{"id":"d80db79f-1075-418a-b18a-5b107587eaaf","arxiv_id":"2411.14992","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Markerless webcam-based biomechanical tracking matches optical motion capture for the shape of stroke drinking-task movement trajectories, but with large systematic angle offsets and weaker agreement on smoothness and coordination outcomes.","lead":"This study compared marker-based optical motion capture with a markerless webcam system that uses differentiable biomechanics to estimate arm movement in 15 stroke patients doing a drinking task. The markerless system matched the shape of movement trajectories closely, but showed large systematic angle offsets and weaker agreement on smoothness and coordination scores.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Agreement metrics are computed after removing per-trial static bias (up to 23.5° for shoulder flexion), so the reported 2–5° RMSE reflects offset-corrected waveform similarity, not standalone absolute joint-angle accuracy.","rationale":"The reader's weakest_assumption correctly identifies the per-trial bias and time-lag removal as the most fragile step. My independent reading agrees: the paper's Section II.D.1 clearly states that all further analyses are conducted on bias-removed trajectories after adding the static offset to the OMC signal, and Table I reports a shoulder flexion bias of -23.55° alongside a post-removal RMSE of 2.39°. This makes the abstract's unqualified 'RMSE 2–5 degrees for joint angles' misleading if read as absolute accuracy. The concern is not that bias removal is never legitimate; it is that the paper does not establish that the offset can be estimated and corrected in a standalone MMC deployment, nor does it report absolute errors for movement quality measures. The discussion acknowledges model differences and marker/keypoint placement as likely causes, but that admission strengthens the need for a calibration test rather than weakening it. The concrete test I propose would directly quantify how much of the reported agreement depends on the removed offsets and whether a practical calibration procedure can recover the accuracy claim. Because the reader already issued a CONDITIONAL verdict with this same concern as the basis, my stress-test does not change the verdict; it sharpens the required condition. I found no independent reason to reject or accept beyond this issue, and I credit the paper for sharing code, reporting exclusion rates, and candidly listing limitations including the OMC-only phase classification.","tokens_in":10959,"tokens_out":2936,"duration_ms":30133,"concrete_test":"Recompute the comparison in two ways with the shared code: (1) the paper's current pipeline with per-trial bias subtraction and time-lag optimization; (2) no bias subtraction and a fixed zero lag (or only a hardware-derived global lag). Report per-trial and per-participant absolute RMSE/MAE and Bland-Altman 95% limits of agreement for shoulder flexion, elbow extension, trunk displacement, and end-effector velocity. If absolute, non-bias-removed RMSE for shoulder flexion stays near 2–5°, the claim stands; if it jumps to roughly 20–25°, the headline must be qualified as offset-corrected waveform agreement. Additionally, test a clinically usable calibration: estimate the per-participant bias from the first five trials of a session and apply it to held-out trials, then compute held-out absolute RMSE.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the abstract and conclusion, is that the markerless pipeline is 'approaching the accuracy of marker-based methods.' This claim rests on the metrics in Table I, but those metrics are computed after subtracting the per-trial mean difference between MMC and OMC and after optimizing a time shift of up to 0.25 s (Section II.D.1). The shoulder flexion bias is -23.55° (IQR roughly -26.6 to -19.5), yet the reported RMSE after bias removal is only 2.39°–2.53°. A reader of the abstract, which reports RMSE of 2–5° for joint angles without mentioning bias removal, would naturally interpret this as absolute error. If the large static offsets reflect genuine differences in joint-angle definitions or model scaling between the OpenSim and MuJoCo models, then the RMSE after offset correction hides a systematic disagreement of tens of degrees. The authors argue the bias is consistent within a session and likely arises from marker/keypoint location differences, but they do not show that a standalone MMC deployment could estimate and remove this offset without OMC. Movement quality measures such as maximum shoulder flexion, elbow extension, and trunk displacement depend on absolute values; without a calibration procedure, these measures would be off by the full bias magnitude. The reported correlations for those measures (e.g., shoulder flexion rs = 0.97, elbow extension rs = 0.98) capture association, not agreement, and Table III already shows weaker results for interjoint coordination (rs = 0.71) and number of movement units (rs = 0.77). The time-lag optimization also removes genuine temporal misalignment, which can inflate correlation and reduce dynamic RMSE. Thus the central accuracy claim is not supported unless the bias and lag are demonstrated to be removable nuisance parameters in a real clinical workflow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript compares a markerless motion capture (MMC) pipeline based on differentiable biomechanics (Cotton et al.'s end-to-end approach) against marker-based optical motion capture (OMC) in 15 stroke patients performing a standardized drinking task. The authors report high agreement between the two systems after per-trial bias removal and time-lag optimization, with median correlations above 0.95 for most kinematic trajectories and RMSE values of 2–5 degrees for joint angles, and conclude that MMC is 'approaching the accuracy of marker-based methods' and could be used in clinical settings. The paper includes a large dataset (1160 trials), a clinically relevant task, and open-source analysis code.","tokens_in":11258,"tokens_out":2320,"duration_ms":23580,"significance":"If the central claim were fully supported, this would be a valuable contribution: it is the first comparison of the end-to-end differentiable MMC approach against OMC in neurological patients, uses a clinically recommended task, and provides reproducible code. The paper also honestly reports several limitations, including OMC phase classification and model differences. However, the load-bearing conclusion rests on agreement metrics computed after subtracting per-trial static bias and optimizing time lag. Because the removed shoulder-flexion bias reaches about 23.5 degrees, the reported RMSE and correlation values do not by themselves establish that MMC can measure absolute joint angles or movement quality measures without an OMC-based calibration. The study's strengths are the scale of the dataset, the clinical relevance, and the transparency about many methodological choices, but the interpretation of the headline accuracy numbers needs substantial revision.","major_comments":[{"comment":"The abstract states that 'median RMSE values ranging from 2-5 degrees for joint angles' were observed, without mentioning that these RMSEs are computed after subtracting a per-trial static bias and optimizing a per-trial time lag of up to 0.25 s. Table I shows a median bias of -23.55 degrees for shoulder flexion and -7.52 degrees for elbow extension. Removing these offsets before computing RMSE makes the reported values measures of waveform similarity after alignment, not absolute joint-angle accuracy. The conclusion that MMC is 'approaching the accuracy of marker-based methods' is therefore not directly supported. The authors should report absolute (offset-unadjusted) RMSE and bias, and discuss whether a standalone MMC deployment could estimate and remove such offsets without an OMC reference. If the large offsets reflect genuine differences in joint-angle definitions or model scaling between OpenSim and MuJoCo, then the offset-corrected RMSE understates the systematic disagreement by tens of degrees.","section":"II.D.1, Table I, Abstract"},{"comment":"Phase classification for both systems is derived exclusively from OMC end-effector velocity. This guarantees that phase boundaries and phase durations are identical for MMC and OMC, which inflates agreement for temporal measures such as time to peak velocity (rs = 0.98, rav = 1.00). The manuscript acknowledges this limitation but still presents these temporal agreement values as evidence of MMC accuracy. The authors should either validate phase classification using MMC alone or clearly re-label these results as reflecting kinematic agreement under OMC-defined phases, and temper the corresponding claims about temporal accuracy.","section":"II.D.2, III.C.2"},{"comment":"It is unclear whether the per-trial bias removal described in Section II.D.1 was also applied to the kinematic trajectories before computing the movement quality measures in Section II.D.2. If the bias is removed before extracting maxima such as shoulder flexion and elbow extension, then the high correlations in Table III (e.g., rs = 0.97 and 0.98) are correlations of offset-corrected waveforms, not agreement of absolute peak values. The manuscript should state explicitly which trajectory version (raw or bias-removed) was used for each movement quality measure, and if bias removal was applied, report the absolute values and their differences.","section":"II.D.1, II.D.2"},{"comment":"The statement that 'the differences in kinematic trajectories between OMC and MMC fall within the session-to-session variability observed with OMC alone [13]' is not supported by the cited reference. Reference [13] (Uchida and Seth) quantifies marker registration and model scaling uncertainty in inverse analyses, not session-to-session repeatability of OMC in a clinical population. The authors should either provide a direct comparison with an actual OMC test-retest study or revise the claim.","section":"IV, Discussion"}],"minor_comments":[{"comment":"There is a typo in 'biomechancical model' in the first paragraph of Section II.C.1.","section":"II.C.1"},{"comment":"Reference [7] contains a typo: 'compensationg' should be 'compensation'.","section":"References"},{"comment":"The table would be easier to read if the rows for each kinematic variable were visually separated or grouped, as the current dense layout makes cross-variable comparisons difficult.","section":"Table I"},{"comment":"The description of the time-lag optimization says 'shifting one kinematic trajectory over the other, up to 0.25 seconds' but does not specify the direction of the shift or the interpolation method used for sub-sample shifts; adding this detail would improve reproducibility.","section":"II.D.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is more ambitious than the presented metrics support, but the underlying dataset and pipeline are valuable and the issues are addressable with re-analysis and revised framing. I would not reject outright; however, the authors must either substantially soften the 'approaching accuracy' claim or provide absolute-error analyses and an OMC-free calibration strategy. The correlation-based movement quality results after bias removal are not sufficient for the clinical-use conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is the first comparison of Cotton's differentiable end-to-end markerless pipeline against optical motion capture in neurological patients, and that comparison is worth having. Fifteen stroke patients, 1160 drinking-task trials, 1.6% reconstruction failures, and shared code. The waveform agreement after offset correction is genuinely encouraging, and the per-trial bias consistency (IQRs of 1-3 degrees for joint angles) is a useful empirical finding for anyone designing calibration procedures. Credit where due: the authors ran a real clinical validation, report failure rates, and their discussion of marker-registration error and model differences is honest. The soft spot is the one the stress test flags, and it lands. The headline RMSE values of 2-5 degrees are computed after subtracting a per-trial static bias and time-shifting each trajectory by up to 0.25 seconds. The shoulder flexion bias is -23.55 degrees, so the 2.39-degree RMSE is not an absolute error; it is a measure of waveform shape similarity after removing a systematic offset the size of the signal. The abstract and conclusion say the system is 'approaching the accuracy of marker-based methods' without mentioning this. That is misleading. The movement-quality correlations in Table III are association, not agreement, and the two weakest measures (interjoint coordination rs=0.71, movement units rs=0.77) are exactly the ones that would matter clinically. Phase classification is OMC-only, so the paper cannot yet support standalone MMC for movement quality measures. I do not think this is a fatal flaw. The authors explicitly acknowledge the bias and discuss its likely sources, and they note that session-to-session bias variability may come from OMC. What they do not do is show that a standalone MMC deployment could estimate and remove that offset without an OMC reference. Without a calibration procedure, the movement quality measures are off by tens of degrees. That gap needs to be stated as a limitation, not buried. The fix is straightforward: report absolute RMSE and bias before offset removal, show Bland-Altman plots for the movement quality measures, and either demonstrate a bias-estimation method that works without OMC or openly state that absolute accuracy is not yet established. The paper deserves a serious referee - the data are real, the clinical question is important, and the methodological issue is tractable. I would bring it to reading group as a case study in how offset correction can inflate apparent agreement. I would cite it if I were working on MMC validation, because it is the first data point for this method in stroke. Send it to review, but the authors should expect a request for major revision.","headline":"Real first validation of end-to-end MMC against OMC in stroke, but the abstract overstates absolute accuracy by hiding per-trial bias removal.","tokens_in":798,"tokens_out":1036,"would_cite":true,"duration_ms":25033,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that markerless motion capture from five synchronized webcams, combined with differentiable biomechanical optimization, recovers upper-limb drinking-task kinematics in stroke patients whose trajectories agree with…","keywords":["markerless motion capture","optical motion capture","differentiable biomechanics","stroke rehabilitation","drinking task","upper limb kinematics","inverse kinematics","movement quality measures"],"falsifier":"Recompute all agreement metrics without subtracting the per-trial mean difference, in particular reporting shoulder-flexion RMSE between MMC and OMC before bias removal; if that error stands near the reported median bias of about 23.5 degrees rather than near 2–3 degrees, the claimed accuracy depends entirely on treating a large systematic difference as removable noise. A complementary test is to record the same participants with dual-plane fluoroscopy alongside both systems and check whether MMC's shoulder-flexion waveform tracks true bone motion as closely as it tracks marker-based angles only after offset removal.","tokens_in":10743,"feed_emoji":"🎥","tokens_out":6833,"duration_ms":62976,"temperature":0.7,"pith_summary":"This paper claims that a markerless motion capture (MMC) pipeline built on differentiable biomechanics can recover upper-limb kinematics in stroke patients with accuracy approaching that of marker-based optical motion capture (OMC). Fifteen stroke patients performed the standardized drinking task while five synchronized webcams recorded them alongside 3D marker cameras; the two systems' kinematic trajectories were then compared. After removing each trial's static offset and optimizing a small time shift, the paper reports median correlations above 0.95 for most joint-angle, trunk, and velocity trajectories, with joint-angle root-mean-square errors of roughly 2–5 degrees. If the claim holds, clinics could obtain quantitative movement-quality measures without marker placement, expensive equipment, or extensive setup, using only consumer webcams. The central requirement for interpreting the result is that the removed static offsets are harmless nuisance parameters rather than real differences in joint-angle definition between the two modeling pipelines.","feed_headline":"Five webcams approach marker-based motion capture for stroke rehab","feed_subtitle":"Trial-by-trial agreement with optical motion capture reaches r > 0.95 and 2–5 degree joint-angle errors.","key_machinery":"The load-bearing mechanism is an end-to-end differentiable biomechanical optimization that replaces the two-stage pipeline of marker trajectories followed by inverse kinematics with a single learnable mapping. An implicit neural function $f_\\phi: t \\to \\theta$ maps time to joint angles, and the parameters $\\phi$, together with per-participant body scaling and marker-offset parameters, are optimized by projecting a forward-kinematic model of the upper limb into calibrated webcam views and minimizing reprojection error. A bilevel optimization passes body-scale parameters across multiple trials, so one participant's scale is learned jointly with all their trajectory reconstructions. Before comparing outputs, the paper removes a per-trial static bias and optimizes a time lag up to 0.25 seconds, which is why the reported agreement describes dynamic waveform shape rather than absolute joint angles.","core_discovery":"The study's central claim is that an end-to-end differentiable markerless motion capture approach is nearing the accuracy of marker-based optical motion capture for measuring upper limb kinematics in stroke patients performing a drinking task. Instead of the conventional two-stage OMC workflow—3D marker trajectories followed by inverse kinematics on a scaled model—the MMC pipeline optimizes a neural implicit function that maps time to joint angles, jointly with body scale and marker offset parameters, by projecting a physics-based biomechanical model into five webcam views. The authors find median Pearson correlation r > 0.95 for a majority of kinematic trajectories and median RMSE values of 2–5 degrees for joint angles, 0.04 m/s for end-effector velocity, and 6 mm for trunk displacement after removing per-trial static bias and aligning time lag. They report that trial-to-trial biases between systems were consistent within participant sessions, with interquartile ranges of about 1–3 degrees for joint angles, 0.01 m/s, and 3 mm, and that reconstruction failed in only 1.6% of trials. The conclusion is that MMC for arm tracking is approaching marker-based accuracy and supports potential use in clinical settings.","pith_inferences":["A decisive test the paper does not run is to compare both OMC and MMC against dual-plane fluoroscopy of bone motion; until then, the reported 'accuracy' is agreement with a marker-based system that itself carries marker-registration and soft-tissue errors.","Because the largest removed bias reaches about 23.5 degrees for shoulder flexion, the headline 2–5 degree RMSE should be read as waveform-shape agreement after offset removal; absolute MMC joint angles are not yet interchangeable with OMC values without a per-session calibration.","The pipeline's reliance on 87 detected keypoints and five calibrated cameras leaves room to test whether monocular or fewer-camera configurations preserve the reported accuracy in everyday clinical environments.","Extending the same differentiable machinery to finger degrees of freedom and to other functional tasks is a natural next step, since the current models lock leg and finger motion."],"forward_implications":["With per-session bias correction, webcam-based MMC could replace marker-based systems for extracting drinking-task movement quality measures in clinical stroke assessments, eliminating marker placement and reducing cost and setup time.","The low and stable trial-to-trial bias implies that a participant-specific offset could be estimated from a single calibration trial and applied to later sessions, enabling longitudinal monitoring of recovery.","Because the reported errors fall within the session-to-session variability previously observed for OMC itself, MMC-based measures may be as clinically interpretable as marker-based ones for patients with mild to moderate impairment.","Metrics such as number of movement units and interjoint coordination need refinement or replacement before smoothness and coordination scores are fully comparable across systems.","Averaging multiple trials, as already recommended for OMC protocols, also improves MMC correlations, so clinical practice could retain multi-trial designs."],"supporting_citations":[{"why":"Introduces the differentiable end-to-end markerless motion capture optimization that this paper validates against optical motion capture.","marker":"[17]"},{"why":"Defines the standardized drinking task and the movement quality measures that structure the comparison.","marker":"[8]"},{"why":"Earlier report of the end-to-end MMC method's low reprojection errors and spatiotemporal gait parameters, motivating the need for validation against OMC.","marker":"[16]"},{"why":"Quantifies how marker registration uncertainty affects inverse kinematics, used to argue the observed MMC-OMC differences fall within OMC's own variability.","marker":"[13]"},{"why":"Provides the baseline biomechanical model on which the OpenSim model for two-stage OMC is built.","marker":"[19]"},{"why":"Supplies the modified upper-limb model used in the physics-based forward kinematics of the MMC pipeline.","marker":"[27]"},{"why":"Supplies the 87 2D keypoints extracted from webcam videos and triangulated for the markerless reconstruction.","marker":"[22]"},{"why":"Provides the bilevel optimization scheme that jointly learns body scale and inverse kinematics across multiple trials.","marker":"[26]"},{"why":"Supplies the GPU-accelerated physics engine that carries the forward kinematic model inside the optimization loop.","marker":"[24]"}],"fun_headline_variants":["Webcam motion capture rivals marker-based accuracy in stroke rehab","Five webcams match optical capture for stroke arm tracking","Markerless motion capture nears gold standard in stroke arm rehab","Webcam-based rehab tracking hits r>0.95 vs marker system","Stroke arm tracking with webcams nearly matches marker precision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the static offset and time lag removed before computing every agreement metric are mere calibration nuisances, not genuine differences in how the two systems define or scale joint angles; if those offsets reflect real biomechanical disagreement, the reported accuracy is overstated.","fun_headline_variants_meta":{"raw":{"variants":["Webcam motion capture rivals marker-based accuracy in stroke rehab","Five webcams match optical capture for stroke arm tracking","Markerless motion capture nears gold standard in stroke arm rehab","Webcam-based rehab tracking hits r>0.95 vs marker system","Stroke arm tracking with webcams nearly matches marker precision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000845,"raw_usage":{"total_tokens":3736,"prompt_tokens":1061,"completion_tokens":2675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":2600}},"tokens_in":677,"tokens_out":2675,"duration_ms":18153,"temperature":1.0,"reasoning_tokens":2600,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:38:32.854009+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute all agreement metrics without subtracting the per-trial mean difference, in particular reporting shoulder-flexion RMSE between MMC and OMC before bias removal; if that error stands near the reported median bias of about 23.5 degrees rather than near 2–3 degrees, the claimed accuracy depends entirely on treating a large systematic difference as removable noise. A complementary test is to record the same participants with dual-plane fluoroscopy alongside both systems and check whether MMC's shoulder-flexion waveform tracks true bone motion as closely as it tracks marker-based angles only after offset removal.","supporting_citations":[{"cited_title":"Kinematic analysis using 3D motion capture of drinking task in people with and without upper-extremity impairments,","cited_arxiv_id":null,"evidence_quote":"Defines the standardized drinking task and the movement quality measures that structure the comparison."},{"cited_title":"Optimizing Trajectories and Inverse Kinematics for Biomechanical Analysis of Markerless Motion Capture Data,","cited_arxiv_id":null,"evidence_quote":"Earlier report of the end-to-end MMC method's low reprojection errors and spatiotemporal gait parameters, motivating the need for validation against OMC."},{"cited_title":"Conclusion or Illusion: Quantifying Un- certainty in Inverse Analyses From Marker-Based Motion Capture due to Errors in Marker Registration and Model Scaling,","cited_arxiv_id":null,"evidence_quote":"Quantifies how marker registration uncertainty affects inverse kinematics, used to argue the observed MMC-OMC differences fall within OMC's own variability."},{"cited_title":"Full-Body Musculoskeletal Model for Muscle- Driven Simulation of Human Gait,","cited_arxiv_id":null,"evidence_quote":"Provides the baseline biomechanical model on which the OpenSim model for two-stage OMC is built."},{"cited_title":"Muscle contributions to propul- sion and support during running,","cited_arxiv_id":null,"evidence_quote":"Supplies the modified upper-limb model used in the physics-based forward kinematics of the MMC pipeline."},{"cited_title":"Learning 3D Human Pose Esti- mation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats,","cited_arxiv_id":null,"evidence_quote":"Supplies the 87 2D keypoints extracted from webcam videos and triangulated for the markerless reconstruction."},{"cited_title":"Rapid bilevel optimization to concurrently solve muscu- loskeletal scaling, marker registration, and inverse kinematic problems for human motion reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides the bilevel optimization scheme that jointly learns body scale and inverse kinematics across multiple trials."},{"cited_title":"MuJoCo: A physics engine for model-based control,","cited_arxiv_id":null,"evidence_quote":"Supplies the GPU-accelerated physics engine that carries the forward kinematic model inside the optimization loop."}],"review_version":1}