{"id":"adc865f6-ad68-4f0e-a7ee-7e4e36ba8f4f","arxiv_id":"2411.18774","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Compared with whole-body marker models, a pelvis-marker-only estimate yields roughly double the force residual in dynamic activities, while Kalman-filter fusion of ground reaction forces did not beat the whole-body marker model.","lead":"This study compared three marker-based ways to estimate the body's center of mass during 14 activities. It found that pelvis-only markers are much less accurate in dynamic movements, while adding force-plate data to a whole-body marker model did not improve accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WMG-vs-WM comparison is circular: the Kalman filter was tuned to match WM, so the observed null result is a fitting artifact, not evidence that GRF fusion cannot help.","rationale":"I agree with the reader's weakest assumption. Section 2.2 explicitly states that the genetic algorithm minimized the difference between WMG and low-pass-filtered WM estimates, so the near-identical WMG/WM results are an expected consequence of the fitting procedure rather than evidence that GRF fusion does not improve CoM estimation. The paper's primary contribution—that the pelvis markerset is much worse than whole-body methods in dynamic activities—is large, robust, and not affected by this tuning issue. However, the secondary conclusion about GRF fusion is the main basis for a conditional verdict. Because the reader already flagged this concern and recommended conditional acceptance pending validation, my stress test does not change that verdict. A retuning or held-out test would either support or overturn the null result. Since the paper's code and data are not released, this test would require the authors to run it or share materials, but it is a concrete way to resolve whether the 'GRFs do not help' claim is real or a fitting artifact.","tokens_in":635,"tokens_out":6733,"duration_ms":83003,"concrete_test":"Retune Vw and Vv without referencing WM: split each subject's trials into tuning and testing sets, and choose the covariances by minimizing the RMS external force residual on tuning trials only (or, better, against synthetic motions with known CoM ground truth). Then recompute the WMG-vs-WM RMS residual comparison on the held-out testing trials. If WMG remains within about 1% of WM, the conclusion survives; if WMG improves or degrades materially, the original null result was an artifact of tuning to WM. To address the residual circularity, also run WMG with u=0 in Eq. 1 and compare; if the GRF input changes the residual substantially, the residual metric is not evaluating WMG independently.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's central negative claim about GRF fusion is undermined by the tuning procedure in Section 2.2. A genetic algorithm specified the Kalman filter noise covariances Vw and Vv by minimizing the difference between the WMG output and the low-pass-filtered WM estimate (6 Hz cutoff). Thus the later result that WMG and WM have nearly identical RMS external force residuals (0.28% difference, d=0.01) is an expected consequence of the fitting objective, not an independent experimental finding. The filter was never tuned against the external force residual or any independent ground truth. Consequently, the recommendation that GRF fusion through the presented Kalman filter does not improve CoM estimation is not supported by the comparison as run. Relatedly, because WMG uses the measured GRFs as control input u in Eq. 1 while Eq. 4 evaluates the same measured GRFs, the residual is not a fully independent accuracy metric for WMG. The pelvic-markerset result (PM vs WM/WMG) is much less affected by these concerns, since PM is not part of the tuning objective; however, the WMG-vs-WM comparison is the sole basis for the 'GRFs do not help' conclusion, and it is confounded.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript compares three center-of-mass (CoM) estimation methods—pelvic marker centroid (PM), whole-body inverse kinematics (WM), and a Kalman filter that fuses WM with ground reaction forces (WMG)—across 14 activities performed by 10 subjects. Accuracy is quantified by the root mean square (RMS) of the external force residual defined in Eq. (4). The authors report no differences between methods in static activities, substantially larger PM residuals in dynamic activities (Cohen's d ≈ 2.9–3.0), and nearly identical WMG and WM residuals. Based on these results, they recommend WM for general use and caution against PM for dynamic CoM estimation.","tokens_in":8963,"tokens_out":8964,"duration_ms":74260,"significance":"If the PM-versus-WM comparison is accepted, the paper provides a useful multi-activity benchmark: the data set covers 14 balance-relevant activities, the statistical analysis reports effect sizes and confidence intervals, and the PM disadvantage in dynamic tasks is large and consistent. The WMG-versus-WM comparison, however, is not an independent test of whether GRF fusion improves CoM estimation, because the Kalman filter covariances were tuned to minimize the difference between WMG and low-pass-filtered WM. The practical recommendation that GRF fusion does not help is therefore conditional on a reanalysis or a restricted claim. The study is valuable mainly for the PM-versus-WM result and for documenting a reproducible comparison protocol.","major_comments":[{"comment":"The Kalman filter covariance matrices Vw and Vv were specified by a genetic algorithm that minimized the difference between the WMG output and the low-pass-filtered WM estimate (6 Hz cutoff). The later observation in §3 that WMG and WM have nearly identical RMS residuals (0.28% difference, d=0.01) is therefore an expected consequence of the tuning objective, not an independent evaluation of whether GRF fusion improves CoM estimation. To support the central negative claim in the Abstract and §4.2, the filter should be tuned on an independent criterion (for example, minimizing the external force residual on held-out trials) or the claim must be weakened to 'no improvement under this tuning protocol'.","section":"§2.2 (after Eq. 3)"},{"comment":"The external force residual is not a fully independent accuracy metric for WMG. WMG uses the measured GRFs as the control input u in Eq. (1), and Eq. (4) evaluates the same measured Fext against the estimated acceleration; a low residual for WMG can reflect that the filter follows the GRF-derived acceleration directly rather than that the kinematic estimate is accurate. This confound should be acknowledged and, if possible, addressed with a metric that does not feed the same measured GRFs into both the estimator and the error signal.","section":"§2.2, Eq. (4)"},{"comment":"The reported WMG−WM post-hoc comparison (p<0.001, d=0.01, 95% CI (−0.82, 0.85)) is internally inconsistent: a 95% confidence interval that spans zero cannot accompany p<0.001 for the same comparison. This contradiction bears directly on the central WMG-versus-WM conclusion and must be resolved; if the CI is correct, the p-value should be non-significant, and if the p-value is correct, the CI or d is misreported.","section":"§3, position results"}],"minor_comments":[{"comment":"The classification of activities into Static and Dynamic is described as qualitative, based on the magnitude of displacement of body segments; please define an operational criterion or report the displacement range used for the split.","section":"§2.1, Table 1"},{"comment":"The smoothing parameter α=300 is said to be found through experimentation; please provide the selection criterion or a sensitivity analysis showing that the main results are robust to this choice.","section":"§2.2, Eq. (5)"},{"comment":"The genetic algorithm that sets Vw and Vv is described in one sentence; algorithmic details such as population size, number of generations, cost function, and convergence criteria are needed for reproducibility.","section":"§2.2, after Eq. (3)"},{"comment":"The One Leg, Squat, and Shoe Lace trials were split into two parts, but the effect of this split on Kalman filter initialization and on the smoothing optimization in Eq. (5) is not explained; please clarify whether the parts were processed separately or concatenated.","section":"§2.1"},{"comment":"For the velocity comparisons, the point estimates d=3.04 and d=2.99 are positive while their reported 95% CIs are negative; this sign inconsistency should be corrected.","section":"§3, velocity results"},{"comment":"The y-axis limits differ across panels, and within panels the axis limits are set to the maximum RMS among methods; please state explicitly that between-panel comparisons of bar heights are not meaningful.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The circular tuning of the Kalman filter and the p-value/CI inconsistency are the two load-bearing issues. The PM-versus-WM comparison is not affected by the tuning problem and is the strongest contribution; with a reanalysis or a restricted WMG claim, the manuscript could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper has one clean result and one circular one. Using a 14-activity dataset with ten subjects, it shows that pelvis-marker CoM estimates have 96–104% higher RMS external-force residuals than whole-body marker estimates in dynamic activities, with huge effect sizes. That result is credible and practically useful: people choosing marker-based CoM methods should be told to avoid pelvis markers when the body is moving a lot. Credit also goes to the diverse activity set and the transparent write-up of the three methods.\n\nThe circular part concerns the Whole-Body Markerset & GRFs (WMG) Kalman filter. In Section 2.2, the noise covariances Vw and Vv are set by a genetic algorithm that minimizes the difference between WMG output and the low-pass-filtered WM estimate. So the later finding that WMG and WM produce nearly identical residuals (0.28% difference, d=0.01) is expected by construction, not evidence that GRF fusion cannot help. The external-force residual metric is also not fully independent for WMG, since the measured GRFs are both the input u and the evaluation target. That does not kill the PM-versus-WM comparison, but it does kill the headline conclusion that GRF fusion offers no benefit.\n\nThe paper ships no data or code, which limits reproducibility, but the methods are described well enough to be followed. The literature is covered reasonably, including recent work like Buurke et al. 2023, though the comparison to that prior work is not deep.\n\nBottom line: this is a methods-comparison paper for biomechanics and wearable robotics. The PM-versus-WM recommendation is a value for practitioners. The GRF-fusion claim should not be published as is. A serious referee could help the authors fix the tuning problem, either by removing the WMG-vs-WM claim or by re-tuning the filter against an independent ground truth or a proper validation split. I would send it to peer review with clear instructions to address the circularity, not desk-reject it.","headline":"PM-vs-WM is a solid, useful result; the WMG-vs-WM null is circular because the Kalman filter was tuned to match WM.","tokens_in":9468,"tokens_out":1516,"would_cite":false,"duration_ms":15585,"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":"Three marker-based CoM estimators separate cleanly by task dynamics: pelvis-only estimates fail in dynamic activities, whole-body kinematics are accurate, and force-plate fusion adds no measurable benefit.","keywords":["center of mass estimation","external force residual","Kalman filter","ground reaction forces","marker-based motion capture","whole-body kinematics","pelvis markerset","balance assessment"],"falsifier":"Re-tune the Kalman filter's covariance matrices Vw and Vv from an independent source—force-plate noise statistics, or cross-validation on activities not used for tuning—and recompute the RMS external force residual in dynamic trials; if the re-tuned GRF-fused estimate beats the whole-body markerset estimate, the paper's central conclusion fails.","tokens_in":8492,"feed_emoji":"🦴","tokens_out":8026,"duration_ms":66303,"temperature":0.7,"pith_summary":"This paper asks which marker-based way of estimating the whole-body center of mass (CoM) and its velocity is accurate enough for balance and movement research. It compares a pelvis-only markerset, a whole-body marker model, and a Kalman filter that fuses whole-body kinematics with ground reaction forces, using the root-mean-square external force residual as the score. In static activities all three methods are statistically indistinguishable; in dynamic activities the pelvis-only method has 96% to 104% higher residuals than the two whole-body methods, with very large effect sizes. The whole-body method and the GRF-fused filter differ by only 0.28% in RMS residual, so the paper concludes the Kalman fusion does not improve whole-body kinematics. If correct, the comparison gives a practical default—whole-body markers—and shows force-plate data are not needed for CoM state estimation in the tested tasks.","feed_headline":"Pelvis-only markers miss dynamic center-of-mass motion by up to 104%","feed_subtitle":"Whole-body kinematics match force-plate fusion across 14 balance tasks; skip the Kalman filter.","key_machinery":"The load-bearing object is the external force residual r(t) = F_ext(t) − m a_est(t), where F_ext is the net external force from the force plates minus gravity, m is body mass, and a_est is CoM acceleration obtained by differentiating the estimated position or velocity; RMS of this residual is the performance metric. The third method is a Kalman filter with state x ∈ $R^{6}$ containing 3D CoM position and velocity, control input u built from GRF-derived acceleration, and measurement y from the whole-body marker position. Its process and measurement covariance matrices Vw and Vv were tuned by a genetic algorithm that minimizes the difference between the Kalman output and the low-pass-filtered whole-body markerset estimate, so the filter's similarity to the whole-body method is partly built into the comparison.","core_discovery":"On its own terms, the paper establishes that method choice matters mainly when the body moves. With ten healthy participants performing 14 Berg Balance Scale-derived activities, the measured external force residual F_ext(t) − m a_est(t) is roughly double for the pelvis-markerset estimate than for either whole-body method in every dynamic activity, and this gap is statistically significant with Cohen's d near 3.0. Static activities show no significant method effect because the CoM stays near the pelvis. The whole-body markerset and the whole-body-plus-GRFs Kalman filter produce nearly equal residuals, so the paper asserts that, at least for the filter's tuning, adding GRFs through a Kalman filter does not improve CoM position or velocity estimates. The authors therefore recommend the whole-body markerset as the default method and restrict the pelvis-only method to static tasks or situations where pelvis markers track whole-body motion.","pith_inferences":["Editorial inference: because the Kalman filter was tuned to match the whole-body markerset estimate on the same data used for evaluation, the claim that GRF fusion adds nothing should be treated as conditional on that tuning; an independent noise model or held-out tuning could change the comparison.","Editorial inference: the external-force residual could serve as a general, force-plate-based validation protocol for other CoM estimators, including inertial-sensor or robot-state estimators, whenever the ground reaction forces are measured.","Editorial inference: the tested activities do not include running, jumping, or impact tasks, so the recommendation may not extend to higher-bandwidth motions where double integration of GRFs is known to drift differently.","Editorial inference: the 96–104% residual gap suggests that using pelvis-only markers in dynamic fall-risk assessments could materially change extrapolated-CoM margin-of-stability values, a hypothesis the paper does not test."],"forward_implications":["In dynamic balance and mobility tasks, pelvis-only marker estimates of CoM position and velocity should be avoided when whole-body marker data are available; they produce roughly twice the external force residual.","Whole-body marker kinematics deliver consistent CoM estimates across all 14 activities, making them a good default when a full markerset is present.","Adding GRF measurements through the tested Kalman filter does not improve the whole-body kinematics estimate, so the added instrumentation and modeling complexity may not be justified for this purpose.","For quiet-stance and low-motion conditions, the pelvis-only method remains an acceptable low-cost option."],"supporting_citations":[{"why":"Supplies the ten-subject dataset and the 14 activity protocol repeated over the instrumented walkway that all methods are evaluated on.","marker":"Alizadeh Noghani and Bolívar-Nieto, 2024"},{"why":"Provides the Berg Balance Scale tasks from which the static and dynamic activities are drawn.","marker":"Berg et al., 1992"},{"why":"Defines the pelvis markerset CoM estimate as the average of the four pelvic markers.","marker":"Saini et al., 1998"},{"why":"Supplies the anthropometric segment inertial parameters used in the whole-body model.","marker":"Dumas et al., 2007"},{"why":"Provides the full-body musculoskeletal model that is modified and used for inverse kinematics.","marker":"Rajagopal et al., 2016"},{"why":"Provides the modeling platform whose inverse kinematics and BodyKinematics tools produce the whole-body CoM trajectory.","marker":"Delp et al., 2007"},{"why":"Original Kalman-filter CoM estimator that fuses whole-body kinematics with GRFs, which the WMG method implements.","marker":"Xinjilefu et al., 2015"}],"fun_headline_variants":["Pelvis markers fail dynamic CoM tracking, whole-body wins","Whole-body markers match force-plate fusion for CoM estimation","Skip the Kalman filter: whole-body markers are enough for CoM","Dynamic CoM errors double with pelvis-only markers","Pelvis-only CoM errors double in motion; whole-body is best"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison's key assumption is that the Kalman filter's noise coefficients were chosen fairly, but they were tuned to reproduce the low-pass-filtered whole-body marker estimate, which biases the force-plate-fusion method toward matching the method it is compared against.","fun_headline_variants_meta":{"raw":{"variants":["Pelvis markers fail dynamic CoM tracking, whole-body wins","Whole-body markers match force-plate fusion for CoM estimation","Skip the Kalman filter: whole-body markers are enough for CoM","Dynamic CoM errors double with pelvis-only markers","Pelvis-only CoM errors double in motion; whole-body is best"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000697,"raw_usage":{"total_tokens":3211,"prompt_tokens":1064,"completion_tokens":2147,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":2060}},"tokens_in":680,"tokens_out":2147,"duration_ms":12836,"temperature":1.0,"reasoning_tokens":2060,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:53:08.324121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-tune the Kalman filter's covariance matrices Vw and Vv from an independent source—force-plate noise statistics, or cross-validation on activities not used for tuning—and recompute the RMS external force residual in dynamic trials; if the re-tuned GRF-fused estimate beats the whole-body markerset estimate, the paper's central conclusion fails.","supporting_citations":[],"review_version":1}