-displacement vectors are mostly within 30 degrees of orthogonal (|cosine| ≤ 0.5) for fewer than half the tracked points. The paper's criterion predicts 2-DOF calibration should fail or couple; if independent 2-DOF control still emerges, the orthogonality assumption is not the deciding factor. Symmetrically, if a placement with the criterion satisfied yields no independent axes, the sufficiency claim is falsified.","tokens_in":12594,"feed_emoji":"🎯","tokens_out":10615,"duration_ms":94256,"temperature":0.7,"pith_summary":"This paper tries to establish that a single ultrasound probe, placed almost anywhere on the body, can give a person continuous control of a cursor after only a three-pose calibration. The method tracks small tissue features in the ultrasound image and maps their aggregate motion — including passive tissue motion, not just muscle contraction — to a continuous 1-DOF signal. All nine participants, including three cervical spinal cord injury survivors, achieved continuous 1-DOF control at all six sensor locations tested, and every participant reached under 5.5% tracking error at least one placement. A preliminary 2-DOF extension, calibrated with the help of an orthogonality check, let all participants modulate a two-dimensional cursor and several complete a letter-drawing task, which the authors call the first location-agnostic multi-DOF continuous sonomyography control. If these claims hold, SMG becomes a rapidly calibratable, high-dimensional control channel for assistive devices that can be tailored to each user's residual function.","feed_headline":"Ultrasound tissue motion yields cursor control at nearly any body site","feed_subtitle":"No per-user training data: a three-pose calibration turns arbitrary tissue motion into 1–2 continuous control axes.","key_machinery":"The load-bearing object is a linear mean-projection control law. For each tracked point, calibration records its neutral, up, and down positions; the live cursor position is the average across points of how far the current point has moved along the neutral-to-up line, normalized by the calibration distance and clamped to [0,1]. For 2-DOF, a computer-guided calibration adds a 'right' pose at which at least roughly half the points move within 30 degrees of orthogonal to their up motion, and the same projection is applied along each axis. This turns raw image motion into a bounded, continuous, unitless signal with no training data and makes the algorithm indifferent to which tissue is moving.","core_discovery":"The central claim: a control signal can be read out of arbitrary tissue deformation with no learned model. Roughly 30 trackable points are selected in the brightness-mode ultrasound image; the user defines neutral, up, and down poses; the live cursor position is the average, across points, of the current position's progress along the neutral-to-up vector, normalized and clamped. The same geometric rule works on biceps, neck, shoulder, and wrist sites, using passive tissue motion when the underlying muscle is not voluntarily active. All nine participants (three with cervical spinal cord injury) achieved continuous 1-DOF control at all six placements, with best errors under 5.5% (often under 4","pith_inferences":["If the geometric readout works at arbitrary sites, a natural next system is a 'body scan' calibration where a clinician sweeps one probe over several regions and software automatically recommends the placement with the largest, most orthogonal feature motions — something the paper's data implicitly enables.","The 2-DOF coupling and 'binding' the authors observe suggest the independence assumption, not the image content, is the bottleneck; a nonlinear decoder or signal-decomposition step trained on the same three-pose calibration could recover more independent axes.","The finding that passive tissue motion carries usable control signals expands the candidate body surface for tetraplegic users beyond innervated muscles; head and neck placements could serve users with no arm function, which the authors flag but did not yet recruit.","The anti-correlation between calibration motion size and error points to a testable refinement: automatically rescaling or amplifying small feature displacements before projection could improve control for users with limited range of motion."],"forward_implications":["Continuous 1-DOF control works at all tested body locations, including placements where the user has no voluntary control of the underlying muscle, because the tracker uses passive tissue motion.","Three-pose calibration makes it practical to quickly test many sensor locations and motions, letting each user find the control site that best fits their residual function and preference.","Since the best placement varied by user and every placement was best for at least one participant, placement-agnosticism is a functional requirement for heterogeneous tetraplegic populations, not just a convenience.","Single-probe 2-DOF control is feasible: all participants could modulate both axes and all three SCI participants traversed the full 2D workspace, supporting high-dimensional control without additional hardware.","Performance is best during slow, held motions and degrades with speed and time due to optical-flow drift, so the current system is a proof point for the approach rather than a finished interface."],"fun_headline_variants":["Three poses, any body site: cursor control from ultrasound","Ultrasound muscle motion becomes a joystick with 3-pose setup","Cursor control from any tissue: ultrasound without per-user training","Minimal calibration, maximal reach: ultrasound cursor control across body","Ultrasound tracks tissue motion to control cursors with 3 poses"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The system's load-bearing premise is that the chosen sensor spot always provides enough trackable image features moving far enough — and, for 2-DOF, in sufficiently orthogonal directions — between the user's calibration poses, and that the resulting linear projection stays accurate for minutes despite tracker drift; the paper itself reports placements where this failed.","fun_headline_variants_meta":{"raw":{"variants":["Three poses, any body site: cursor control from ultrasound","Ultrasound muscle motion becomes a joystick with 3-pose setup","Cursor control from any tissue: ultrasound without per-user training","Minimal calibration, maximal reach: ultrasound cursor control across body","Ultrasound tracks tissue motion to control cursors with 3 poses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2761,"prompt_tokens":884,"completion_tokens":1877,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1787}},"tokens_in":628,"tokens_out":1877,"duration_ms":16771,"temperature":1.0,"reasoning_tokens":1787,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:39:12.891493+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sensor placement that passes the 1-DOF calibration but whose 'up' and 'right' feature-displacement vectors are mostly within 30 degrees of orthogonal (|cosine| ≤ 0.5) for fewer than half the tracked points. The paper's criterion predicts 2-DOF calibration should fail or couple; if independent 2-DOF control still emerges, the orthogonality assumption is not the deciding factor. Symmetrically, if a placement with the criterion satisfied yields no independent axes, the sufficiency claim is falsified.","supporting_citations":[],"review_version":1}